diff --git a/.gitignore b/.gitignore new file mode 100644 index 0000000..0d8e3a3 --- /dev/null +++ b/.gitignore @@ -0,0 +1,72 @@ +# JPEG +*.jpg +*.jpeg +*.jpe +*.jif +*.jfif +*.jfi + +# JPEG 2000 +*.jp2 +*.j2k +*.jpf +*.jpx +*.jpm +*.mj2 + +# JPEG XR +*.jxr +*.hdp +*.wdp + +# Graphics Interchange Format +*.gif + +# RAW +*.raw + +# Web P +*.webp + +# Portable Network Graphics +*.png + +# Animated Portable Network Graphics +*.apng + +# Multiple-image Network Graphics +*.mng + +# Tagged Image File Format +*.tiff +*.tif + +# Scalable Vector Graphics +*.svg +*.svgz + +# Portable Document Format +*.pdf +*.docx + +# X BitMap +*.xbm + +# BMP +*.bmp +*.dib + +# ICO +*.ico + +# network +*.pth +*.pt +# matlabs +*.mat +# numpy +*.npy +# 3D Images +*.3dm +*.max +*.xlsx diff --git a/MOT16_eval/eval.sh b/MOT16_eval/eval.sh new file mode 100644 index 0000000..5d1249b --- /dev/null +++ b/MOT16_eval/eval.sh @@ -0,0 +1,109 @@ +#!/bin/bash + +set +e + + +# start from clean slate +for i in data.zip MOT16.zip +do + zip -T $i + if [ $? -eq 0 ] + then + echo 'zip is ok' + + else + echo 'zip corrupted, deleting' + rm -rf $i + fi +done + + +# create output folder if it doesn't exist +if [ ! -d ./inference/output ] +then + mkdir -p ./inference/output + echo 'inference output folder created' +fi + + + +# clone evaluation repo if it does not exist +if [ ! -d ./MOT16_eval/TrackEval ] +then + echo 'Cloning official MOT16 evaluation repo' + git clone https://github.com/JonathonLuiten/TrackEval ./MOT16_eval/TrackEval + # download quick start data folder if it does not exist + if [ ! -d ./MOT16_eval/TrackEval/data ] + then + # download data + wget -nc https://omnomnom.vision.rwth-aachen.de/data/TrackEval/data.zip -O ./data.zip + # unzip + unzip -q ./data.zip -d ./MOT16_eval/TrackEval/ + # delete zip + #rm data.zip + fi +fi + + +# create MOT16 folder if it doesn't exist +if [ ! -d ./MOT16_eval/TrackEval/data/MOT16 ] +then + mkdir -p ./MOT16_eval/TrackEval/data/MOT16 +fi + + +# if MOT16 data not unziped, then download, unzip and lastly remove zip MOT16 data +if [[ ! -d ./MOT16_eval/TrackEval/data/MOT16/train ]] && [[ ! -d ./MOT16_eval/TrackEval/data/MOT16/test ]] +then + # download data + wget -nc https://motchallenge.net/data/MOT16.zip -O ./MOT16.zip + # unzip + unzip -q MOT16.zip -d ./MOT16_eval/TrackEval/data/MOT16/ + # delete zip + #rm MOT16.zip +fi + + +# create folder to place tracking results for this method +mkdir -p ./MOT16_eval/TrackEval/data/trackers/mot_challenge/MOT16-train/ch_yolov5m_deep_sort/data/ + +# inference on 4 MOT16 video sequences at the same time +# suits a 4GB GRAM GPU, feel free to increase if you have more memory +N=4 + +# generate tracking results for each sequence +for i in MOT16-02 MOT16-04 MOT16-05 MOT16-09 MOT16-10 MOT16-11 MOT16-13 +do + ( + # change name to inference source so that each thread write to its own .txt file + if [ ! -d ./MOT16_eval/TrackEval/data/MOT16/train/$i/$i ] + then + mv ./MOT16_eval/TrackEval/data/MOT16/train/$i/img1/ ./MOT16_eval/TrackEval/data/MOT16/train/$i/$i + fi + # run inference on sequence frames + python3 track.py --source ./MOT16_eval/TrackEval/data/MOT16/train/$i/$i --save-txt --yolo-weights yolov5/weights/crowdhuman_yolov5m.pt --classes 0 --exist-ok --imgsz 1280 + # move generated results to evaluation repo + ) & + # https://unix.stackexchange.com/questions/103920/parallelize-a-bash-for-loop + # allow to execute up to $N jobs in parallel + if [[ $(jobs -r -p | wc -l) -ge $N ]] + then + # now there are $N jobs already running, so wait here for any job + # to be finished so there is a place to start next one. + wait -n + fi +done + +# no more jobs to be started but wait for pending jobs +# (all need to be finished) +wait +echo "Inference on all MOT16 sequences DONE" + +echo "Moving data from experiment folder to MOT16" +mv ./runs/track/exp/* \ + ./MOT16_eval/TrackEval/data/trackers/mot_challenge/MOT16-train/ch_yolov5m_deep_sort/data/ + +# run the evaluation +python ./MOT16_eval/TrackEval/scripts/run_mot_challenge.py --BENCHMARK MOT16 \ + --TRACKERS_TO_EVAL ch_yolov5m_deep_sort --SPLIT_TO_EVAL train --METRICS CLEAR Identity \ + --USE_PARALLEL False --NUM_PARALLEL_CORES 4 diff --git a/Vehicle Recognition/Car Runner.py b/Vehicle Recognition/Car Runner.py new file mode 100644 index 0000000..9f0dc96 --- /dev/null +++ b/Vehicle Recognition/Car Runner.py @@ -0,0 +1,264 @@ +from torchreid.models import build_model +from torchreid import utils +import torch, os +import torch.nn as nn +from torch.utils.data import DataLoader +from torch.utils.data import Dataset +from torchvision import transforms +from PIL import Image +import numpy as np +import glob + +softmax = torch.nn.Softmax(dim=-1) +device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') + +def f1_loss(y_true, y_pred): #should change + + tp = (y_true * y_pred).sum().to(torch.float32) + tn = ((1 - y_true) * (1 - y_pred)).sum().to(torch.float32) + fp = ((1 - y_true) * y_pred).sum().to(torch.float32) + fn = (y_true * (1 - y_pred)).sum().to(torch.float32) + + epsilon = 1e-7 + + precision = tp / (tp + fp + epsilon) + recall = tp / (tp + fn + epsilon) + + f1 = 2* (precision*recall) / (precision + recall + epsilon) + return f1 + +class CA_Loader(Dataset): + def __init__(self,img_path, + attr, + resolution, + transform=None): + + # images variables: + self.img_path = img_path + self.img_names = attr['img_names'] + self.resolution = resolution + + self.attr = attr['attributes'] + + if transform: + self.transform = transform + else: + self.transform = None + + transform_list = [] + transform_list += [transforms.ToTensor()] + transform_list += [transforms.Resize(resolution)] #should change + transform_list += [transforms.Normalize(mean=[0.4611, 0.4658, 0.4728], std=[0.2552, 0.2502, 0.2520])] + preprocess = transforms.Compose(transform_list) + self.preprocess = preprocess + + def __len__(self): + return len(self.img_names) + + def __getitem__(self,idx): + + img = Image.open(self.img_names[idx]).convert('RGB') + if self.transform: + img = self.transform(img) + img = self.preprocess(img) + out = {'img' : img} + out.update({'attributes':self.attr[idx]}) + + return out + +def load_attributes(path_attr): + attr_vec_np = np.load(path_attr)# loading attributes + # attributes + attr_vec_np = attr_vec_np.astype(np.int32) + return torch.from_numpy(attr_vec_np) + +def load_image_names(main_path): + paths = glob.glob(main_path+'/*/*.jpg') + paths2 = [] + for i in paths: + ii = (i.split('\\'))[-1].split('.jpg')[0] + label = i.split('\\')[-2] + paths2.append((ii, label)) + paths = sorted(paths2,key=lambda x:float(x[0])) + paths = [os.path.join(main_path, label , s+'.jpg') for s, label in paths] + return np.array(paths) + +def data_delivery(main_path, path_attr=None): + output = {} + attr_vec = load_attributes(path_attr) # numpy array + output.update({'attributes':attr_vec}) + img_names = load_image_names(main_path) + output.update({'img_names':img_names}) + + output.update({'names' : ['206','207i','405','Arisun','Dena','HcCross','JackS5','KaraMazdaPickup','L90','MVM315H', + 'MVMX22','NeissanVanet','Pars','PeykanSavari','PeykanVanet','Pride131nasimsaba', + 'Pride132and111','Pride141','PrideVanet151','Quik','RenaultPK','RioSD','Runna','Saina', + 'Samand','SamandSoren','Shahin','Tiba','Xantia']}) #should change + + return output + +#IranianCarsDataset bama +#CarDataset camera +train_img_path = './Car/Data/SVID/train' +test_img_path = './Car/Data/SVID/test' +path_attr_train = './Car/Data/SVID/train.npy' +path_attr_test = './Car/Data/SVID/test.npy' +saving_path = './Car/Results/' + +attr_train = data_delivery(train_img_path, path_attr_train) #should check for new dataset +attr_test = data_delivery(test_img_path, path_attr_test) + +train_transform = transforms.Compose([transforms.RandomRotation(degrees=(0,45)), + transforms.RandomHorizontalFlip(), + transforms.GaussianBlur(kernel_size=(5, 9), sigma=(0.2, 2)), + transforms.ColorJitter(saturation=[0.7,1.4], brightness = (0.8, 1.2), contrast = (0.8, 1.2)), + transforms.RandomPerspective(distortion_scale=0.3, p=0.4), + transforms.RandomAffine(degrees=(1, 15), scale=(0.92, 0.99)), + #transforms.RandomInvert(), + #transforms.RandomPosterize(bits=2), + #transforms.RandomSolarize(threshold=192.0) + ]) + +train_data = CA_Loader(img_path=train_img_path, #check for new dataset + attr=attr_train, + resolution=(256,256), + transform=train_transform) + +test_data = CA_Loader(img_path=test_img_path, + attr=attr_test, + resolution=(256,256), + transform=None) + +loss = nn.CrossEntropyLoss().to(device) + +batch_size = 32 #can increase +train_loader = DataLoader(train_data,batch_size=batch_size,shuffle=True) +test_loader = DataLoader(test_data,batch_size=100,shuffle=False) + +model = build_model(name='resnet50', + num_classes = attr_train['attributes'].shape[1], #check this for new dataset + loss='softmax', + pretrained=True + ) + +'''def get_n_params(model): + pp=0 + for p in list(model.parameters()): + nn=1 + for s in list(p.size()): + nn = nn*s + pp += nn + return pp +print(get_n_params(model))''' + +#pretrained = './checkpoints/osnet_x1_0_msmt17.pth' +trained = './best_attr_net.pth' #set the best pretrained +utils.load_pretrained_weights(model, trained) + +'''params = model.parameters() +for idx, param in enumerate(params): + param.requires_grad = False''' + +#attr_net = attributes_model(model, feature_dim = 512, attr_dim = attr_train['attributes'].shape[1]) +attr_net = model.to(device) + +params = attr_net.parameters() + +optimizer = torch.optim.AdamW(params, lr=3e-5, betas=(0.9, 0.99), eps=1e-08) #check the lr and milestones +scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=[20, 60, 100, 140], gamma=0.7) + +train_loss = [] +test_loss = [] +F1_train = [] +F1_test = [] +Acc_train = [] +Acc_test = [] +recall_test = [] +precision_test = [] + +num_epoch = 200 #check this + +attr_loss_train = torch.zeros((num_epoch)) +attr_loss_test = torch.zeros((num_epoch)) + +loss_min = 10000 +f1_best = 0 + +from sklearn.metrics import f1_score, precision_recall_curve + +for epoch in range(1, num_epoch + 1): + attr_net = attr_net.to(device) + attr_net.train() + + loss_e = [] + loss_t = [] + ft_train = [] + ft_test = [] + acc_train = [] + acc_test = [] + ret_test = [] + prt_test = [] + + for idx, data in enumerate(train_loader): + for key, _ in data.items(): + data[key] = data[key].to(device) + # forward step + optimizer.zero_grad() + out_data = attr_net.forward(data['img']) + + loss_part = loss(out_data, data['attributes'].float()) + + loss_e.append(loss_part.item()) + print(loss_part.item()) + loss_part.backward() + + optimizer.step() + + train_loss.append(np.mean(loss_e)) + attr_net.eval() + + with torch.no_grad(): + for idx, data in enumerate(test_loader): + for key, _ in data.items(): + data[key] = data[key].to(device) + + out_data = attr_net.forward_attr_eval(data['img']) + + loss_part = loss(out_data, data['attributes'].float()) + + out_data = softmax(out_data) + _, preds = torch.max(out_data, 1) + out_data = torch.zeros(out_data.shape[0], out_data.shape[1]).to(device) + + for i, pred in enumerate(preds): + out_data[i][pred] = 1 + + f1_sc = f1_loss(data['attributes'], out_data) + #for i in range(data['attributes'].shape[1]): lst.append(f1_score(data['attributes'][i].cpu().data, out_data[i].cpu())) + ft_test.append(f1_sc.item()) + #ret_test.append(recall_sc.item()) + #prt_test.append(precision_sc.item()) + #acc_test.append(test_attr_metrics[-3]) + loss_t.append(loss_part.item()) + + test_loss.append(np.mean(loss_t)) + F1_test.append(np.mean(ft_test)) + #recall_test.append(np.mean(ret_test)) + #precision_test.append(np.mean(prt_test)) + #Acc_test.append(np.mean(acc_test)) + + print('Epoch: {}\ntrain loss: {:.6f}\ntest loss: {:.6f}\nF1:{:.6f}\n'.format( + epoch,train_loss[-1],test_loss[-1], F1_test[-1])) + + #print('Epoch: {}\ntrain loss: {:.6f}\ntest loss: {:.6f}\n\nF1 train: {:.4f}\nF1 test: {:.4f}\n\nacc_train: {:.4f}\nacc_test: {:.4f}\n'.format( + # epoch,train_loss[-1],test_loss[-1],F1_train[-1],F1_test[-1],Acc_train[-1],Acc_test[-1])) + + scheduler.step() + + d = 0 + if test_loss[-1] < loss_min: + loss_min = min(test_loss) + d += 1 + torch.save(attr_net.state_dict(), os.path.join(saving_path, 'best_attr_net.pth')) + torch.save(epoch, os.path.join(saving_path, 'best_epoch.pth')) + print('test loss improved') \ No newline at end of file diff --git a/Vehicle Recognition/Car/Data/SVID/OSikesakhtammiomadbenevisam.py b/Vehicle Recognition/Car/Data/SVID/OSikesakhtammiomadbenevisam.py new file mode 100644 index 0000000..6504cdb --- /dev/null +++ b/Vehicle Recognition/Car/Data/SVID/OSikesakhtammiomadbenevisam.py @@ -0,0 +1,6 @@ +import os +import glob +paths = glob.glob('./train/*/*.jpg') + +for path in paths: + os.replace(path, "./train/"+path.split('\\')[-1]) diff --git a/Vehicle Recognition/Car/Data/SVID/checklabelchonseyedocddare.py b/Vehicle Recognition/Car/Data/SVID/checklabelchonseyedocddare.py new file mode 100644 index 0000000..437421b --- /dev/null +++ b/Vehicle Recognition/Car/Data/SVID/checklabelchonseyedocddare.py @@ -0,0 +1,35 @@ +import numpy as np +import glob +import matplotlib.pyplot as plt +from PIL import Image +import cv2, os + +names = ['206','207i','405','Arisun','Dena','HcCross','JackS5','KaraMazdaPickup','L90','MVM315H', + 'MVMX22','NeissanVanet','Pars','PeykanSavari','PeykanVanet','Pride131nasimsaba', + 'Pride132and111','Pride141','PrideVanet151','Quik','RenaultPK','RioSD','Runna','Saina', + 'Samand','SamandSoren','Shahin','Tiba','Xantia'] + +paths = glob.glob('.\\train\\*\\*.jpg') +paths2 = [] + +for i in paths: + ii = (i.split('\\'))[-1].split('.jpg')[0] + label = i.split('\\')[-2] + paths2.append((ii, label)) + +paths = sorted(paths2,key=lambda x:float(x[0])) + +paths = [os.path.join('.\\train\\', label , s+'.jpg') for s, label in paths] + +train = np.load('train.npy') +i = 29362 + +print(train[i]) +print(names[np.argmax(train[i])]) + +im = cv2.imread(paths[i]) + +cv2.imshow('kir', im) + +cv2.waitKey(0) +print() \ No newline at end of file diff --git a/Vehicle Recognition/Car/Data/SVID/normstd_calc.py b/Vehicle Recognition/Car/Data/SVID/normstd_calc.py new file mode 100644 index 0000000..4a1789c --- /dev/null +++ b/Vehicle Recognition/Car/Data/SVID/normstd_calc.py @@ -0,0 +1,36 @@ +import glob +import torchvision.transforms as T +import numpy as np +from PIL import Image + +'''m1 = [0.47365347, 0.4684487, 0.46542493] +std1 = [0.2530077, 0.24997568, 0.25495428] +m2 = [0.47200933, 0.46324202, 0.45690387] +std2 = [0.25115883, 0.25053665, 0.25555712]''' + +transforms = [] +transforms += [T.Resize((256, 256))] +transforms += [T.ToTensor()] +preprocess = T.Compose(transforms) + +path = glob.glob('.\\train\\*.jpg') + +p1 = path[:len(path)//2] +p2 = path[len(path)//2:] + +images = [] +images = [preprocess(Image.open(i)) for i in p1] # generator comprehension +images = np.stack(images) # this takes time +mean1 = [np.mean(images[:,0,:,:]),np.mean(images[:,1,:,:]),np.mean(images[:,2,:,:])] +std1 = [np.std(images[:,0,:,:]),np.std(images[:,1,:,:]),np.std(images[:,2,:,:])] + +images = [] +images = [preprocess(Image.open(i)) for i in p2] # generator comprehension +images = np.stack(images) # this takes time +mean2 = [np.mean(images[:,0,:,:]),np.mean(images[:,1,:,:]),np.mean(images[:,2,:,:])] +std2 = [np.std(images[:,0,:,:]),np.std(images[:,1,:,:]),np.std(images[:,2,:,:])] + +print(mean1) +print(std1) +print(mean2) +print(std2) diff --git a/Vehicle Recognition/Car/Data/SVID/osikesakhtambudbaratest.py b/Vehicle Recognition/Car/Data/SVID/osikesakhtambudbaratest.py new file mode 100644 index 0000000..bbad489 --- /dev/null +++ b/Vehicle Recognition/Car/Data/SVID/osikesakhtambudbaratest.py @@ -0,0 +1,6 @@ +import os +import glob +paths = glob.glob('./test/*/*.jpg') + +for path in paths: + os.replace(path, "./test/"+path.split('\\')[-1]) diff --git a/Vehicle Recognition/Car/Data/SVID/test.py b/Vehicle Recognition/Car/Data/SVID/test.py new file mode 100644 index 0000000..bbfe962 --- /dev/null +++ b/Vehicle Recognition/Car/Data/SVID/test.py @@ -0,0 +1,43 @@ +import glob +import os +paths = glob.glob('./all/test/*/*.jpg') + +data = {} + +paths2 = [] + +for i in paths: + ii = (i.split('\\'))[-1].split('.jpg')[0] + label = i.split('\\')[-2] + paths2.append((ii, label)) + +paths = sorted(paths2,key=lambda x:float(x[0])) + +paths = [os.path.join('.\\test\\', label , s+'.jpg') for s, label in paths] + +for i, path in enumerate(paths): + label = path.split("\\")[2] + filename = path.split("\\")[3] + data.update({filename:label}) + + #os.rename(path, os.path.join('./All', str(i)+'.jpg')) + #os.remove(path) + #im = Image.open(path).convert("RGB") + #im.save(path.split(".webp")[0]+".jpg","jpeg" + +classes = list(data.values()) + +categories = glob.glob('./all/test/*') +categories = [i.split('\\')[-1] for i in categories] + +import numpy as np + +labels = np.zeros((len(classes), len(categories))) + +for i, element in enumerate(classes): + idx = categories.index(element) + labels[i][idx] = 1 + +np.save('test.npy', labels) + +print() \ No newline at end of file diff --git a/Vehicle Recognition/Car/Data/SVID/train.py b/Vehicle Recognition/Car/Data/SVID/train.py new file mode 100644 index 0000000..1b4203a --- /dev/null +++ b/Vehicle Recognition/Car/Data/SVID/train.py @@ -0,0 +1,43 @@ +import glob +import os +paths = glob.glob('./all/train/*/*.jpg') + +data = {} + +paths2 = [] + +for i in paths: + ii = (i.split('\\'))[-1].split('.jpg')[0] + label = i.split('\\')[-2] + paths2.append((ii, label)) + +paths = sorted(paths2,key=lambda x:float(x[0])) + +paths = [os.path.join('.\\train\\', label , s+'.jpg') for s, label in paths] + +for i, path in enumerate(paths): + label = path.split("\\")[2] + filename = path.split("\\")[3] + data.update({filename:label}) + + #os.rename(path, os.path.join('./All', str(i)+'.jpg')) + #os.remove(path) + #im = Image.open(path).convert("RGB") + #im.save(path.split(".webp")[0]+".jpg","jpeg" + +classes = list(data.values()) + +categories = glob.glob('./all/train/*') +categories = [i.split('\\')[-1] for i in categories] + +import numpy as np + +labels = np.zeros((len(classes), len(categories))) + +for i, element in enumerate(classes): + idx = categories.index(element) + labels[i][idx] = 1 + +np.save('train.npy', labels) + +print() \ No newline at end of file diff --git a/Vehicle Recognition/errors.pkl b/Vehicle Recognition/errors.pkl new file mode 100644 index 0000000..80c124d Binary files /dev/null and b/Vehicle Recognition/errors.pkl differ diff --git a/Vehicle Recognition/seyed_bullshit.py b/Vehicle Recognition/seyed_bullshit.py new file mode 100644 index 0000000..fda0f25 --- /dev/null +++ b/Vehicle Recognition/seyed_bullshit.py @@ -0,0 +1,60 @@ +import os +from glob import glob +import numpy as np +from random import shuffle + + + +directory = r'carScrape\SVID' + +folders = glob(directory+ r'\*') + +images = glob(directory + r'\*' + r'\*.jpg') + + +for i,image in enumerate(images): + os.rename(image, os.path.join(os.path.join(image.split('\\')[0], image.split('\\')[1], image.split('\\')[2]), f"{i}.jpg")) + + + +def train_test_split(directory): + folders = glob(directory + r'\*') + for folder in folders: + images = glob(os.path.join(folder, '*.jpg')) + + + + + + +from PIL import Image +import glob, os +paths = glob.glob('./*/*.jpg') + +data = {} + +for i, path in enumerate(paths): + label = path.split("\\")[1] + filename = path.split("\\")[1] + '-' + path.split("\\")[2] + data.update({filename:label}) + + #os.rename(path, os.path.join('./All', str(i)+'.jpg')) + #os.remove(path) + #im = Image.open(path).convert("RGB") + #im.save(path.split(".webp")[0]+".jpg","jpeg") + +classes = list(data.values()) + +categories = list(dict.fromkeys(classes)) + +import numpy as np + +labels = np.zeros((len(classes), len(categories))) + +for i, element in enumerate(classes): + idx = categories.index(element) + labels[i][idx] = 1 + +print() + + diff --git a/Vehicle Recognition/test_cnn_single_image.py b/Vehicle Recognition/test_cnn_single_image.py new file mode 100644 index 0000000..27e684b --- /dev/null +++ b/Vehicle Recognition/test_cnn_single_image.py @@ -0,0 +1,45 @@ +from torchreid.models import build_model +import torch +from PIL import Image +import torchvision.transforms as T + +names = ['206','207i','405','Arisun','Dena','HcCross','JackS5','KaraMazdaPickup','L90','MVM315H', + 'MVMX22','NeissanVanet','Pars','PeykanSavari','PeykanVanet','Pride131nasimsaba', + 'Pride132and111','Pride141','PrideVanet151','Quik','RenaultPK','RioSD','Runna','Saina', + 'Samand','SamandSoren','Shahin','Tiba','Xantia'] + +model = build_model( + name='resnet50', + num_classes=29, + loss='softmax', + pretrained=False + ) + +trained_net = torch.load("./best_attr_net.pth") +model.load_state_dict(trained_net) + +model.eval() + +model.to('cuda') + +image = Image.open("./ssss.jpg").convert('RGB') + +transforms = [] +transforms += [T.Resize((256, 256))] +transforms += [T.ToTensor()] +transforms += [T.Normalize(mean=[0.4611, 0.4658, 0.4728], std=[0.2552, 0.2502, 0.2520])] +preprocess = T.Compose(transforms) + +image = preprocess(image) + +images = image.unsqueeze(0).to('cuda') + +res = model.forward_attr_eval(images) + +softmax = torch.nn.Softmax(dim=-1) +out_data = softmax(res) +_, preds = torch.max(out_data, 1) + +print(names[preds.item()]) + +print() \ No newline at end of file diff --git a/Vehicle Recognition/test_cnn_test_set.py b/Vehicle Recognition/test_cnn_test_set.py new file mode 100644 index 0000000..d947e0f --- /dev/null +++ b/Vehicle Recognition/test_cnn_test_set.py @@ -0,0 +1,63 @@ +from torchreid.models import build_model +import torch +from PIL import Image +import torchvision.transforms as T +from glob import glob +import os +import numpy as np +import pickle +with open('errors.pkl', 'rb') as f: + errors = pickle.load(f) + +npy = np.load('./Car/Data/SVID/test.npy') +images = glob('./Car/Data/SVID/test/*/*.jpg') + +paths2 = [] + +for i in images: + ii = (i.split('\\'))[-1].split('.jpg')[0] + label = i.split('\\')[-2] + paths2.append((ii, label)) + +paths = sorted(paths2,key=lambda x:float(x[0])) + +paths = [os.path.join('.\\Car\\Data\\SVID\\test\\', label , s+'.jpg') for s, label in paths] + +model = build_model( + name='resnet50', + num_classes=29, + loss='softmax', + pretrained=False + ) + +trained_net = torch.load("./best_attr_net.pth") +model.load_state_dict(trained_net) + +model.eval() + +model.to('cuda') +os.makedirs('errors', exist_ok=True) +errors = dict() +for i, img in enumerate(paths): + image = Image.open(img).convert('RGB') + + transforms = [] + transforms += [T.Resize((256, 256))] + transforms += [T.ToTensor()] + transforms += [T.Normalize(mean=[0.4611, 0.4658, 0.4728], std=[0.2552, 0.2502, 0.2520])] + preprocess = T.Compose(transforms) + + image = preprocess(image) + + image = image.unsqueeze(0).to('cuda') + + res = model.forward_attr_eval(image) + + softmax = torch.nn.Softmax(dim=-1) + out_data = softmax(res) + conf, preds = torch.max(out_data, 1) + if preds != np.argmax(npy[i]): #### change this to match test images and npy + errors[img] = (preds.item(), npy[i], conf.item()) #### I'm not sure that my dataset is ok + del image + +print() \ No newline at end of file diff --git a/Vehicle Recognition/transformcheck.py b/Vehicle Recognition/transformcheck.py new file mode 100644 index 0000000..388e43b --- /dev/null +++ b/Vehicle Recognition/transformcheck.py @@ -0,0 +1,22 @@ +from torchvision import transforms as T +import matplotlib.pyplot as plt +from PIL import Image +import numpy as np + + +transforms = [] +transforms += [T.ToTensor()] +transforms += [T.RandomPosterize(bits=2)] +preprocess = T.Compose(transforms) + +image = Image.open('./aaa.jpg').convert('RGB') + +image = preprocess(image).permute((1,2,0)) + +image = image.cpu().detach().numpy() + + +plt.imshow(image) +plt.show() + +print() \ No newline at end of file diff --git a/aaa.mp4 b/aaa.mp4 new file mode 100644 index 0000000..9fdfedd Binary files /dev/null and b/aaa.mp4 differ diff --git a/eee.mp4 b/eee.mp4 new file mode 100644 index 0000000..ff72f02 Binary files /dev/null and b/eee.mp4 differ diff --git a/ggg.mp4 b/ggg.mp4 new file mode 100644 index 0000000..d8dab6f Binary files /dev/null and b/ggg.mp4 differ diff --git a/kkk.mp4 b/kkk.mp4 new file mode 100644 index 0000000..982b2bb Binary files /dev/null and b/kkk.mp4 differ diff --git a/mb_models.py b/mb_models.py new file mode 100644 index 0000000..a390035 --- /dev/null +++ b/mb_models.py @@ -0,0 +1,524 @@ +#!/usr/bin/env python3 +# -*- coding: utf-8 -*- +""" +Created on Wed Feb 17 14:03:10 2021 + +@author: hossein + +here we can find different types of models +that are define for person-attribute detection. +this is Hossein Bodaghies thesis +""" + +import torch.nn as nn +import torch +import copy +#%% +from torchreid.models.osnet import Conv1x1, OSBlock +device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') + +blocks = [OSBlock, OSBlock, OSBlock] +layers = [2, 2, 2] +channels = [16, 64, 96, 128] # channels are the only difference between os_net_x_1 and others + +def _make_layer( + block, + layer, + in_channels, + out_channels, + reduce_spatial_size, + IN=False +): + layers = [] + + layers.append(block(in_channels, out_channels, IN=IN)) + for i in range(1, layer): + layers.append(block(out_channels, out_channels, IN=IN)) + + if reduce_spatial_size: + layers.append( + nn.Sequential( + Conv1x1(out_channels, out_channels), + nn.AvgPool2d(2, stride=2) + ) + ) + + return nn.Sequential(*layers) + +#%% + +class CD_builder(nn.Module): + + def __init__(self, + model, + num_id, + feature_dim = channels[3], + attr_feat_dim = channels[1], + attr_dim = 46, + dropout_p = 0.3): + + super().__init__() + + self.feature_dim = feature_dim + self.attr_feat_dim = attr_feat_dim + self.dropout_p = dropout_p + self.global_avgpool = nn.AdaptiveAvgPool2d(1) + self.softmax = nn.Softmax(dim=1) + self.sigmoid = nn.Sigmoid() + self.model = model + + self.fc = self._construct_fc_layer(self.attr_feat_dim, channels[-1], dropout_p=dropout_p) + + self.attr_clf = nn.Linear(self.attr_feat_dim, attr_dim) + + + def _construct_fc_layer(self, fc_dims, input_dim, dropout_p=None): + if fc_dims is None or fc_dims < 0: + self.feature_dim = input_dim + return None + + if isinstance(fc_dims, int): + fc_dims = [fc_dims] + + layers = [] + for dim in fc_dims: + layers.append(nn.Linear(input_dim, dim)) + layers.append(nn.BatchNorm1d(dim)) + layers.append(nn.ReLU(inplace=True)) + if dropout_p is not None: + layers.append(nn.Dropout(p=dropout_p)) + input_dim = dim + + self.feature_dim = fc_dims[-1] + + return nn.Sequential(*layers) + + def get_feature(self, x, get_attr=True, get_feature=True, get_collection=False): + + out_conv4 = self.out_layers_extractor(x, 'out_conv4') + # The path for multi-branches for attributes + out_head = self.attr_branch(out_conv4, self.conv_head, self.head_fc, self.head_clf, need_feature=True) + out_body = self.attr_branch(out_conv4, self.conv_body, self.body_fc, self.body_clf, need_feature=True) + out_body_type = self.attr_branch(out_conv4, self.conv_body_type, self.body_type_fc, self.body_type_clf, need_feature=True) + out_leg = self.attr_branch(out_conv4, self.conv_leg ,self.leg_fc, self.leg_clf, need_feature=True) + out_foot = self.attr_branch(out_conv4, self.conv_foot, self.foot_fc, self.foot_clf, need_feature=True) + out_gender = self.attr_branch(out_conv4, self.conv_gender, self.gender_fc, self.gender_clf, need_feature=True) + out_bags = self.attr_branch(out_conv4, self.conv_bags, self.bags_fc, self.bags_clf, need_feature=True) + out_body_colour = self.attr_branch(out_conv4, self.conv_body_color, self.body_color_fc, self.body_color_clf, need_feature=True) + out_leg_colour = self.attr_branch(out_conv4, self.conv_leg_color, self.leg_color_fc, self.leg_color_clf, need_feature=True) + out_foot_colour = self.attr_branch(out_conv4, self.conv_foot_color, self.foot_color_fc, self.foot_color_clf, need_feature=True) + + # The path for person re-id: + del out_conv4 + x = self.out_layers_extractor(x, 'out_fc') + x = [out_head, out_body, out_body_type, out_leg, + out_foot, out_gender, out_bags, out_body_colour, + out_leg_colour, out_foot_colour, x] + outputs = torch.cat(x, dim=1) + return outputs + + + def vector_features(self, x): + features = self.model(x) + out_attr = self.attr_lin(features) + out_features = torch.cat(features, out_attr, dim=1) + return out_features + + def out_layers_extractor(self, x, layer): + out_os_layers = self.model.layer_extractor(x, layer) + return out_os_layers + + def attr_branch(self, x, conv_layer, fc_layer, clf_layer, need_feature=False): + x = conv_layer(x) + x = self.global_avgpool(x) + x = x.view(x.size(0), -1) + x = fc_layer(x) + if need_feature: + return x + else: + x = clf_layer(x) + return x + + def forward(self, x): + features = self.out_layers_extractor(x, 'out_globalavg') + features = features.view(features.size(0), -1) + features = self.fc(features) + out_attr = self.attr_clf(features) + + return {'attr':out_attr} + + def save_baseline(self, saving_path): + torch.save(self.model.state_dict(), saving_path) + print('baseline model save to {}'.format(saving_path)) + +#%% + +class attributes_model(nn.Module): + + ''' + a model for training whole attributes + ''' + def __init__(self, + model, + feature_dim = 512, + attr_dim = 79): + + super().__init__() + self.feature_dim = feature_dim + self.model = model + self.attr_lin = nn.Linear(in_features=feature_dim , out_features=attr_dim) + + def out_layers_extractor(self, x, layer): + out_os_layers = self.model.layer_extractor(x, layer) + return out_os_layers + + def forward(self, x, get_features = False): + + features = self.out_layers_extractor(x, 'fc') + if get_features: + return features + else: + return {'attributes':self.attr_lin(features)} + + def save_baseline(self, saving_path): + torch.save(self.model.state_dict(), saving_path) + print('baseline model save to {}'.format(saving_path)) + +#%% +from torchvision import transforms +class Loss_weighting(nn.Module): + + ''' + a model for training weights of loss functions + ''' + def __init__(self, weights_dim=48): + + super().__init__() + self.weights_dim = weights_dim + + self.weights_lin1 = nn.Linear(in_features=weights_dim , out_features=weights_dim) + self.weights_lin2 = nn.Linear(in_features=weights_dim , out_features=weights_dim) + self.relu = nn.ReLU() + def forward(self, weights): + weights = self.weights_lin1(weights) + weights = self.relu(weights) + weights = self.weights_lin2(weights) + weights = torch.sigmoid(weights) + return weights + + def save_baseline(self, saving_path): + torch.save(self.weights_lin.state_dict(), saving_path) + print('loss_weights saved to {}'.format(saving_path)) + + +class mb_CA_auto_build_model(nn.Module): + + def __init__(self, + model, + main_cov_size = 512, + attr_dim = 128, + dropout_p = 0.3, + sep_conv_size = 64, + branch_names = None, + feature_selection = None): + + super().__init__() + self.feat_indices = feature_selection + self.feature_dim = main_cov_size + if self.feature_dim != 384 and self.feature_dim != 512: + raise Exception('main_cov_size should be 384 or 512') + self.dropout_p = dropout_p + self.global_avgpool = nn.AdaptiveAvgPool2d(1) + self.softmax = nn.Softmax(dim=1) + self.sigmoid = nn.Sigmoid() + self.model = model + self.sep_conv_size = sep_conv_size + self.attr_dim = attr_dim + self.branch_names = branch_names + + if self.feat_indices is not None: + self.feature_dim = 25 + + self.attr_feat_dim = sep_conv_size + self.branches = {} + for k in self.branch_names.keys(): + # convs + setattr(self, 'conv_'+k, _make_layer(blocks[2], + layers[2], + self.feature_dim, + self.sep_conv_size, + reduce_spatial_size=False + )) + + # fully connecteds + setattr(self, 'fc_'+k, self._construct_fc_layer(self.attr_dim, self.attr_feat_dim, dropout_p=dropout_p)) + # classifiers + setattr(self, 'clf_'+k, nn.Linear(self.attr_dim, branch_names[k])) + + + def _construct_fc_layer(self, fc_dims, input_dim, dropout_p=None): + + if isinstance(fc_dims, int): + fc_dims = [fc_dims] + + layers = [] + for dim in fc_dims: + layers.append(nn.Linear(input_dim, dim)) + layers.append(nn.BatchNorm1d(dim)) + layers.append(nn.ReLU(inplace=True)) + if dropout_p is not None: + layers.append(nn.Dropout(p=dropout_p)) + input_dim = dim + + return nn.Sequential(*layers) + + def get_feature(self, x, get_attr=True, get_feature=True, method='both', get_collection=False): + + if self.feature_dim == 512: + out_conv4 = self.out_layers_extractor(x, 'out_conv4') + elif self.feature_dim == 384: + out_conv4 = self.out_layers_extractor(x, 'out_conv3') + else: + raise Exception('main_cov_size should be 384 or 512') + + out_features = {} + + for k in self.branch_names.keys(): + out_features.setdefault(k, self.attr_branch(out_conv4 if self.feat_indices == None else torch.index_select(out_conv4, 1, self.feat_indices[0]), + fc_layer = getattr(self,'fc_'+k), + clf_layer = getattr(self,'clf_'+k), + conv_layer = getattr(self,'conv_'+k), need_feature = True) + ) + + del out_conv4 + out_fc_branches = [item[0] for item in list(out_features.values())] + outputs_clfs = {} + for k, v in out_features.items(): + outputs_clfs.update({k: v[1]}) + + x = self.out_layers_extractor(x, 'out_fc') + out_fc_branches = torch.cat(out_fc_branches, dim=1) + if method == 'both': + outputs_fcs = torch.cat((out_fc_branches,x), dim=1) + elif method == 'baseline': + outputs_fcs = x + elif method == 'branches': + outputs_fcs = out_fc_branches + return outputs_fcs, outputs_clfs + + + def vector_features(self, x): + features = self.model(x) + out_attr = self.attr_lin(features) + out_features = torch.cat(features, out_attr, dim=1) + return out_features + + def out_layers_extractor(self, x, layer): + out_os_layers = self.model.layer_extractor(x, layer) + return out_os_layers + + def attr_branch(self, x, fc_layer, clf_layer, + conv_layer=None, need_feature=False): + ''' fc_layer should be a list of fully connecteds + clf_layer hould be a list of classifiers + ''' + # handling conv layer + if conv_layer: + x = conv_layer(x) + x = self.global_avgpool(x) + x = x.view(x.size(0), -1) + x = fc_layer(x) + if need_feature: + out = clf_layer(x) + return x, out + out = clf_layer(x) + return out + + def forward(self, x, need_feature=False): + if self.feature_dim == 512: + out_conv4 = self.out_layers_extractor(x, 'out_conv4') + elif self.feature_dim == 384: + out_conv4 = self.out_layers_extractor(x, 'out_conv3') + else: + raise Exception('main_cov_size should be 384 or 512') + + out_attributes = {} + + for k in self.branch_names.keys(): + out_attributes.setdefault(k, self.attr_branch(out_conv4 if self.feat_indices == None else torch.index_select(out_conv4, 1, self.feat_indices[0]), + fc_layer = getattr(self,'fc_'+k), + clf_layer = getattr(self,'clf_'+k), + conv_layer = getattr(self,'conv_'+k), need_feature = need_feature) + ) + + return out_attributes + + def save_baseline(self, saving_path): + torch.save(self.model.state_dict(), saving_path) + print('baseline model save to {}'.format(saving_path)) + + +branch_channels = [16, 64, 96, 128] +base_channels = [64,64,256,384,512] +class mb_CA_auto_same_depth_build_model(nn.Module): + + def __init__(self, + model, + branch_place, + dropout_p = 0.3, + branch_names = None, + feature_selection = None): + super().__init__() + self.branch_place = branch_place + self.feat_indices = feature_selection + self.dropout_p = dropout_p + self.global_avgpool = nn.AdaptiveAvgPool2d(1) + self.softmax = nn.Softmax(dim=1) + self.sigmoid = nn.Sigmoid() + self.model = model + self.branch_fcs = branch_names + + self.layer_list = ['conv1', 'maxpool', 'conv2', 'conv3', 'conv4'] + self.layer_init_dim = [64,64,256,384,512] + + if self.feat_indices is not None: + self.feature_dim = 25 + + branches = {k:[] for k,v in self.branch_fcs.items()} + for k in self.branch_fcs.keys(): + # convs + if branch_place not in ['conv5', 'conv4']: + + idx = self.layer_list.index(branch_place)-1 + + for i, layer in enumerate(self.layer_list[self.layer_list.index(branch_place)+1:]): + + branches[k].append(_make_layer( + blocks[idx], + layers[idx], + self.layer_init_dim[idx+1] if i==0 else branch_channels[idx], + branch_channels[idx+1], + reduce_spatial_size=False if layer=='conv4' else True + )) + idx += 1 + + # classifiers + # if branch_place == 'conv4' or branch_place == 'conv5': + idx = self.layer_list.index('conv4')-1 + if branch_place != 'conv5': + branches[k].append(Conv1x1(base_channels[idx+1] if branch_place=='conv4' else branch_channels[idx] , + branch_channels[idx])) + branches[k].append(nn.AdaptiveAvgPool2d(1)) + branches[k].append(self._construct_fc_layer(branch_channels[idx], + branch_channels[idx], dropout_p=None)) + else: + branches[k].append(nn.AdaptiveAvgPool2d(1)) + branches[k].append(self._construct_fc_layer(branch_channels[idx], + base_channels[idx+1], dropout_p=None)) + branches[k].append(nn.Linear(branch_channels[idx], branch_names[k])) + setattr(self, 'branch_'+k, nn.Sequential(*branches[k])) + '''# convs + for layer in self.layer_list[self.layer_list.index(branch_place)+1:]: + branches[k].append(copy.deepcopy(getattr(model, layer))) + branches[k][-1].load_state_dict(getattr(model, layer).state_dict()) + + # classifiers + branches[k].append(nn.Linear(self.attr_dim, branch_names[k])) + setattr(self, 'branch_'+k, nn.Sequential(*branches[k])) + self.layer_list.append('clf')''' + + def _construct_fc_layer(self, fc_dims, input_dim, dropout_p=None): + + if isinstance(fc_dims, int): + fc_dims = [fc_dims] + + layers = [] + for dim in fc_dims: + layers.append(nn.Linear(input_dim, dim)) + layers.append(nn.BatchNorm1d(dim)) + layers.append(nn.ReLU(inplace=True)) + if dropout_p is not None: + layers.append(nn.Dropout(p=dropout_p)) + input_dim = dim + + return nn.Sequential(*layers) + + def get_feature(self, x, get_attr=True, get_feature=True, method='both', get_collection=False): + + out_baseline = self.out_layers_extractor(x, self.branch_place) + out_attributes = {} + + for k in self.branch_fcs.keys(): + out_attributes.setdefault(k, self.attr_branch(out_baseline if self.feat_indices == None else torch.index_select(out_baseline, 1, self.feat_indices[0]), + getattr(self,'branch_'+k), + need_feature = False) + ) + out_baseline = self.out_layers_extractor(x, 'fc') + + return out_attributes, out_baseline + + def get_all_branch_features(self, x, need_feature=True, baseline='conv4'): + + out_baseline = self.out_layers_extractor(x, self.branch_place) + out_attributes = {} + + for k in self.branch_fcs.keys(): + out_attributes.setdefault(k, self.attr_branch(out_baseline if self.feat_indices == None else torch.index_select(out_baseline, 1, self.feat_indices[0]), + getattr(self,'branch_'+k), + need_feature = need_feature) + ) + + return out_attributes + + def vector_features(self, x): + features = self.model(x) + out_attr = self.attr_lin(features) + out_features = torch.cat(features, out_attr, dim=1) + return out_features + + def out_layers_extractor(self, x, layer): + baseline, attention_point = self.model.layer_extractor(x, 'fc') + return baseline, attention_point + + def attr_branch(self, x, branch_layers, need_feature=False): + ''' fc_layer should be a list of fully connecteds + clf_layer hould be a list of classifiers + ''' + # handling conv layer + if self.branch_place != 'conv5': + start_point = self.layer_list.index(self.branch_place) + else: + start_point = self.layer_list.index('conv4') + + features = [] + + for idx, layer in enumerate(branch_layers): + + if layer == branch_layers[-2]: + x = x.view(x.size(0), -1) + + if need_feature: + x = layer(x) + features.append(x) + else: + x = layer(x) + features = x + + return features + + def forward(self, x, need_feature=False): + + out_baseline, out_attention = self.out_layers_extractor(x, self.branch_place) + out_attributes = {} + + for k in self.branch_fcs.keys(): + out_attributes.setdefault(k, self.attr_branch(out_attention if self.feat_indices == None else torch.index_select(out_attention, 1, self.feat_indices[0]), + getattr(self,'branch_'+k), + need_feature = need_feature) + ) + + return out_baseline, out_attributes + + def save_baseline(self, saving_path): + torch.save(self.model.state_dict(), saving_path) + print('baseline model save to {}'.format(saving_path)) \ No newline at end of file diff --git a/ooo.mp4 b/ooo.mp4 new file mode 100644 index 0000000..44f63f8 Binary files /dev/null and b/ooo.mp4 differ diff --git a/qqq.mp4 b/qqq.mp4 new file mode 100644 index 0000000..a000a95 Binary files /dev/null and b/qqq.mp4 differ diff --git a/requirements.txt b/requirements.txt new file mode 100644 index 0000000..df8ef02 --- /dev/null +++ b/requirements.txt @@ -0,0 +1,37 @@ +# pip install -r requirements.txt + +# base ---------------------------------------- + +matplotlib>=3.2.2 +numpy>=1.18.5 +opencv-python>=4.1.2 +Pillow>=7.1.2 +PyYAML>=5.3.1 +requests>=2.23.0 +scipy>=1.4.1 +torch>=1.7.0 +torchvision>=0.8.1 +tqdm>=4.41.0 + +# plotting ------------------------------------ + +pandas>=1.1.4 +seaborn>=0.11.0 + +# deep_sort ----------------------------------- + +easydict + +# torchreid + +Cython +h5py +six +tb-nightly +future +yacs +gdown +flake8 +yapf +isort==4.3.21 +imageio \ No newline at end of file diff --git a/rrr.mp4 b/rrr.mp4 new file mode 100644 index 0000000..6fad21e Binary files /dev/null and b/rrr.mp4 differ diff --git a/sss.mp4 b/sss.mp4 new file mode 100644 index 0000000..db53495 Binary files /dev/null and b/sss.mp4 differ diff --git a/strong_sort/.gitignore b/strong_sort/.gitignore new file mode 100644 index 0000000..37ed2f4 --- /dev/null +++ b/strong_sort/.gitignore @@ -0,0 +1,13 @@ +# Folders +__pycache__/ +build/ +*.egg-info + + +# Files +*.weights +*.t7 +*.mp4 +*.avi +*.so +*.txt diff --git a/strong_sort/LICENSE b/strong_sort/LICENSE new file mode 100644 index 0000000..92a1ed5 --- /dev/null +++ b/strong_sort/LICENSE @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2020 Ziqiang + +Permission is hereby granted, free of charge, to any person obtaining a copy +of this software and associated documentation files (the "Software"), to deal +in the Software without restriction, including without limitation the rights +to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +SOFTWARE. \ No newline at end of file diff --git a/strong_sort/README.md b/strong_sort/README.md new file mode 100644 index 0000000..6073f80 --- /dev/null +++ b/strong_sort/README.md @@ -0,0 +1,137 @@ +# Deep Sort with PyTorch + +![](demo/demo.gif) + +## Update(1-1-2020) +Changes +- fix bugs +- refactor code +- accerate detection by adding nms on gpu + +## Latest Update(07-22) +Changes +- bug fix (Thanks @JieChen91 and @yingsen1 for bug reporting). +- using batch for feature extracting for each frame, which lead to a small speed up. +- code improvement. + +Futher improvement direction +- Train detector on specific dataset rather than the official one. +- Retrain REID model on pedestrain dataset for better performance. +- Replace YOLOv3 detector with advanced ones. + +**Any contributions to this repository is welcome!** + + +## Introduction +This is an implement of MOT tracking algorithm deep sort. Deep sort is basicly the same with sort but added a CNN model to extract features in image of human part bounded by a detector. This CNN model is indeed a RE-ID model and the detector used in [PAPER](https://arxiv.org/abs/1703.07402) is FasterRCNN , and the original source code is [HERE](https://github.com/nwojke/deep_sort). +However in original code, the CNN model is implemented with tensorflow, which I'm not familier with. SO I re-implemented the CNN feature extraction model with PyTorch, and changed the CNN model a little bit. Also, I use **YOLOv3** to generate bboxes instead of FasterRCNN. + +## Dependencies +- python 3 (python2 not sure) +- numpy +- scipy +- opencv-python +- sklearn +- torch >= 0.4 +- torchvision >= 0.1 +- pillow +- vizer +- edict + +## Quick Start +0. Check all dependencies installed +```bash +pip install -r requirements.txt +``` +for user in china, you can specify pypi source to accelerate install like: +```bash +pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple +``` + +1. Clone this repository +``` +git clone git@github.com:ZQPei/deep_sort_pytorch.git +``` + +2. Download YOLOv3 parameters +``` +cd detector/YOLOv3/weight/ +wget https://pjreddie.com/media/files/yolov3.weights +wget https://pjreddie.com/media/files/yolov3-tiny.weights +cd ../../../ +``` + +3. Download deepsort parameters ckpt.t7 +``` +cd deep_sort/deep/checkpoint +# download ckpt.t7 from +https://drive.google.com/drive/folders/1xhG0kRH1EX5B9_Iz8gQJb7UNnn_riXi6 to this folder +cd ../../../ +``` + +4. Compile nms module +```bash +cd detector/YOLOv3/nms +sh build.sh +cd ../../.. +``` + +Notice: +If compiling failed, the simplist way is to **Upgrade your pytorch >= 1.1 and torchvision >= 0.3" and you can avoid the troublesome compiling problems which are most likely caused by either `gcc version too low` or `libraries missing`. + +5. Run demo +``` +usage: python yolov3_deepsort.py VIDEO_PATH + [--help] + [--frame_interval FRAME_INTERVAL] + [--config_detection CONFIG_DETECTION] + [--config_deepsort CONFIG_DEEPSORT] + [--display] + [--display_width DISPLAY_WIDTH] + [--display_height DISPLAY_HEIGHT] + [--save_path SAVE_PATH] + [--cpu] + +# yolov3 + deepsort +python yolov3_deepsort.py [VIDEO_PATH] + +# yolov3_tiny + deepsort +python yolov3_deepsort.py [VIDEO_PATH] --config_detection ./configs/yolov3_tiny.yaml + +# yolov3 + deepsort on webcam +python3 yolov3_deepsort.py /dev/video0 --camera 0 + +# yolov3_tiny + deepsort on webcam +python3 yolov3_deepsort.py /dev/video0 --config_detection ./configs/yolov3_tiny.yaml --camera 0 +``` +Use `--display` to enable display. +Results will be saved to `./output/results.avi` and `./output/results.txt`. + +All files above can also be accessed from BaiduDisk! +linker：[BaiduDisk](https://pan.baidu.com/s/1YJ1iPpdFTlUyLFoonYvozg) +passwd：fbuw + +## Training the RE-ID model +The original model used in paper is in original_model.py, and its parameter here [original_ckpt.t7](https://drive.google.com/drive/folders/1xhG0kRH1EX5B9_Iz8gQJb7UNnn_riXi6). + +To train the model, first you need download [Market1501](http://www.liangzheng.com.cn/Project/project_reid.html) dataset or [Mars](http://www.liangzheng.com.cn/Project/project_mars.html) dataset. + +Then you can try [train.py](deep_sort/deep/train.py) to train your own parameter and evaluate it using [test.py](deep_sort/deep/test.py) and [evaluate.py](deep_sort/deep/evalute.py). +![train.jpg](deep_sort/deep/train.jpg) + +## Demo videos and images +[demo.avi](https://drive.google.com/drive/folders/1xhG0kRH1EX5B9_Iz8gQJb7UNnn_riXi6) +[demo2.avi](https://drive.google.com/drive/folders/1xhG0kRH1EX5B9_Iz8gQJb7UNnn_riXi6) + +![1.jpg](demo/1.jpg) +![2.jpg](demo/2.jpg) + + +## References +- paper: [Simple Online and Realtime Tracking with a Deep Association Metric](https://arxiv.org/abs/1703.07402) + +- code: [nwojke/deep_sort](https://github.com/nwojke/deep_sort) + +- paper: [YOLOv3](https://pjreddie.com/media/files/papers/YOLOv3.pdf) + +- code: [Joseph Redmon/yolov3](https://pjreddie.com/darknet/yolo/) diff --git a/strong_sort/__init__.py b/strong_sort/__init__.py new file mode 100644 index 0000000..2a942db --- /dev/null +++ b/strong_sort/__init__.py @@ -0,0 +1,11 @@ +from .strong_sort import StrongSORT + + +__all__ = ['StrongSORT', 'build_tracker'] + + +def build_tracker(cfg, use_cuda): + return StrongSORT(cfg.STRONGSORT.REID_CKPT, + max_dist=cfg.STRONGSORT.MAX_DIST, min_confidence=cfg.STRONGSORT.MIN_CONFIDENCE, + nms_max_overlap=cfg.STRONGSORT.NMS_MAX_OVERLAP, max_iou_distance=cfg.STRONGSORT.MAX_IOU_DISTANCE, + max_age=cfg.STRONGSORT.MAX_AGE, n_init=cfg.STRONGSORT.N_INIT, nn_budget=cfg.STRONGSORT.NN_BUDGET, use_cuda=use_cuda) diff --git a/strong_sort/configs/strong_sort.yaml b/strong_sort/configs/strong_sort.yaml new file mode 100644 index 0000000..8f36793 --- /dev/null +++ b/strong_sort/configs/strong_sort.yaml @@ -0,0 +1,10 @@ +STRONGSORT: + ECC: True # activate camera motion compensation + MC_LAMBDA: 0.995 # matching with both appearance (1 - MC_LAMBDA) and motion cost + EMA_ALPHA: 0.9 # updates appearance state in an exponential moving average manner + MAX_DIST: 0.2 # The matching threshold. Samples with larger distance are considered an invalid match + MAX_IOU_DISTANCE: 0.7 # Gating threshold. Associations with cost larger than this value are disregarded. + MAX_AGE: 30 # Maximum number of missed misses before a track is deleted + N_INIT: 3 # Number of frames that a track remains in initialization phase + NN_BUDGET: 100 # Maximum size of the appearance descriptors gallery + diff --git a/strong_sort/deep/__init__.py b/strong_sort/deep/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/strong_sort/deep/checkpoint/.gitkeep b/strong_sort/deep/checkpoint/.gitkeep new file mode 100644 index 0000000..e69de29 diff --git a/strong_sort/deep/reid/.flake8 b/strong_sort/deep/reid/.flake8 new file mode 100644 index 0000000..4fc103c --- /dev/null +++ b/strong_sort/deep/reid/.flake8 @@ -0,0 +1,18 @@ +[flake8] +ignore = + # At least two spaces before inline comment + E261, + # Line lengths are recommended to be no greater than 79 characters + E501, + # Missing whitespace around arithmetic operator + E226, + # Blank line contains whitespace + W293, + # Do not use bare 'except' + E722, + # Line break after binary operator + W504, + # isort found an import in the wrong position + I001 +max-line-length = 79 +exclude = __init__.py, build, torchreid/metrics/rank_cylib/ \ No newline at end of file diff --git a/strong_sort/deep/reid/.gitignore b/strong_sort/deep/reid/.gitignore new file mode 100644 index 0000000..6c093a6 --- /dev/null +++ b/strong_sort/deep/reid/.gitignore @@ -0,0 +1,140 @@ +# Byte-compiled / optimized / DLL files +__pycache__/ +*.py[cod] +*$py.class + +# C extensions +*.so + +# Distribution / packaging +.Python +build/ +develop-eggs/ +dist/ +downloads/ +eggs/ +.eggs/ +lib/ +lib64/ +parts/ +sdist/ +var/ +wheels/ +pip-wheel-metadata/ +share/python-wheels/ +*.egg-info/ +.installed.cfg +*.egg +MANIFEST + +# PyInstaller +# Usually these files are written by a python script from a template +# before PyInstaller builds the exe, so as to inject date/other infos into it. +*.manifest +*.spec + +# Installer logs +pip-log.txt +pip-delete-this-directory.txt + +# Unit test / coverage reports +htmlcov/ +.tox/ +.nox/ +.coverage +.coverage.* +.cache +nosetests.xml +coverage.xml +*.cover +.hypothesis/ +.pytest_cache/ + +# Translations +*.mo +*.pot + +# Django stuff: +*.log +local_settings.py +db.sqlite3 + +# Flask stuff: +instance/ +.webassets-cache + +# Scrapy stuff: +.scrapy + +# Sphinx documentation +docs/_build/ + +# PyBuilder +target/ + +# Jupyter Notebook +.ipynb_checkpoints + +# IPython +profile_default/ +ipython_config.py + +# pyenv +.python-version + +# pipenv +# According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control. +# However, in case of collaboration, if having platform-specific dependencies or dependencies +# having no cross-platform support, pipenv may install dependencies that don't work, or not +# install all needed dependencies. +#Pipfile.lock + +# celery beat schedule file +celerybeat-schedule + +# SageMath parsed files +*.sage.py + +# Environments +.env +.venv +env/ +venv/ +ENV/ +env.bak/ +venv.bak/ + +# Spyder project settings +.spyderproject +.spyproject + +# Rope project settings +.ropeproject + +# mkdocs documentation +/site + +# mypy +.mypy_cache/ +.dmypy.json +dmypy.json + +# Pyre type checker +.pyre/ + +# Cython eval code +*.c +*.html + +# OS X +.DS_Store +.Spotlight-V100 +.Trashes +._* + +# ReID +reid-data/ +log/ +saved-models/ +model-zoo/ +debug* diff --git a/strong_sort/deep/reid/.isort.cfg b/strong_sort/deep/reid/.isort.cfg new file mode 100644 index 0000000..8039326 --- /dev/null +++ b/strong_sort/deep/reid/.isort.cfg @@ -0,0 +1,10 @@ +[isort] +line_length=79 +multi_line_output=3 +length_sort=true +known_standard_library=numpy,setuptools +known_myself=torchreid +known_third_party=matplotlib,cv2,torch,torchvision,PIL,yacs +no_lines_before=STDLIB,THIRDPARTY +sections=FUTURE,STDLIB,THIRDPARTY,myself,FIRSTPARTY,LOCALFOLDER +default_section=FIRSTPARTY \ No newline at end of file diff --git a/strong_sort/deep/reid/.style.yapf b/strong_sort/deep/reid/.style.yapf new file mode 100644 index 0000000..29d8e52 --- /dev/null +++ b/strong_sort/deep/reid/.style.yapf @@ -0,0 +1,7 @@ +[style] +BASED_ON_STYLE = pep8 +BLANK_LINE_BEFORE_NESTED_CLASS_OR_DEF = true +SPLIT_BEFORE_EXPRESSION_AFTER_OPENING_PAREN = true +DEDENT_CLOSING_BRACKETS = true +SPACES_BEFORE_COMMENT = 1 +ARITHMETIC_PRECEDENCE_INDICATION = true \ No newline at end of file diff --git a/strong_sort/deep/reid/LICENSE b/strong_sort/deep/reid/LICENSE new file mode 100644 index 0000000..d2bcb88 --- /dev/null +++ b/strong_sort/deep/reid/LICENSE @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2018 Kaiyang Zhou + +Permission is hereby granted, free of charge, to any person obtaining a copy +of this software and associated documentation files (the "Software"), to deal +in the Software without restriction, including without limitation the rights +to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +SOFTWARE. diff --git a/strong_sort/deep/reid/README.rst b/strong_sort/deep/reid/README.rst new file mode 100644 index 0000000..57be7a8 --- /dev/null +++ b/strong_sort/deep/reid/README.rst @@ -0,0 +1,317 @@ +Torchreid +=========== +Torchreid is a library for deep-learning person re-identification, written in `PyTorch `_ and developed for our ICCV'19 project, `Omni-Scale Feature Learning for Person Re-Identification `_. + +It features: + +- multi-GPU training +- support both image- and video-reid +- end-to-end training and evaluation +- incredibly easy preparation of reid datasets +- multi-dataset training +- cross-dataset evaluation +- standard protocol used by most research papers +- highly extensible (easy to add models, datasets, training methods, etc.) +- implementations of state-of-the-art deep reid models +- access to pretrained reid models +- advanced training techniques +- visualization tools (tensorboard, ranks, etc.) + + +Code: https://github.com/KaiyangZhou/deep-person-reid. + +Documentation: https://kaiyangzhou.github.io/deep-person-reid/. + +How-to instructions: https://kaiyangzhou.github.io/deep-person-reid/user_guide. + +Model zoo: https://kaiyangzhou.github.io/deep-person-reid/MODEL_ZOO. + +Tech report: https://arxiv.org/abs/1910.10093. + +You can find some research projects that are built on top of Torchreid `here `_. + + +What's new +------------ +- [Aug 2021] We have released the ImageNet-pretrained models of ``osnet_ain_x0_75``, ``osnet_ain_x0_5`` and ``osnet_ain_x0_25``. The pretraining setup follows `pycls `_. +- [Apr 2021] We have updated the appendix in the `TPAMI version of OSNet `_ to include results in the multi-source domain generalization setting. The trained models can be found in the `Model Zoo `_. +- [Apr 2021] We have added a script to automate the process of calculating average results over multiple splits. For more details please see ``tools/parse_test_res.py``. +- [Apr 2021] ``v1.4.0``: We added the person search dataset, `CUHK-SYSU `_. Please see the `documentation `_ regarding how to download the dataset (it contains cropped person images). +- [Apr 2021] All models in the model zoo have been moved to google drive. Please raise an issue if any model's performance is inconsistent with the numbers shown in the model zoo page (could be caused by wrong links). +- [Mar 2021] `OSNet `_ will appear in the TPAMI journal! Compared with the conference version, which focuses on discriminative feature learning using the omni-scale building block, this journal extension further considers generalizable feature learning by integrating `instance normalization layers `_ with the OSNet architecture. We hope this journal paper can motivate more future work to taclke the generalization issue in cross-dataset re-ID. +- [Mar 2021] Generalization across domains (datasets) in person re-ID is crucial in real-world applications, which is closely related to the topic of *domain generalization*. Interested in learning how the field of domain generalization has developed over the last decade? Check our recent survey in this topic at https://arxiv.org/abs/2103.02503, with coverage on the history, datasets, related problems, methodologies, potential directions, and so on (*methods designed for generalizable re-ID are also covered*!). +- [Feb 2021] ``v1.3.6`` Added `University-1652 `_, a new dataset for multi-view multi-source geo-localization (credit to `Zhedong Zheng `_). +- [Feb 2021] ``v1.3.5``: Now the `cython code `_ works on Windows (credit to `lablabla `_). +- [Jan 2021] Our recent work, `MixStyle `_ (mixing instance-level feature statistics of samples of different domains for improving domain generalization), has been accepted to ICLR'21. The code has been released at https://github.com/KaiyangZhou/mixstyle-release where the person re-ID part is based on Torchreid. +- [Jan 2021] A new evaluation metric called `mean Inverse Negative Penalty (mINP)` for person re-ID has been introduced in `Deep Learning for Person Re-identification: A Survey and Outlook (TPAMI 2021) `_. Their code can be accessed at ``_. +- [Aug 2020] ``v1.3.3``: Fixed bug in ``visrank`` (caused by not unpacking ``dsetid``). +- [Aug 2020] ``v1.3.2``: Added ``_junk_pids`` to ``grid`` and ``prid``. This avoids using mislabeled gallery images for training when setting ``combineall=True``. +- [Aug 2020] ``v1.3.0``: (1) Added ``dsetid`` to the existing 3-tuple data source, resulting in ``(impath, pid, camid, dsetid)``. This variable denotes the dataset ID and is useful when combining multiple datasets for training (as a dataset indicator). E.g., when combining ``market1501`` and ``cuhk03``, the former will be assigned ``dsetid=0`` while the latter will be assigned ``dsetid=1``. (2) Added ``RandomDatasetSampler``. Analogous to ``RandomDomainSampler``, ``RandomDatasetSampler`` samples a certain number of images (``batch_size // num_datasets``) from each of specified datasets (the amount is determined by ``num_datasets``). +- [Aug 2020] ``v1.2.6``: Added ``RandomDomainSampler`` (it samples ``num_cams`` cameras each with ``batch_size // num_cams`` images to form a mini-batch). +- [Jun 2020] ``v1.2.5``: (1) Dataloader's output from ``__getitem__`` has been changed from ``list`` to ``dict``. Previously, an element, e.g. image tensor, was fetched with ``imgs=data[0]``. Now it should be obtained by ``imgs=data['img']``. See this `commit `_ for detailed changes. (2) Added ``k_tfm`` as an option to image data loader, which allows data augmentation to be applied ``k_tfm`` times *independently* to an image. If ``k_tfm > 1``, ``imgs=data['img']`` returns a list with ``k_tfm`` image tensors. +- [May 2020] Added the person attribute recognition code used in `Omni-Scale Feature Learning for Person Re-Identification (ICCV'19) `_. See ``projects/attribute_recognition/``. +- [May 2020] ``v1.2.1``: Added a simple API for feature extraction (``torchreid/utils/feature_extractor.py``). See the `documentation `_ for the instruction. +- [Apr 2020] Code for reproducing the experiments of `deep mutual learning `_ in the `OSNet paper `__ (Supp. B) has been released at ``projects/DML``. +- [Apr 2020] Upgraded to ``v1.2.0``. The engine class has been made more model-agnostic to improve extensibility. See `Engine `_ and `ImageSoftmaxEngine `_ for more details. Credit to `Dassl.pytorch `_. +- [Dec 2019] Our `OSNet paper `_ has been updated, with additional experiments (in section B of the supplementary) showing some useful techniques for improving OSNet's performance in practice. +- [Nov 2019] ``ImageDataManager`` can load training data from target datasets by setting ``load_train_targets=True``, and the train-loader can be accessed with ``train_loader_t = datamanager.train_loader_t``. This feature is useful for domain adaptation research. + + +Installation +--------------- + +Make sure `conda `_ is installed. + + +.. code-block:: bash + + # cd to your preferred directory and clone this repo + git clone https://github.com/KaiyangZhou/deep-person-reid.git + + # create environment + cd deep-person-reid/ + conda create --name torchreid python=3.7 + conda activate torchreid + + # install dependencies + # make sure `which python` and `which pip` point to the correct path + pip install -r requirements.txt + + # install torch and torchvision (select the proper cuda version to suit your machine) + conda install pytorch torchvision cudatoolkit=9.0 -c pytorch + + # install torchreid (don't need to re-build it if you modify the source code) + python setup.py develop + + +Get started: 30 seconds to Torchreid +------------------------------------- +1. Import ``torchreid`` + +.. code-block:: python + + import torchreid + +2. Load data manager + +.. code-block:: python + + datamanager = torchreid.data.ImageDataManager( + root="reid-data", + sources="market1501", + targets="market1501", + height=256, + width=128, + batch_size_train=32, + batch_size_test=100, + transforms=["random_flip", "random_crop"] + ) + +3 Build model, optimizer and lr_scheduler + +.. code-block:: python + + model = torchreid.models.build_model( + name="resnet50", + num_classes=datamanager.num_train_pids, + loss="softmax", + pretrained=True + ) + + model = model.cuda() + + optimizer = torchreid.optim.build_optimizer( + model, + optim="adam", + lr=0.0003 + ) + + scheduler = torchreid.optim.build_lr_scheduler( + optimizer, + lr_scheduler="single_step", + stepsize=20 + ) + +4. Build engine + +.. code-block:: python + + engine = torchreid.engine.ImageSoftmaxEngine( + datamanager, + model, + optimizer=optimizer, + scheduler=scheduler, + label_smooth=True + ) + +5. Run training and test + +.. code-block:: python + + engine.run( + save_dir="log/resnet50", + max_epoch=60, + eval_freq=10, + print_freq=10, + test_only=False + ) + + +A unified interface +----------------------- +In "deep-person-reid/scripts/", we provide a unified interface to train and test a model. See "scripts/main.py" and "scripts/default_config.py" for more details. The folder "configs/" contains some predefined configs which you can use as a starting point. + +Below we provide an example to train and test `OSNet (Zhou et al. ICCV'19) `_. Assume :code:`PATH_TO_DATA` is the directory containing reid datasets. The environmental variable :code:`CUDA_VISIBLE_DEVICES` is omitted, which you need to specify if you have a pool of gpus and want to use a specific set of them. + +Conventional setting +^^^^^^^^^^^^^^^^^^^^^ + +To train OSNet on Market1501, do + +.. code-block:: bash + + python scripts/main.py \ + --config-file configs/im_osnet_x1_0_softmax_256x128_amsgrad_cosine.yaml \ + --transforms random_flip random_erase \ + --root $PATH_TO_DATA + + +The config file sets Market1501 as the default dataset. If you wanna use DukeMTMC-reID, do + +.. code-block:: bash + + python scripts/main.py \ + --config-file configs/im_osnet_x1_0_softmax_256x128_amsgrad_cosine.yaml \ + -s dukemtmcreid \ + -t dukemtmcreid \ + --transforms random_flip random_erase \ + --root $PATH_TO_DATA \ + data.save_dir log/osnet_x1_0_dukemtmcreid_softmax_cosinelr + +The code will automatically (download and) load the ImageNet pretrained weights. After the training is done, the model will be saved as "log/osnet_x1_0_market1501_softmax_cosinelr/model.pth.tar-250". Under the same folder, you can find the `tensorboard `_ file. To visualize the learning curves using tensorboard, you can run :code:`tensorboard --logdir=log/osnet_x1_0_market1501_softmax_cosinelr` in the terminal and visit :code:`http://localhost:6006/` in your web browser. + +Evaluation is automatically performed at the end of training. To run the test again using the trained model, do + +.. code-block:: bash + + python scripts/main.py \ + --config-file configs/im_osnet_x1_0_softmax_256x128_amsgrad_cosine.yaml \ + --root $PATH_TO_DATA \ + model.load_weights log/osnet_x1_0_market1501_softmax_cosinelr/model.pth.tar-250 \ + test.evaluate True + + +Cross-domain setting +^^^^^^^^^^^^^^^^^^^^^ + +Suppose you wanna train OSNet on DukeMTMC-reID and test its performance on Market1501, you can do + +.. code-block:: bash + + python scripts/main.py \ + --config-file configs/im_osnet_x1_0_softmax_256x128_amsgrad.yaml \ + -s dukemtmcreid \ + -t market1501 \ + --transforms random_flip color_jitter \ + --root $PATH_TO_DATA + +Here we only test the cross-domain performance. However, if you also want to test the performance on the source dataset, i.e. DukeMTMC-reID, you can set :code:`-t dukemtmcreid market1501`, which will evaluate the model on the two datasets separately. + +Different from the same-domain setting, here we replace :code:`random_erase` with :code:`color_jitter`. This can improve the generalization performance on the unseen target dataset. + +Pretrained models are available in the `Model Zoo `_. + + +Datasets +-------- + +Image-reid datasets +^^^^^^^^^^^^^^^^^^^^^ +- `Market1501 `_ +- `CUHK03 `_ +- `DukeMTMC-reID `_ +- `MSMT17 `_ +- `VIPeR `_ +- `GRID `_ +- `CUHK01 `_ +- `SenseReID `_ +- `QMUL-iLIDS `_ +- `PRID `_ + +Geo-localization datasets +^^^^^^^^^^^^^^^^^^^^^^^^^^^ +- `University-1652 `_ + +Video-reid datasets +^^^^^^^^^^^^^^^^^^^^^^^ +- `MARS `_ +- `iLIDS-VID `_ +- `PRID2011 `_ +- `DukeMTMC-VideoReID `_ + + +Models +------- + +ImageNet classification models +^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +- `ResNet `_ +- `ResNeXt `_ +- `SENet `_ +- `DenseNet `_ +- `Inception-ResNet-V2 `_ +- `Inception-V4 `_ +- `Xception `_ +- `IBN-Net `_ + +Lightweight models +^^^^^^^^^^^^^^^^^^^ +- `NASNet `_ +- `MobileNetV2 `_ +- `ShuffleNet `_ +- `ShuffleNetV2 `_ +- `SqueezeNet `_ + +ReID-specific models +^^^^^^^^^^^^^^^^^^^^^^ +- `MuDeep `_ +- `ResNet-mid `_ +- `HACNN `_ +- `PCB `_ +- `MLFN `_ +- `OSNet `_ +- `OSNet-AIN `_ + + +Useful links +------------- +- `OSNet-IBN1-Lite (test-only code with lite docker container) `_ +- `Deep Learning for Person Re-identification: A Survey and Outlook `_ + + +Citation +--------- +If you use this code or the models in your research, please give credit to the following papers: + +.. code-block:: bash + + @article{torchreid, + title={Torchreid: A Library for Deep Learning Person Re-Identification in Pytorch}, + author={Zhou, Kaiyang and Xiang, Tao}, + journal={arXiv preprint arXiv:1910.10093}, + year={2019} + } + + @inproceedings{zhou2019osnet, + title={Omni-Scale Feature Learning for Person Re-Identification}, + author={Zhou, Kaiyang and Yang, Yongxin and Cavallaro, Andrea and Xiang, Tao}, + booktitle={ICCV}, + year={2019} + } + + @article{zhou2021osnet, + title={Learning Generalisable Omni-Scale Representations for Person Re-Identification}, + author={Zhou, Kaiyang and Yang, Yongxin and Cavallaro, Andrea and Xiang, Tao}, + journal={TPAMI}, + year={2021} + } diff --git a/strong_sort/deep/reid/configs/im_osnet_ain_x1_0_softmax_256x128_amsgrad_cosine.yaml b/strong_sort/deep/reid/configs/im_osnet_ain_x1_0_softmax_256x128_amsgrad_cosine.yaml new file mode 100644 index 0000000..c8b63ec --- /dev/null +++ b/strong_sort/deep/reid/configs/im_osnet_ain_x1_0_softmax_256x128_amsgrad_cosine.yaml @@ -0,0 +1,35 @@ +model: + name: 'osnet_ain_x1_0' + pretrained: True + +data: + type: 'image' + sources: ['market1501'] + targets: ['market1501', 'dukemtmcreid'] + height: 256 + width: 128 + combineall: False + transforms: ['random_flip', 'color_jitter'] + save_dir: 'log/osnet_ain_x1_0_market1501_softmax_cosinelr' + +loss: + name: 'softmax' + softmax: + label_smooth: True + +train: + optim: 'amsgrad' + lr: 0.0015 + max_epoch: 100 + batch_size: 64 + fixbase_epoch: 10 + open_layers: ['classifier'] + lr_scheduler: 'cosine' + +test: + batch_size: 300 + dist_metric: 'cosine' + normalize_feature: False + evaluate: False + eval_freq: -1 + rerank: False \ No newline at end of file diff --git a/strong_sort/deep/reid/configs/im_osnet_ibn_x1_0_softmax_256x128_amsgrad.yaml b/strong_sort/deep/reid/configs/im_osnet_ibn_x1_0_softmax_256x128_amsgrad.yaml new file mode 100644 index 0000000..b4f8903 --- /dev/null +++ b/strong_sort/deep/reid/configs/im_osnet_ibn_x1_0_softmax_256x128_amsgrad.yaml @@ -0,0 +1,36 @@ +model: + name: 'osnet_ibn_x1_0' + pretrained: True + +data: + type: 'image' + sources: ['market1501'] + targets: ['dukemtmcreid'] + height: 256 + width: 128 + combineall: False + transforms: ['random_flip', 'color_jitter'] + save_dir: 'log/osnet_ibn_x1_0_market2duke_softmax' + +loss: + name: 'softmax' + softmax: + label_smooth: True + +train: + optim: 'amsgrad' + lr: 0.0015 + max_epoch: 150 + batch_size: 64 + fixbase_epoch: 10 + open_layers: ['classifier'] + lr_scheduler: 'single_step' + stepsize: [60] + +test: + batch_size: 300 + dist_metric: 'euclidean' + normalize_feature: False + evaluate: False + eval_freq: -1 + rerank: False \ No newline at end of file diff --git a/strong_sort/deep/reid/configs/im_osnet_x0_25_softmax_256x128_amsgrad.yaml b/strong_sort/deep/reid/configs/im_osnet_x0_25_softmax_256x128_amsgrad.yaml new file mode 100644 index 0000000..d256a6b --- /dev/null +++ b/strong_sort/deep/reid/configs/im_osnet_x0_25_softmax_256x128_amsgrad.yaml @@ -0,0 +1,36 @@ +model: + name: 'osnet_x0_25' + pretrained: True + +data: + type: 'image' + sources: ['market1501'] + targets: ['market1501'] + height: 256 + width: 128 + combineall: False + transforms: ['random_flip'] + save_dir: 'log/osnet_x0_25_market1501_softmax' + +loss: + name: 'softmax' + softmax: + label_smooth: True + +train: + optim: 'amsgrad' + lr: 0.003 + max_epoch: 180 + batch_size: 128 + fixbase_epoch: 10 + open_layers: ['classifier'] + lr_scheduler: 'single_step' + stepsize: [80] + +test: + batch_size: 300 + dist_metric: 'euclidean' + normalize_feature: False + evaluate: False + eval_freq: -1 + rerank: False \ No newline at end of file diff --git a/strong_sort/deep/reid/configs/im_osnet_x0_5_softmax_256x128_amsgrad.yaml b/strong_sort/deep/reid/configs/im_osnet_x0_5_softmax_256x128_amsgrad.yaml new file mode 100644 index 0000000..467305f --- /dev/null +++ b/strong_sort/deep/reid/configs/im_osnet_x0_5_softmax_256x128_amsgrad.yaml @@ -0,0 +1,36 @@ +model: + name: 'osnet_x0_5' + pretrained: True + +data: + type: 'image' + sources: ['market1501'] + targets: ['market1501'] + height: 256 + width: 128 + combineall: False + transforms: ['random_flip'] + save_dir: 'log/osnet_x0_5_market1501_softmax' + +loss: + name: 'softmax' + softmax: + label_smooth: True + +train: + optim: 'amsgrad' + lr: 0.003 + max_epoch: 180 + batch_size: 128 + fixbase_epoch: 10 + open_layers: ['classifier'] + lr_scheduler: 'single_step' + stepsize: [80] + +test: + batch_size: 300 + dist_metric: 'euclidean' + normalize_feature: False + evaluate: False + eval_freq: -1 + rerank: False \ No newline at end of file diff --git a/strong_sort/deep/reid/configs/im_osnet_x0_75_softmax_256x128_amsgrad.yaml b/strong_sort/deep/reid/configs/im_osnet_x0_75_softmax_256x128_amsgrad.yaml new file mode 100644 index 0000000..04203fd --- /dev/null +++ b/strong_sort/deep/reid/configs/im_osnet_x0_75_softmax_256x128_amsgrad.yaml @@ -0,0 +1,36 @@ +model: + name: 'osnet_x0_75' + pretrained: True + +data: + type: 'image' + sources: ['market1501'] + targets: ['market1501'] + height: 256 + width: 128 + combineall: False + transforms: ['random_flip'] + save_dir: 'log/osnet_x0_75_market1501_softmax' + +loss: + name: 'softmax' + softmax: + label_smooth: True + +train: + optim: 'amsgrad' + lr: 0.0015 + max_epoch: 150 + batch_size: 64 + fixbase_epoch: 10 + open_layers: ['classifier'] + lr_scheduler: 'single_step' + stepsize: [60] + +test: + batch_size: 300 + dist_metric: 'euclidean' + normalize_feature: False + evaluate: False + eval_freq: -1 + rerank: False \ No newline at end of file diff --git a/strong_sort/deep/reid/configs/im_osnet_x1_0_softmax_256x128_amsgrad.yaml b/strong_sort/deep/reid/configs/im_osnet_x1_0_softmax_256x128_amsgrad.yaml new file mode 100644 index 0000000..f1c970f --- /dev/null +++ b/strong_sort/deep/reid/configs/im_osnet_x1_0_softmax_256x128_amsgrad.yaml @@ -0,0 +1,36 @@ +model: + name: 'osnet_x1_0' + pretrained: True + +data: + type: 'image' + sources: ['market1501'] + targets: ['market1501'] + height: 256 + width: 128 + combineall: False + transforms: ['random_flip'] + save_dir: 'log/osnet_x1_0_market1501_softmax' + +loss: + name: 'softmax' + softmax: + label_smooth: True + +train: + optim: 'amsgrad' + lr: 0.0015 + max_epoch: 150 + batch_size: 64 + fixbase_epoch: 10 + open_layers: ['classifier'] + lr_scheduler: 'single_step' + stepsize: [60] + +test: + batch_size: 300 + dist_metric: 'euclidean' + normalize_feature: False + evaluate: False + eval_freq: -1 + rerank: False \ No newline at end of file diff --git a/strong_sort/deep/reid/configs/im_osnet_x1_0_softmax_256x128_amsgrad_cosine.yaml b/strong_sort/deep/reid/configs/im_osnet_x1_0_softmax_256x128_amsgrad_cosine.yaml new file mode 100644 index 0000000..29750bc --- /dev/null +++ b/strong_sort/deep/reid/configs/im_osnet_x1_0_softmax_256x128_amsgrad_cosine.yaml @@ -0,0 +1,35 @@ +model: + name: 'osnet_x1_0' + pretrained: True + +data: + type: 'image' + sources: ['market1501'] + targets: ['market1501'] + height: 256 + width: 128 + combineall: False + transforms: ['random_flip'] + save_dir: 'log/osnet_x1_0_market1501_softmax_cosinelr' + +loss: + name: 'softmax' + softmax: + label_smooth: True + +train: + optim: 'amsgrad' + lr: 0.0015 + max_epoch: 250 + batch_size: 64 + fixbase_epoch: 10 + open_layers: ['classifier'] + lr_scheduler: 'cosine' + +test: + batch_size: 300 + dist_metric: 'euclidean' + normalize_feature: False + evaluate: False + eval_freq: -1 + rerank: False \ No newline at end of file diff --git a/strong_sort/deep/reid/configs/im_r50_softmax_256x128_amsgrad.yaml b/strong_sort/deep/reid/configs/im_r50_softmax_256x128_amsgrad.yaml new file mode 100644 index 0000000..6173589 --- /dev/null +++ b/strong_sort/deep/reid/configs/im_r50_softmax_256x128_amsgrad.yaml @@ -0,0 +1,36 @@ +model: + name: 'resnet50_fc512' + pretrained: True + +data: + type: 'image' + sources: ['market1501'] + targets: ['market1501'] + height: 256 + width: 128 + combineall: False + transforms: ['random_flip'] + save_dir: 'log/resnet50_market1501_softmax' + +loss: + name: 'softmax' + softmax: + label_smooth: True + +train: + optim: 'amsgrad' + lr: 0.0003 + max_epoch: 60 + batch_size: 32 + fixbase_epoch: 5 + open_layers: ['classifier'] + lr_scheduler: 'single_step' + stepsize: [20] + +test: + batch_size: 100 + dist_metric: 'euclidean' + normalize_feature: False + evaluate: False + eval_freq: -1 + rerank: False \ No newline at end of file diff --git a/strong_sort/deep/reid/configs/im_r50fc512_softmax_256x128_amsgrad.yaml b/strong_sort/deep/reid/configs/im_r50fc512_softmax_256x128_amsgrad.yaml new file mode 100644 index 0000000..5ec2b70 --- /dev/null +++ b/strong_sort/deep/reid/configs/im_r50fc512_softmax_256x128_amsgrad.yaml @@ -0,0 +1,36 @@ +model: + name: 'resnet50_fc512' + pretrained: True + +data: + type: 'image' + sources: ['market1501'] + targets: ['market1501'] + height: 256 + width: 128 + combineall: False + transforms: ['random_flip'] + save_dir: 'log/resnet50_fc512_market1501_softmax' + +loss: + name: 'softmax' + softmax: + label_smooth: True + +train: + optim: 'amsgrad' + lr: 0.0003 + max_epoch: 60 + batch_size: 32 + fixbase_epoch: 5 + open_layers: ['fc', 'classifier'] + lr_scheduler: 'single_step' + stepsize: [20] + +test: + batch_size: 100 + dist_metric: 'euclidean' + normalize_feature: False + evaluate: False + eval_freq: -1 + rerank: False \ No newline at end of file diff --git a/strong_sort/deep/reid/docs/AWESOME_REID.md b/strong_sort/deep/reid/docs/AWESOME_REID.md new file mode 100644 index 0000000..3121535 --- /dev/null +++ b/strong_sort/deep/reid/docs/AWESOME_REID.md @@ -0,0 +1,69 @@ +# Awesome-ReID +Here is a collection of ReID-related research with links to papers and code. You are welcome to submit [PR](https://help.github.com/articles/creating-a-pull-request/)s if you find something missing. + + +- [TPAMI21] Learning Generalisable Omni-Scale Representations for Person Re-Identification [[paper](https://arxiv.org/abs/1910.06827)][[code](https://github.com/KaiyangZhou/deep-person-reid)] + +- [TPAMI21] Deep Learning for Person Re-identification: A Survey and Outlook [[paper](https://arxiv.org/abs/2001.04193)] [[code](https://github.com/mangye16/ReID-Survey)] + +- [ICCV19] RGB-Infrared Cross-Modality Person Re-Identification via Joint Pixel and Feature Alignment. [[paper](http://openaccess.thecvf.com/content_ICCV_2019/papers/Wang_RGB-Infrared_Cross-Modality_Person_Re-Identification_via_Joint_Pixel_and_Feature_Alignment_ICCV_2019_paper.pdf)] [[code](https://github.com/wangguanan/AlignGAN)] + +- [ICCV19] Unsupervised Graph Association for Person Re-identification. [[paper](https://github.com/yichuan9527/Unsupervised-Graph-Association-for-Person-Re-identification)] [[code](https://github.com/yichuan9527/Unsupervised-Graph-Association-for-Person-Re-identification)] + +- [ICCV19] Self-similarity Grouping: A Simple Unsupervised Cross Domain Adaptation Approach for Person Re-identification. [[paper](http://openaccess.thecvf.com/content_ICCV_2019/papers/Fu_Self-Similarity_Grouping_A_Simple_Unsupervised_Cross_Domain_Adaptation_Approach_for_ICCV_2019_paper.pdf)] [[code](https://github.com/OasisYang/SSG)] + +- [ICCV19] Spectral Feature Transformation for Person Re-Identification. [[paper](http://openaccess.thecvf.com/content_ICCV_2019/papers/Luo_Spectral_Feature_Transformation_for_Person_Re-Identification_ICCV_2019_paper.pdf)] [[code](https://github.com/LuckyDC/SFT_REID)] + +- [ICCV19] Beyond Human Parts: Dual Part-Aligned Representations for Person Re-Identification. [[paper](http://openaccess.thecvf.com/content_ICCV_2019/papers/Guo_Beyond_Human_Parts_Dual_Part-Aligned_Representations_for_Person_Re-Identification_ICCV_2019_paper.pdf)] [[code](https://github.com/ggjy/P2Net.pytorch)] + +- [ICCV19] Co-segmentation Inspired Attention Networks for Video-based Person Re-identification. [[paper](http://openaccess.thecvf.com/content_ICCV_2019/papers/Subramaniam_Co-Segmentation_Inspired_Attention_Networks_for_Video-Based_Person_Re-Identification_ICCV_2019_paper.pdf)][[code](https://github.com/InnovArul/vidreid_cosegmentation)] + +- [ICCV19] Mixed High-Order Attention Network for Person Re-Identification. [[paper](https://arxiv.org/abs/1908.05819)][[code](https://github.com/chenbinghui1/MHN)] + +- [ICCV19] ABD-Net: Attentive but Diverse Person Re-Identification. [[paper](https://arxiv.org/abs/1908.01114)] [[code](https://github.com/TAMU-VITA/ABD-Net)] + +- [ICCV19] Omni-Scale Feature Learning for Person Re-Identification. [[paper](https://arxiv.org/abs/1905.00953)] [[code](https://github.com/KaiyangZhou/deep-person-reid)] + +- [CVPR19] Joint Discriminative and Generative Learning for Person Re-identification. [[paper](https://arxiv.org/abs/1904.07223)][[code](https://github.com/NVlabs/DG-Net)] +- [CVPR19] Invariance Matters: Exemplar Memory for Domain Adaptive Person Re-identification. [[paper](https://arxiv.org/abs/1904.01990)][[code](https://github.com/zhunzhong07/ECN)] +- [CVPR19] Dissecting Person Re-identification from the Viewpoint of Viewpoint. [[paper](https://arxiv.org/abs/1812.02162)][[code](https://github.com/sxzrt/Dissecting-Person-Re-ID-from-the-Viewpoint-of-Viewpoint)] +- [CVPR19] Unsupervised Person Re-identification by Soft Multilabel Learning. [[paper](https://arxiv.org/abs/1903.06325)][[code](https://github.com/KovenYu/MAR)] +- [CVPR19] Patch-based Discriminative Feature Learning for Unsupervised Person Re-identification. [[paper](https://kovenyu.com/publication/2019-cvpr-pedal/)][[code](https://github.com/QizeYang/PAUL)] + +- [AAAI19] Spatial and Temporal Mutual Promotion for Video-based Person Re-identification. [[paper](https://arxiv.org/abs/1812.10305)][[code](https://github.com/yolomax/person-reid-lib)] + +- [AAAI19] Spatial-Temporal Person Re-identification. [[paper](https://arxiv.org/abs/1812.03282)][[code](https://github.com/Wanggcong/Spatial-Temporal-Re-identification)] + +- [AAAI19] Horizontal Pyramid Matching for Person Re-identification. [[paper](https://arxiv.org/abs/1804.05275)][[code](https://github.com/OasisYang/HPM)] + +- [AAAI19] Backbone Can Not be Trained at Once: Rolling Back to Pre-trained Network for Person Re-identification. [[paper](https://arxiv.org/abs/1901.06140)][[code](https://github.com/youngminPIL/rollback)] + +- [AAAI19] A Bottom-Up Clustering Approach to Unsupervised Person Re-identification. [[paper](https://vana77.github.io/vana77.github.io/images/AAAI19.pdf)][[code](https://github.com/vana77/Bottom-up-Clustering-Person-Re-identification)] + +- [NIPS18] FD-GAN: Pose-guided Feature Distilling GAN for Robust Person Re-identification. [[paper](https://arxiv.org/abs/1810.02936)][[code](https://github.com/yxgeee/FD-GAN)] + +- [ECCV18] Generalizing A Person Retrieval Model Hetero- and Homogeneously. [[paper](http://openaccess.thecvf.com/content_ECCV_2018/papers/Zhun_Zhong_Generalizing_A_Person_ECCV_2018_paper.pdf)][[code](https://github.com/zhunzhong07/HHL)] + +- [ECCV18] Pose-Normalized Image Generation for Person Re-identification. [[paper](https://arxiv.org/abs/1712.02225)][[code](https://github.com/naiq/PN_GAN)] + +- [CVPR18] Camera Style Adaptation for Person Re-Identification. [[paper](https://arxiv.org/abs/1711.10295)][[code](https://github.com/zhunzhong07/CamStyle)] + +- [CVPR18] Deep Group-Shuffling Random Walk for Person Re-Identification. [[paper](https://arxiv.org/abs/1807.11178)][[code](https://github.com/YantaoShen/kpm_rw_person_reid)] + +- [CVPR18] End-to-End Deep Kronecker-Product Matching for Person Re-identification. [[paper](https://arxiv.org/abs/1807.11182)][[code](https://github.com/YantaoShen/kpm_rw_person_reid)] + +- [CVPR18] Features for Multi-Target Multi-Camera Tracking and Re-Identification. [[paper](https://arxiv.org/abs/1803.10859)][[code](https://github.com/ergysr/DeepCC)] + +- [CVPR18] Group Consistent Similarity Learning via Deep CRF for Person Re-Identification. [[paper](http://openaccess.thecvf.com/content_cvpr_2018/papers/Chen_Group_Consistent_Similarity_CVPR_2018_paper.pdf)][[code](https://github.com/dapengchen123/crf_affinity)] + +- [CVPR18] Harmonious Attention Network for Person Re-Identification. [[paper](https://arxiv.org/abs/1802.08122)][[code](https://github.com/KaiyangZhou/deep-person-reid)] + +- [CVPR18] Human Semantic Parsing for Person Re-Identification. [[paper](https://arxiv.org/abs/1804.00216)][[code](https://github.com/emrahbasaran/SPReID)] + +- [CVPR18] Multi-Level Factorisation Net for Person Re-Identification. [[paper](https://arxiv.org/abs/1803.09132)][[code](https://github.com/KaiyangZhou/deep-person-reid)] + +- [CVPR18] Resource Aware Person Re-identification across Multiple Resolutions. [[paper](https://arxiv.org/abs/1805.08805)][[code](https://github.com/mileyan/DARENet)] + +- [CVPR18] Exploit the Unknown Gradually: One-Shot Video-Based Person Re-Identification by Stepwise Learning. [[paper](https://yu-wu.net/pdf/CVPR2018_Exploit-Unknown-Gradually.pdf)][[code](https://github.com/Yu-Wu/Exploit-Unknown-Gradually)] + +- [ArXiv18] Revisiting Temporal Modeling for Video-based Person ReID. [[paper](https://arxiv.org/abs/1805.02104)][[code](https://github.com/jiyanggao/Video-Person-ReID)] diff --git a/strong_sort/deep/reid/docs/MODEL_ZOO.md b/strong_sort/deep/reid/docs/MODEL_ZOO.md new file mode 100644 index 0000000..8a9306f --- /dev/null +++ b/strong_sort/deep/reid/docs/MODEL_ZOO.md @@ -0,0 +1,93 @@ +# Model Zoo + +- Results are presented in the format of **. +- When computing model size and FLOPs, only layers that are used at test time are considered (see `torchreid.utils.compute_model_complexity`). +- Asterisk (\*) means the model is trained from scratch. +- `combineall=True` means all images in the dataset are used for model training. +- Why not use heavy data augmentation like [random erasing](https://arxiv.org/abs/1708.04896) for model training? It's because heavy data augmentation might harm the cross-dataset generalization performance (see [this paper](https://arxiv.org/abs/1708.04896)). + + +## ImageNet pretrained models + + +| Model | Download | +| :--- | :---: | +| shufflenet | [model](https://drive.google.com/file/d/1RFnYcHK1TM-yt3yLsNecaKCoFO4Yb6a-/view?usp=sharing) | +| mobilenetv2_x1_0 | [model](https://drive.google.com/file/d/1K7_CZE_L_Tf-BRY6_vVm0G-0ZKjVWh3R/view?usp=sharing) | +| mobilenetv2_x1_4 | [model](https://drive.google.com/file/d/10c0ToIGIVI0QZTx284nJe8QfSJl5bIta/view?usp=sharing) | +| mlfn | [model](https://drive.google.com/file/d/1PP8Eygct5OF4YItYRfA3qypYY9xiqHuV/view?usp=sharing) | +| osnet_x1_0 | [model](https://drive.google.com/file/d/1LaG1EJpHrxdAxKnSCJ_i0u-nbxSAeiFY/view?usp=sharing) | +| osnet_x0_75 | [model](https://drive.google.com/file/d/1uwA9fElHOk3ZogwbeY5GkLI6QPTX70Hq/view?usp=sharing) | +| osnet_x0_5 | [model](https://drive.google.com/file/d/16DGLbZukvVYgINws8u8deSaOqjybZ83i/view?usp=sharing) | +| osnet_x0_25 | [model](https://drive.google.com/file/d/1rb8UN5ZzPKRc_xvtHlyDh-cSz88YX9hs/view?usp=sharing) | +| osnet_ibn_x1_0 | [model](https://drive.google.com/file/d/1sr90V6irlYYDd4_4ISU2iruoRG8J__6l/view?usp=sharing) | +| osnet_ain_x1_0 | [model](https://drive.google.com/file/d/1-CaioD9NaqbHK_kzSMW8VE4_3KcsRjEo/view?usp=sharing) | +| osnet_ain_x0_75 | [model](https://drive.google.com/file/d/1apy0hpsMypqstfencdH-jKIUEFOW4xoM/view?usp=sharing) | +| osnet_ain_x0_5 | [model](https://drive.google.com/file/d/1KusKvEYyKGDTUBVRxRiz55G31wkihB6l/view?usp=sharing) | +| osnet_ain_x0_25 | [model](https://drive.google.com/file/d/1SxQt2AvmEcgWNhaRb2xC4rP6ZwVDP0Wt/view?usp=sharing) | + + +## Same-domain ReID + + +| Model | # Param (10^6) | GFLOPs | Loss | Input | Transforms | Distance | market1501 | dukemtmcreid | msmt17 | +| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | +| resnet50 | 23.5 | 2.7 | softmax | (256, 128) | `random_flip`, `random_crop` | `euclidean` | [87.9 (70.4)](https://drive.google.com/file/d/1dUUZ4rHDWohmsQXCRe2C_HbYkzz94iBV/view?usp=sharing) | [78.3 (58.9)](https://drive.google.com/file/d/17ymnLglnc64NRvGOitY3BqMRS9UWd1wg/view?usp=sharing) | [63.2 (33.9)](https://drive.google.com/file/d/1ep7RypVDOthCRIAqDnn4_N-UhkkFHJsj/view?usp=sharing) | +| resnet50_fc512 | 24.6 | 4.1 | softmax | (256, 128) | `random_flip`, `random_crop` | `euclidean` | [90.8 (75.3)](https://drive.google.com/file/d/1kv8l5laX_YCdIGVCetjlNdzKIA3NvsSt/view?usp=sharing) | [81.0 (64.0)](https://drive.google.com/file/d/13QN8Mp3XH81GK4BPGXobKHKyTGH50Rtx/view?usp=sharing) | [69.6 (38.4)](https://drive.google.com/file/d/1fDJLcz4O5wxNSUvImIIjoaIF9u1Rwaud/view?usp=sharing) | +| mlfn | 32.5 | 2.8 | softmax | (256, 128) | `random_flip`, `random_crop` | `euclidean` | [90.1 (74.3)](https://drive.google.com/file/d/1wXcvhA_b1kpDfrt9s2Pma-MHxtj9pmvS/view?usp=sharing) | [81.1 (63.2)](https://drive.google.com/file/d/1rExgrTNb0VCIcOnXfMsbwSUW1h2L1Bum/view?usp=sharing) | [66.4 (37.2)](https://drive.google.com/file/d/18JzsZlJb3Wm7irCbZbZ07TN4IFKvR6p-/view?usp=sharing) | +| hacnn* | 4.5 | 0.5 | softmax | (160, 64) | `random_flip`, `random_crop` | `euclidean` | [90.9 (75.6)](https://drive.google.com/file/d/1LRKIQduThwGxMDQMiVkTScBwR7WidmYF/view?usp=sharing) | [80.1 (63.2)](https://drive.google.com/file/d/1zNm6tP4ozFUCUQ7Sv1Z98EAJWXJEhtYH/view?usp=sharing) | [64.7 (37.2)](https://drive.google.com/file/d/1MsKRtPM5WJ3_Tk2xC0aGOO7pM3VaFDNZ/view?usp=sharing) | +| mobilenetv2_x1_0 | 2.2 | 0.2 | softmax | (256, 128) | `random_flip`, `random_crop` | `euclidean` | [85.6 (67.3)](https://drive.google.com/file/d/18DgHC2ZJkjekVoqBWszD8_Xiikz-fewp/view?usp=sharing) | [74.2 (54.7)](https://drive.google.com/file/d/1q1WU2FETRJ3BXcpVtfJUuqq4z3psetds/view?usp=sharing) | [57.4 (29.3)](https://drive.google.com/file/d/1j50Hv14NOUAg7ZeB3frzfX-WYLi7SrhZ/view?usp=sharing) | +| mobilenetv2_x1_4 | 4.3 | 0.4 | softmax | (256, 128) | `random_flip`, `random_crop` | `euclidean` | [87.0 (68.5)](https://drive.google.com/file/d/1t6JCqphJG-fwwPVkRLmGGyEBhGOf2GO5/view?usp=sharing) | [76.2 (55.8)](https://drive.google.com/file/d/12uD5FeVqLg9-AFDju2L7SQxjmPb4zpBN/view?usp=sharing) | [60.1 (31.5)](https://drive.google.com/file/d/1ZY5P2Zgm-3RbDpbXM0kIBMPvspeNIbXz/view?usp=sharing) | +| osnet_x1_0 | 2.2 | 0.98 | softmax | (256, 128) | `random_flip` | `euclidean` | [94.2 (82.6)](https://drive.google.com/file/d/1vduhq5DpN2q1g4fYEZfPI17MJeh9qyrA/view?usp=sharing) | [87.0 (70.2)](https://drive.google.com/file/d/1QZO_4sNf4hdOKKKzKc-TZU9WW1v6zQbq/view?usp=sharing) | [74.9 (43.8)](https://drive.google.com/file/d/112EMUfBPYeYg70w-syK6V6Mx8-Qb9Q1M/view?usp=sharing) | +| osnet_x0_75 | 1.3 | 0.57 | softmax | (256, 128) | `random_flip` | `euclidean` | [93.7 (81.2)](https://drive.google.com/file/d/1ozRaDSQw_EQ8_93OUmjDbvLXw9TnfPer/view?usp=sharing) | [85.8 (69.8)](https://drive.google.com/file/d/1IE3KRaTPp4OUa6PGTFL_d5_KQSJbP0Or/view?usp=sharing) | [72.8 (41.4)](https://drive.google.com/file/d/1QEGO6WnJ-BmUzVPd3q9NoaO_GsPNlmWc/view?usp=sharing) | +| osnet_x0_5 | 0.6 | 0.27 | softmax | (256, 128) | `random_flip` | `euclidean` | [92.5 (79.8)](https://drive.google.com/file/d/1PLB9rgqrUM7blWrg4QlprCuPT7ILYGKT/view?usp=sharing) | [85.1 (67.4)](https://drive.google.com/file/d/1KoUVqmiST175hnkALg9XuTi1oYpqcyTu/view?usp=sharing) | [69.7 (37.5)](https://drive.google.com/file/d/1UT3AxIaDvS2PdxzZmbkLmjtiqq7AIKCv/view?usp=sharing) | +| osnet_x0_25 | 0.2 | 0.08 | softmax | (256, 128) | `random_flip` | `euclidean` | [91.2 (75.0)](https://drive.google.com/file/d/1z1UghYvOTtjx7kEoRfmqSMu-z62J6MAj/view?usp=sharing) | [82.0 (61.4)](https://drive.google.com/file/d/1eumrtiXT4NOspjyEV4j8cHmlOaaCGk5l/view?usp=sharing) | [61.4 (29.5)](https://drive.google.com/file/d/1sSwXSUlj4_tHZequ_iZ8w_Jh0VaRQMqF/view?usp=sharing) | + + +## Cross-domain ReID + +#### Market1501 -> DukeMTMC-reID + + +| Model | # Param (10^6) | GFLOPs | Loss | Input | Transforms | Distance | Rank-1 | Rank-5 | Rank-10 | mAP | Download | +| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | +| osnet_ibn_x1_0 | 2.2 | 0.98 | softmax | (256, 128) | `random_flip`, `color_jitter` | `euclidean` | 48.5 | 62.3 | 67.4 | 26.7 | [model](https://drive.google.com/file/d/1uWW7_z_IcUmRNPqQOrEBdsvic94fWH37/view?usp=sharing) | +| osnet_ain_x1_0 | 2.2 | 0.98 | softmax | (256, 128) | `random_flip`, `color_jitter` | `cosine` | 52.4 | 66.1 | 71.2 | 30.5 | [model](https://drive.google.com/file/d/14bNFGm0FhwHEkEpYKqKiDWjLNhXywFAd/view?usp=sharing) | + + +#### DukeMTMC-reID -> Market1501 + + +| Model | # Param (10^6) | GFLOPs | Loss | Input | Transforms | Distance | Rank-1 | Rank-5 | Rank-10 | mAP | Download | +| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | +| osnet_ibn_x1_0 | 2.2 | 0.98 | softmax | (256, 128) | `random_flip`, `color_jitter` | `euclidean` | 57.7 | 73.7 | 80.0 | 26.1 | [model](https://drive.google.com/file/d/1CNxL1IP0BjcE1TSttiVOID1VNipAjiF3/view?usp=sharing) | +| osnet_ain_x1_0 | 2.2 | 0.98 | softmax | (256, 128) | `random_flip`, `color_jitter` | `cosine` | 61.0 | 77.0 | 82.5 | 30.6 | [model](https://drive.google.com/file/d/1hypJvq8G04SOby6jvF337GEkg5K_bmCw/view?usp=sharing) | + + +#### MSMT17 (`combineall=True`) -> Market1501 & DukeMTMC-reID + + +| Model | # Param (10^6) | GFLOPs | Loss | Input | Transforms | Distance | msmt17 -> market1501 | msmt17 -> dukemtmcreid | Download | +| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | +| resnet50 | 23.5 | 2.7 | softmax | (256, 128) | `random_flip`, `color_jitter` | `euclidean` | 46.3 (22.8) | 52.3 (32.1) | [model](https://drive.google.com/file/d/1yiBteqgIZoOeywE8AhGmEQl7FTVwrQmf/view?usp=sharing) | +| osnet_x1_0 | 2.2 | 0.98 | softmax | (256, 128) | `random_flip`, `color_jitter` | `euclidean` | 66.6 (37.5) | 66.0 (45.3) | [model](https://drive.google.com/file/d/1IosIFlLiulGIjwW3H8uMRmx3MzPwf86x/view?usp=sharing) | +| osnet_x0_75 | 1.3 | 0.57 | softmax | (256, 128) | `random_flip`, `color_jitter` | `euclidean` | 63.6 (35.5) | 65.3 (44.5) | [model](https://drive.google.com/file/d/1fhjSS_7SUGCioIf2SWXaRGPqIY9j7-uw/view?usp=sharing) | +| osnet_x0_5 | 0.6 | 0.27 | softmax | (256, 128) | `random_flip`, `color_jitter` | `euclidean` | 64.3 (34.9) | 65.2 (43.3) | [model](https://drive.google.com/file/d/1DHgmb6XV4fwG3n-CnCM0zdL9nMsZ9_RF/view?usp=sharing) | +| osnet_x0_25 | 0.2 | 0.08 | softmax | (256, 128) | `random_flip`, `color_jitter` | `euclidean` | 59.9 (31.0) | 61.5 (39.6) | [model](https://drive.google.com/file/d/1Kkx2zW89jq_NETu4u42CFZTMVD5Hwm6e/view?usp=sharing) | +| osnet_ibn_x1_0 | 2.2 | 0.98 | softmax | (256, 128) | `random_flip`, `color_jitter` | `euclidean` | 66.5 (37.2) | 67.4 (45.6) | [model](https://drive.google.com/file/d/1q3Sj2ii34NlfxA4LvmHdWO_75NDRmECJ/view?usp=sharing) | +| osnet_ain_x1_0 | 2.2 | 0.98 | softmax | (256, 128) | `random_flip`, `color_jitter` | `cosine` | 70.1 (43.3) | 71.1 (52.7) | [model](https://drive.google.com/file/d/1SigwBE6mPdqiJMqhuIY4aqC7--5CsMal/view?usp=sharing) | + + +#### Multi-source domain generalization + +The models below are trained using multiple source datasets, as described in [Zhou et al. TPAMI'21](https://arxiv.org/abs/1910.06827). + +Regarding the abbreviations, MS is MSMT17; M is Market1501; D is DukeMTMC-reID; and C is CUHK03. + +All models were trained with [im_osnet_ain_x1_0_softmax_256x128_amsgrad_cosine.yaml](https://github.com/KaiyangZhou/deep-person-reid/blob/master/configs/im_osnet_ain_x1_0_softmax_256x128_amsgrad_cosine.yaml) and `max_epoch=50`. + +| Model | # Param (10^6) | GFLOPs | Loss | Input | Transforms | Distance | MS+D+C->M | MS+M+C->D | MS+D+M->C |D+M+C->MS | +| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | +| osnet_x1_0 | 2.2 | 0.98 | softmax | (256, 128) | `random_flip`, `color_jitter` | `cosine` | [72.5 (44.2)](https://drive.google.com/file/d/1tuYY1vQXReEd8N8_npUkc7npPDDmjNCV/view?usp=sharing) | [65.2 (47.0)](https://drive.google.com/file/d/1UxUI4NsE108UCvcy3O1Ufe73nIVPKCiu/view?usp=sharing) | [23.9 (23.3)](https://drive.google.com/file/d/1kAA6qHJvbaJtyh1b39ZyEqWROwUgWIhl/view?usp=sharing) | [33.2 (12.6)](https://drive.google.com/file/d/1wAHuYVTzj8suOwqCNcEmu6YdbVnHDvA2/view?usp=sharing) | +| osnet_ibn_x1_0 | 2.2 | 0.98 | softmax | (256, 128) | `random_flip`, `color_jitter` | `cosine` | [73.0 (44.9)](https://drive.google.com/file/d/14sH6yZwuNHPTElVoEZ26zozOOZIej5Mf/view?usp=sharing) | [64.6 (45.7)](https://drive.google.com/file/d/1Sk-2SSwKAF8n1Z4p_Lm_pl0E6v2WlIBn/view?usp=sharing) | [25.7 (25.4)](https://drive.google.com/file/d/1actHP7byqWcK4eBE1ojnspSMdo7k2W4G/view?usp=sharing) | [39.8 (16.2)](https://drive.google.com/file/d/1BGOSdLdZgqHe2qFafatb-5sPY40JlYfp/view?usp=sharing) | +| osnet_ain_x1_0 | 2.2 | 0.98 | softmax | (256, 128) | `random_flip`, `color_jitter` | `cosine` | [73.3 (45.8)](https://drive.google.com/file/d/1nIrszJVYSHf3Ej8-j6DTFdWz8EnO42PB/view?usp=sharing) | [65.6 (47.2)](https://drive.google.com/file/d/1YjJ1ZprCmaKG6MH2P9nScB9FL_Utf9t1/view?usp=sharing) | [27.4 (27.1)](https://drive.google.com/file/d/1IxIg5P0cei3KPOJQ9ZRWDE_Mdrz01ha2/view?usp=sharing) | [40.2 (16.2)](https://drive.google.com/file/d/1KcoUKzLmsUoGHI7B6as_Z2fXL50gzexS/view?usp=sharing) | diff --git a/strong_sort/deep/reid/docs/Makefile b/strong_sort/deep/reid/docs/Makefile new file mode 100644 index 0000000..298ea9e --- /dev/null +++ b/strong_sort/deep/reid/docs/Makefile @@ -0,0 +1,19 @@ +# Minimal makefile for Sphinx documentation +# + +# You can set these variables from the command line. +SPHINXOPTS = +SPHINXBUILD = sphinx-build +SOURCEDIR = . +BUILDDIR = _build + +# Put it first so that "make" without argument is like "make help". +help: + @$(SPHINXBUILD) -M help "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O) + +.PHONY: help Makefile + +# Catch-all target: route all unknown targets to Sphinx using the new +# "make mode" option. $(O) is meant as a shortcut for $(SPHINXOPTS). +%: Makefile + @$(SPHINXBUILD) -M $@ "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O) \ No newline at end of file diff --git a/strong_sort/deep/reid/docs/conf.py b/strong_sort/deep/reid/docs/conf.py new file mode 100644 index 0000000..4d27eed --- /dev/null +++ b/strong_sort/deep/reid/docs/conf.py @@ -0,0 +1,181 @@ +# -*- coding: utf-8 -*- +# +# Configuration file for the Sphinx documentation builder. +# +# This file does only contain a selection of the most common options. For a +# full list see the documentation: +# http://www.sphinx-doc.org/en/master/config + +# -- Path setup -------------------------------------------------------------- + +# If extensions (or modules to document with autodoc) are in another directory, +# add these directories to sys.path here. If the directory is relative to the +# documentation root, use os.path.abspath to make it absolute, like shown here. +# +import os +import sys + +sys.path.insert(0, os.path.abspath('..')) + +# -- Project information ----------------------------------------------------- + +project = u'torchreid' +copyright = u'2019, Kaiyang Zhou' +author = u'Kaiyang Zhou' + +version_file = '../torchreid/__init__.py' +with open(version_file, 'r') as f: + exec(compile(f.read(), version_file, 'exec')) +__version__ = locals()['__version__'] + +# The short X.Y version +version = __version__ +# The full version, including alpha/beta/rc tags +release = __version__ + +# -- General configuration --------------------------------------------------- + +# If your documentation needs a minimal Sphinx version, state it here. +# +# needs_sphinx = '1.0' + +# Add any Sphinx extension module names here, as strings. They can be +# extensions coming with Sphinx (named 'sphinx.ext.*') or your custom +# ones. +extensions = [ + 'sphinx.ext.autodoc', + 'sphinxcontrib.napoleon', + 'sphinx.ext.viewcode', + 'sphinx.ext.githubpages', + 'sphinx_markdown_tables', +] + +# Add any paths that contain templates here, relative to this directory. +templates_path = ['_templates'] + +# The suffix(es) of source filenames. +# You can specify multiple suffix as a list of string: +# +source_suffix = ['.rst', '.md'] +# source_suffix = '.rst' +source_parsers = {'.md': 'recommonmark.parser.CommonMarkParser'} + +# The master toctree document. +master_doc = 'index' + +# The language for content autogenerated by Sphinx. Refer to documentation +# for a list of supported languages. +# +# This is also used if you do content translation via gettext catalogs. +# Usually you set "language" from the command line for these cases. +language = None + +# List of patterns, relative to source directory, that match files and +# directories to ignore when looking for source files. +# This pattern also affects html_static_path and html_extra_path. +exclude_patterns = [u'_build', 'Thumbs.db', '.DS_Store'] + +# The name of the Pygments (syntax highlighting) style to use. +pygments_style = None + +# -- Options for HTML output ------------------------------------------------- + +# The theme to use for HTML and HTML Help pages. See the documentation for +# a list of builtin themes. +# +html_theme = 'sphinx_rtd_theme' + +# Theme options are theme-specific and customize the look and feel of a theme +# further. For a list of options available for each theme, see the +# documentation. +# +# html_theme_options = {} + +# Add any paths that contain custom static files (such as style sheets) here, +# relative to this directory. They are copied after the builtin static files, +# so a file named "default.css" will overwrite the builtin "default.css". +html_static_path = ['_static'] + +# Custom sidebar templates, must be a dictionary that maps document names +# to template names. +# +# The default sidebars (for documents that don't match any pattern) are +# defined by theme itself. Builtin themes are using these templates by +# default: ``['localtoc.html', 'relations.html', 'sourcelink.html', +# 'searchbox.html']``. +# +# html_sidebars = {} + +# -- Options for HTMLHelp output --------------------------------------------- + +# Output file base name for HTML help builder. +htmlhelp_basename = 'torchreiddoc' + +# -- Options for LaTeX output ------------------------------------------------ + +latex_elements = { + # The paper size ('letterpaper' or 'a4paper'). + # + # 'papersize': 'letterpaper', + + # The font size ('10pt', '11pt' or '12pt'). + # + # 'pointsize': '10pt', + + # Additional stuff for the LaTeX preamble. + # + # 'preamble': '', + + # Latex figure (float) alignment + # + # 'figure_align': 'htbp', +} + +# Grouping the document tree into LaTeX files. List of tuples +# (source start file, target name, title, +# author, documentclass [howto, manual, or own class]). +latex_documents = [ + ( + master_doc, 'torchreid.tex', u'torchreid Documentation', + u'Kaiyang Zhou', 'manual' + ), +] + +# -- Options for manual page output ------------------------------------------ + +# One entry per manual page. List of tuples +# (source start file, name, description, authors, manual section). +man_pages = [ + (master_doc, 'torchreid', u'torchreid Documentation', [author], 1) +] + +# -- Options for Texinfo output ---------------------------------------------- + +# Grouping the document tree into Texinfo files. List of tuples +# (source start file, target name, title, author, +# dir menu entry, description, category) +texinfo_documents = [ + ( + master_doc, 'torchreid', u'torchreid Documentation', author, + 'torchreid', 'One line description of project.', 'Miscellaneous' + ), +] + +# -- Options for Epub output ------------------------------------------------- + +# Bibliographic Dublin Core info. +epub_title = project + +# The unique identifier of the text. This can be a ISBN number +# or the project homepage. +# +# epub_identifier = '' + +# A unique identification for the text. +# +# epub_uid = '' + +# A list of files that should not be packed into the epub file. +epub_exclude_files = ['search.html'] + +# -- Extension configuration ------------------------------------------------- diff --git a/strong_sort/deep/reid/docs/datasets.rst b/strong_sort/deep/reid/docs/datasets.rst new file mode 100644 index 0000000..31f8611 --- /dev/null +++ b/strong_sort/deep/reid/docs/datasets.rst @@ -0,0 +1,264 @@ +.. _datasets: + +Datasets +========= + +Here we provide a comprehensive guide on how to prepare the datasets. + +Suppose you want to store the reid data in a directory called "path/to/reid-data/", you need to specify the ``root`` as *root='path/to/reid-data/'* when initializing ``DataManager``. Below we use ``$REID`` to denote "path/to/reid-data". + +Please refer to :ref:`torchreid_data` for details regarding the arguments. + + +.. note:: + Dataset with a :math:`\dagger` symbol means that the process is automated, so you can directly call the dataset in ``DataManager`` (which automatically downloads the dataset and organizes the data structure). However, we also provide a way below to help the manual setup in case the automation fails. + + +.. note:: + The keys to use specific datasets are enclosed in the parantheses beside the datasets' names. + + +.. note:: + You are suggested to use the provided names for dataset folders such as "market1501" for Market1501 and "dukemtmcreid" for DukeMTMC-reID when doing the manual setup, otherwise you need to modify the source code accordingly (i.e. the ``dataset_dir`` attribute). + +.. note:: + Some download links provided by the original authors might not work. You can email `Kaiyang Zhou `_ to reqeust new links. Please do provide your full name, institution, and purpose of using the data in the email (best use your work email address). + +.. contents:: + :local: + + +Image Datasets +-------------- + +Market1501 :math:`^\dagger` (``market1501``) +^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +- Create a directory named "market1501" under ``$REID``. +- Download the dataset to "market1501" from http://www.liangzheng.org/Project/project_reid.html and extract the files. +- The data structure should look like + +.. code-block:: none + + market1501/ + Market-1501-v15.09.15/ + query/ + bounding_box_train/ + bounding_box_test/ + +- To use the extra 500K distractors (i.e. Market1501 + 500K), go to the **Market-1501+500k Dataset** section at http://www.liangzheng.org/Project/project_reid.html, download the zip file "distractors_500k.zip" and extract it under "market1501/Market-1501-v15.09.15". The argument to use these 500K distrctors is ``market1501_500k`` in ``ImageDataManager``. + + +CUHK03 (``cuhk03``) +^^^^^^^^^^^^^^^^^^^^^ +- Create a folder named "cuhk03" under ``$REID``. +- Download the dataset to "cuhk03/" from http://www.ee.cuhk.edu.hk/~xgwang/CUHK_identification.html and extract "cuhk03_release.zip", resulting in "cuhk03/cuhk03_release/". +- Download the new split (767/700) from `person-re-ranking `_. What you need are "cuhk03_new_protocol_config_detected.mat" and "cuhk03_new_protocol_config_labeled.mat". Put these two mat files under "cuhk03/". +- The data structure should look like + +.. code-block:: none + + cuhk03/ + cuhk03_release/ + cuhk03_new_protocol_config_detected.mat + cuhk03_new_protocol_config_labeled.mat + + +- In the default mode, we load data using the new split (767/700). If you wanna use the original (20) splits (1367/100), please set ``cuhk03_classic_split`` to True in ``ImageDataManager``. As the CMC is computed differently from Market1501 for the 1367/100 split (see `here `_), you need to enable ``use_metric_cuhk03`` in ``ImageDataManager`` to activate the *single-gallery-shot* metric for fair comparison with some methods that adopt the old splits (*do not need to report mAP*). In addition, we support both *labeled* and *detected* modes. The default mode loads *detected* images. Enable ``cuhk03_labeled`` in ``ImageDataManager`` if you wanna train and test on *labeled* images. + +.. note:: + The code will extract images in "cuhk-03.mat" and save them under "cuhk03/images_detected" and "cuhk03/images_labeled". Also, four json files will be automatically generated, i.e. "splits_classic_detected.json", "splits_classic_labeled.json", "splits_new_detected.json" and "splits_new_labeled.json". If the parent path of ``$REID`` is changed, these json files should be manually deleted. The code can automatically generate new json files to match the new path. + + +DukeMTMC-reID :math:`^\dagger` (``dukemtmcreid``) +^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +- Create a directory called "dukemtmc-reid" under ``$REID``. +- Download "DukeMTMC-reID" from http://vision.cs.duke.edu/DukeMTMC/ and extract it under "dukemtmc-reid". +- The data structure should look like + +.. code-block:: none + + dukemtmc-reid/ + DukeMTMC-reID/ + query/ + bounding_box_train/ + bounding_box_test/ + ... + +MSMT17 (``msmt17``) +^^^^^^^^^^^^^^^^^^^^^ +- Create a directory called "msmt17" under ``$REID``. +- Download the dataset from http://www.pkuvmc.com/publications/msmt17.html to "msmt17" and extract the files. +- The data structure should look like + +.. code-block:: none + + msmt17/ + MSMT17_V1/ # or MSMT17_V2 + train/ + test/ + list_train.txt + list_query.txt + list_gallery.txt + list_val.txt + +VIPeR :math:`^\dagger` (``viper``) +^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +- The download link is http://users.soe.ucsc.edu/~manduchi/VIPeR.v1.0.zip. +- Organize the dataset in a folder named "viper" as follows + +.. code-block:: none + + viper/ + VIPeR/ + cam_a/ + cam_b/ + +GRID :math:`^\dagger` (``grid``) +^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +- The download link is http://personal.ie.cuhk.edu.hk/~ccloy/files/datasets/underground_reid.zip. +- Organize the dataset in a folder named "grid" as follows + +.. code-block:: none + + grid/ + underground_reid/ + probe/ + gallery/ + ... + +CUHK01 (``cuhk01``) +^^^^^^^^^^^^^^^^^^^^^^^^ +- Create a folder named "cuhk01" under ``$REID``. +- Download "CUHK01.zip" from http://www.ee.cuhk.edu.hk/~xgwang/CUHK_identification.html and place it under "cuhk01/". +- The code can automatically extract the files, or you can do it yourself. +- The data structure should look like + +.. code-block:: none + + cuhk01/ + campus/ + +SenseReID (``sensereid``) +^^^^^^^^^^^^^^^^^^^^^^^^^^^ +- Create "sensereid" under ``$REID``. +- Download the dataset from this `link `_ and extract it to "sensereid". +- Organize the data to be like + +.. code-block:: none + + sensereid/ + SenseReID/ + test_probe/ + test_gallery/ + +QMUL-iLIDS :math:`^\dagger` (``ilids``) +^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +- Create a folder named "ilids" under ``$REID``. +- Download the dataset from http://www.eecs.qmul.ac.uk/~jason/data/i-LIDS_Pedestrian.tgz and organize it to look like + +.. code-block:: none + + ilids/ + i-LIDS_Pedestrian/ + Persons/ + +PRID (``prid``) +^^^^^^^^^^^^^^^^^^^ +- Create a directory named "prid2011" under ``$REID``. +- Download the dataset from https://www.tugraz.at/institute/icg/research/team-bischof/lrs/downloads/PRID11/ and extract it under "prid2011". +- The data structure should end up with + +.. code-block:: none + + prid2011/ + prid_2011/ + single_shot/ + multi_shot/ + +CUHK02 (``cuhk02``) +^^^^^^^^^^^^^^^^^^^^^ +- Create a folder named "cuhk02" under ``$REID``. +- Download the data from http://www.ee.cuhk.edu.hk/~xgwang/CUHK_identification.html and put it under "cuhk02/". +- Extract the file so the data structure looks like + +.. code-block:: none + + cuhk02/ + Dataset/ + P1/ + P2/ + P3/ + P4/ + P5/ + +CUHKSYSU (``cuhksysu``) +^^^^^^^^^^^^^^^^^^^^^^^^^^ +- Create a folder named "cuhksysu" under ``$REID``. +- Download the data to "cuhksysu/" from this `google drive link `_. +- Extract the zip file under "cuhksysu/". +- The data structure should look like + +.. code-block:: none + + cuhksysu/ + cropped_images + + +Video Datasets +-------------- + +MARS (``mars``) +^^^^^^^^^^^^^^^^^ +- Create "mars/" under ``$REID``. +- Download the dataset from http://www.liangzheng.com.cn/Project/project_mars.html and place it in "mars/". +- Extract "bbox_train.zip" and "bbox_test.zip". +- Download the split metadata from https://github.com/liangzheng06/MARS-evaluation/tree/master/info and put "info/" in "mars/". +- The data structure should end up with + +.. code-block:: none + + mars/ + bbox_test/ + bbox_train/ + info/ + +iLIDS-VID :math:`^\dagger` (``ilidsvid``) +^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +- Create "ilids-vid" under ``$REID``. +- Download the dataset from https://xiatian-zhu.github.io/downloads_qmul_iLIDS-VID_ReID_dataset.html to "ilids-vid". +- Organize the data structure to match + +.. code-block:: none + + ilids-vid/ + i-LIDS-VID/ + train-test people splits/ + +PRID2011 (``prid2011``) +^^^^^^^^^^^^^^^^^^^^^^^^^ +- Create a directory named "prid2011" under ``$REID``. +- Download the dataset from https://www.tugraz.at/institute/icg/research/team-bischof/lrs/downloads/PRID11/ and extract it under "prid2011". +- Download the split created by *iLIDS-VID* from `this google drive `_ and put it under "prid2011/". Following the standard protocol, only 178 persons whose sequences are more than a threshold are used. +- The data structure should end up with + +.. code-block:: none + + prid2011/ + splits_prid2011.json + prid_2011/ + single_shot/ + multi_shot/ + +DukeMTMC-VideoReID :math:`^\dagger` (``dukemtmcvidreid``) +^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +- Create "dukemtmc-vidreid" under ``$REID``. +- Download "DukeMTMC-VideoReID" from http://vision.cs.duke.edu/DukeMTMC/ and unzip the file to "dukemtmc-vidreid/". +- The data structure should look like + +.. code-block:: none + + dukemtmc-vidreid/ + DukeMTMC-VideoReID/ + train/ + query/ + gallery/ diff --git a/strong_sort/deep/reid/docs/evaluation.rst b/strong_sort/deep/reid/docs/evaluation.rst new file mode 100644 index 0000000..979ec75 --- /dev/null +++ b/strong_sort/deep/reid/docs/evaluation.rst @@ -0,0 +1,21 @@ +Evaluation +========== + +Image ReID +----------- +- **Market1501**, **DukeMTMC-reID**, **CUHK03 (767/700 split)** and **MSMT17** have fixed split so keeping ``split_id=0`` is fine. +- **CUHK03 (classic split)** has 20 fixed splits, so do ``split_id=0~19``. +- **VIPeR** contains 632 identities each with 2 images under two camera views. Evaluation should be done for 10 random splits. Each split randomly divides 632 identities to 316 train ids (632 images) and the other 316 test ids (632 images). Note that, in each random split, there are two sub-splits, one using camera-A as query and camera-B as gallery while the other one using camera-B as query and camera-A as gallery. Thus, there are totally 20 splits generated with ``split_id`` starting from 0 to 19. Models can be trained on ``split_id=[0, 2, 4, 6, 8, 10, 12, 14, 16, 18]`` (because ``split_id=0`` and ``split_id=1`` share the same train set, and so on and so forth.). At test time, models trained on ``split_id=0`` can be directly evaluated on ``split_id=1``, models trained on ``split_id=2`` can be directly evaluated on ``split_id=3``, and so on and so forth. +- **CUHK01** is similar to VIPeR in the split generation. +- **GRID** , **iLIDS** and **PRID** have 10 random splits, so evaluation should be done by varying ``split_id`` from 0 to 9. +- **SenseReID** has no training images and is used for evaluation only. + + +.. note:: + The ``split_id`` argument is defined in ``ImageDataManager`` and ``VideoDataManager``. Please refer to :ref:`torchreid_data`. + + +Video ReID +----------- +- **MARS** and **DukeMTMC-VideoReID** have fixed single split so using ``split_id=0`` is ok. +- **iLIDS-VID** and **PRID2011** have 10 predefined splits so evaluation should be done by varying ``split_id`` from 0 to 9. \ No newline at end of file diff --git a/strong_sort/deep/reid/docs/index.rst b/strong_sort/deep/reid/docs/index.rst new file mode 100644 index 0000000..c437dea --- /dev/null +++ b/strong_sort/deep/reid/docs/index.rst @@ -0,0 +1,35 @@ +.. include:: ../README.rst + + +.. toctree:: + :hidden: + + user_guide + datasets + evaluation + +.. toctree:: + :caption: Package Reference + :hidden: + + pkg/data + pkg/engine + pkg/losses + pkg/metrics + pkg/models + pkg/optim + pkg/utils + +.. toctree:: + :caption: Resources + :hidden: + + AWESOME_REID.md + MODEL_ZOO.md + + +Indices and tables +================== + +* :ref:`genindex` +* :ref:`modindex` \ No newline at end of file diff --git a/strong_sort/deep/reid/docs/pkg/data.rst b/strong_sort/deep/reid/docs/pkg/data.rst new file mode 100644 index 0000000..3dc47d6 --- /dev/null +++ b/strong_sort/deep/reid/docs/pkg/data.rst @@ -0,0 +1,86 @@ +.. _torchreid_data: + +torchreid.data +============== + + +Data Manager +--------------------------- + +.. automodule:: torchreid.data.datamanager + :members: + + +Sampler +----------------------- + +.. automodule:: torchreid.data.sampler + :members: + + +Transforms +--------------------------- + +.. automodule:: torchreid.data.transforms + :members: + + +Dataset +--------------------------- + +.. automodule:: torchreid.data.datasets.dataset + :members: + + +.. automodule:: torchreid.data.datasets.__init__ + :members: + + +Image Datasets +------------------------------ + +.. automodule:: torchreid.data.datasets.image.market1501 + :members: + +.. automodule:: torchreid.data.datasets.image.cuhk03 + :members: + +.. automodule:: torchreid.data.datasets.image.dukemtmcreid + :members: + +.. automodule:: torchreid.data.datasets.image.msmt17 + :members: + +.. automodule:: torchreid.data.datasets.image.viper + :members: + +.. automodule:: torchreid.data.datasets.image.grid + :members: + +.. automodule:: torchreid.data.datasets.image.cuhk01 + :members: + +.. automodule:: torchreid.data.datasets.image.ilids + :members: + +.. automodule:: torchreid.data.datasets.image.sensereid + :members: + +.. automodule:: torchreid.data.datasets.image.prid + :members: + + +Video Datasets +------------------------------ + +.. automodule:: torchreid.data.datasets.video.mars + :members: + +.. automodule:: torchreid.data.datasets.video.ilidsvid + :members: + +.. automodule:: torchreid.data.datasets.video.prid2011 + :members: + +.. automodule:: torchreid.data.datasets.video.dukemtmcvidreid + :members: \ No newline at end of file diff --git a/strong_sort/deep/reid/docs/pkg/engine.rst b/strong_sort/deep/reid/docs/pkg/engine.rst new file mode 100644 index 0000000..ae2bc68 --- /dev/null +++ b/strong_sort/deep/reid/docs/pkg/engine.rst @@ -0,0 +1,31 @@ +.. _torchreid_engine: + +torchreid.engine +================== + + +Base Engine +------------ + +.. autoclass:: torchreid.engine.engine.Engine + :members: + + +Image Engines +------------- + +.. autoclass:: torchreid.engine.image.softmax.ImageSoftmaxEngine + :members: + + +.. autoclass:: torchreid.engine.image.triplet.ImageTripletEngine + :members: + + +Video Engines +------------- + +.. autoclass:: torchreid.engine.video.softmax.VideoSoftmaxEngine + + +.. autoclass:: torchreid.engine.video.triplet.VideoTripletEngine \ No newline at end of file diff --git a/strong_sort/deep/reid/docs/pkg/losses.rst b/strong_sort/deep/reid/docs/pkg/losses.rst new file mode 100644 index 0000000..33fd9bc --- /dev/null +++ b/strong_sort/deep/reid/docs/pkg/losses.rst @@ -0,0 +1,18 @@ +.. _torchreid_losses: + +torchreid.losses +================= + + +Softmax +-------- + +.. automodule:: torchreid.losses.cross_entropy_loss + :members: + + +Triplet +------- + +.. automodule:: torchreid.losses.hard_mine_triplet_loss + :members: \ No newline at end of file diff --git a/strong_sort/deep/reid/docs/pkg/metrics.rst b/strong_sort/deep/reid/docs/pkg/metrics.rst new file mode 100644 index 0000000..5a52a90 --- /dev/null +++ b/strong_sort/deep/reid/docs/pkg/metrics.rst @@ -0,0 +1,25 @@ +.. _torchreid_metrics: + +torchreid.metrics +================= + + +Distance +--------- + +.. automodule:: torchreid.metrics.distance + :members: + + +Accuracy +-------- + +.. automodule:: torchreid.metrics.accuracy + :members: + + +Rank +----- + +.. automodule:: torchreid.metrics.rank + :members: evaluate_rank \ No newline at end of file diff --git a/strong_sort/deep/reid/docs/pkg/models.rst b/strong_sort/deep/reid/docs/pkg/models.rst new file mode 100644 index 0000000..685bc73 --- /dev/null +++ b/strong_sort/deep/reid/docs/pkg/models.rst @@ -0,0 +1,43 @@ +.. _torchreid_models: + +torchreid.models +================= + +Interface +--------- + +.. automodule:: torchreid.models.__init__ + :members: + + +ImageNet Classification Models +------------------------------- + +.. autoclass:: torchreid.models.resnet.ResNet +.. autoclass:: torchreid.models.senet.SENet +.. autoclass:: torchreid.models.densenet.DenseNet +.. autoclass:: torchreid.models.inceptionresnetv2.InceptionResNetV2 +.. autoclass:: torchreid.models.inceptionv4.InceptionV4 +.. autoclass:: torchreid.models.xception.Xception + + +Lightweight Models +------------------ + +.. autoclass:: torchreid.models.nasnet.NASNetAMobile +.. autoclass:: torchreid.models.mobilenetv2.MobileNetV2 +.. autoclass:: torchreid.models.shufflenet.ShuffleNet +.. autoclass:: torchreid.models.squeezenet.SqueezeNet +.. autoclass:: torchreid.models.shufflenetv2.ShuffleNetV2 + + +ReID-specific Models +-------------------- + +.. autoclass:: torchreid.models.mudeep.MuDeep +.. autoclass:: torchreid.models.resnetmid.ResNetMid +.. autoclass:: torchreid.models.hacnn.HACNN +.. autoclass:: torchreid.models.pcb.PCB +.. autoclass:: torchreid.models.mlfn.MLFN +.. autoclass:: torchreid.models.osnet.OSNet +.. autoclass:: torchreid.models.osnet_ain.OSNet \ No newline at end of file diff --git a/strong_sort/deep/reid/docs/pkg/optim.rst b/strong_sort/deep/reid/docs/pkg/optim.rst new file mode 100644 index 0000000..5601623 --- /dev/null +++ b/strong_sort/deep/reid/docs/pkg/optim.rst @@ -0,0 +1,18 @@ +.. _torchreid_optim: + +torchreid.optim +================= + + +Optimizer +---------- + +.. automodule:: torchreid.optim.optimizer + :members: build_optimizer + + +LR Scheduler +------------- + +.. automodule:: torchreid.optim.lr_scheduler + :members: build_lr_scheduler \ No newline at end of file diff --git a/strong_sort/deep/reid/docs/pkg/utils.rst b/strong_sort/deep/reid/docs/pkg/utils.rst new file mode 100644 index 0000000..1545bc2 --- /dev/null +++ b/strong_sort/deep/reid/docs/pkg/utils.rst @@ -0,0 +1,41 @@ +.. _torchreid_utils: + +torchreid.utils +================= + +Average Meter +-------------- + +.. automodule:: torchreid.utils.avgmeter + :members: + + +Loggers +------- + +.. automodule:: torchreid.utils.loggers + :members: + + +Generic Tools +--------------- +.. automodule:: torchreid.utils.tools + :members: + + +ReID Tools +---------- + +.. automodule:: torchreid.utils.reidtools + :members: + + +Torch Tools +------------ + +.. automodule:: torchreid.utils.torchtools + :members: + + +.. automodule:: torchreid.utils.model_complexity + :members: diff --git a/strong_sort/deep/reid/docs/user_guide.rst b/strong_sort/deep/reid/docs/user_guide.rst new file mode 100644 index 0000000..5415109 --- /dev/null +++ b/strong_sort/deep/reid/docs/user_guide.rst @@ -0,0 +1,351 @@ +How-to +============ + +.. contents:: + :local: + + +Prepare datasets +----------------- +See :ref:`datasets`. + + +Find model keys +----------------- +Keys are listed under the *Public keys* section within each model class in :ref:`torchreid_models`. + + +Show available models +---------------------- + +.. code-block:: python + + import torchreid + torchreid.models.show_avai_models() + + +Change the training sampler +----------------------------- +The default ``train_sampler`` is "RandomSampler". You can give the specific sampler name as input to ``train_sampler``, e.g. ``train_sampler='RandomIdentitySampler'`` for triplet loss. + + +Choose an optimizer/lr_scheduler +---------------------------------- +Please refer to the source code of ``build_optimizer``/``build_lr_scheduler`` in :ref:`torchreid_optim` for details. + + +Resume training +---------------- +Suppose the checkpoint is saved in "log/resnet50/model.pth.tar-30", you can do + +.. code-block:: python + + start_epoch = torchreid.utils.resume_from_checkpoint( + 'log/resnet50/model.pth.tar-30', + model, + optimizer + ) + + engine.run( + save_dir='log/resnet50', + max_epoch=60, + start_epoch=start_epoch + ) + + +Compute model complexity +-------------------------- +We provide a tool in ``torchreid.utils.model_complexity.py`` to automatically compute the model complexity, i.e. number of parameters and FLOPs. + +.. code-block:: python + + from torchreid import models, utils + + model = models.build_model(name='resnet50', num_classes=1000) + num_params, flops = utils.compute_model_complexity(model, (1, 3, 256, 128)) + + # show detailed complexity for each module + utils.compute_model_complexity(model, (1, 3, 256, 128), verbose=True) + + # count flops for all layers including ReLU and BatchNorm + utils.compute_model_complexity(model, (1, 3, 256, 128), verbose=True, only_conv_linear=False) + +Note that (1) this function only provides an estimate of the theoretical time complexity rather than the actual running time which depends on implementations and hardware; (2) the FLOPs is only counted for layers that are used at test time. This means that redundant layers such as person ID classification layer will be ignored. The inference graph depends on how you define the computations in ``forward()``. + + +Combine multiple datasets +--------------------------- +Easy. Just give whatever datasets (keys) you want to the ``sources`` argument when instantiating a data manager. For example, + +.. code-block:: python + + datamanager = torchreid.data.ImageDataManager( + root='reid-data', + sources=['market1501', 'dukemtmcreid', 'cuhk03', 'msmt17'], + height=256, + width=128, + batch_size=32 + ) + +In this example, the target datasets are Market1501, DukeMTMC-reID, CUHK03 and MSMT17 as the ``targets`` argument is not specified. Please refer to ``Engine.test()`` in :ref:`torchreid_engine` for details regarding how evaluation is performed. + + +Do cross-dataset evaluation +----------------------------- +Easy. Just give whatever datasets (keys) you want to the argument ``targets``, like + +.. code-block:: python + + datamanager = torchreid.data.ImageDataManager( + root='reid-data', + sources='market1501', + targets='dukemtmcreid', # or targets='cuhk03' or targets=['dukemtmcreid', 'cuhk03'] + height=256, + width=128, + batch_size=32 + ) + + +Combine train, query and gallery +--------------------------------- +This can be easily done by setting ``combineall=True`` when instantiating a data manager. Below is an example of using Market1501, + +.. code-block:: python + + datamanager = torchreid.data.ImageDataManager( + root='reid-data', + sources='market1501', + height=256, + width=128, + batch_size=32, + market1501_500k=False, + combineall=True # it's me, here + ) + +More specifically, with ``combineall=False``, you will get + +.. code-block:: none + + => Loaded Market1501 + ---------------------------------------- + subset | # ids | # images | # cameras + ---------------------------------------- + train | 751 | 12936 | 6 + query | 750 | 3368 | 6 + gallery | 751 | 15913 | 6 + --------------------------------------- + +with ``combineall=True``, you will get + +.. code-block:: none + + => Loaded Market1501 + ---------------------------------------- + subset | # ids | # images | # cameras + ---------------------------------------- + train | 1501 | 29419 | 6 + query | 750 | 3368 | 6 + gallery | 751 | 15913 | 6 + --------------------------------------- + + +Optimize layers with different learning rates +----------------------------------------------- +A common practice for fine-tuning pretrained models is to use a smaller learning rate for base layers and a large learning rate for randomly initialized layers (referred to as ``new_layers``). ``torchreid.optim.optimizer`` has implemented such feature. What you need to do is to set ``staged_lr=True`` and give the names of ``new_layers`` such as "classifier". + +Below is an example of setting different learning rates for base layers and new layers in ResNet50, + +.. code-block:: python + + # New layer "classifier" has a learning rate of 0.01 + # The base layers have a learning rate of 0.001 + optimizer = torchreid.optim.build_optimizer( + model, + optim='sgd', + lr=0.01, + staged_lr=True, + new_layers='classifier', + base_lr_mult=0.1 + ) + +Please refer to :ref:`torchreid_optim` for more details. + + +Do two-stepped transfer learning +------------------------------------- +To prevent the pretrained layers from being damaged by harmful gradients back-propagated from randomly initialized layers, one can adopt the *two-stepped transfer learning strategy* presented in `Deep Transfer Learning for Person Re-identification `_. The basic idea is to pretrain the randomly initialized layers for few epochs while keeping the base layers frozen before training all layers end-to-end. + +This has been implemented in ``Engine.train()`` (see :ref:`torchreid_engine`). The arguments related to this feature are ``fixbase_epoch`` and ``open_layers``. Intuitively, ``fixbase_epoch`` denotes the number of epochs to keep the base layers frozen; ``open_layers`` means which layer is open for training. + +For example, say you want to pretrain the classification layer named "classifier" in ResNet50 for 5 epochs before training all layers, you can do + +.. code-block:: python + + engine.run( + save_dir='log/resnet50', + max_epoch=60, + eval_freq=10, + print_freq=10, + test_only=False, + fixbase_epoch=5, + open_layers='classifier' + ) + # or open_layers=['fc', 'classifier'] if there is another fc layer that + # is randomly initialized, like resnet50_fc512 + +Note that ``fixbase_epoch`` is counted into ``max_epoch``. In the above example, the base network will be fixed for 5 epochs and then open for training for 55 epochs. Thus, if you want to freeze some layers throughout the training, what you can do is to set ``fixbase_epoch`` equal to ``max_epoch`` and put the layer names in ``open_layers`` which you want to train. + + +Test a trained model +---------------------- +You can load a trained model using :code:`torchreid.utils.load_pretrained_weights(model, weight_path)` and set ``test_only=True`` in ``engine.run()``. + + +Fine-tune a model pre-trained on reid datasets +----------------------------------------------- +Use :code:`torchreid.utils.load_pretrained_weights(model, weight_path)` to load the pre-trained weights and then fine-tune on the dataset you want. + + +Visualize learning curves with tensorboard +-------------------------------------------- +The ``SummaryWriter()`` for tensorboard will be automatically initialized in ``engine.run()`` when you are training your model. Therefore, you do not need to do extra jobs. After the training is done, the ``*tf.events*`` file will be saved in ``save_dir``. Then, you just call ``tensorboard --logdir=your_save_dir`` in your terminal and visit ``http://localhost:6006/`` in a web browser. See `pytorch tensorboard `_ for further information. + + +Visualize ranking results +--------------------------- +This can be achieved by setting ``visrank`` to true in ``engine.run()``. ``visrank_topk`` determines the top-k images to be visualized (Default is ``visrank_topk=10``). Note that ``visrank`` can only be used in test mode, i.e. ``test_only=True`` in ``engine.run()``. The output will be saved under ``save_dir/visrank_DATASETNAME`` where each plot contains the top-k similar gallery images given a query. An example is shown below where red and green denote incorrect and correct matches respectively. + +.. image:: figures/ranking_results.jpg + :width: 800px + :align: center + + +Visualize activation maps +-------------------------- +To understand where the CNN focuses on to extract features for ReID, you can visualize the activation maps as in `OSNet `_. This is implemented in ``tools/visualize_actmap.py`` (check the code for more details). An example running command is + +.. code-block:: shell + + python tools/visualize_actmap.py \ + --root $DATA/reid \ + -d market1501 \ + -m osnet_x1_0 \ + --weights PATH_TO_PRETRAINED_WEIGHTS \ + --save-dir log/visactmap_osnet_x1_0_market1501 + +The output will look like (from left to right: image, activation map, overlapped image) + +.. image:: figures/actmap.jpg + :width: 300px + :align: center + + +.. note:: + In order to visualize activation maps, the CNN needs to output the last convolutional feature maps at eval mode. See ``torchreid/models/osnet.py`` for example. + + +Use your own dataset +---------------------- +1. Write your own dataset class. Below is a template for image dataset. However, it can also be applied to a video dataset class, for which you simply change ``ImageDataset`` to ``VideoDataset``. + +.. code-block:: python + + from __future__ import absolute_import + from __future__ import print_function + from __future__ import division + + import sys + import os + import os.path as osp + + from torchreid.data import ImageDataset + + + class NewDataset(ImageDataset): + dataset_dir = 'new_dataset' + + def __init__(self, root='', **kwargs): + self.root = osp.abspath(osp.expanduser(root)) + self.dataset_dir = osp.join(self.root, self.dataset_dir) + + # All you need to do here is to generate three lists, + # which are train, query and gallery. + # Each list contains tuples of (img_path, pid, camid), + # where + # - img_path (str): absolute path to an image. + # - pid (int): person ID, e.g. 0, 1. + # - camid (int): camera ID, e.g. 0, 1. + # Note that + # - pid and camid should be 0-based. + # - query and gallery should share the same pid scope (e.g. + # pid=0 in query refers to the same person as pid=0 in gallery). + # - train, query and gallery share the same camid scope (e.g. + # camid=0 in train refers to the same camera as camid=0 + # in query/gallery). + train = ... + query = ... + gallery = ... + + super(NewDataset, self).__init__(train, query, gallery, **kwargs) + + +2. Register your dataset. + +.. code-block:: python + + import torchreid + torchreid.data.register_image_dataset('new_dataset', NewDataset) + + +3. Initialize a data manager with your dataset. + +.. code-block:: python + + # use your own dataset only + datamanager = torchreid.data.ImageDataManager( + root='reid-data', + sources='new_dataset' + ) + # combine with other datasets + datamanager = torchreid.data.ImageDataManager( + root='reid-data', + sources=['new_dataset', 'dukemtmcreid'] + ) + # cross-dataset evaluation + datamanager = torchreid.data.ImageDataManager( + root='reid-data', + sources=['new_dataset', 'dukemtmcreid'], + targets='market1501' # or targets=['market1501', 'cuhk03'] + ) + + + +Design your own Engine +------------------------ +A new Engine should be designed if you have your own loss function. The base Engine class ``torchreid.engine.Engine`` has implemented some generic methods which you can inherit to avoid re-writing. Please refer to the source code for more details. You are suggested to see how ``ImageSoftmaxEngine`` and ``ImageTripletEngine`` are constructed (also ``VideoSoftmaxEngine`` and ``VideoTripletEngine``). All you need to implement might be just a ``forward_backward()`` function. + + +Use Torchreid as a feature extractor in your projects +------------------------------------------------------- +We have provided a simple API for feature extraction, which accepts input of various types such as a list of image paths or numpy arrays. More details can be found in the code at ``torchreid/utils/feature_extractor.py``. Here we show a simple example of how to extract features given a list of image paths. + +.. code-block:: python + + from torchreid.utils import FeatureExtractor + + extractor = FeatureExtractor( + model_name='osnet_x1_0', + model_path='a/b/c/model.pth.tar', + device='cuda' + ) + + image_list = [ + 'a/b/c/image001.jpg', + 'a/b/c/image002.jpg', + 'a/b/c/image003.jpg', + 'a/b/c/image004.jpg', + 'a/b/c/image005.jpg' + ] + + features = extractor(image_list) + print(features.shape) # output (5, 512) diff --git a/strong_sort/deep/reid/linter.sh b/strong_sort/deep/reid/linter.sh new file mode 100644 index 0000000..9db34f9 --- /dev/null +++ b/strong_sort/deep/reid/linter.sh @@ -0,0 +1,11 @@ +echo "Running isort" +isort -y -sp . +echo "Done" + +echo "Running yapf" +yapf -i -r -vv -e build . +echo "Done" + +echo "Running flake8" +flake8 . +echo "Done" \ No newline at end of file diff --git a/strong_sort/deep/reid/projects/DML/README.md b/strong_sort/deep/reid/projects/DML/README.md new file mode 100644 index 0000000..e81be7f --- /dev/null +++ b/strong_sort/deep/reid/projects/DML/README.md @@ -0,0 +1,16 @@ +# Deep mutual learning + +This repo implements [Deep Mutual Learning (CVPR'18)](https://zpascal.net/cvpr2018/Zhang_Deep_Mutual_Learning_CVPR_2018_paper.pdf) (DML) for person re-id. + +We used this code in our [OSNet](https://arxiv.org/pdf/1905.00953.pdf) paper (see Supp. B). The training command to reproduce the result of "triplet + DML" (Table 12f in the paper) is +```bash +python main.py \ +--config-file im_osnet_x1_0_dml_256x128_amsgrad_cosine.yaml \ +--root $DATA +``` + +`$DATA` corresponds to the path to your dataset folder. + +Change `model.deploy` to `both` if you wanna enable model ensembling. + +If you have any questions, please raise an issue in the Issues area. \ No newline at end of file diff --git a/strong_sort/deep/reid/projects/DML/default_config.py b/strong_sort/deep/reid/projects/DML/default_config.py new file mode 100644 index 0000000..9b15e35 --- /dev/null +++ b/strong_sort/deep/reid/projects/DML/default_config.py @@ -0,0 +1,207 @@ +from yacs.config import CfgNode as CN + + +def get_default_config(): + cfg = CN() + + # model + cfg.model = CN() + cfg.model.name = 'resnet50' + cfg.model.pretrained = True # automatically load pretrained model weights if available + cfg.model.load_weights1 = '' # path to model-1 weights + cfg.model.load_weights2 = '' # path to model-2 weights + cfg.model.resume1 = '' # path to checkpoint for resume training + cfg.model.resume2 = '' # path to checkpoint for resume training + cfg.model.deploy = 'model1' # model1, model2 or both + + # data + cfg.data = CN() + cfg.data.type = 'image' + cfg.data.root = 'reid-data' + cfg.data.sources = ['market1501'] + cfg.data.targets = ['market1501'] + cfg.data.workers = 4 # number of data loading workers + cfg.data.split_id = 0 # split index + cfg.data.height = 256 # image height + cfg.data.width = 128 # image width + cfg.data.combineall = False # combine train, query and gallery for training + cfg.data.transforms = ['random_flip'] # data augmentation + cfg.data.norm_mean = [0.485, 0.456, 0.406] # default is imagenet mean + cfg.data.norm_std = [0.229, 0.224, 0.225] # default is imagenet std + cfg.data.save_dir = 'log' # path to save log + cfg.data.load_train_targets = False + + # specific datasets + cfg.market1501 = CN() + cfg.market1501.use_500k_distractors = False # add 500k distractors to the gallery set for market1501 + cfg.cuhk03 = CN() + cfg.cuhk03.labeled_images = False # use labeled images, if False, use detected images + cfg.cuhk03.classic_split = False # use classic split by Li et al. CVPR14 + cfg.cuhk03.use_metric_cuhk03 = False # use cuhk03's metric for evaluation + + # sampler + cfg.sampler = CN() + cfg.sampler.train_sampler = 'RandomSampler' + cfg.sampler.num_instances = 4 # number of instances per identity for RandomIdentitySampler + + # video reid setting + cfg.video = CN() + cfg.video.seq_len = 15 # number of images to sample in a tracklet + cfg.video.sample_method = 'evenly' # how to sample images from a tracklet + cfg.video.pooling_method = 'avg' # how to pool features over a tracklet + + # train + cfg.train = CN() + cfg.train.optim = 'adam' + cfg.train.lr = 0.0003 + cfg.train.weight_decay = 5e-4 + cfg.train.max_epoch = 60 + cfg.train.start_epoch = 0 + cfg.train.batch_size = 32 + cfg.train.fixbase_epoch = 0 # number of epochs to fix base layers + cfg.train.open_layers = [ + 'classifier' + ] # layers for training while keeping others frozen + cfg.train.staged_lr = False # set different lr to different layers + cfg.train.new_layers = ['classifier'] # newly added layers with default lr + cfg.train.base_lr_mult = 0.1 # learning rate multiplier for base layers + cfg.train.lr_scheduler = 'single_step' + cfg.train.stepsize = [20] # stepsize to decay learning rate + cfg.train.gamma = 0.1 # learning rate decay multiplier + cfg.train.print_freq = 20 # print frequency + cfg.train.seed = 1 # random seed + + # optimizer + cfg.sgd = CN() + cfg.sgd.momentum = 0.9 # momentum factor for sgd and rmsprop + cfg.sgd.dampening = 0. # dampening for momentum + cfg.sgd.nesterov = False # Nesterov momentum + cfg.rmsprop = CN() + cfg.rmsprop.alpha = 0.99 # smoothing constant + cfg.adam = CN() + cfg.adam.beta1 = 0.9 # exponential decay rate for first moment + cfg.adam.beta2 = 0.999 # exponential decay rate for second moment + + # loss + cfg.loss = CN() + cfg.loss.name = 'triplet' + cfg.loss.softmax = CN() + cfg.loss.softmax.label_smooth = True # use label smoothing regularizer + cfg.loss.triplet = CN() + cfg.loss.triplet.margin = 0.3 # distance margin + cfg.loss.triplet.weight_t = 1. # weight to balance hard triplet loss + cfg.loss.triplet.weight_x = 0. # weight to balance cross entropy loss + cfg.loss.dml = CN() + cfg.loss.dml.weight_ml = 1. # weight for mutual learning loss + + # test + cfg.test = CN() + cfg.test.batch_size = 100 + cfg.test.dist_metric = 'euclidean' # distance metric, ['euclidean', 'cosine'] + cfg.test.normalize_feature = False # normalize feature vectors before computing distance + cfg.test.ranks = [1, 5, 10, 20] # cmc ranks + cfg.test.evaluate = False # test only + cfg.test.eval_freq = -1 # evaluation frequency (-1 means to only test after training) + cfg.test.start_eval = 0 # start to evaluate after a specific epoch + cfg.test.rerank = False # use person re-ranking + cfg.test.visrank = False # visualize ranked results (only available when cfg.test.evaluate=True) + cfg.test.visrank_topk = 10 # top-k ranks to visualize + + return cfg + + +def imagedata_kwargs(cfg): + return { + 'root': cfg.data.root, + 'sources': cfg.data.sources, + 'targets': cfg.data.targets, + 'height': cfg.data.height, + 'width': cfg.data.width, + 'transforms': cfg.data.transforms, + 'norm_mean': cfg.data.norm_mean, + 'norm_std': cfg.data.norm_std, + 'use_gpu': cfg.use_gpu, + 'split_id': cfg.data.split_id, + 'combineall': cfg.data.combineall, + 'load_train_targets': cfg.data.load_train_targets, + 'batch_size_train': cfg.train.batch_size, + 'batch_size_test': cfg.test.batch_size, + 'workers': cfg.data.workers, + 'num_instances': cfg.sampler.num_instances, + 'train_sampler': cfg.sampler.train_sampler, + # image + 'cuhk03_labeled': cfg.cuhk03.labeled_images, + 'cuhk03_classic_split': cfg.cuhk03.classic_split, + 'market1501_500k': cfg.market1501.use_500k_distractors, + } + + +def videodata_kwargs(cfg): + return { + 'root': cfg.data.root, + 'sources': cfg.data.sources, + 'targets': cfg.data.targets, + 'height': cfg.data.height, + 'width': cfg.data.width, + 'transforms': cfg.data.transforms, + 'norm_mean': cfg.data.norm_mean, + 'norm_std': cfg.data.norm_std, + 'use_gpu': cfg.use_gpu, + 'split_id': cfg.data.split_id, + 'combineall': cfg.data.combineall, + 'batch_size_train': cfg.train.batch_size, + 'batch_size_test': cfg.test.batch_size, + 'workers': cfg.data.workers, + 'num_instances': cfg.sampler.num_instances, + 'train_sampler': cfg.sampler.train_sampler, + # video + 'seq_len': cfg.video.seq_len, + 'sample_method': cfg.video.sample_method + } + + +def optimizer_kwargs(cfg): + return { + 'optim': cfg.train.optim, + 'lr': cfg.train.lr, + 'weight_decay': cfg.train.weight_decay, + 'momentum': cfg.sgd.momentum, + 'sgd_dampening': cfg.sgd.dampening, + 'sgd_nesterov': cfg.sgd.nesterov, + 'rmsprop_alpha': cfg.rmsprop.alpha, + 'adam_beta1': cfg.adam.beta1, + 'adam_beta2': cfg.adam.beta2, + 'staged_lr': cfg.train.staged_lr, + 'new_layers': cfg.train.new_layers, + 'base_lr_mult': cfg.train.base_lr_mult + } + + +def lr_scheduler_kwargs(cfg): + return { + 'lr_scheduler': cfg.train.lr_scheduler, + 'stepsize': cfg.train.stepsize, + 'gamma': cfg.train.gamma, + 'max_epoch': cfg.train.max_epoch + } + + +def engine_run_kwargs(cfg): + return { + 'save_dir': cfg.data.save_dir, + 'max_epoch': cfg.train.max_epoch, + 'start_epoch': cfg.train.start_epoch, + 'fixbase_epoch': cfg.train.fixbase_epoch, + 'open_layers': cfg.train.open_layers, + 'start_eval': cfg.test.start_eval, + 'eval_freq': cfg.test.eval_freq, + 'test_only': cfg.test.evaluate, + 'print_freq': cfg.train.print_freq, + 'dist_metric': cfg.test.dist_metric, + 'normalize_feature': cfg.test.normalize_feature, + 'visrank': cfg.test.visrank, + 'visrank_topk': cfg.test.visrank_topk, + 'use_metric_cuhk03': cfg.cuhk03.use_metric_cuhk03, + 'ranks': cfg.test.ranks, + 'rerank': cfg.test.rerank + } diff --git a/strong_sort/deep/reid/projects/DML/dml.py b/strong_sort/deep/reid/projects/DML/dml.py new file mode 100644 index 0000000..546e573 --- /dev/null +++ b/strong_sort/deep/reid/projects/DML/dml.py @@ -0,0 +1,149 @@ +from __future__ import division, print_function, absolute_import +import torch +from torch.nn import functional as F + +from torchreid.utils import open_all_layers, open_specified_layers +from torchreid.engine import Engine +from torchreid.losses import TripletLoss, CrossEntropyLoss + + +class ImageDMLEngine(Engine): + + def __init__( + self, + datamanager, + model1, + optimizer1, + scheduler1, + model2, + optimizer2, + scheduler2, + margin=0.3, + weight_t=0.5, + weight_x=1., + weight_ml=1., + use_gpu=True, + label_smooth=True, + deploy='model1' + ): + super(ImageDMLEngine, self).__init__(datamanager, use_gpu) + + self.model1 = model1 + self.optimizer1 = optimizer1 + self.scheduler1 = scheduler1 + self.register_model('model1', model1, optimizer1, scheduler1) + + self.model2 = model2 + self.optimizer2 = optimizer2 + self.scheduler2 = scheduler2 + self.register_model('model2', model2, optimizer2, scheduler2) + + self.weight_t = weight_t + self.weight_x = weight_x + self.weight_ml = weight_ml + + assert deploy in ['model1', 'model2', 'both'] + self.deploy = deploy + + self.criterion_t = TripletLoss(margin=margin) + self.criterion_x = CrossEntropyLoss( + num_classes=self.datamanager.num_train_pids, + use_gpu=self.use_gpu, + label_smooth=label_smooth + ) + + def forward_backward(self, data): + imgs, pids = self.parse_data_for_train(data) + + if self.use_gpu: + imgs = imgs.cuda() + pids = pids.cuda() + + outputs1, features1 = self.model1(imgs) + loss1_x = self.compute_loss(self.criterion_x, outputs1, pids) + loss1_t = self.compute_loss(self.criterion_t, features1, pids) + + outputs2, features2 = self.model2(imgs) + loss2_x = self.compute_loss(self.criterion_x, outputs2, pids) + loss2_t = self.compute_loss(self.criterion_t, features2, pids) + + loss1_ml = self.compute_kl_div( + outputs2.detach(), outputs1, is_logit=True + ) + loss2_ml = self.compute_kl_div( + outputs1.detach(), outputs2, is_logit=True + ) + + loss1 = 0 + loss1 += loss1_x * self.weight_x + loss1 += loss1_t * self.weight_t + loss1 += loss1_ml * self.weight_ml + + loss2 = 0 + loss2 += loss2_x * self.weight_x + loss2 += loss2_t * self.weight_t + loss2 += loss2_ml * self.weight_ml + + self.optimizer1.zero_grad() + loss1.backward() + self.optimizer1.step() + + self.optimizer2.zero_grad() + loss2.backward() + self.optimizer2.step() + + loss_dict = { + 'loss1_x': loss1_x.item(), + 'loss1_t': loss1_t.item(), + 'loss1_ml': loss1_ml.item(), + 'loss2_x': loss1_x.item(), + 'loss2_t': loss1_t.item(), + 'loss2_ml': loss1_ml.item() + } + + return loss_dict + + @staticmethod + def compute_kl_div(p, q, is_logit=True): + if is_logit: + p = F.softmax(p, dim=1) + q = F.softmax(q, dim=1) + return -(p * torch.log(q + 1e-8)).sum(1).mean() + + def two_stepped_transfer_learning( + self, epoch, fixbase_epoch, open_layers, model=None + ): + """Two stepped transfer learning. + + The idea is to freeze base layers for a certain number of epochs + and then open all layers for training. + + Reference: https://arxiv.org/abs/1611.05244 + """ + model1 = self.model1 + model2 = self.model2 + + if (epoch + 1) <= fixbase_epoch and open_layers is not None: + print( + '* Only train {} (epoch: {}/{})'.format( + open_layers, epoch + 1, fixbase_epoch + ) + ) + open_specified_layers(model1, open_layers) + open_specified_layers(model2, open_layers) + else: + open_all_layers(model1) + open_all_layers(model2) + + def extract_features(self, input): + if self.deploy == 'model1': + return self.model1(input) + + elif self.deploy == 'model2': + return self.model2(input) + + else: + features = [] + features.append(self.model1(input)) + features.append(self.model2(input)) + return torch.cat(features, 1) diff --git a/strong_sort/deep/reid/projects/DML/im_osnet_x1_0_dml_256x128_amsgrad_cosine.yaml b/strong_sort/deep/reid/projects/DML/im_osnet_x1_0_dml_256x128_amsgrad_cosine.yaml new file mode 100644 index 0000000..3a240b1 --- /dev/null +++ b/strong_sort/deep/reid/projects/DML/im_osnet_x1_0_dml_256x128_amsgrad_cosine.yaml @@ -0,0 +1,42 @@ +model: + name: 'osnet_x1_0' + pretrained: True + deploy: 'model1' + +data: + type: 'image' + sources: ['market1501'] + targets: ['market1501'] + height: 256 + width: 128 + combineall: False + transforms: ['random_flip', 'random_erase'] + save_dir: 'log/osnet_x1_0_market1501_dml_cosinelr' + +loss: + name: 'triplet' + softmax: + label_smooth: True + triplet: + margin: 0.3 + weight_t: 0.5 + weight_x: 1. + dml: + weight_ml: 1. + +train: + optim: 'amsgrad' + lr: 0.0015 + max_epoch: 250 + batch_size: 64 + fixbase_epoch: 10 + open_layers: ['classifier'] + lr_scheduler: 'cosine' + +test: + batch_size: 300 + dist_metric: 'cosine' + normalize_feature: False + evaluate: False + eval_freq: -1 + rerank: False \ No newline at end of file diff --git a/strong_sort/deep/reid/projects/DML/main.py b/strong_sort/deep/reid/projects/DML/main.py new file mode 100644 index 0000000..9af33d9 --- /dev/null +++ b/strong_sort/deep/reid/projects/DML/main.py @@ -0,0 +1,166 @@ +import sys +import copy +import time +import os.path as osp +import argparse +import torch +import torch.nn as nn + +import torchreid +from torchreid.utils import ( + Logger, check_isfile, set_random_seed, collect_env_info, + resume_from_checkpoint, load_pretrained_weights, compute_model_complexity +) + +from dml import ImageDMLEngine +from default_config import ( + imagedata_kwargs, optimizer_kwargs, engine_run_kwargs, get_default_config, + lr_scheduler_kwargs +) + + +def reset_config(cfg, args): + if args.root: + cfg.data.root = args.root + if args.sources: + cfg.data.sources = args.sources + if args.targets: + cfg.data.targets = args.targets + if args.transforms: + cfg.data.transforms = args.transforms + + +def main(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + parser.add_argument( + '--config-file', type=str, default='', help='path to config file' + ) + parser.add_argument( + '-s', + '--sources', + type=str, + nargs='+', + help='source datasets (delimited by space)' + ) + parser.add_argument( + '-t', + '--targets', + type=str, + nargs='+', + help='target datasets (delimited by space)' + ) + parser.add_argument( + '--transforms', type=str, nargs='+', help='data augmentation' + ) + parser.add_argument( + '--root', type=str, default='', help='path to data root' + ) + parser.add_argument( + 'opts', + default=None, + nargs=argparse.REMAINDER, + help='Modify config options using the command-line' + ) + args = parser.parse_args() + + cfg = get_default_config() + cfg.use_gpu = torch.cuda.is_available() + if args.config_file: + cfg.merge_from_file(args.config_file) + reset_config(cfg, args) + cfg.merge_from_list(args.opts) + set_random_seed(cfg.train.seed) + + log_name = 'test.log' if cfg.test.evaluate else 'train.log' + log_name += time.strftime('-%Y-%m-%d-%H-%M-%S') + sys.stdout = Logger(osp.join(cfg.data.save_dir, log_name)) + + print('Show configuration\n{}\n'.format(cfg)) + print('Collecting env info ...') + print('** System info **\n{}\n'.format(collect_env_info())) + + if cfg.use_gpu: + torch.backends.cudnn.benchmark = True + + datamanager = torchreid.data.ImageDataManager(**imagedata_kwargs(cfg)) + + print('Building model-1: {}'.format(cfg.model.name)) + model1 = torchreid.models.build_model( + name=cfg.model.name, + num_classes=datamanager.num_train_pids, + loss=cfg.loss.name, + pretrained=cfg.model.pretrained, + use_gpu=cfg.use_gpu + ) + num_params, flops = compute_model_complexity( + model1, (1, 3, cfg.data.height, cfg.data.width) + ) + print('Model complexity: params={:,} flops={:,}'.format(num_params, flops)) + + print('Copying model-1 to model-2') + model2 = copy.deepcopy(model1) + + if cfg.model.load_weights1 and check_isfile(cfg.model.load_weights1): + load_pretrained_weights(model1, cfg.model.load_weights1) + + if cfg.model.load_weights2 and check_isfile(cfg.model.load_weights2): + load_pretrained_weights(model2, cfg.model.load_weights2) + + if cfg.use_gpu: + model1 = nn.DataParallel(model1).cuda() + model2 = nn.DataParallel(model2).cuda() + + optimizer1 = torchreid.optim.build_optimizer( + model1, **optimizer_kwargs(cfg) + ) + scheduler1 = torchreid.optim.build_lr_scheduler( + optimizer1, **lr_scheduler_kwargs(cfg) + ) + + optimizer2 = torchreid.optim.build_optimizer( + model2, **optimizer_kwargs(cfg) + ) + scheduler2 = torchreid.optim.build_lr_scheduler( + optimizer2, **lr_scheduler_kwargs(cfg) + ) + + if cfg.model.resume1 and check_isfile(cfg.model.resume1): + cfg.train.start_epoch = resume_from_checkpoint( + cfg.model.resume1, + model1, + optimizer=optimizer1, + scheduler=scheduler1 + ) + + if cfg.model.resume2 and check_isfile(cfg.model.resume2): + resume_from_checkpoint( + cfg.model.resume2, + model2, + optimizer=optimizer2, + scheduler=scheduler2 + ) + + print('Building DML-engine for image-reid') + engine = ImageDMLEngine( + datamanager, + model1, + optimizer1, + scheduler1, + model2, + optimizer2, + scheduler2, + margin=cfg.loss.triplet.margin, + weight_t=cfg.loss.triplet.weight_t, + weight_x=cfg.loss.triplet.weight_x, + weight_ml=cfg.loss.dml.weight_ml, + use_gpu=cfg.use_gpu, + label_smooth=cfg.loss.softmax.label_smooth, + deploy=cfg.model.deploy + ) + engine.run(**engine_run_kwargs(cfg)) + + +if __name__ == '__main__': + main() diff --git a/strong_sort/deep/reid/projects/OSNet_AIN/README.md b/strong_sort/deep/reid/projects/OSNet_AIN/README.md new file mode 100644 index 0000000..b313183 --- /dev/null +++ b/strong_sort/deep/reid/projects/OSNet_AIN/README.md @@ -0,0 +1,38 @@ +# Differentiable NAS for OSNet-AIN + +## Introduction +This repository contains the neural architecture search (NAS) code (based on [Torchreid](https://arxiv.org/abs/1910.10093)) for [OSNet-AIN](https://arxiv.org/abs/1910.06827), an extension of [OSNet](https://arxiv.org/abs/1905.00953) that achieves strong performance on cross-domain person re-identification (re-ID) benchmarks (*without using any target data*). OSNet-AIN builds on the idea of using [instance normalisation](https://arxiv.org/abs/1607.08022) (IN) layers to eliminate instance-specific contrast in images for domain-generalisable representation learning. This is inspired by the [neural style transfer](https://arxiv.org/abs/1703.06868) works that use IN to remove image styles. Though IN naturally suits the cross-domain person re-ID task, it still remains unclear that where to insert IN to a re-ID CNN can maximise the performance gain. To avoid exhaustively evaluating all possible designs, OSNet-AIN learns to search for the optimal OSNet+IN design from data using a differentiable NAS algorithm. For technical details, please refer to our paper at https://arxiv.org/abs/1910.06827. + +
+ +
+ +## Training +Assume the reid data is stored at `$DATA`. Run +``` +python main.py --config-file nas.yaml --root $DATA +``` + +The structure of the found architecture will be shown at the end of training. + +The default config was designed for 8 Tesla V100 32GB GPUs. You can modify the batch size based on your device memory. + +**Note** that the test result obtained at the end of architecture search is not meaningful (due to the stochastic sampling layers). Therefore, do not rely on the result to judge the model performance. Instead, you should construct the found architecture in `osnet_child.py` and re-train and evaluate the model on the reid datasets. + +## Citation +If you find this code useful to your research, please consider citing the following papers. +``` +@article{zhou2021osnet, + title={Learning Generalisable Omni-Scale Representations for Person Re-Identification}, + author={Zhou, Kaiyang and Yang, Yongxin and Cavallaro, Andrea and Xiang, Tao}, + journal={TPAMI}, + year={2021} +} + +@inproceedings{zhou2019osnet, + title={Omni-Scale Feature Learning for Person Re-Identification}, + author={Zhou, Kaiyang and Yang, Yongxin and Cavallaro, Andrea and Xiang, Tao}, + booktitle={ICCV}, + year={2019} +} +``` \ No newline at end of file diff --git a/strong_sort/deep/reid/projects/OSNet_AIN/default_config.py b/strong_sort/deep/reid/projects/OSNet_AIN/default_config.py new file mode 100644 index 0000000..733a9f0 --- /dev/null +++ b/strong_sort/deep/reid/projects/OSNet_AIN/default_config.py @@ -0,0 +1,210 @@ +from yacs.config import CfgNode as CN + + +def get_default_config(): + cfg = CN() + + # model + cfg.model = CN() + cfg.model.name = 'resnet50' + cfg.model.pretrained = True # automatically load pretrained model weights if available + cfg.model.load_weights = '' # path to model weights + cfg.model.resume = '' # path to checkpoint for resume training + + # NAS + cfg.nas = CN() + cfg.nas.mc_iter = 1 # Monte Carlo sampling + cfg.nas.init_lmda = 10. # initial lambda value + cfg.nas.min_lmda = 1. # minimum lambda value + cfg.nas.lmda_decay_step = 20 # decay step for lambda + cfg.nas.lmda_decay_rate = 0.5 # decay rate for lambda + cfg.nas.fixed_lmda = False # keep lambda unchanged + + # data + cfg.data = CN() + cfg.data.type = 'image' + cfg.data.root = 'reid-data' + cfg.data.sources = ['market1501'] + cfg.data.targets = ['market1501'] + cfg.data.workers = 4 # number of data loading workers + cfg.data.split_id = 0 # split index + cfg.data.height = 256 # image height + cfg.data.width = 128 # image width + cfg.data.combineall = False # combine train, query and gallery for training + cfg.data.transforms = ['random_flip'] # data augmentation + cfg.data.norm_mean = [0.485, 0.456, 0.406] # default is imagenet mean + cfg.data.norm_std = [0.229, 0.224, 0.225] # default is imagenet std + cfg.data.save_dir = 'log' # path to save log + + # specific datasets + cfg.market1501 = CN() + cfg.market1501.use_500k_distractors = False # add 500k distractors to the gallery set for market1501 + cfg.cuhk03 = CN() + cfg.cuhk03.labeled_images = False # use labeled images, if False, use detected images + cfg.cuhk03.classic_split = False # use classic split by Li et al. CVPR14 + cfg.cuhk03.use_metric_cuhk03 = False # use cuhk03's metric for evaluation + + # sampler + cfg.sampler = CN() + cfg.sampler.train_sampler = 'RandomSampler' + cfg.sampler.num_instances = 4 # number of instances per identity for RandomIdentitySampler + + # video reid setting + cfg.video = CN() + cfg.video.seq_len = 15 # number of images to sample in a tracklet + cfg.video.sample_method = 'evenly' # how to sample images from a tracklet + cfg.video.pooling_method = 'avg' # how to pool features over a tracklet + + # train + cfg.train = CN() + cfg.train.optim = 'adam' + cfg.train.lr = 0.0003 + cfg.train.weight_decay = 5e-4 + cfg.train.max_epoch = 60 + cfg.train.start_epoch = 0 + cfg.train.batch_size = 32 + cfg.train.fixbase_epoch = 0 # number of epochs to fix base layers + cfg.train.open_layers = [ + 'classifier' + ] # layers for training while keeping others frozen + cfg.train.staged_lr = False # set different lr to different layers + cfg.train.new_layers = ['classifier'] # newly added layers with default lr + cfg.train.base_lr_mult = 0.1 # learning rate multiplier for base layers + cfg.train.lr_scheduler = 'single_step' + cfg.train.stepsize = [20] # stepsize to decay learning rate + cfg.train.gamma = 0.1 # learning rate decay multiplier + cfg.train.print_freq = 20 # print frequency + cfg.train.seed = 1 # random seed + + # optimizer + cfg.sgd = CN() + cfg.sgd.momentum = 0.9 # momentum factor for sgd and rmsprop + cfg.sgd.dampening = 0. # dampening for momentum + cfg.sgd.nesterov = False # Nesterov momentum + cfg.rmsprop = CN() + cfg.rmsprop.alpha = 0.99 # smoothing constant + cfg.adam = CN() + cfg.adam.beta1 = 0.9 # exponential decay rate for first moment + cfg.adam.beta2 = 0.999 # exponential decay rate for second moment + + # loss + cfg.loss = CN() + cfg.loss.name = 'softmax' + cfg.loss.softmax = CN() + cfg.loss.softmax.label_smooth = True # use label smoothing regularizer + cfg.loss.triplet = CN() + cfg.loss.triplet.margin = 0.3 # distance margin + cfg.loss.triplet.weight_t = 1. # weight to balance hard triplet loss + cfg.loss.triplet.weight_x = 0. # weight to balance cross entropy loss + + # test + cfg.test = CN() + cfg.test.batch_size = 100 + cfg.test.dist_metric = 'euclidean' # distance metric, ['euclidean', 'cosine'] + cfg.test.normalize_feature = False # normalize feature vectors before computing distance + cfg.test.ranks = [1, 5, 10, 20] # cmc ranks + cfg.test.evaluate = False # test only + cfg.test.eval_freq = -1 # evaluation frequency (-1 means to only test after training) + cfg.test.start_eval = 0 # start to evaluate after a specific epoch + cfg.test.rerank = False # use person re-ranking + cfg.test.visrank = False # visualize ranked results (only available when cfg.test.evaluate=True) + cfg.test.visrank_topk = 10 # top-k ranks to visualize + cfg.test.visactmap = False # visualize CNN activation maps + + return cfg + + +def imagedata_kwargs(cfg): + return { + 'root': cfg.data.root, + 'sources': cfg.data.sources, + 'targets': cfg.data.targets, + 'height': cfg.data.height, + 'width': cfg.data.width, + 'transforms': cfg.data.transforms, + 'norm_mean': cfg.data.norm_mean, + 'norm_std': cfg.data.norm_std, + 'use_gpu': cfg.use_gpu, + 'split_id': cfg.data.split_id, + 'combineall': cfg.data.combineall, + 'batch_size_train': cfg.train.batch_size, + 'batch_size_test': cfg.test.batch_size, + 'workers': cfg.data.workers, + 'num_instances': cfg.sampler.num_instances, + 'train_sampler': cfg.sampler.train_sampler, + # image + 'cuhk03_labeled': cfg.cuhk03.labeled_images, + 'cuhk03_classic_split': cfg.cuhk03.classic_split, + 'market1501_500k': cfg.market1501.use_500k_distractors, + } + + +def videodata_kwargs(cfg): + return { + 'root': cfg.data.root, + 'sources': cfg.data.sources, + 'targets': cfg.data.targets, + 'height': cfg.data.height, + 'width': cfg.data.width, + 'transforms': cfg.data.transforms, + 'norm_mean': cfg.data.norm_mean, + 'norm_std': cfg.data.norm_std, + 'use_gpu': cfg.use_gpu, + 'split_id': cfg.data.split_id, + 'combineall': cfg.data.combineall, + 'batch_size_train': cfg.train.batch_size, + 'batch_size_test': cfg.test.batch_size, + 'workers': cfg.data.workers, + 'num_instances': cfg.sampler.num_instances, + 'train_sampler': cfg.sampler.train_sampler, + # video + 'seq_len': cfg.video.seq_len, + 'sample_method': cfg.video.sample_method + } + + +def optimizer_kwargs(cfg): + return { + 'optim': cfg.train.optim, + 'lr': cfg.train.lr, + 'weight_decay': cfg.train.weight_decay, + 'momentum': cfg.sgd.momentum, + 'sgd_dampening': cfg.sgd.dampening, + 'sgd_nesterov': cfg.sgd.nesterov, + 'rmsprop_alpha': cfg.rmsprop.alpha, + 'adam_beta1': cfg.adam.beta1, + 'adam_beta2': cfg.adam.beta2, + 'staged_lr': cfg.train.staged_lr, + 'new_layers': cfg.train.new_layers, + 'base_lr_mult': cfg.train.base_lr_mult + } + + +def lr_scheduler_kwargs(cfg): + return { + 'lr_scheduler': cfg.train.lr_scheduler, + 'stepsize': cfg.train.stepsize, + 'gamma': cfg.train.gamma, + 'max_epoch': cfg.train.max_epoch + } + + +def engine_run_kwargs(cfg): + return { + 'save_dir': cfg.data.save_dir, + 'max_epoch': cfg.train.max_epoch, + 'start_epoch': cfg.train.start_epoch, + 'fixbase_epoch': cfg.train.fixbase_epoch, + 'open_layers': cfg.train.open_layers, + 'start_eval': cfg.test.start_eval, + 'eval_freq': cfg.test.eval_freq, + 'test_only': cfg.test.evaluate, + 'print_freq': cfg.train.print_freq, + 'dist_metric': cfg.test.dist_metric, + 'normalize_feature': cfg.test.normalize_feature, + 'visrank': cfg.test.visrank, + 'visrank_topk': cfg.test.visrank_topk, + 'use_metric_cuhk03': cfg.cuhk03.use_metric_cuhk03, + 'ranks': cfg.test.ranks, + 'rerank': cfg.test.rerank + } diff --git a/strong_sort/deep/reid/projects/OSNet_AIN/main.py b/strong_sort/deep/reid/projects/OSNet_AIN/main.py new file mode 100644 index 0000000..f591770 --- /dev/null +++ b/strong_sort/deep/reid/projects/OSNet_AIN/main.py @@ -0,0 +1,145 @@ +import os +import sys +import time +import os.path as osp +import argparse +import torch +import torch.nn as nn + +import torchreid +from torchreid.utils import ( + Logger, check_isfile, set_random_seed, collect_env_info, + resume_from_checkpoint, compute_model_complexity +) + +import osnet_search as osnet_models +from softmax_nas import ImageSoftmaxNASEngine +from default_config import ( + imagedata_kwargs, optimizer_kwargs, engine_run_kwargs, get_default_config, + lr_scheduler_kwargs +) + + +def reset_config(cfg, args): + if args.root: + cfg.data.root = args.root + if args.sources: + cfg.data.sources = args.sources + if args.targets: + cfg.data.targets = args.targets + if args.transforms: + cfg.data.transforms = args.transforms + + +def main(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + parser.add_argument( + '--config-file', type=str, default='', help='path to config file' + ) + parser.add_argument( + '-s', + '--sources', + type=str, + nargs='+', + help='source datasets (delimited by space)' + ) + parser.add_argument( + '-t', + '--targets', + type=str, + nargs='+', + help='target datasets (delimited by space)' + ) + parser.add_argument( + '--transforms', type=str, nargs='+', help='data augmentation' + ) + parser.add_argument( + '--root', type=str, default='', help='path to data root' + ) + parser.add_argument( + '--gpu-devices', + type=str, + default='', + ) + parser.add_argument( + 'opts', + default=None, + nargs=argparse.REMAINDER, + help='Modify config options using the command-line' + ) + args = parser.parse_args() + + cfg = get_default_config() + cfg.use_gpu = torch.cuda.is_available() + if args.config_file: + cfg.merge_from_file(args.config_file) + reset_config(cfg, args) + cfg.merge_from_list(args.opts) + set_random_seed(cfg.train.seed) + + if cfg.use_gpu and args.gpu_devices: + # if gpu_devices is not specified, all available gpus will be used + os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_devices + log_name = 'test.log' if cfg.test.evaluate else 'train.log' + log_name += time.strftime('-%Y-%m-%d-%H-%M-%S') + sys.stdout = Logger(osp.join(cfg.data.save_dir, log_name)) + + print('Show configuration\n{}\n'.format(cfg)) + print('Collecting env info ...') + print('** System info **\n{}\n'.format(collect_env_info())) + + if cfg.use_gpu: + torch.backends.cudnn.benchmark = True + + datamanager = torchreid.data.ImageDataManager(**imagedata_kwargs(cfg)) + + print('Building model: {}'.format(cfg.model.name)) + model = osnet_models.build_model( + cfg.model.name, num_classes=datamanager.num_train_pids + ) + num_params, flops = compute_model_complexity( + model, (1, 3, cfg.data.height, cfg.data.width) + ) + print('Model complexity: params={:,} flops={:,}'.format(num_params, flops)) + + if cfg.use_gpu: + model = nn.DataParallel(model).cuda() + + optimizer = torchreid.optim.build_optimizer(model, **optimizer_kwargs(cfg)) + scheduler = torchreid.optim.build_lr_scheduler( + optimizer, **lr_scheduler_kwargs(cfg) + ) + + if cfg.model.resume and check_isfile(cfg.model.resume): + cfg.train.start_epoch = resume_from_checkpoint( + cfg.model.resume, model, optimizer=optimizer + ) + + print('Building NAS engine') + engine = ImageSoftmaxNASEngine( + datamanager, + model, + optimizer, + scheduler=scheduler, + use_gpu=cfg.use_gpu, + label_smooth=cfg.loss.softmax.label_smooth, + mc_iter=cfg.nas.mc_iter, + init_lmda=cfg.nas.init_lmda, + min_lmda=cfg.nas.min_lmda, + lmda_decay_step=cfg.nas.lmda_decay_step, + lmda_decay_rate=cfg.nas.lmda_decay_rate, + fixed_lmda=cfg.nas.fixed_lmda + ) + engine.run(**engine_run_kwargs(cfg)) + + print('*** Display the found architecture ***') + if cfg.use_gpu: + model.module.build_child_graph() + else: + model.build_child_graph() + + +if __name__ == '__main__': + main() diff --git a/strong_sort/deep/reid/projects/OSNet_AIN/nas.yaml b/strong_sort/deep/reid/projects/OSNet_AIN/nas.yaml new file mode 100644 index 0000000..507b2ec --- /dev/null +++ b/strong_sort/deep/reid/projects/OSNet_AIN/nas.yaml @@ -0,0 +1,44 @@ +model: + name: 'osnet_nas' + pretrained: False + +nas: + mc_iter: 1 + init_lmda: 10. + min_lmda: 1. + lmda_decay_step: 20 + lmda_decay_rate: 0.5 + fixed_lmda: False + +data: + type: 'image' + sources: ['msmt17'] + targets: ['market1501'] + height: 256 + width: 128 + combineall: True + transforms: ['random_flip', 'color_jitter'] + save_dir: 'log/osnet_nas' + +loss: + name: 'softmax' + softmax: + label_smooth: True + +train: + optim: 'sgd' + lr: 0.1 + max_epoch: 120 + batch_size: 512 + fixbase_epoch: 0 + open_layers: ['classifier'] + lr_scheduler: 'cosine' + +test: + batch_size: 300 + dist_metric: 'cosine' + normalize_feature: False + evaluate: False + eval_freq: -1 + rerank: False + visactmap: False \ No newline at end of file diff --git a/strong_sort/deep/reid/projects/OSNet_AIN/osnet_child.py b/strong_sort/deep/reid/projects/OSNet_AIN/osnet_child.py new file mode 100644 index 0000000..b47d747 --- /dev/null +++ b/strong_sort/deep/reid/projects/OSNet_AIN/osnet_child.py @@ -0,0 +1,535 @@ +from __future__ import division, absolute_import +from torch import nn +from torch.nn import functional as F + + +########## +# Basic layers +########## +class ConvLayer(nn.Module): + """Convolution layer (conv + bn + relu).""" + + def __init__( + self, + in_channels, + out_channels, + kernel_size, + stride=1, + padding=0, + groups=1, + IN=False + ): + super(ConvLayer, self).__init__() + self.conv = nn.Conv2d( + in_channels, + out_channels, + kernel_size, + stride=stride, + padding=padding, + bias=False, + groups=groups + ) + if IN: + self.bn = nn.InstanceNorm2d(out_channels, affine=True) + else: + self.bn = nn.BatchNorm2d(out_channels) + self.relu = nn.ReLU(inplace=True) + + def forward(self, x): + x = self.conv(x) + x = self.bn(x) + return self.relu(x) + + +class Conv1x1(nn.Module): + """1x1 convolution + bn + relu.""" + + def __init__(self, in_channels, out_channels, stride=1, groups=1): + super(Conv1x1, self).__init__() + self.conv = nn.Conv2d( + in_channels, + out_channels, + 1, + stride=stride, + padding=0, + bias=False, + groups=groups + ) + self.bn = nn.BatchNorm2d(out_channels) + self.relu = nn.ReLU(inplace=True) + + def forward(self, x): + x = self.conv(x) + x = self.bn(x) + return self.relu(x) + + +class Conv1x1Linear(nn.Module): + """1x1 convolution + bn (w/o non-linearity).""" + + def __init__(self, in_channels, out_channels, stride=1, bn=True): + super(Conv1x1Linear, self).__init__() + self.conv = nn.Conv2d( + in_channels, out_channels, 1, stride=stride, padding=0, bias=False + ) + self.bn = None + if bn: + self.bn = nn.BatchNorm2d(out_channels) + + def forward(self, x): + x = self.conv(x) + if self.bn is not None: + x = self.bn(x) + return x + + +class Conv3x3(nn.Module): + """3x3 convolution + bn + relu.""" + + def __init__(self, in_channels, out_channels, stride=1, groups=1): + super(Conv3x3, self).__init__() + self.conv = nn.Conv2d( + in_channels, + out_channels, + 3, + stride=stride, + padding=1, + bias=False, + groups=groups + ) + self.bn = nn.BatchNorm2d(out_channels) + self.relu = nn.ReLU(inplace=True) + + def forward(self, x): + x = self.conv(x) + x = self.bn(x) + return self.relu(x) + + +class LightConv3x3(nn.Module): + """Lightweight 3x3 convolution. + + 1x1 (linear) + dw 3x3 (nonlinear). + """ + + def __init__(self, in_channels, out_channels): + super(LightConv3x3, self).__init__() + self.conv1 = nn.Conv2d( + in_channels, out_channels, 1, stride=1, padding=0, bias=False + ) + self.conv2 = nn.Conv2d( + out_channels, + out_channels, + 3, + stride=1, + padding=1, + bias=False, + groups=out_channels + ) + self.bn = nn.BatchNorm2d(out_channels) + self.relu = nn.ReLU(inplace=True) + + def forward(self, x): + x = self.conv1(x) + x = self.conv2(x) + x = self.bn(x) + return self.relu(x) + + +class LightConvStream(nn.Module): + """Lightweight convolution stream.""" + + def __init__(self, in_channels, out_channels, depth): + super(LightConvStream, self).__init__() + assert depth >= 1, 'depth must be equal to or larger than 1, but got {}'.format( + depth + ) + layers = [] + layers += [LightConv3x3(in_channels, out_channels)] + for i in range(depth - 1): + layers += [LightConv3x3(out_channels, out_channels)] + self.layers = nn.Sequential(*layers) + + def forward(self, x): + return self.layers(x) + + +########## +# Building blocks for omni-scale feature learning +########## +class ChannelGate(nn.Module): + """A mini-network that generates channel-wise gates conditioned on input tensor.""" + + def __init__( + self, + in_channels, + num_gates=None, + return_gates=False, + gate_activation='sigmoid', + reduction=16, + layer_norm=False + ): + super(ChannelGate, self).__init__() + if num_gates is None: + num_gates = in_channels + self.return_gates = return_gates + self.global_avgpool = nn.AdaptiveAvgPool2d(1) + self.fc1 = nn.Conv2d( + in_channels, + in_channels // reduction, + kernel_size=1, + bias=True, + padding=0 + ) + self.norm1 = None + if layer_norm: + self.norm1 = nn.LayerNorm((in_channels // reduction, 1, 1)) + self.relu = nn.ReLU(inplace=True) + self.fc2 = nn.Conv2d( + in_channels // reduction, + num_gates, + kernel_size=1, + bias=True, + padding=0 + ) + if gate_activation == 'sigmoid': + self.gate_activation = nn.Sigmoid() + elif gate_activation == 'relu': + self.gate_activation = nn.ReLU(inplace=True) + elif gate_activation == 'linear': + self.gate_activation = None + else: + raise RuntimeError( + "Unknown gate activation: {}".format(gate_activation) + ) + + def forward(self, x): + input = x + x = self.global_avgpool(x) + x = self.fc1(x) + if self.norm1 is not None: + x = self.norm1(x) + x = self.relu(x) + x = self.fc2(x) + if self.gate_activation is not None: + x = self.gate_activation(x) + if self.return_gates: + return x + return input * x + + +class OSBlock(nn.Module): + """Omni-scale feature learning block.""" + + def __init__(self, in_channels, out_channels, reduction=4, T=4, **kwargs): + super(OSBlock, self).__init__() + assert T >= 1 + assert out_channels >= reduction and out_channels % reduction == 0 + mid_channels = out_channels // reduction + + self.conv1 = Conv1x1(in_channels, mid_channels) + self.conv2 = nn.ModuleList() + for t in range(1, T + 1): + self.conv2 += [LightConvStream(mid_channels, mid_channels, t)] + self.gate = ChannelGate(mid_channels) + self.conv3 = Conv1x1Linear(mid_channels, out_channels) + self.downsample = None + if in_channels != out_channels: + self.downsample = Conv1x1Linear(in_channels, out_channels) + + def forward(self, x): + identity = x + x1 = self.conv1(x) + x2 = 0 + for conv2_t in self.conv2: + x2_t = conv2_t(x1) + x2 = x2 + self.gate(x2_t) + x3 = self.conv3(x2) + if self.downsample is not None: + identity = self.downsample(identity) + out = x3 + identity + return F.relu(out) + + +class OSBlockINv1(nn.Module): + """Omni-scale feature learning block with instance normalization.""" + + def __init__(self, in_channels, out_channels, reduction=4, T=4, **kwargs): + super(OSBlockINv1, self).__init__() + assert T >= 1 + assert out_channels >= reduction and out_channels % reduction == 0 + mid_channels = out_channels // reduction + + self.conv1 = Conv1x1(in_channels, mid_channels) + self.conv2 = nn.ModuleList() + for t in range(1, T + 1): + self.conv2 += [LightConvStream(mid_channels, mid_channels, t)] + self.gate = ChannelGate(mid_channels) + self.conv3 = Conv1x1Linear(mid_channels, out_channels, bn=False) + self.downsample = None + if in_channels != out_channels: + self.downsample = Conv1x1Linear(in_channels, out_channels) + self.IN = nn.InstanceNorm2d(out_channels, affine=True) + + def forward(self, x): + identity = x + x1 = self.conv1(x) + x2 = 0 + for conv2_t in self.conv2: + x2_t = conv2_t(x1) + x2 = x2 + self.gate(x2_t) + x3 = self.conv3(x2) + x3 = self.IN(x3) # IN inside residual + if self.downsample is not None: + identity = self.downsample(identity) + out = x3 + identity + return F.relu(out) + + +class OSBlockINv2(nn.Module): + """Omni-scale feature learning block with instance normalization.""" + + def __init__(self, in_channels, out_channels, reduction=4, T=4, **kwargs): + super(OSBlockINv2, self).__init__() + assert T >= 1 + assert out_channels >= reduction and out_channels % reduction == 0 + mid_channels = out_channels // reduction + + self.conv1 = Conv1x1(in_channels, mid_channels) + self.conv2 = nn.ModuleList() + for t in range(1, T + 1): + self.conv2 += [LightConvStream(mid_channels, mid_channels, t)] + self.gate = ChannelGate(mid_channels) + self.conv3 = Conv1x1Linear(mid_channels, out_channels) + self.downsample = None + if in_channels != out_channels: + self.downsample = Conv1x1Linear(in_channels, out_channels) + self.IN = nn.InstanceNorm2d(out_channels, affine=True) + + def forward(self, x): + identity = x + x1 = self.conv1(x) + x2 = 0 + for conv2_t in self.conv2: + x2_t = conv2_t(x1) + x2 = x2 + self.gate(x2_t) + x3 = self.conv3(x2) + if self.downsample is not None: + identity = self.downsample(identity) + out = x3 + identity + out = self.IN(out) # IN outside residual + return F.relu(out) + + +class OSBlockINv3(nn.Module): + """Omni-scale feature learning block with instance normalization.""" + + def __init__(self, in_channels, out_channels, reduction=4, T=4, **kwargs): + super(OSBlockINv3, self).__init__() + assert T >= 1 + assert out_channels >= reduction and out_channels % reduction == 0 + mid_channels = out_channels // reduction + + self.conv1 = Conv1x1(in_channels, mid_channels) + self.conv2 = nn.ModuleList() + for t in range(1, T + 1): + self.conv2 += [LightConvStream(mid_channels, mid_channels, t)] + self.gate = ChannelGate(mid_channels) + self.conv3 = Conv1x1Linear(mid_channels, out_channels, bn=False) + self.downsample = None + if in_channels != out_channels: + self.downsample = Conv1x1Linear(in_channels, out_channels) + self.IN_in = nn.InstanceNorm2d(out_channels, affine=True) + self.IN_out = nn.InstanceNorm2d(out_channels, affine=True) + + def forward(self, x): + identity = x + x1 = self.conv1(x) + x2 = 0 + for conv2_t in self.conv2: + x2_t = conv2_t(x1) + x2 = x2 + self.gate(x2_t) + x3 = self.conv3(x2) + x3 = self.IN_in(x3) # IN inside residual + if self.downsample is not None: + identity = self.downsample(identity) + out = x3 + identity + out = self.IN_out(out) # IN outside residual + return F.relu(out) + + +########## +# Network architecture +########## +class OSNet(nn.Module): + """Omni-Scale Network. + + Reference: + - Zhou et al. Omni-Scale Feature Learning for Person Re-Identification. ICCV, 2019. + - Zhou et al. Learning Generalisable Omni-Scale Representations + for Person Re-Identification. TPAMI, 2021. + """ + + def __init__( + self, + num_classes, + blocks, + layers, + channels, + feature_dim=512, + loss='softmax', + conv1_IN=True, + **kwargs + ): + super(OSNet, self).__init__() + num_blocks = len(blocks) + assert num_blocks == len(layers) + assert num_blocks == len(channels) - 1 + self.loss = loss + self.feature_dim = feature_dim + + # convolutional backbone + self.conv1 = ConvLayer( + 3, channels[0], 7, stride=2, padding=3, IN=conv1_IN + ) + self.maxpool = nn.MaxPool2d(3, stride=2, padding=1) + self.conv2 = self._make_layer( + blocks[0], layers[0], channels[0], channels[1] + ) + self.pool2 = nn.Sequential( + Conv1x1(channels[1], channels[1]), nn.AvgPool2d(2, stride=2) + ) + self.conv3 = self._make_layer( + blocks[1], layers[1], channels[1], channels[2] + ) + self.pool3 = nn.Sequential( + Conv1x1(channels[2], channels[2]), nn.AvgPool2d(2, stride=2) + ) + self.conv4 = self._make_layer( + blocks[2], layers[2], channels[2], channels[3] + ) + self.conv5 = Conv1x1(channels[3], channels[3]) + self.global_avgpool = nn.AdaptiveAvgPool2d(1) + # fully connected layer + self.fc = self._construct_fc_layer( + self.feature_dim, channels[3], dropout_p=None + ) + # identity classification layer + self.classifier = nn.Linear(self.feature_dim, num_classes) + + self._init_params() + + def _make_layer(self, blocks, layer, in_channels, out_channels): + layers = [] + layers += [blocks[0](in_channels, out_channels)] + for i in range(1, len(blocks)): + layers += [blocks[i](out_channels, out_channels)] + return nn.Sequential(*layers) + + def _construct_fc_layer(self, fc_dims, input_dim, dropout_p=None): + if fc_dims is None or fc_dims < 0: + self.feature_dim = input_dim + return None + + if isinstance(fc_dims, int): + fc_dims = [fc_dims] + + layers = [] + for dim in fc_dims: + layers.append(nn.Linear(input_dim, dim)) + layers.append(nn.BatchNorm1d(dim)) + layers.append(nn.ReLU(inplace=True)) + if dropout_p is not None: + layers.append(nn.Dropout(p=dropout_p)) + input_dim = dim + + self.feature_dim = fc_dims[-1] + + return nn.Sequential(*layers) + + def _init_params(self): + for m in self.modules(): + if isinstance(m, nn.Conv2d): + nn.init.kaiming_normal_( + m.weight, mode='fan_out', nonlinearity='relu' + ) + if m.bias is not None: + nn.init.constant_(m.bias, 0) + + elif isinstance(m, nn.BatchNorm2d): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + + elif isinstance(m, nn.BatchNorm1d): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + + elif isinstance(m, nn.InstanceNorm2d): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + + elif isinstance(m, nn.Linear): + nn.init.normal_(m.weight, 0, 0.01) + if m.bias is not None: + nn.init.constant_(m.bias, 0) + + def featuremaps(self, x): + x = self.conv1(x) + x = self.maxpool(x) + x = self.conv2(x) + x = self.pool2(x) + x = self.conv3(x) + x = self.pool3(x) + x = self.conv4(x) + return self.conv5(x) + + def forward(self, x, return_featuremaps=False, **kwargs): + x = self.featuremaps(x) + if return_featuremaps: + return x + v = self.global_avgpool(x) + v = v.view(v.size(0), -1) + if self.fc is not None: + v = self.fc(v) + if not self.training: + return v + y = self.classifier(v) + if self.loss == 'softmax': + return y + elif self.loss == 'triplet': + return y, v + else: + raise KeyError("Unsupported loss: {}".format(self.loss)) + + +########## +# Instantiation +########## +def osnet_ain_x1_0( + num_classes=1000, pretrained=True, loss='softmax', **kwargs +): + model = OSNet( + num_classes, + blocks=[ + [OSBlockINv1, OSBlockINv1], [OSBlock, OSBlockINv1], + [OSBlockINv1, OSBlock] + ], + layers=[2, 2, 2], + channels=[64, 256, 384, 512], + loss=loss, + conv1_IN=True, + **kwargs + ) + return model + + +__models = {'osnet_ain_x1_0': osnet_ain_x1_0} + + +def build_model(name, num_classes=100): + avai_models = list(__models.keys()) + if name not in avai_models: + raise KeyError( + 'Unknown model: {}. Must be one of {}'.format(name, avai_models) + ) + return __models[name](num_classes=num_classes) diff --git a/strong_sort/deep/reid/projects/OSNet_AIN/osnet_search.py b/strong_sort/deep/reid/projects/OSNet_AIN/osnet_search.py new file mode 100644 index 0000000..1820144 --- /dev/null +++ b/strong_sort/deep/reid/projects/OSNet_AIN/osnet_search.py @@ -0,0 +1,584 @@ +from __future__ import division, absolute_import +import torch +from torch import nn +from torch.nn import functional as F + +EPS = 1e-12 +NORM_AFFINE = False # enable affine transformations for normalization layer + + +########## +# Basic layers +########## +class IBN(nn.Module): + """Instance + Batch Normalization.""" + + def __init__(self, num_channels): + super(IBN, self).__init__() + half1 = int(num_channels / 2) + self.half = half1 + half2 = num_channels - half1 + self.IN = nn.InstanceNorm2d(half1, affine=NORM_AFFINE) + self.BN = nn.BatchNorm2d(half2, affine=NORM_AFFINE) + + def forward(self, x): + split = torch.split(x, self.half, 1) + out1 = self.IN(split[0].contiguous()) + out2 = self.BN(split[1].contiguous()) + return torch.cat((out1, out2), 1) + + +class ConvLayer(nn.Module): + """Convolution layer (conv + bn + relu).""" + + def __init__( + self, + in_channels, + out_channels, + kernel_size, + stride=1, + padding=0, + groups=1, + IN=False + ): + super(ConvLayer, self).__init__() + self.conv = nn.Conv2d( + in_channels, + out_channels, + kernel_size, + stride=stride, + padding=padding, + bias=False, + groups=groups + ) + if IN: + self.bn = nn.InstanceNorm2d(out_channels, affine=NORM_AFFINE) + else: + self.bn = nn.BatchNorm2d(out_channels, affine=NORM_AFFINE) + self.relu = nn.ReLU(inplace=True) + + def forward(self, x): + x = self.conv(x) + x = self.bn(x) + return self.relu(x) + + +class Conv1x1(nn.Module): + """1x1 convolution + bn + relu.""" + + def __init__( + self, in_channels, out_channels, stride=1, groups=1, ibn=False + ): + super(Conv1x1, self).__init__() + self.conv = nn.Conv2d( + in_channels, + out_channels, + 1, + stride=stride, + padding=0, + bias=False, + groups=groups + ) + if ibn: + self.bn = IBN(out_channels) + else: + self.bn = nn.BatchNorm2d(out_channels, affine=NORM_AFFINE) + self.relu = nn.ReLU(inplace=True) + + def forward(self, x): + x = self.conv(x) + x = self.bn(x) + return self.relu(x) + + +class Conv1x1Linear(nn.Module): + """1x1 convolution + bn (w/o non-linearity).""" + + def __init__(self, in_channels, out_channels, stride=1, bn=True): + super(Conv1x1Linear, self).__init__() + self.conv = nn.Conv2d( + in_channels, out_channels, 1, stride=stride, padding=0, bias=False + ) + self.bn = None + if bn: + self.bn = nn.BatchNorm2d(out_channels, affine=NORM_AFFINE) + + def forward(self, x): + x = self.conv(x) + if self.bn is not None: + x = self.bn(x) + return x + + +class Conv3x3(nn.Module): + """3x3 convolution + bn + relu.""" + + def __init__(self, in_channels, out_channels, stride=1, groups=1): + super(Conv3x3, self).__init__() + self.conv = nn.Conv2d( + in_channels, + out_channels, + 3, + stride=stride, + padding=1, + bias=False, + groups=groups + ) + self.bn = nn.BatchNorm2d(out_channels, affine=NORM_AFFINE) + self.relu = nn.ReLU(inplace=True) + + def forward(self, x): + x = self.conv(x) + x = self.bn(x) + return self.relu(x) + + +class LightConv3x3(nn.Module): + """Lightweight 3x3 convolution. + + 1x1 (linear) + dw 3x3 (nonlinear). + """ + + def __init__(self, in_channels, out_channels): + super(LightConv3x3, self).__init__() + self.conv1 = nn.Conv2d( + in_channels, out_channels, 1, stride=1, padding=0, bias=False + ) + self.conv2 = nn.Conv2d( + out_channels, + out_channels, + 3, + stride=1, + padding=1, + bias=False, + groups=out_channels + ) + self.bn = nn.BatchNorm2d(out_channels, affine=NORM_AFFINE) + self.relu = nn.ReLU(inplace=True) + + def forward(self, x): + x = self.conv1(x) + x = self.conv2(x) + x = self.bn(x) + return self.relu(x) + + +class LightConvStream(nn.Module): + """Lightweight convolution stream.""" + + def __init__(self, in_channels, out_channels, depth): + super(LightConvStream, self).__init__() + assert depth >= 1, 'depth must be equal to or larger than 1, but got {}'.format( + depth + ) + layers = [] + layers += [LightConv3x3(in_channels, out_channels)] + for i in range(depth - 1): + layers += [LightConv3x3(out_channels, out_channels)] + self.layers = nn.Sequential(*layers) + + def forward(self, x): + return self.layers(x) + + +########## +# Building blocks for omni-scale feature learning +########## +class ChannelGate(nn.Module): + """A mini-network that generates channel-wise gates conditioned on input tensor.""" + + def __init__( + self, + in_channels, + num_gates=None, + return_gates=False, + gate_activation='sigmoid', + reduction=16, + layer_norm=False + ): + super(ChannelGate, self).__init__() + if num_gates is None: + num_gates = in_channels + self.return_gates = return_gates + self.global_avgpool = nn.AdaptiveAvgPool2d(1) + self.fc1 = nn.Conv2d( + in_channels, + in_channels // reduction, + kernel_size=1, + bias=True, + padding=0 + ) + self.norm1 = None + if layer_norm: + self.norm1 = nn.LayerNorm((in_channels // reduction, 1, 1)) + self.relu = nn.ReLU(inplace=True) + self.fc2 = nn.Conv2d( + in_channels // reduction, + num_gates, + kernel_size=1, + bias=True, + padding=0 + ) + if gate_activation == 'sigmoid': + self.gate_activation = nn.Sigmoid() + elif gate_activation == 'relu': + self.gate_activation = nn.ReLU(inplace=True) + elif gate_activation == 'linear': + self.gate_activation = None + else: + raise RuntimeError( + "Unknown gate activation: {}".format(gate_activation) + ) + + def forward(self, x): + input = x + x = self.global_avgpool(x) + x = self.fc1(x) + if self.norm1 is not None: + x = self.norm1(x) + x = self.relu(x) + x = self.fc2(x) + if self.gate_activation is not None: + x = self.gate_activation(x) + if self.return_gates: + return x + return input * x + + +class OSBlock(nn.Module): + """Omni-scale feature learning block.""" + + def __init__(self, in_channels, out_channels, reduction=4, T=4, **kwargs): + super(OSBlock, self).__init__() + assert T >= 1 + assert out_channels >= reduction and out_channels % reduction == 0 + mid_channels = out_channels // reduction + + self.conv1 = Conv1x1(in_channels, mid_channels) + self.conv2 = nn.ModuleList() + for t in range(1, T + 1): + self.conv2 += [LightConvStream(mid_channels, mid_channels, t)] + self.gate = ChannelGate(mid_channels) + self.conv3 = Conv1x1Linear(mid_channels, out_channels) + self.downsample = None + if in_channels != out_channels: + self.downsample = Conv1x1Linear(in_channels, out_channels) + + def forward(self, x): + identity = x + x1 = self.conv1(x) + x2 = 0 + for conv2_t in self.conv2: + x2_t = conv2_t(x1) + x2 = x2 + self.gate(x2_t) + x3 = self.conv3(x2) + if self.downsample is not None: + identity = self.downsample(identity) + out = x3 + identity + return F.relu(out) + + +class OSBlockINv1(nn.Module): + """Omni-scale feature learning block with instance normalization.""" + + def __init__(self, in_channels, out_channels, reduction=4, T=4, **kwargs): + super(OSBlockINv1, self).__init__() + assert T >= 1 + assert out_channels >= reduction and out_channels % reduction == 0 + mid_channels = out_channels // reduction + + self.conv1 = Conv1x1(in_channels, mid_channels) + self.conv2 = nn.ModuleList() + for t in range(1, T + 1): + self.conv2 += [LightConvStream(mid_channels, mid_channels, t)] + self.gate = ChannelGate(mid_channels) + self.conv3 = Conv1x1Linear(mid_channels, out_channels, bn=False) + self.downsample = None + if in_channels != out_channels: + self.downsample = Conv1x1Linear(in_channels, out_channels) + self.IN = nn.InstanceNorm2d(out_channels, affine=NORM_AFFINE) + + def forward(self, x): + identity = x + x1 = self.conv1(x) + x2 = 0 + for conv2_t in self.conv2: + x2_t = conv2_t(x1) + x2 = x2 + self.gate(x2_t) + x3 = self.conv3(x2) + x3 = self.IN(x3) # IN inside residual + if self.downsample is not None: + identity = self.downsample(identity) + out = x3 + identity + return F.relu(out) + + +class OSBlockINv2(nn.Module): + """Omni-scale feature learning block with instance normalization.""" + + def __init__(self, in_channels, out_channels, reduction=4, T=4, **kwargs): + super(OSBlockINv2, self).__init__() + assert T >= 1 + assert out_channels >= reduction and out_channels % reduction == 0 + mid_channels = out_channels // reduction + + self.conv1 = Conv1x1(in_channels, mid_channels) + self.conv2 = nn.ModuleList() + for t in range(1, T + 1): + self.conv2 += [LightConvStream(mid_channels, mid_channels, t)] + self.gate = ChannelGate(mid_channels) + self.conv3 = Conv1x1Linear(mid_channels, out_channels) + self.downsample = None + if in_channels != out_channels: + self.downsample = Conv1x1Linear(in_channels, out_channels) + self.IN = nn.InstanceNorm2d(out_channels, affine=NORM_AFFINE) + + def forward(self, x): + identity = x + x1 = self.conv1(x) + x2 = 0 + for conv2_t in self.conv2: + x2_t = conv2_t(x1) + x2 = x2 + self.gate(x2_t) + x3 = self.conv3(x2) + if self.downsample is not None: + identity = self.downsample(identity) + out = x3 + identity + out = self.IN(out) # IN outside residual + return F.relu(out) + + +class OSBlockINv3(nn.Module): + """Omni-scale feature learning block with instance normalization.""" + + def __init__(self, in_channels, out_channels, reduction=4, T=4, **kwargs): + super(OSBlockINv3, self).__init__() + assert T >= 1 + assert out_channels >= reduction and out_channels % reduction == 0 + mid_channels = out_channels // reduction + + self.conv1 = Conv1x1(in_channels, mid_channels) + self.conv2 = nn.ModuleList() + for t in range(1, T + 1): + self.conv2 += [LightConvStream(mid_channels, mid_channels, t)] + self.gate = ChannelGate(mid_channels) + self.conv3 = Conv1x1Linear(mid_channels, out_channels, bn=False) + self.downsample = None + if in_channels != out_channels: + self.downsample = Conv1x1Linear(in_channels, out_channels) + self.IN_in = nn.InstanceNorm2d(out_channels, affine=NORM_AFFINE) + self.IN_out = nn.InstanceNorm2d(out_channels, affine=NORM_AFFINE) + + def forward(self, x): + identity = x + x1 = self.conv1(x) + x2 = 0 + for conv2_t in self.conv2: + x2_t = conv2_t(x1) + x2 = x2 + self.gate(x2_t) + x3 = self.conv3(x2) + x3 = self.IN_in(x3) # inside residual + if self.downsample is not None: + identity = self.downsample(identity) + out = x3 + identity + out = self.IN_out(out) # IN outside residual + return F.relu(out) + + +class NASBlock(nn.Module): + """Neural architecture search layer.""" + + def __init__(self, in_channels, out_channels, search_space=None): + super(NASBlock, self).__init__() + self._is_child_graph = False + self.search_space = search_space + if self.search_space is None: + raise ValueError('search_space is None') + + self.os_block = nn.ModuleList() + for block in self.search_space: + self.os_block += [block(in_channels, out_channels)] + self.weights = nn.Parameter(torch.ones(len(self.search_space))) + + def build_child_graph(self): + if self._is_child_graph: + raise RuntimeError('build_child_graph() can only be called once') + + idx = self.weights.data.max(dim=0)[1].item() + self.os_block = self.os_block[idx] + self.weights = None + self._is_child_graph = True + return self.search_space[idx] + + def forward(self, x, lmda=1.): + if self._is_child_graph: + return self.os_block(x) + + uniform = torch.rand_like(self.weights) + gumbel = -torch.log(-torch.log(uniform + EPS)) + nonneg_weights = F.relu(self.weights) + logits = torch.log(nonneg_weights + EPS) + gumbel + exp = torch.exp(logits / lmda) + weights_softmax = exp / (exp.sum() + EPS) + + output = 0 + for i, weight in enumerate(weights_softmax): + output = output + weight * self.os_block[i](x) + return output + + +########## +# Network architecture +########## +class OSNet(nn.Module): + """Omni-Scale Network. + + Reference: + - Zhou et al. Omni-Scale Feature Learning for Person Re-Identification. ICCV, 2019. + - Zhou et al. Learning Generalisable Omni-Scale Representations + for Person Re-Identification. TPAMI, 2021. + """ + + def __init__( + self, + num_classes, + blocks, + layers, + channels, + feature_dim=512, + loss='softmax', + search_space=None, + **kwargs + ): + super(OSNet, self).__init__() + num_blocks = len(blocks) + assert num_blocks == len(layers) + assert num_blocks == len(channels) - 1 + # no matter what loss is specified, the model only returns the ID predictions + self.loss = loss + self.feature_dim = feature_dim + + # convolutional backbone + self.conv1 = ConvLayer(3, channels[0], 7, stride=2, padding=3, IN=True) + self.maxpool = nn.MaxPool2d(3, stride=2, padding=1) + self.conv2 = self._make_layer( + blocks[0], layers[0], channels[0], channels[1], search_space + ) + self.pool2 = nn.Sequential( + Conv1x1(channels[1], channels[1]), nn.AvgPool2d(2, stride=2) + ) + self.conv3 = self._make_layer( + blocks[1], layers[1], channels[1], channels[2], search_space + ) + self.pool3 = nn.Sequential( + Conv1x1(channels[2], channels[2]), nn.AvgPool2d(2, stride=2) + ) + self.conv4 = self._make_layer( + blocks[2], layers[2], channels[2], channels[3], search_space + ) + self.conv5 = Conv1x1(channels[3], channels[3]) + self.global_avgpool = nn.AdaptiveAvgPool2d(1) + # fully connected layer + self.fc = self._construct_fc_layer( + self.feature_dim, channels[3], dropout_p=None + ) + # identity classification layer + self.classifier = nn.Linear(self.feature_dim, num_classes) + + def _make_layer( + self, block, layer, in_channels, out_channels, search_space + ): + layers = nn.ModuleList() + layers += [block(in_channels, out_channels, search_space=search_space)] + for i in range(1, layer): + layers += [ + block(out_channels, out_channels, search_space=search_space) + ] + return layers + + def _construct_fc_layer(self, fc_dims, input_dim, dropout_p=None): + if fc_dims is None or fc_dims < 0: + self.feature_dim = input_dim + return None + + if isinstance(fc_dims, int): + fc_dims = [fc_dims] + + layers = [] + for dim in fc_dims: + layers.append(nn.Linear(input_dim, dim)) + layers.append(nn.BatchNorm1d(dim, affine=NORM_AFFINE)) + layers.append(nn.ReLU(inplace=True)) + if dropout_p is not None: + layers.append(nn.Dropout(p=dropout_p)) + input_dim = dim + + self.feature_dim = fc_dims[-1] + + return nn.Sequential(*layers) + + def build_child_graph(self): + print('Building child graph') + for i, conv in enumerate(self.conv2): + block = conv.build_child_graph() + print('- conv2-{} Block={}'.format(i + 1, block.__name__)) + for i, conv in enumerate(self.conv3): + block = conv.build_child_graph() + print('- conv3-{} Block={}'.format(i + 1, block.__name__)) + for i, conv in enumerate(self.conv4): + block = conv.build_child_graph() + print('- conv4-{} Block={}'.format(i + 1, block.__name__)) + + def featuremaps(self, x, lmda): + x = self.conv1(x) + x = self.maxpool(x) + for conv in self.conv2: + x = conv(x, lmda) + x = self.pool2(x) + for conv in self.conv3: + x = conv(x, lmda) + x = self.pool3(x) + for conv in self.conv4: + x = conv(x, lmda) + return self.conv5(x) + + def forward(self, x, lmda=1., return_featuremaps=False): + # lmda (float): temperature parameter for concrete distribution + x = self.featuremaps(x, lmda) + if return_featuremaps: + return x + v = self.global_avgpool(x) + v = v.view(v.size(0), -1) + if self.fc is not None: + v = self.fc(v) + if not self.training: + return v + return self.classifier(v) + + +########## +# Instantiation +########## +def osnet_nas(num_classes=1000, loss='softmax', **kwargs): + # standard size (width x1.0) + return OSNet( + num_classes, + blocks=[NASBlock, NASBlock, NASBlock], + layers=[2, 2, 2], + channels=[64, 256, 384, 512], + loss=loss, + search_space=[OSBlock, OSBlockINv1, OSBlockINv2, OSBlockINv3], + **kwargs + ) + + +__NAS_models = {'osnet_nas': osnet_nas} + + +def build_model(name, num_classes=100): + avai_models = list(__NAS_models.keys()) + if name not in avai_models: + raise KeyError( + 'Unknown model: {}. Must be one of {}'.format(name, avai_models) + ) + return __NAS_models[name](num_classes=num_classes) diff --git a/strong_sort/deep/reid/projects/OSNet_AIN/softmax_nas.py b/strong_sort/deep/reid/projects/OSNet_AIN/softmax_nas.py new file mode 100644 index 0000000..e2a03b6 --- /dev/null +++ b/strong_sort/deep/reid/projects/OSNet_AIN/softmax_nas.py @@ -0,0 +1,73 @@ +from __future__ import division, print_function, absolute_import + +from torchreid import metrics +from torchreid.engine import Engine +from torchreid.losses import CrossEntropyLoss + + +class ImageSoftmaxNASEngine(Engine): + + def __init__( + self, + datamanager, + model, + optimizer, + scheduler=None, + use_gpu=False, + label_smooth=True, + mc_iter=1, + init_lmda=1., + min_lmda=1., + lmda_decay_step=20, + lmda_decay_rate=0.5, + fixed_lmda=False + ): + super(ImageSoftmaxNASEngine, self).__init__(datamanager, use_gpu) + self.mc_iter = mc_iter + self.init_lmda = init_lmda + self.min_lmda = min_lmda + self.lmda_decay_step = lmda_decay_step + self.lmda_decay_rate = lmda_decay_rate + self.fixed_lmda = fixed_lmda + + self.model = model + self.optimizer = optimizer + self.scheduler = scheduler + self.register_model('model', model, optimizer, scheduler) + + self.criterion = CrossEntropyLoss( + num_classes=self.datamanager.num_train_pids, + use_gpu=self.use_gpu, + label_smooth=label_smooth + ) + + def forward_backward(self, data): + imgs, pids = self.parse_data_for_train(data) + + if self.use_gpu: + imgs = imgs.cuda() + pids = pids.cuda() + + # softmax temporature + if self.fixed_lmda or self.lmda_decay_step == -1: + lmda = self.init_lmda + else: + lmda = self.init_lmda * self.lmda_decay_rate**( + self.epoch // self.lmda_decay_step + ) + if lmda < self.min_lmda: + lmda = self.min_lmda + + for k in range(self.mc_iter): + outputs = self.model(imgs, lmda=lmda) + loss = self.compute_loss(self.criterion, outputs, pids) + self.optimizer.zero_grad() + loss.backward() + self.optimizer.step() + + loss_dict = { + 'loss': loss.item(), + 'acc': metrics.accuracy(outputs, pids)[0].item() + } + + return loss_dict diff --git a/strong_sort/deep/reid/projects/README.md b/strong_sort/deep/reid/projects/README.md new file mode 100644 index 0000000..fa63919 --- /dev/null +++ b/strong_sort/deep/reid/projects/README.md @@ -0,0 +1,5 @@ +Here are some research projects built on [Torchreid](https://arxiv.org/abs/1910.10093). + ++ `OSNet_AIN`: [Learning Generalisable Omni-Scale Representations for Person Re-Identification](https://arxiv.org/abs/1910.06827) ++ `DML`: [Deep Mutual Learning (CVPR'18)](https://arxiv.org/abs/1706.00384) ++ `attribute_recognition`: [Omni-Scale Feature Learning for Person Re-Identification (ICCV'19)](https://arxiv.org/abs/1905.00953) \ No newline at end of file diff --git a/strong_sort/deep/reid/projects/attribute_recognition/README.md b/strong_sort/deep/reid/projects/attribute_recognition/README.md new file mode 100644 index 0000000..a20b6e7 --- /dev/null +++ b/strong_sort/deep/reid/projects/attribute_recognition/README.md @@ -0,0 +1,18 @@ +# Person Attribute Recognition +This code was developed for the experiment of person attribute recognition in [Omni-Scale Feature Learning for Person Re-Identification (ICCV'19)](https://arxiv.org/abs/1905.00953). + +## Download data +Download the PA-100K dataset from [https://github.com/xh-liu/HydraPlus-Net](https://github.com/xh-liu/HydraPlus-Net), and extract the file under the folder where you store your data (say $DATASET). The folder structure should look like +```bash +$DATASET/ + pa100k/ + data/ # images + annotation/ + annotation.mat +``` + +## Train +The training command is provided in `train.sh`. Run `bash train.sh $DATASET` to start training. + +## Test +To test a pretrained model, add the following two arguments to `train.sh`: `--load-weights $PATH_TO_WEIGHTS --evaluate`. \ No newline at end of file diff --git a/strong_sort/deep/reid/projects/attribute_recognition/datasets/__init__.py b/strong_sort/deep/reid/projects/attribute_recognition/datasets/__init__.py new file mode 100644 index 0000000..664598c --- /dev/null +++ b/strong_sort/deep/reid/projects/attribute_recognition/datasets/__init__.py @@ -0,0 +1,15 @@ +from __future__ import division, print_function, absolute_import + +from .pa100k import PA100K + +__datasets = {'pa100k': PA100K} + + +def init_dataset(name, **kwargs): + avai_datasets = list(__datasets.keys()) + if name not in avai_datasets: + raise ValueError( + 'Invalid dataset name. Received "{}", ' + 'but expected to be one of {}'.format(name, avai_datasets) + ) + return __datasets[name](**kwargs) diff --git a/strong_sort/deep/reid/projects/attribute_recognition/datasets/dataset.py b/strong_sort/deep/reid/projects/attribute_recognition/datasets/dataset.py new file mode 100644 index 0000000..7ce4ebf --- /dev/null +++ b/strong_sort/deep/reid/projects/attribute_recognition/datasets/dataset.py @@ -0,0 +1,87 @@ +from __future__ import division, print_function, absolute_import +import os.path as osp + +from torchreid.utils import read_image + + +class Dataset(object): + + def __init__( + self, + train, + val, + test, + attr_dict, + transform=None, + mode='train', + verbose=True, + **kwargs + ): + self.train = train + self.val = val + self.test = test + self._attr_dict = attr_dict + self._num_attrs = len(self.attr_dict) + self.transform = transform + + if mode == 'train': + self.data = self.train + elif mode == 'val': + self.data = self.val + else: + self.data = self.test + + if verbose: + self.show_summary() + + @property + def num_attrs(self): + return self._num_attrs + + @property + def attr_dict(self): + return self._attr_dict + + def __len__(self): + return len(self.data) + + def __getitem__(self, index): + img_path, attrs = self.data[index] + img = read_image(img_path) + if self.transform is not None: + img = self.transform(img) + return img, attrs, img_path + + def check_before_run(self, required_files): + """Checks if required files exist before going deeper. + Args: + required_files (str or list): string file name(s). + """ + if isinstance(required_files, str): + required_files = [required_files] + + for fpath in required_files: + if not osp.exists(fpath): + raise RuntimeError('"{}" is not found'.format(fpath)) + + def show_summary(self): + num_train = len(self.train) + num_val = len(self.val) + num_test = len(self.test) + num_total = num_train + num_val + num_test + + print('=> Loaded {}'.format(self.__class__.__name__)) + print(" ------------------------------") + print(" subset | # images") + print(" ------------------------------") + print(" train | {:8d}".format(num_train)) + print(" val | {:8d}".format(num_val)) + print(" test | {:8d}".format(num_test)) + print(" ------------------------------") + print(" total | {:8d}".format(num_total)) + print(" ------------------------------") + print(" # attributes: {}".format(len(self.attr_dict))) + print(" attributes:") + for label, attr in self.attr_dict.items(): + print(' {:3d}: {}'.format(label, attr)) + print(" ------------------------------") diff --git a/strong_sort/deep/reid/projects/attribute_recognition/datasets/pa100k.py b/strong_sort/deep/reid/projects/attribute_recognition/datasets/pa100k.py new file mode 100644 index 0000000..61dd26c --- /dev/null +++ b/strong_sort/deep/reid/projects/attribute_recognition/datasets/pa100k.py @@ -0,0 +1,59 @@ +from __future__ import division, print_function, absolute_import +import numpy as np +import os.path as osp +from scipy.io import loadmat + +from .dataset import Dataset + + +class PA100K(Dataset): + """Pedestrian attribute dataset. + + 80k training images + 20k test images. + + The folder structure should be: + pa100k/ + data/ # images + annotation/ + annotation.mat + """ + dataset_dir = 'pa100k' + + def __init__(self, root='', **kwargs): + self.root = osp.abspath(osp.expanduser(root)) + self.dataset_dir = osp.join(self.root, self.dataset_dir) + self.data_dir = osp.join(self.dataset_dir, 'data') + self.anno_mat_path = osp.join( + self.dataset_dir, 'annotation', 'annotation.mat' + ) + + required_files = [self.data_dir, self.anno_mat_path] + self.check_before_run(required_files) + + train, val, test, attr_dict = self.extract_data() + super(PA100K, self).__init__(train, val, test, attr_dict, **kwargs) + + def extract_data(self): + # anno_mat is a dictionary with keys: ['test_images_name', 'val_images_name', + # 'train_images_name', 'val_label', 'attributes', 'test_label', 'train_label'] + anno_mat = loadmat(self.anno_mat_path) + + def _extract(key_name, key_label): + names = anno_mat[key_name] + labels = anno_mat[key_label] + num_imgs = names.shape[0] + data = [] + for i in range(num_imgs): + name = names[i, 0][0] + attrs = labels[i, :].astype(np.float32) + img_path = osp.join(self.data_dir, name) + data.append((img_path, attrs)) + return data + + train = _extract('train_images_name', 'train_label') + val = _extract('val_images_name', 'val_label') + test = _extract('test_images_name', 'test_label') + attrs = anno_mat['attributes'] + attr_dict = {i: str(attr[0][0]) for i, attr in enumerate(attrs)} + + return train, val, test, attr_dict diff --git a/strong_sort/deep/reid/projects/attribute_recognition/default_parser.py b/strong_sort/deep/reid/projects/attribute_recognition/default_parser.py new file mode 100644 index 0000000..d19a8be --- /dev/null +++ b/strong_sort/deep/reid/projects/attribute_recognition/default_parser.py @@ -0,0 +1,243 @@ +from __future__ import print_function, absolute_import +import argparse + + +def init_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + # ************************************************************ + # Datasets + # ************************************************************ + parser.add_argument( + '--root', + type=str, + default='', + required=True, + help='root path to data directory' + ) + parser.add_argument( + '-d', + '--dataset', + type=str, + required=True, + help='which dataset to choose' + ) + parser.add_argument( + '-j', + '--workers', + type=int, + default=4, + help='number of data loading workers (tips: 4 or 8 times number of gpus)' + ) + parser.add_argument( + '--height', type=int, default=256, help='height of an image' + ) + parser.add_argument( + '--width', type=int, default=128, help='width of an image' + ) + + # ************************************************************ + # Optimization options + # ************************************************************ + parser.add_argument( + '--optim', + type=str, + default='adam', + help='optimization algorithm (see optimizers.py)' + ) + parser.add_argument( + '--lr', type=float, default=0.0003, help='initial learning rate' + ) + parser.add_argument( + '--weight-decay', type=float, default=5e-04, help='weight decay' + ) + # sgd + parser.add_argument( + '--momentum', + type=float, + default=0.9, + help='momentum factor for sgd and rmsprop' + ) + parser.add_argument( + '--sgd-dampening', + type=float, + default=0, + help='sgd\'s dampening for momentum' + ) + parser.add_argument( + '--sgd-nesterov', + action='store_true', + help='whether to enable sgd\'s Nesterov momentum' + ) + # rmsprop + parser.add_argument( + '--rmsprop-alpha', + type=float, + default=0.99, + help='rmsprop\'s smoothing constant' + ) + # adam/amsgrad + parser.add_argument( + '--adam-beta1', + type=float, + default=0.9, + help='exponential decay rate for adam\'s first moment' + ) + parser.add_argument( + '--adam-beta2', + type=float, + default=0.999, + help='exponential decay rate for adam\'s second moment' + ) + + # ************************************************************ + # Training hyperparameters + # ************************************************************ + parser.add_argument( + '--max-epoch', type=int, default=60, help='maximum epochs to run' + ) + parser.add_argument( + '--start-epoch', + type=int, + default=0, + help='manual epoch number (useful when restart)' + ) + parser.add_argument( + '--batch-size', type=int, default=32, help='batch size' + ) + + parser.add_argument( + '--fixbase-epoch', + type=int, + default=0, + help='number of epochs to fix base layers' + ) + parser.add_argument( + '--open-layers', + type=str, + nargs='+', + default=['classifier'], + help='open specified layers for training while keeping others frozen' + ) + + parser.add_argument( + '--staged-lr', + action='store_true', + help='set different lr to different layers' + ) + parser.add_argument( + '--new-layers', + type=str, + nargs='+', + default=['classifier'], + help='newly added layers with default lr' + ) + parser.add_argument( + '--base-lr-mult', + type=float, + default=0.1, + help='learning rate multiplier for base layers' + ) + + # ************************************************************ + # Learning rate scheduler options + # ************************************************************ + parser.add_argument( + '--lr-scheduler', + type=str, + default='multi_step', + help='learning rate scheduler (see lr_schedulers.py)' + ) + parser.add_argument( + '--stepsize', + type=int, + default=[20, 40], + nargs='+', + help='stepsize to decay learning rate' + ) + parser.add_argument( + '--gamma', type=float, default=0.1, help='learning rate decay' + ) + + # ************************************************************ + # Architecture + # ************************************************************ + parser.add_argument( + '-a', '--arch', type=str, default='', help='model architecture' + ) + parser.add_argument( + '--no-pretrained', + action='store_true', + help='do not load pretrained weights' + ) + + # ************************************************************ + # Loss + # ************************************************************ + parser.add_argument( + '--weighted-bce', action='store_true', help='use weighted BCELoss' + ) + + # ************************************************************ + # Test settings + # ************************************************************ + parser.add_argument( + '--load-weights', type=str, default='', help='load pretrained weights' + ) + parser.add_argument( + '--evaluate', action='store_true', help='evaluate only' + ) + parser.add_argument( + '--save-prediction', action='store_true', help='save prediction' + ) + + # ************************************************************ + # Miscs + # ************************************************************ + parser.add_argument( + '--print-freq', type=int, default=20, help='print frequency' + ) + parser.add_argument('--seed', type=int, default=1, help='manual seed') + parser.add_argument( + '--resume', + type=str, + default='', + metavar='PATH', + help='resume from a checkpoint' + ) + parser.add_argument( + '--save-dir', + type=str, + default='log', + help='path to save log and model weights' + ) + parser.add_argument('--use-cpu', action='store_true', help='use cpu') + + return parser + + +def optimizer_kwargs(parsed_args): + return { + 'optim': parsed_args.optim, + 'lr': parsed_args.lr, + 'weight_decay': parsed_args.weight_decay, + 'momentum': parsed_args.momentum, + 'sgd_dampening': parsed_args.sgd_dampening, + 'sgd_nesterov': parsed_args.sgd_nesterov, + 'rmsprop_alpha': parsed_args.rmsprop_alpha, + 'adam_beta1': parsed_args.adam_beta1, + 'adam_beta2': parsed_args.adam_beta2, + 'staged_lr': parsed_args.staged_lr, + 'new_layers': parsed_args.new_layers, + 'base_lr_mult': parsed_args.base_lr_mult + } + + +def lr_scheduler_kwargs(parsed_args): + return { + 'lr_scheduler': parsed_args.lr_scheduler, + 'stepsize': parsed_args.stepsize, + 'gamma': parsed_args.gamma + } diff --git a/strong_sort/deep/reid/projects/attribute_recognition/main.py b/strong_sort/deep/reid/projects/attribute_recognition/main.py new file mode 100644 index 0000000..1fb4238 --- /dev/null +++ b/strong_sort/deep/reid/projects/attribute_recognition/main.py @@ -0,0 +1,399 @@ +from __future__ import division, print_function +import sys +import copy +import time +import numpy as np +import os.path as osp +import datetime +import warnings +import torch +import torch.nn as nn + +import torchreid +from torchreid.utils import ( + Logger, AverageMeter, check_isfile, open_all_layers, save_checkpoint, + set_random_seed, collect_env_info, open_specified_layers, + load_pretrained_weights, compute_model_complexity +) +from torchreid.data.transforms import ( + Resize, Compose, ToTensor, Normalize, Random2DTranslation, + RandomHorizontalFlip +) + +import models +import datasets +from default_parser import init_parser, optimizer_kwargs, lr_scheduler_kwargs + +parser = init_parser() +args = parser.parse_args() + + +def init_dataset(use_gpu): + normalize = Normalize( + mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225] + ) + + transform_tr = Compose( + [ + Random2DTranslation(args.height, args.width, p=0.5), + RandomHorizontalFlip(), + ToTensor(), normalize + ] + ) + + transform_te = Compose( + [Resize([args.height, args.width]), + ToTensor(), normalize] + ) + + trainset = datasets.init_dataset( + args.dataset, + root=args.root, + transform=transform_tr, + mode='train', + verbose=True + ) + + valset = datasets.init_dataset( + args.dataset, + root=args.root, + transform=transform_te, + mode='val', + verbose=False + ) + + testset = datasets.init_dataset( + args.dataset, + root=args.root, + transform=transform_te, + mode='test', + verbose=False + ) + + num_attrs = trainset.num_attrs + attr_dict = trainset.attr_dict + + trainloader = torch.utils.data.DataLoader( + trainset, + batch_size=args.batch_size, + shuffle=True, + num_workers=args.workers, + pin_memory=use_gpu, + drop_last=True + ) + + valloader = torch.utils.data.DataLoader( + valset, + batch_size=args.batch_size, + shuffle=False, + num_workers=args.workers, + pin_memory=use_gpu, + drop_last=False + ) + + testloader = torch.utils.data.DataLoader( + testset, + batch_size=args.batch_size, + shuffle=False, + num_workers=args.workers, + pin_memory=use_gpu, + drop_last=False + ) + + return trainloader, valloader, testloader, num_attrs, attr_dict + + +def main(): + global args + + set_random_seed(args.seed) + use_gpu = torch.cuda.is_available() and not args.use_cpu + log_name = 'test.log' if args.evaluate else 'train.log' + sys.stdout = Logger(osp.join(args.save_dir, log_name)) + + print('** Arguments **') + arg_keys = list(args.__dict__.keys()) + arg_keys.sort() + for key in arg_keys: + print('{}: {}'.format(key, args.__dict__[key])) + print('\n') + print('Collecting env info ...') + print('** System info **\n{}\n'.format(collect_env_info())) + + if use_gpu: + torch.backends.cudnn.benchmark = True + else: + warnings.warn( + 'Currently using CPU, however, GPU is highly recommended' + ) + + dataset_vars = init_dataset(use_gpu) + trainloader, valloader, testloader, num_attrs, attr_dict = dataset_vars + + if args.weighted_bce: + print('Use weighted binary cross entropy') + print('Computing the weights ...') + bce_weights = torch.zeros(num_attrs, dtype=torch.float) + for _, attrs, _ in trainloader: + bce_weights += attrs.sum(0) # sum along the batch dim + bce_weights /= len(trainloader) * args.batch_size + print('Sample ratio for each attribute: {}'.format(bce_weights)) + bce_weights = torch.exp(-1 * bce_weights) + print('BCE weights: {}'.format(bce_weights)) + bce_weights = bce_weights.expand(args.batch_size, num_attrs) + criterion = nn.BCEWithLogitsLoss(weight=bce_weights) + + else: + print('Use plain binary cross entropy') + criterion = nn.BCEWithLogitsLoss() + + print('Building model: {}'.format(args.arch)) + model = models.build_model( + args.arch, + num_attrs, + pretrained=not args.no_pretrained, + use_gpu=use_gpu + ) + num_params, flops = compute_model_complexity( + model, (1, 3, args.height, args.width) + ) + print('Model complexity: params={:,} flops={:,}'.format(num_params, flops)) + + if args.load_weights and check_isfile(args.load_weights): + load_pretrained_weights(model, args.load_weights) + + if use_gpu: + model = nn.DataParallel(model).cuda() + criterion = criterion.cuda() + + if args.evaluate: + test(model, testloader, attr_dict, use_gpu) + return + + optimizer = torchreid.optim.build_optimizer( + model, **optimizer_kwargs(args) + ) + scheduler = torchreid.optim.build_lr_scheduler( + optimizer, **lr_scheduler_kwargs(args) + ) + + start_epoch = args.start_epoch + best_result = -np.inf + if args.resume and check_isfile(args.resume): + checkpoint = torch.load(args.resume) + model.load_state_dict(checkpoint['state_dict']) + optimizer.load_state_dict(checkpoint['optimizer']) + start_epoch = checkpoint['epoch'] + best_result = checkpoint['label_mA'] + print('Loaded checkpoint from "{}"'.format(args.resume)) + print('- start epoch: {}'.format(start_epoch)) + print('- label_mA: {}'.format(best_result)) + + time_start = time.time() + + for epoch in range(start_epoch, args.max_epoch): + train( + epoch, model, criterion, optimizer, scheduler, trainloader, use_gpu + ) + test_outputs = test(model, testloader, attr_dict, use_gpu) + label_mA = test_outputs[0] + is_best = label_mA > best_result + if is_best: + best_result = label_mA + + save_checkpoint( + { + 'state_dict': model.state_dict(), + 'epoch': epoch + 1, + 'label_mA': label_mA, + 'optimizer': optimizer.state_dict(), + }, + args.save_dir, + is_best=is_best + ) + + elapsed = round(time.time() - time_start) + elapsed = str(datetime.timedelta(seconds=elapsed)) + print('Elapsed {}'.format(elapsed)) + + +def train(epoch, model, criterion, optimizer, scheduler, trainloader, use_gpu): + losses = AverageMeter() + batch_time = AverageMeter() + data_time = AverageMeter() + model.train() + + if (epoch + 1) <= args.fixbase_epoch and args.open_layers is not None: + print( + '* Only train {} (epoch: {}/{})'.format( + args.open_layers, epoch + 1, args.fixbase_epoch + ) + ) + open_specified_layers(model, args.open_layers) + else: + open_all_layers(model) + + end = time.time() + for batch_idx, data in enumerate(trainloader): + data_time.update(time.time() - end) + + imgs, attrs = data[0], data[1] + if use_gpu: + imgs = imgs.cuda() + attrs = attrs.cuda() + + optimizer.zero_grad() + outputs = model(imgs) + loss = criterion(outputs, attrs) + loss.backward() + optimizer.step() + + batch_time.update(time.time() - end) + + losses.update(loss.item(), imgs.size(0)) + + if (batch_idx+1) % args.print_freq == 0: + # estimate remaining time + num_batches = len(trainloader) + eta_seconds = batch_time.avg * ( + num_batches - (batch_idx+1) + (args.max_epoch - + (epoch+1)) * num_batches + ) + eta_str = str(datetime.timedelta(seconds=int(eta_seconds))) + print( + 'Epoch: [{0}/{1}][{2}/{3}]\t' + 'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t' + 'Data {data_time.val:.3f} ({data_time.avg:.3f})\t' + 'Loss {loss.val:.4f} ({loss.avg:.4f})\t' + 'Lr {lr:.6f}\t' + 'Eta {eta}'.format( + epoch + 1, + args.max_epoch, + batch_idx + 1, + len(trainloader), + batch_time=batch_time, + data_time=data_time, + loss=losses, + lr=optimizer.param_groups[0]['lr'], + eta=eta_str + ) + ) + + end = time.time() + + scheduler.step() + + +@torch.no_grad() +def test(model, testloader, attr_dict, use_gpu): + batch_time = AverageMeter() + model.eval() + + num_persons = 0 + prob_thre = 0.5 + ins_acc = 0 + ins_prec = 0 + ins_rec = 0 + mA_history = { + 'correct_pos': 0, + 'real_pos': 0, + 'correct_neg': 0, + 'real_neg': 0 + } + + print('Testing ...') + + for batch_idx, data in enumerate(testloader): + imgs, attrs, img_paths = data + if use_gpu: + imgs = imgs.cuda() + + end = time.time() + orig_outputs = model(imgs) + batch_time.update(time.time() - end) + + orig_outputs = orig_outputs.data.cpu().numpy() + attrs = attrs.data.numpy() + + # transform raw outputs to attributes (binary codes) + outputs = copy.deepcopy(orig_outputs) + outputs[outputs < prob_thre] = 0 + outputs[outputs >= prob_thre] = 1 + + # compute label-based metric + overlaps = outputs * attrs + mA_history['correct_pos'] += overlaps.sum(0) + mA_history['real_pos'] += attrs.sum(0) + inv_overlaps = (1-outputs) * (1-attrs) + mA_history['correct_neg'] += inv_overlaps.sum(0) + mA_history['real_neg'] += (1 - attrs).sum(0) + + outputs = outputs.astype(bool) + attrs = attrs.astype(bool) + + # compute instabce-based accuracy + intersect = (outputs & attrs).astype(float) + union = (outputs | attrs).astype(float) + ins_acc += (intersect.sum(1) / union.sum(1)).sum() + ins_prec += (intersect.sum(1) / outputs.astype(float).sum(1)).sum() + ins_rec += (intersect.sum(1) / attrs.astype(float).sum(1)).sum() + + num_persons += imgs.size(0) + + if (batch_idx+1) % args.print_freq == 0: + print( + 'Processed batch {}/{}'.format(batch_idx + 1, len(testloader)) + ) + + if args.save_prediction: + txtfile = open(osp.join(args.save_dir, 'prediction.txt'), 'a') + for idx in range(imgs.size(0)): + img_path = img_paths[idx] + probs = orig_outputs[idx, :] + labels = attrs[idx, :] + txtfile.write('{}\n'.format(img_path)) + txtfile.write('*** Correct prediction ***\n') + for attr_idx, (label, prob) in enumerate(zip(labels, probs)): + if label: + attr_name = attr_dict[attr_idx] + info = '{}: {:.1%} '.format(attr_name, prob) + txtfile.write(info) + txtfile.write('\n*** Incorrect prediction ***\n') + for attr_idx, (label, prob) in enumerate(zip(labels, probs)): + if not label and prob > 0.5: + attr_name = attr_dict[attr_idx] + info = '{}: {:.1%} '.format(attr_name, prob) + txtfile.write(info) + txtfile.write('\n\n') + txtfile.close() + + print( + '=> BatchTime(s)/BatchSize(img): {:.4f}/{}'.format( + batch_time.avg, args.batch_size + ) + ) + + ins_acc /= num_persons + ins_prec /= num_persons + ins_rec /= num_persons + ins_f1 = (2*ins_prec*ins_rec) / (ins_prec+ins_rec) + + term1 = mA_history['correct_pos'] / mA_history['real_pos'] + term2 = mA_history['correct_neg'] / mA_history['real_neg'] + label_mA_verbose = (term1+term2) * 0.5 + label_mA = label_mA_verbose.mean() + + print('* Results *') + print(' # test persons: {}'.format(num_persons)) + print(' (instance-based) accuracy: {:.1%}'.format(ins_acc)) + print(' (instance-based) precition: {:.1%}'.format(ins_prec)) + print(' (instance-based) recall: {:.1%}'.format(ins_rec)) + print(' (instance-based) f1-score: {:.1%}'.format(ins_f1)) + print(' (label-based) mean accuracy: {:.1%}'.format(label_mA)) + print(' mA for each attribute: {}'.format(label_mA_verbose)) + + return label_mA, ins_acc, ins_prec, ins_rec, ins_f1 + + +if __name__ == '__main__': + main() diff --git a/strong_sort/deep/reid/projects/attribute_recognition/models/__init__.py b/strong_sort/deep/reid/projects/attribute_recognition/models/__init__.py new file mode 100644 index 0000000..ff0f0ed --- /dev/null +++ b/strong_sort/deep/reid/projects/attribute_recognition/models/__init__.py @@ -0,0 +1,17 @@ +from __future__ import absolute_import + +from .osnet import * + +__model_factory = { + 'osnet_avgpool': osnet_avgpool, + 'osnet_maxpool': osnet_maxpool +} + + +def build_model(name, num_classes, pretrained=True, use_gpu=True): + avai_models = list(__model_factory.keys()) + if name not in avai_models: + raise KeyError + return __model_factory[name]( + num_classes=num_classes, pretrained=pretrained, use_gpu=use_gpu + ) diff --git a/strong_sort/deep/reid/projects/attribute_recognition/models/osnet.py b/strong_sort/deep/reid/projects/attribute_recognition/models/osnet.py new file mode 100644 index 0000000..12569da --- /dev/null +++ b/strong_sort/deep/reid/projects/attribute_recognition/models/osnet.py @@ -0,0 +1,414 @@ +from __future__ import division, absolute_import +import torch +from torch import nn +from torch.nn import functional as F + +__all__ = ['osnet_avgpool', 'osnet_maxpool'] + + +########## +# Basic layers +########## +class ConvLayer(nn.Module): + """Convolution layer.""" + + def __init__( + self, + in_channels, + out_channels, + kernel_size, + stride=1, + padding=0, + groups=1 + ): + super(ConvLayer, self).__init__() + self.conv = nn.Conv2d( + in_channels, + out_channels, + kernel_size, + stride=stride, + padding=padding, + bias=False, + groups=groups + ) + self.bn = nn.BatchNorm2d(out_channels) + self.relu = nn.ReLU(inplace=True) + + def forward(self, x): + x = self.conv(x) + x = self.bn(x) + x = self.relu(x) + return x + + +class Conv1x1(nn.Module): + """1x1 convolution.""" + + def __init__(self, in_channels, out_channels, stride=1, groups=1): + super(Conv1x1, self).__init__() + self.conv = nn.Conv2d( + in_channels, + out_channels, + 1, + stride=stride, + padding=0, + bias=False, + groups=groups + ) + self.bn = nn.BatchNorm2d(out_channels) + self.relu = nn.ReLU(inplace=True) + + def forward(self, x): + x = self.conv(x) + x = self.bn(x) + x = self.relu(x) + return x + + +class Conv1x1Linear(nn.Module): + """1x1 convolution without non-linearity.""" + + def __init__(self, in_channels, out_channels, stride=1): + super(Conv1x1Linear, self).__init__() + self.conv = nn.Conv2d( + in_channels, out_channels, 1, stride=stride, padding=0, bias=False + ) + self.bn = nn.BatchNorm2d(out_channels) + + def forward(self, x): + x = self.conv(x) + x = self.bn(x) + return x + + +class Conv3x3(nn.Module): + """3x3 convolution.""" + + def __init__(self, in_channels, out_channels, stride=1, groups=1): + super(Conv3x3, self).__init__() + self.conv = nn.Conv2d( + in_channels, + out_channels, + 3, + stride=stride, + padding=1, + bias=False, + groups=groups + ) + self.bn = nn.BatchNorm2d(out_channels) + self.relu = nn.ReLU(inplace=True) + + def forward(self, x): + x = self.conv(x) + x = self.bn(x) + x = self.relu(x) + return x + + +class LightConv3x3(nn.Module): + """Lightweight 3x3 convolution. + + 1x1 (linear) + dw 3x3 (nonlinear). + """ + + def __init__(self, in_channels, out_channels): + super(LightConv3x3, self).__init__() + self.conv1 = nn.Conv2d( + in_channels, out_channels, 1, stride=1, padding=0, bias=False + ) + self.conv2 = nn.Conv2d( + out_channels, + out_channels, + 3, + stride=1, + padding=1, + bias=False, + groups=out_channels + ) + self.bn = nn.BatchNorm2d(out_channels) + self.relu = nn.ReLU(inplace=True) + + def forward(self, x): + x = self.conv1(x) + x = self.conv2(x) + x = self.bn(x) + x = self.relu(x) + return x + + +########## +# Building blocks for omni-scale feature learning +########## +class ChannelGate(nn.Module): + """A mini-network that generates channel-wise gates conditioned on input.""" + + def __init__( + self, + in_channels, + num_gates=None, + return_gates=False, + gate_activation='sigmoid', + reduction=16, + layer_norm=False + ): + super(ChannelGate, self).__init__() + if num_gates is None: + num_gates = in_channels + self.return_gates = return_gates + self.global_avgpool = nn.AdaptiveAvgPool2d(1) + self.fc1 = nn.Conv2d( + in_channels, + in_channels // reduction, + kernel_size=1, + bias=True, + padding=0 + ) + self.norm1 = None + if layer_norm: + self.norm1 = nn.LayerNorm((in_channels // reduction, 1, 1)) + self.relu = nn.ReLU(inplace=True) + self.fc2 = nn.Conv2d( + in_channels // reduction, + num_gates, + kernel_size=1, + bias=True, + padding=0 + ) + if gate_activation == 'sigmoid': + self.gate_activation = nn.Sigmoid() + elif gate_activation == 'relu': + self.gate_activation = nn.ReLU(inplace=True) + elif gate_activation == 'linear': + self.gate_activation = None + else: + raise RuntimeError( + "Unknown gate activation: {}".format(gate_activation) + ) + + def forward(self, x): + input = x + x = self.global_avgpool(x) + x = self.fc1(x) + if self.norm1 is not None: + x = self.norm1(x) + x = self.relu(x) + x = self.fc2(x) + if self.gate_activation is not None: + x = self.gate_activation(x) + if self.return_gates: + return x + return input * x + + +class OSBlock(nn.Module): + """Omni-scale feature learning block.""" + + def __init__(self, in_channels, out_channels, **kwargs): + super(OSBlock, self).__init__() + mid_channels = out_channels // 4 + self.conv1 = Conv1x1(in_channels, mid_channels) + self.conv2a = LightConv3x3(mid_channels, mid_channels) + self.conv2b = nn.Sequential( + LightConv3x3(mid_channels, mid_channels), + LightConv3x3(mid_channels, mid_channels), + ) + self.conv2c = nn.Sequential( + LightConv3x3(mid_channels, mid_channels), + LightConv3x3(mid_channels, mid_channels), + LightConv3x3(mid_channels, mid_channels), + ) + self.conv2d = nn.Sequential( + LightConv3x3(mid_channels, mid_channels), + LightConv3x3(mid_channels, mid_channels), + LightConv3x3(mid_channels, mid_channels), + LightConv3x3(mid_channels, mid_channels), + ) + self.gate = ChannelGate(mid_channels) + self.conv3 = Conv1x1Linear(mid_channels, out_channels) + self.downsample = None + if in_channels != out_channels: + self.downsample = Conv1x1Linear(in_channels, out_channels) + + def forward(self, x): + residual = x + x1 = self.conv1(x) + x2a = self.conv2a(x1) + x2b = self.conv2b(x1) + x2c = self.conv2c(x1) + x2d = self.conv2d(x1) + x2 = self.gate(x2a) + self.gate(x2b) + self.gate(x2c) + self.gate(x2d) + x3 = self.conv3(x2) + if self.downsample is not None: + residual = self.downsample(residual) + out = x3 + residual + return F.relu(out) + + +########## +# Network architecture +########## +class BaseNet(nn.Module): + + def _make_layer( + self, block, layer, in_channels, out_channels, reduce_spatial_size + ): + layers = [] + + layers.append(block(in_channels, out_channels)) + for i in range(1, layer): + layers.append(block(out_channels, out_channels)) + + if reduce_spatial_size: + layers.append( + nn.Sequential( + Conv1x1(out_channels, out_channels), + nn.AvgPool2d(2, stride=2) + ) + ) + + return nn.Sequential(*layers) + + def _construct_fc_layer(self, fc_dims, input_dim, dropout_p=None): + if fc_dims is None or fc_dims < 0: + self.feature_dim = input_dim + return None + + if isinstance(fc_dims, int): + fc_dims = [fc_dims] + + layers = [] + for dim in fc_dims: + layers.append(nn.Linear(input_dim, dim)) + layers.append(nn.BatchNorm1d(dim)) + layers.append(nn.ReLU(inplace=True)) + if dropout_p is not None: + layers.append(nn.Dropout(p=dropout_p)) + input_dim = dim + + self.feature_dim = fc_dims[-1] + + return nn.Sequential(*layers) + + def init_params(self): + for m in self.modules(): + if isinstance(m, nn.Conv2d): + nn.init.kaiming_normal_( + m.weight, mode='fan_out', nonlinearity='relu' + ) + if m.bias is not None: + nn.init.constant_(m.bias, 0) + + elif isinstance(m, nn.BatchNorm2d): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + + elif isinstance(m, nn.BatchNorm1d): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + + elif isinstance(m, nn.Linear): + nn.init.normal_(m.weight, 0, 0.01) + if m.bias is not None: + nn.init.constant_(m.bias, 0) + + +class OSNet(BaseNet): + + def __init__( + self, + num_classes, + blocks, + layers, + channels, + feature_dim=512, + loss='softmax', + pool='avg', + **kwargs + ): + super(OSNet, self).__init__() + num_blocks = len(blocks) + assert num_blocks == len(layers) + assert num_blocks == len(channels) - 1 + self.loss = loss + + # convolutional backbone + self.conv1 = ConvLayer(3, channels[0], 7, stride=2, padding=3) + self.maxpool = nn.MaxPool2d(3, stride=2, padding=1) + self.conv2 = self._make_layer( + blocks[0], + layers[0], + channels[0], + channels[1], + reduce_spatial_size=True + ) + self.conv3 = self._make_layer( + blocks[1], + layers[1], + channels[1], + channels[2], + reduce_spatial_size=True + ) + self.conv4 = self._make_layer( + blocks[2], + layers[2], + channels[2], + channels[3], + reduce_spatial_size=False + ) + self.conv5 = Conv1x1(channels[3], channels[3]) + if pool == 'avg': + self.global_pool = nn.AdaptiveAvgPool2d(1) + else: + self.global_pool = nn.AdaptiveMaxPool2d(1) + # fully connected layer + self.fc = self._construct_fc_layer( + feature_dim, channels[3], dropout_p=None + ) + # classification layer + self.classifier = nn.Linear(self.feature_dim, num_classes) + + self.init_params() + + def featuremaps(self, x): + x = self.conv1(x) + x = self.maxpool(x) + x = self.conv2(x) + x = self.conv3(x) + x = self.conv4(x) + x = self.conv5(x) + return x + + def forward(self, x): + x = self.featuremaps(x) + v = self.global_pool(x) + v = v.view(v.size(0), -1) + if self.fc is not None: + v = self.fc(v) + y = self.classifier(v) + if not self.training: + y = torch.sigmoid(y) + return y + + +def osnet_avgpool(num_classes=1000, loss='softmax', **kwargs): + return OSNet( + num_classes, + blocks=[OSBlock, OSBlock, OSBlock], + layers=[2, 2, 2], + channels=[64, 256, 384, 512], + loss=loss, + pool='avg', + **kwargs + ) + + +def osnet_maxpool(num_classes=1000, loss='softmax', **kwargs): + return OSNet( + num_classes, + blocks=[OSBlock, OSBlock, OSBlock], + layers=[2, 2, 2], + channels=[64, 256, 384, 512], + loss=loss, + pool='max', + **kwargs + ) diff --git a/strong_sort/deep/reid/projects/attribute_recognition/train.sh b/strong_sort/deep/reid/projects/attribute_recognition/train.sh new file mode 100644 index 0000000..9080a66 --- /dev/null +++ b/strong_sort/deep/reid/projects/attribute_recognition/train.sh @@ -0,0 +1,16 @@ +#!/bin/bash + +# DATASET points to the directory containing pa100k/ +DATASET=$1 + +python main.py \ +--root ${DATASET} \ +-d pa100k \ +-a osnet_maxpool \ +--max-epoch 50 \ +--stepsize 30 40 \ +--batch-size 32 \ +--lr 0.065 \ +--optim sgd \ +--weighted-bce \ +--save-dir log/pa100k-osnet_maxpool \ No newline at end of file diff --git a/strong_sort/deep/reid/scripts/default_config.py b/strong_sort/deep/reid/scripts/default_config.py new file mode 100644 index 0000000..da448e3 --- /dev/null +++ b/strong_sort/deep/reid/scripts/default_config.py @@ -0,0 +1,212 @@ +from yacs.config import CfgNode as CN + + +def get_default_config(): + cfg = CN() + + # model + cfg.model = CN() + cfg.model.name = 'resnet50' + cfg.model.pretrained = True # automatically load pretrained model weights if available + cfg.model.load_weights = '' # path to model weights + cfg.model.resume = '' # path to checkpoint for resume training + + # data + cfg.data = CN() + cfg.data.type = 'image' + cfg.data.root = 'reid-data' + cfg.data.sources = ['market1501'] + cfg.data.targets = ['market1501'] + cfg.data.workers = 4 # number of data loading workers + cfg.data.split_id = 0 # split index + cfg.data.height = 256 # image height + cfg.data.width = 128 # image width + cfg.data.combineall = False # combine train, query and gallery for training + cfg.data.transforms = ['random_flip'] # data augmentation + cfg.data.k_tfm = 1 # number of times to apply augmentation to an image independently + cfg.data.norm_mean = [0.485, 0.456, 0.406] # default is imagenet mean + cfg.data.norm_std = [0.229, 0.224, 0.225] # default is imagenet std + cfg.data.save_dir = 'log' # path to save log + cfg.data.load_train_targets = False # load training set from target dataset + + # specific datasets + cfg.market1501 = CN() + cfg.market1501.use_500k_distractors = False # add 500k distractors to the gallery set for market1501 + cfg.cuhk03 = CN() + cfg.cuhk03.labeled_images = False # use labeled images, if False, use detected images + cfg.cuhk03.classic_split = False # use classic split by Li et al. CVPR14 + cfg.cuhk03.use_metric_cuhk03 = False # use cuhk03's metric for evaluation + + # sampler + cfg.sampler = CN() + cfg.sampler.train_sampler = 'RandomSampler' # sampler for source train loader + cfg.sampler.train_sampler_t = 'RandomSampler' # sampler for target train loader + cfg.sampler.num_instances = 4 # number of instances per identity for RandomIdentitySampler + cfg.sampler.num_cams = 1 # number of cameras to sample in a batch (for RandomDomainSampler) + cfg.sampler.num_datasets = 1 # number of datasets to sample in a batch (for RandomDatasetSampler) + + # video reid setting + cfg.video = CN() + cfg.video.seq_len = 15 # number of images to sample in a tracklet + cfg.video.sample_method = 'evenly' # how to sample images from a tracklet + cfg.video.pooling_method = 'avg' # how to pool features over a tracklet + + # train + cfg.train = CN() + cfg.train.optim = 'adam' + cfg.train.lr = 0.0003 + cfg.train.weight_decay = 5e-4 + cfg.train.max_epoch = 60 + cfg.train.start_epoch = 0 + cfg.train.batch_size = 32 + cfg.train.fixbase_epoch = 0 # number of epochs to fix base layers + cfg.train.open_layers = [ + 'classifier' + ] # layers for training while keeping others frozen + cfg.train.staged_lr = False # set different lr to different layers + cfg.train.new_layers = ['classifier'] # newly added layers with default lr + cfg.train.base_lr_mult = 0.1 # learning rate multiplier for base layers + cfg.train.lr_scheduler = 'single_step' + cfg.train.stepsize = [20] # stepsize to decay learning rate + cfg.train.gamma = 0.1 # learning rate decay multiplier + cfg.train.print_freq = 20 # print frequency + cfg.train.seed = 1 # random seed + + # optimizer + cfg.sgd = CN() + cfg.sgd.momentum = 0.9 # momentum factor for sgd and rmsprop + cfg.sgd.dampening = 0. # dampening for momentum + cfg.sgd.nesterov = False # Nesterov momentum + cfg.rmsprop = CN() + cfg.rmsprop.alpha = 0.99 # smoothing constant + cfg.adam = CN() + cfg.adam.beta1 = 0.9 # exponential decay rate for first moment + cfg.adam.beta2 = 0.999 # exponential decay rate for second moment + + # loss + cfg.loss = CN() + cfg.loss.name = 'softmax' + cfg.loss.softmax = CN() + cfg.loss.softmax.label_smooth = True # use label smoothing regularizer + cfg.loss.triplet = CN() + cfg.loss.triplet.margin = 0.3 # distance margin + cfg.loss.triplet.weight_t = 1. # weight to balance hard triplet loss + cfg.loss.triplet.weight_x = 0. # weight to balance cross entropy loss + + # test + cfg.test = CN() + cfg.test.batch_size = 100 + cfg.test.dist_metric = 'euclidean' # distance metric, ['euclidean', 'cosine'] + cfg.test.normalize_feature = False # normalize feature vectors before computing distance + cfg.test.ranks = [1, 5, 10, 20] # cmc ranks + cfg.test.evaluate = False # test only + cfg.test.eval_freq = -1 # evaluation frequency (-1 means to only test after training) + cfg.test.start_eval = 0 # start to evaluate after a specific epoch + cfg.test.rerank = False # use person re-ranking + cfg.test.visrank = False # visualize ranked results (only available when cfg.test.evaluate=True) + cfg.test.visrank_topk = 10 # top-k ranks to visualize + + return cfg + + +def imagedata_kwargs(cfg): + return { + 'root': cfg.data.root, + 'sources': cfg.data.sources, + 'targets': cfg.data.targets, + 'height': cfg.data.height, + 'width': cfg.data.width, + 'transforms': cfg.data.transforms, + 'k_tfm': cfg.data.k_tfm, + 'norm_mean': cfg.data.norm_mean, + 'norm_std': cfg.data.norm_std, + 'use_gpu': cfg.use_gpu, + 'split_id': cfg.data.split_id, + 'combineall': cfg.data.combineall, + 'load_train_targets': cfg.data.load_train_targets, + 'batch_size_train': cfg.train.batch_size, + 'batch_size_test': cfg.test.batch_size, + 'workers': cfg.data.workers, + 'num_instances': cfg.sampler.num_instances, + 'num_cams': cfg.sampler.num_cams, + 'num_datasets': cfg.sampler.num_datasets, + 'train_sampler': cfg.sampler.train_sampler, + 'train_sampler_t': cfg.sampler.train_sampler_t, + # image dataset specific + 'cuhk03_labeled': cfg.cuhk03.labeled_images, + 'cuhk03_classic_split': cfg.cuhk03.classic_split, + 'market1501_500k': cfg.market1501.use_500k_distractors, + } + + +def videodata_kwargs(cfg): + return { + 'root': cfg.data.root, + 'sources': cfg.data.sources, + 'targets': cfg.data.targets, + 'height': cfg.data.height, + 'width': cfg.data.width, + 'transforms': cfg.data.transforms, + 'norm_mean': cfg.data.norm_mean, + 'norm_std': cfg.data.norm_std, + 'use_gpu': cfg.use_gpu, + 'split_id': cfg.data.split_id, + 'combineall': cfg.data.combineall, + 'batch_size_train': cfg.train.batch_size, + 'batch_size_test': cfg.test.batch_size, + 'workers': cfg.data.workers, + 'num_instances': cfg.sampler.num_instances, + 'num_cams': cfg.sampler.num_cams, + 'num_datasets': cfg.sampler.num_datasets, + 'train_sampler': cfg.sampler.train_sampler, + # video dataset specific + 'seq_len': cfg.video.seq_len, + 'sample_method': cfg.video.sample_method + } + + +def optimizer_kwargs(cfg): + return { + 'optim': cfg.train.optim, + 'lr': cfg.train.lr, + 'weight_decay': cfg.train.weight_decay, + 'momentum': cfg.sgd.momentum, + 'sgd_dampening': cfg.sgd.dampening, + 'sgd_nesterov': cfg.sgd.nesterov, + 'rmsprop_alpha': cfg.rmsprop.alpha, + 'adam_beta1': cfg.adam.beta1, + 'adam_beta2': cfg.adam.beta2, + 'staged_lr': cfg.train.staged_lr, + 'new_layers': cfg.train.new_layers, + 'base_lr_mult': cfg.train.base_lr_mult + } + + +def lr_scheduler_kwargs(cfg): + return { + 'lr_scheduler': cfg.train.lr_scheduler, + 'stepsize': cfg.train.stepsize, + 'gamma': cfg.train.gamma, + 'max_epoch': cfg.train.max_epoch + } + + +def engine_run_kwargs(cfg): + return { + 'save_dir': cfg.data.save_dir, + 'max_epoch': cfg.train.max_epoch, + 'start_epoch': cfg.train.start_epoch, + 'fixbase_epoch': cfg.train.fixbase_epoch, + 'open_layers': cfg.train.open_layers, + 'start_eval': cfg.test.start_eval, + 'eval_freq': cfg.test.eval_freq, + 'test_only': cfg.test.evaluate, + 'print_freq': cfg.train.print_freq, + 'dist_metric': cfg.test.dist_metric, + 'normalize_feature': cfg.test.normalize_feature, + 'visrank': cfg.test.visrank, + 'visrank_topk': cfg.test.visrank_topk, + 'use_metric_cuhk03': cfg.cuhk03.use_metric_cuhk03, + 'ranks': cfg.test.ranks, + 'rerank': cfg.test.rerank + } diff --git a/strong_sort/deep/reid/scripts/main.py b/strong_sort/deep/reid/scripts/main.py new file mode 100644 index 0000000..61aa49d --- /dev/null +++ b/strong_sort/deep/reid/scripts/main.py @@ -0,0 +1,191 @@ +import sys +import time +import os.path as osp +import argparse +import torch +import torch.nn as nn + +import torchreid +from torchreid.utils import ( + Logger, check_isfile, set_random_seed, collect_env_info, + resume_from_checkpoint, load_pretrained_weights, compute_model_complexity +) + +from default_config import ( + imagedata_kwargs, optimizer_kwargs, videodata_kwargs, engine_run_kwargs, + get_default_config, lr_scheduler_kwargs +) + + +def build_datamanager(cfg): + if cfg.data.type == 'image': + return torchreid.data.ImageDataManager(**imagedata_kwargs(cfg)) + else: + return torchreid.data.VideoDataManager(**videodata_kwargs(cfg)) + + +def build_engine(cfg, datamanager, model, optimizer, scheduler): + if cfg.data.type == 'image': + if cfg.loss.name == 'softmax': + engine = torchreid.engine.ImageSoftmaxEngine( + datamanager, + model, + optimizer=optimizer, + scheduler=scheduler, + use_gpu=cfg.use_gpu, + label_smooth=cfg.loss.softmax.label_smooth + ) + + else: + engine = torchreid.engine.ImageTripletEngine( + datamanager, + model, + optimizer=optimizer, + margin=cfg.loss.triplet.margin, + weight_t=cfg.loss.triplet.weight_t, + weight_x=cfg.loss.triplet.weight_x, + scheduler=scheduler, + use_gpu=cfg.use_gpu, + label_smooth=cfg.loss.softmax.label_smooth + ) + + else: + if cfg.loss.name == 'softmax': + engine = torchreid.engine.VideoSoftmaxEngine( + datamanager, + model, + optimizer=optimizer, + scheduler=scheduler, + use_gpu=cfg.use_gpu, + label_smooth=cfg.loss.softmax.label_smooth, + pooling_method=cfg.video.pooling_method + ) + + else: + engine = torchreid.engine.VideoTripletEngine( + datamanager, + model, + optimizer=optimizer, + margin=cfg.loss.triplet.margin, + weight_t=cfg.loss.triplet.weight_t, + weight_x=cfg.loss.triplet.weight_x, + scheduler=scheduler, + use_gpu=cfg.use_gpu, + label_smooth=cfg.loss.softmax.label_smooth + ) + + return engine + + +def reset_config(cfg, args): + if args.root: + cfg.data.root = args.root + if args.sources: + cfg.data.sources = args.sources + if args.targets: + cfg.data.targets = args.targets + if args.transforms: + cfg.data.transforms = args.transforms + + +def check_cfg(cfg): + if cfg.loss.name == 'triplet' and cfg.loss.triplet.weight_x == 0: + assert cfg.train.fixbase_epoch == 0, \ + 'The output of classifier is not included in the computational graph' + + +def main(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + parser.add_argument( + '--config-file', type=str, default='', help='path to config file' + ) + parser.add_argument( + '-s', + '--sources', + type=str, + nargs='+', + help='source datasets (delimited by space)' + ) + parser.add_argument( + '-t', + '--targets', + type=str, + nargs='+', + help='target datasets (delimited by space)' + ) + parser.add_argument( + '--transforms', type=str, nargs='+', help='data augmentation' + ) + parser.add_argument( + '--root', type=str, default='', help='path to data root' + ) + parser.add_argument( + 'opts', + default=None, + nargs=argparse.REMAINDER, + help='Modify config options using the command-line' + ) + args = parser.parse_args() + + cfg = get_default_config() + cfg.use_gpu = torch.cuda.is_available() + if args.config_file: + cfg.merge_from_file(args.config_file) + reset_config(cfg, args) + cfg.merge_from_list(args.opts) + set_random_seed(cfg.train.seed) + check_cfg(cfg) + + log_name = 'test.log' if cfg.test.evaluate else 'train.log' + log_name += time.strftime('-%Y-%m-%d-%H-%M-%S') + sys.stdout = Logger(osp.join(cfg.data.save_dir, log_name)) + + print('Show configuration\n{}\n'.format(cfg)) + print('Collecting env info ...') + print('** System info **\n{}\n'.format(collect_env_info())) + + if cfg.use_gpu: + torch.backends.cudnn.benchmark = True + + datamanager = build_datamanager(cfg) + + print('Building model: {}'.format(cfg.model.name)) + model = torchreid.models.build_model( + name=cfg.model.name, + num_classes=datamanager.num_train_pids, + loss=cfg.loss.name, + pretrained=cfg.model.pretrained, + use_gpu=cfg.use_gpu + ) + num_params, flops = compute_model_complexity( + model, (1, 3, cfg.data.height, cfg.data.width) + ) + print('Model complexity: params={:,} flops={:,}'.format(num_params, flops)) + + if cfg.model.load_weights and check_isfile(cfg.model.load_weights): + load_pretrained_weights(model, cfg.model.load_weights) + + if cfg.use_gpu: + model = nn.DataParallel(model).cuda() + + optimizer = torchreid.optim.build_optimizer(model, **optimizer_kwargs(cfg)) + scheduler = torchreid.optim.build_lr_scheduler( + optimizer, **lr_scheduler_kwargs(cfg) + ) + + if cfg.model.resume and check_isfile(cfg.model.resume): + cfg.train.start_epoch = resume_from_checkpoint( + cfg.model.resume, model, optimizer=optimizer, scheduler=scheduler + ) + + print( + 'Building {}-engine for {}-reid'.format(cfg.loss.name, cfg.data.type) + ) + engine = build_engine(cfg, datamanager, model, optimizer, scheduler) + engine.run(**engine_run_kwargs(cfg)) + + +if __name__ == '__main__': + main() diff --git a/strong_sort/deep/reid/setup.py b/strong_sort/deep/reid/setup.py new file mode 100644 index 0000000..a8ee83e --- /dev/null +++ b/strong_sort/deep/reid/setup.py @@ -0,0 +1,57 @@ +import numpy as np +import os.path as osp +from setuptools import setup, find_packages +from distutils.extension import Extension +from Cython.Build import cythonize + + +def readme(): + with open('README.rst') as f: + content = f.read() + return content + + +def find_version(): + version_file = 'torchreid/__init__.py' + with open(version_file, 'r') as f: + exec(compile(f.read(), version_file, 'exec')) + return locals()['__version__'] + + +def numpy_include(): + try: + numpy_include = np.get_include() + except AttributeError: + numpy_include = np.get_numpy_include() + return numpy_include + + +ext_modules = [ + Extension( + 'torchreid.metrics.rank_cylib.rank_cy', + ['torchreid/metrics/rank_cylib/rank_cy.pyx'], + include_dirs=[numpy_include()], + ) +] + + +def get_requirements(filename='requirements.txt'): + here = osp.dirname(osp.realpath(__file__)) + with open(osp.join(here, filename), 'r') as f: + requires = [line.replace('\n', '') for line in f.readlines()] + return requires + + +setup( + name='torchreid', + version=find_version(), + description='A library for deep learning person re-ID in PyTorch', + author='Kaiyang Zhou', + license='MIT', + long_description=readme(), + url='https://github.com/KaiyangZhou/deep-person-reid', + packages=find_packages(), + install_requires=get_requirements(), + keywords=['Person Re-Identification', 'Deep Learning', 'Computer Vision'], + ext_modules=cythonize(ext_modules) +) diff --git a/strong_sort/deep/reid/tools/compute_mean_std.py b/strong_sort/deep/reid/tools/compute_mean_std.py new file mode 100644 index 0000000..e0a5dbe --- /dev/null +++ b/strong_sort/deep/reid/tools/compute_mean_std.py @@ -0,0 +1,59 @@ +""" +Compute channel-wise mean and standard deviation of a dataset. + +Usage: +$ python compute_mean_std.py DATASET_ROOT DATASET_KEY + +- The first argument points to the root path where you put the datasets. +- The second argument means the specific dataset key. + +For instance, your datasets are put under $DATA and you wanna +compute the statistics of Market1501, do +$ python compute_mean_std.py $DATA market1501 +""" +import argparse + +import torchreid + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument('root', type=str) + parser.add_argument('sources', type=str) + args = parser.parse_args() + + datamanager = torchreid.data.ImageDataManager( + root=args.root, + sources=args.sources, + targets=None, + height=256, + width=128, + batch_size_train=100, + batch_size_test=100, + transforms=None, + norm_mean=[0., 0., 0.], + norm_std=[1., 1., 1.], + train_sampler='SequentialSampler' + ) + train_loader = datamanager.train_loader + + print('Computing mean and std ...') + mean = 0. + std = 0. + n_samples = 0. + for data in train_loader: + data = data['img'] + batch_size = data.size(0) + data = data.view(batch_size, data.size(1), -1) + mean += data.mean(2).sum(0) + std += data.std(2).sum(0) + n_samples += batch_size + + mean /= n_samples + std /= n_samples + print('Mean: {}'.format(mean)) + print('Std: {}'.format(std)) + + +if __name__ == '__main__': + main() diff --git a/strong_sort/deep/reid/tools/parse_test_res.py b/strong_sort/deep/reid/tools/parse_test_res.py new file mode 100644 index 0000000..fd5b018 --- /dev/null +++ b/strong_sort/deep/reid/tools/parse_test_res.py @@ -0,0 +1,103 @@ +""" +This script aims to automate the process of calculating average results +stored in the test.log files over multiple splits. + +How to use: +For example, you have done evaluation over 20 splits on VIPeR, leading to +the following file structure + +log/ + eval_viper/ + split_0/ + test.log-xxxx + split_1/ + test.log-xxxx + split_2/ + test.log-xxxx + ... + +You can run the following command in your terminal to get the average performance: +$ python tools/parse_test_res.py log/eval_viper +""" +import os +import re +import glob +import numpy as np +import argparse +from collections import defaultdict + +from torchreid.utils import check_isfile, listdir_nohidden + + +def parse_file(filepath, regex_mAP, regex_r1, regex_r5, regex_r10, regex_r20): + results = {} + + with open(filepath, 'r') as f: + lines = f.readlines() + + for line in lines: + line = line.strip() + + match_mAP = regex_mAP.search(line) + if match_mAP: + mAP = float(match_mAP.group(1)) + results['mAP'] = mAP + + match_r1 = regex_r1.search(line) + if match_r1: + r1 = float(match_r1.group(1)) + results['r1'] = r1 + + match_r5 = regex_r5.search(line) + if match_r5: + r5 = float(match_r5.group(1)) + results['r5'] = r5 + + match_r10 = regex_r10.search(line) + if match_r10: + r10 = float(match_r10.group(1)) + results['r10'] = r10 + + match_r20 = regex_r20.search(line) + if match_r20: + r20 = float(match_r20.group(1)) + results['r20'] = r20 + + return results + + +def main(args): + regex_mAP = re.compile(r'mAP: ([\.\deE+-]+)%') + regex_r1 = re.compile(r'Rank-1 : ([\.\deE+-]+)%') + regex_r5 = re.compile(r'Rank-5 : ([\.\deE+-]+)%') + regex_r10 = re.compile(r'Rank-10 : ([\.\deE+-]+)%') + regex_r20 = re.compile(r'Rank-20 : ([\.\deE+-]+)%') + + final_res = defaultdict(list) + + directories = listdir_nohidden(args.directory, sort=True) + num_dirs = len(directories) + for directory in directories: + fullpath = os.path.join(args.directory, directory) + filepath = glob.glob(os.path.join(fullpath, 'test.log*'))[0] + check_isfile(filepath) + print(f'Parsing {filepath}') + res = parse_file( + filepath, regex_mAP, regex_r1, regex_r5, regex_r10, regex_r20 + ) + for key, value in res.items(): + final_res[key].append(value) + + print('Finished parsing') + print(f'The average results over {num_dirs} splits are shown below') + + for key, values in final_res.items(): + mean_val = np.mean(values) + print(f'{key}: {mean_val:.1f}') + + +if __name__ == '__main__': + parser = argparse.ArgumentParser() + parser.add_argument('directory', type=str, help='Path to directory') + args = parser.parse_args() + main(args) diff --git a/strong_sort/deep/reid/tools/visualize_actmap.py b/strong_sort/deep/reid/tools/visualize_actmap.py new file mode 100644 index 0000000..ae69991 --- /dev/null +++ b/strong_sort/deep/reid/tools/visualize_actmap.py @@ -0,0 +1,173 @@ +"""Visualizes CNN activation maps to see where the CNN focuses on to extract features. + +Reference: + - Zagoruyko and Komodakis. Paying more attention to attention: Improving the + performance of convolutional neural networks via attention transfer. ICLR, 2017 + - Zhou et al. Omni-Scale Feature Learning for Person Re-Identification. ICCV, 2019. +""" +import numpy as np +import os.path as osp +import argparse +import cv2 +import torch +from torch.nn import functional as F + +import torchreid +from torchreid.utils import ( + check_isfile, mkdir_if_missing, load_pretrained_weights +) + +IMAGENET_MEAN = [0.485, 0.456, 0.406] +IMAGENET_STD = [0.229, 0.224, 0.225] +GRID_SPACING = 10 + + +@torch.no_grad() +def visactmap( + model, + test_loader, + save_dir, + width, + height, + use_gpu, + img_mean=None, + img_std=None +): + if img_mean is None or img_std is None: + # use imagenet mean and std + img_mean = IMAGENET_MEAN + img_std = IMAGENET_STD + + model.eval() + + for target in list(test_loader.keys()): + data_loader = test_loader[target]['query'] # only process query images + # original images and activation maps are saved individually + actmap_dir = osp.join(save_dir, 'actmap_' + target) + mkdir_if_missing(actmap_dir) + print('Visualizing activation maps for {} ...'.format(target)) + + for batch_idx, data in enumerate(data_loader): + imgs, paths = data['img'], data['impath'] + if use_gpu: + imgs = imgs.cuda() + + # forward to get convolutional feature maps + try: + outputs = model(imgs, return_featuremaps=True) + except TypeError: + raise TypeError( + 'forward() got unexpected keyword argument "return_featuremaps". ' + 'Please add return_featuremaps as an input argument to forward(). When ' + 'return_featuremaps=True, return feature maps only.' + ) + + if outputs.dim() != 4: + raise ValueError( + 'The model output is supposed to have ' + 'shape of (b, c, h, w), i.e. 4 dimensions, but got {} dimensions. ' + 'Please make sure you set the model output at eval mode ' + 'to be the last convolutional feature maps'.format( + outputs.dim() + ) + ) + + # compute activation maps + outputs = (outputs**2).sum(1) + b, h, w = outputs.size() + outputs = outputs.view(b, h * w) + outputs = F.normalize(outputs, p=2, dim=1) + outputs = outputs.view(b, h, w) + + if use_gpu: + imgs, outputs = imgs.cpu(), outputs.cpu() + + for j in range(outputs.size(0)): + # get image name + path = paths[j] + imname = osp.basename(osp.splitext(path)[0]) + + # RGB image + img = imgs[j, ...] + for t, m, s in zip(img, img_mean, img_std): + t.mul_(s).add_(m).clamp_(0, 1) + img_np = np.uint8(np.floor(img.numpy() * 255)) + img_np = img_np.transpose((1, 2, 0)) # (c, h, w) -> (h, w, c) + + # activation map + am = outputs[j, ...].numpy() + am = cv2.resize(am, (width, height)) + am = 255 * (am - np.min(am)) / ( + np.max(am) - np.min(am) + 1e-12 + ) + am = np.uint8(np.floor(am)) + am = cv2.applyColorMap(am, cv2.COLORMAP_JET) + + # overlapped + overlapped = img_np*0.3 + am*0.7 + overlapped[overlapped > 255] = 255 + overlapped = overlapped.astype(np.uint8) + + # save images in a single figure (add white spacing between images) + # from left to right: original image, activation map, overlapped image + grid_img = 255 * np.ones( + (height, 3*width + 2*GRID_SPACING, 3), dtype=np.uint8 + ) + grid_img[:, :width, :] = img_np[:, :, ::-1] + grid_img[:, + width + GRID_SPACING:2*width + GRID_SPACING, :] = am + grid_img[:, 2*width + 2*GRID_SPACING:, :] = overlapped + cv2.imwrite(osp.join(actmap_dir, imname + '.jpg'), grid_img) + + if (batch_idx+1) % 10 == 0: + print( + '- done batch {}/{}'.format( + batch_idx + 1, len(data_loader) + ) + ) + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument('--root', type=str) + parser.add_argument('-d', '--dataset', type=str, default='market1501') + parser.add_argument('-m', '--model', type=str, default='osnet_x1_0') + parser.add_argument('--weights', type=str) + parser.add_argument('--save-dir', type=str, default='log') + parser.add_argument('--height', type=int, default=256) + parser.add_argument('--width', type=int, default=128) + args = parser.parse_args() + + use_gpu = torch.cuda.is_available() + + datamanager = torchreid.data.ImageDataManager( + root=args.root, + sources=args.dataset, + height=args.height, + width=args.width, + batch_size_train=100, + batch_size_test=100, + transforms=None, + train_sampler='SequentialSampler' + ) + test_loader = datamanager.test_loader + + model = torchreid.models.build_model( + name=args.model, + num_classes=datamanager.num_train_pids, + use_gpu=use_gpu + ) + + if use_gpu: + model = model.cuda() + + if args.weights and check_isfile(args.weights): + load_pretrained_weights(model, args.weights) + + visactmap( + model, test_loader, args.save_dir, args.width, args.height, use_gpu + ) + + +if __name__ == '__main__': + main() diff --git a/strong_sort/deep/reid/torchreid/__init__.py b/strong_sort/deep/reid/torchreid/__init__.py new file mode 100644 index 0000000..35eeca6 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/__init__.py @@ -0,0 +1,9 @@ +from __future__ import print_function, absolute_import + +from torchreid import data, optim, utils, engine, losses, models, metrics + +__version__ = '1.4.0' +__author__ = 'Kaiyang Zhou' +__homepage__ = 'https://kaiyangzhou.github.io/' +__description__ = 'Deep learning person re-identification in PyTorch' +__url__ = 'https://github.com/KaiyangZhou/deep-person-reid' diff --git a/strong_sort/deep/reid/torchreid/data/__init__.py b/strong_sort/deep/reid/torchreid/data/__init__.py new file mode 100644 index 0000000..5318a16 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/data/__init__.py @@ -0,0 +1,7 @@ +from __future__ import print_function, absolute_import + +from .datasets import ( + Dataset, ImageDataset, VideoDataset, register_image_dataset, + register_video_dataset +) +from .datamanager import ImageDataManager, VideoDataManager diff --git a/strong_sort/deep/reid/torchreid/data/datamanager.py b/strong_sort/deep/reid/torchreid/data/datamanager.py new file mode 100644 index 0000000..7ae28cb --- /dev/null +++ b/strong_sort/deep/reid/torchreid/data/datamanager.py @@ -0,0 +1,554 @@ +from __future__ import division, print_function, absolute_import +import torch + +from torchreid.data.sampler import build_train_sampler +from torchreid.data.datasets import init_image_dataset, init_video_dataset +from torchreid.data.transforms import build_transforms + + +class DataManager(object): + r"""Base data manager. + + Args: + sources (str or list): source dataset(s). + targets (str or list, optional): target dataset(s). If not given, + it equals to ``sources``. + height (int, optional): target image height. Default is 256. + width (int, optional): target image width. Default is 128. + transforms (str or list of str, optional): transformations applied to model training. + Default is 'random_flip'. + norm_mean (list or None, optional): data mean. Default is None (use imagenet mean). + norm_std (list or None, optional): data std. Default is None (use imagenet std). + use_gpu (bool, optional): use gpu. Default is True. + """ + + def __init__( + self, + sources=None, + targets=None, + height=256, + width=128, + transforms='random_flip', + norm_mean=None, + norm_std=None, + use_gpu=False + ): + self.sources = sources + self.targets = targets + self.height = height + self.width = width + + if self.sources is None: + raise ValueError('sources must not be None') + + if isinstance(self.sources, str): + self.sources = [self.sources] + + if self.targets is None: + self.targets = self.sources + + if isinstance(self.targets, str): + self.targets = [self.targets] + + self.transform_tr, self.transform_te = build_transforms( + self.height, + self.width, + transforms=transforms, + norm_mean=norm_mean, + norm_std=norm_std + ) + + self.use_gpu = (torch.cuda.is_available() and use_gpu) + + @property + def num_train_pids(self): + """Returns the number of training person identities.""" + return self._num_train_pids + + @property + def num_train_cams(self): + """Returns the number of training cameras.""" + return self._num_train_cams + + def fetch_test_loaders(self, name): + """Returns query and gallery of a test dataset, each containing + tuples of (img_path(s), pid, camid). + + Args: + name (str): dataset name. + """ + query_loader = self.test_dataset[name]['query'] + gallery_loader = self.test_dataset[name]['gallery'] + return query_loader, gallery_loader + + def preprocess_pil_img(self, img): + """Transforms a PIL image to torch tensor for testing.""" + return self.transform_te(img) + + +class ImageDataManager(DataManager): + r"""Image data manager. + + Args: + root (str): root path to datasets. + sources (str or list): source dataset(s). + targets (str or list, optional): target dataset(s). If not given, + it equals to ``sources``. + height (int, optional): target image height. Default is 256. + width (int, optional): target image width. Default is 128. + transforms (str or list of str, optional): transformations applied to model training. + Default is 'random_flip'. + k_tfm (int): number of times to apply augmentation to an image + independently. If k_tfm > 1, the transform function will be + applied k_tfm times to an image. This variable will only be + useful for training and is currently valid for image datasets only. + norm_mean (list or None, optional): data mean. Default is None (use imagenet mean). + norm_std (list or None, optional): data std. Default is None (use imagenet std). + use_gpu (bool, optional): use gpu. Default is True. + split_id (int, optional): split id (*0-based*). Default is 0. + combineall (bool, optional): combine train, query and gallery in a dataset for + training. Default is False. + load_train_targets (bool, optional): construct train-loader for target datasets. + Default is False. This is useful for domain adaptation research. + batch_size_train (int, optional): number of images in a training batch. Default is 32. + batch_size_test (int, optional): number of images in a test batch. Default is 32. + workers (int, optional): number of workers. Default is 4. + num_instances (int, optional): number of instances per identity in a batch. + Default is 4. + num_cams (int, optional): number of cameras to sample in a batch (when using + ``RandomDomainSampler``). Default is 1. + num_datasets (int, optional): number of datasets to sample in a batch (when + using ``RandomDatasetSampler``). Default is 1. + train_sampler (str, optional): sampler. Default is RandomSampler. + train_sampler_t (str, optional): sampler for target train loader. Default is RandomSampler. + cuhk03_labeled (bool, optional): use cuhk03 labeled images. + Default is False (defaul is to use detected images). + cuhk03_classic_split (bool, optional): use the classic split in cuhk03. + Default is False. + market1501_500k (bool, optional): add 500K distractors to the gallery + set in market1501. Default is False. + + Examples:: + + datamanager = torchreid.data.ImageDataManager( + root='path/to/reid-data', + sources='market1501', + height=256, + width=128, + batch_size_train=32, + batch_size_test=100 + ) + + # return train loader of source data + train_loader = datamanager.train_loader + + # return test loader of target data + test_loader = datamanager.test_loader + + # return train loader of target data + train_loader_t = datamanager.train_loader_t + """ + data_type = 'image' + + def __init__( + self, + root='', + sources=None, + targets=None, + height=256, + width=128, + transforms='random_flip', + k_tfm=1, + norm_mean=None, + norm_std=None, + use_gpu=True, + split_id=0, + combineall=False, + load_train_targets=False, + batch_size_train=32, + batch_size_test=32, + workers=4, + num_instances=4, + num_cams=1, + num_datasets=1, + train_sampler='RandomSampler', + train_sampler_t='RandomSampler', + cuhk03_labeled=False, + cuhk03_classic_split=False, + market1501_500k=False + ): + + super(ImageDataManager, self).__init__( + sources=sources, + targets=targets, + height=height, + width=width, + transforms=transforms, + norm_mean=norm_mean, + norm_std=norm_std, + use_gpu=use_gpu + ) + + print('=> Loading train (source) dataset') + trainset = [] + for name in self.sources: + trainset_ = init_image_dataset( + name, + transform=self.transform_tr, + k_tfm=k_tfm, + mode='train', + combineall=combineall, + root=root, + split_id=split_id, + cuhk03_labeled=cuhk03_labeled, + cuhk03_classic_split=cuhk03_classic_split, + market1501_500k=market1501_500k + ) + trainset.append(trainset_) + trainset = sum(trainset) + + self._num_train_pids = trainset.num_train_pids + self._num_train_cams = trainset.num_train_cams + + self.train_loader = torch.utils.data.DataLoader( + trainset, + sampler=build_train_sampler( + trainset.train, + train_sampler, + batch_size=batch_size_train, + num_instances=num_instances, + num_cams=num_cams, + num_datasets=num_datasets + ), + batch_size=batch_size_train, + shuffle=False, + num_workers=workers, + pin_memory=self.use_gpu, + drop_last=True + ) + + self.train_loader_t = None + if load_train_targets: + # check if sources and targets are identical + assert len(set(self.sources) & set(self.targets)) == 0, \ + 'sources={} and targets={} must not have overlap'.format(self.sources, self.targets) + + print('=> Loading train (target) dataset') + trainset_t = [] + for name in self.targets: + trainset_t_ = init_image_dataset( + name, + transform=self.transform_tr, + k_tfm=k_tfm, + mode='train', + combineall=False, # only use the training data + root=root, + split_id=split_id, + cuhk03_labeled=cuhk03_labeled, + cuhk03_classic_split=cuhk03_classic_split, + market1501_500k=market1501_500k + ) + trainset_t.append(trainset_t_) + trainset_t = sum(trainset_t) + + self.train_loader_t = torch.utils.data.DataLoader( + trainset_t, + sampler=build_train_sampler( + trainset_t.train, + train_sampler_t, + batch_size=batch_size_train, + num_instances=num_instances, + num_cams=num_cams, + num_datasets=num_datasets + ), + batch_size=batch_size_train, + shuffle=False, + num_workers=workers, + pin_memory=self.use_gpu, + drop_last=True + ) + + print('=> Loading test (target) dataset') + self.test_loader = { + name: { + 'query': None, + 'gallery': None + } + for name in self.targets + } + self.test_dataset = { + name: { + 'query': None, + 'gallery': None + } + for name in self.targets + } + + for name in self.targets: + # build query loader + queryset = init_image_dataset( + name, + transform=self.transform_te, + mode='query', + combineall=combineall, + root=root, + split_id=split_id, + cuhk03_labeled=cuhk03_labeled, + cuhk03_classic_split=cuhk03_classic_split, + market1501_500k=market1501_500k + ) + self.test_loader[name]['query'] = torch.utils.data.DataLoader( + queryset, + batch_size=batch_size_test, + shuffle=False, + num_workers=workers, + pin_memory=self.use_gpu, + drop_last=False + ) + + # build gallery loader + galleryset = init_image_dataset( + name, + transform=self.transform_te, + mode='gallery', + combineall=combineall, + verbose=False, + root=root, + split_id=split_id, + cuhk03_labeled=cuhk03_labeled, + cuhk03_classic_split=cuhk03_classic_split, + market1501_500k=market1501_500k + ) + self.test_loader[name]['gallery'] = torch.utils.data.DataLoader( + galleryset, + batch_size=batch_size_test, + shuffle=False, + num_workers=workers, + pin_memory=self.use_gpu, + drop_last=False + ) + + self.test_dataset[name]['query'] = queryset.query + self.test_dataset[name]['gallery'] = galleryset.gallery + + print('\n') + print(' **************** Summary ****************') + print(' source : {}'.format(self.sources)) + print(' # source datasets : {}'.format(len(self.sources))) + print(' # source ids : {}'.format(self.num_train_pids)) + print(' # source images : {}'.format(len(trainset))) + print(' # source cameras : {}'.format(self.num_train_cams)) + if load_train_targets: + print( + ' # target images : {} (unlabeled)'.format(len(trainset_t)) + ) + print(' target : {}'.format(self.targets)) + print(' *****************************************') + print('\n') + + +class VideoDataManager(DataManager): + r"""Video data manager. + + Args: + root (str): root path to datasets. + sources (str or list): source dataset(s). + targets (str or list, optional): target dataset(s). If not given, + it equals to ``sources``. + height (int, optional): target image height. Default is 256. + width (int, optional): target image width. Default is 128. + transforms (str or list of str, optional): transformations applied to model training. + Default is 'random_flip'. + norm_mean (list or None, optional): data mean. Default is None (use imagenet mean). + norm_std (list or None, optional): data std. Default is None (use imagenet std). + use_gpu (bool, optional): use gpu. Default is True. + split_id (int, optional): split id (*0-based*). Default is 0. + combineall (bool, optional): combine train, query and gallery in a dataset for + training. Default is False. + batch_size_train (int, optional): number of tracklets in a training batch. Default is 3. + batch_size_test (int, optional): number of tracklets in a test batch. Default is 3. + workers (int, optional): number of workers. Default is 4. + num_instances (int, optional): number of instances per identity in a batch. + Default is 4. + num_cams (int, optional): number of cameras to sample in a batch (when using + ``RandomDomainSampler``). Default is 1. + num_datasets (int, optional): number of datasets to sample in a batch (when + using ``RandomDatasetSampler``). Default is 1. + train_sampler (str, optional): sampler. Default is RandomSampler. + seq_len (int, optional): how many images to sample in a tracklet. Default is 15. + sample_method (str, optional): how to sample images in a tracklet. Default is "evenly". + Choices are ["evenly", "random", "all"]. "evenly" and "random" will sample ``seq_len`` + images in a tracklet while "all" samples all images in a tracklet, where the batch size + needs to be set to 1. + + Examples:: + + datamanager = torchreid.data.VideoDataManager( + root='path/to/reid-data', + sources='mars', + height=256, + width=128, + batch_size_train=3, + batch_size_test=3, + seq_len=15, + sample_method='evenly' + ) + + # return train loader of source data + train_loader = datamanager.train_loader + + # return test loader of target data + test_loader = datamanager.test_loader + + .. note:: + The current implementation only supports image-like training. Therefore, each image in a + sampled tracklet will undergo independent transformation functions. To achieve tracklet-aware + training, you need to modify the transformation functions for video reid such that each function + applies the same operation to all images in a tracklet to keep consistency. + """ + data_type = 'video' + + def __init__( + self, + root='', + sources=None, + targets=None, + height=256, + width=128, + transforms='random_flip', + norm_mean=None, + norm_std=None, + use_gpu=True, + split_id=0, + combineall=False, + batch_size_train=3, + batch_size_test=3, + workers=4, + num_instances=4, + num_cams=1, + num_datasets=1, + train_sampler='RandomSampler', + seq_len=15, + sample_method='evenly' + ): + + super(VideoDataManager, self).__init__( + sources=sources, + targets=targets, + height=height, + width=width, + transforms=transforms, + norm_mean=norm_mean, + norm_std=norm_std, + use_gpu=use_gpu + ) + + print('=> Loading train (source) dataset') + trainset = [] + for name in self.sources: + trainset_ = init_video_dataset( + name, + transform=self.transform_tr, + mode='train', + combineall=combineall, + root=root, + split_id=split_id, + seq_len=seq_len, + sample_method=sample_method + ) + trainset.append(trainset_) + trainset = sum(trainset) + + self._num_train_pids = trainset.num_train_pids + self._num_train_cams = trainset.num_train_cams + + train_sampler = build_train_sampler( + trainset.train, + train_sampler, + batch_size=batch_size_train, + num_instances=num_instances, + num_cams=num_cams, + num_datasets=num_datasets + ) + + self.train_loader = torch.utils.data.DataLoader( + trainset, + sampler=train_sampler, + batch_size=batch_size_train, + shuffle=False, + num_workers=workers, + pin_memory=self.use_gpu, + drop_last=True + ) + + print('=> Loading test (target) dataset') + self.test_loader = { + name: { + 'query': None, + 'gallery': None + } + for name in self.targets + } + self.test_dataset = { + name: { + 'query': None, + 'gallery': None + } + for name in self.targets + } + + for name in self.targets: + # build query loader + queryset = init_video_dataset( + name, + transform=self.transform_te, + mode='query', + combineall=combineall, + root=root, + split_id=split_id, + seq_len=seq_len, + sample_method=sample_method + ) + self.test_loader[name]['query'] = torch.utils.data.DataLoader( + queryset, + batch_size=batch_size_test, + shuffle=False, + num_workers=workers, + pin_memory=self.use_gpu, + drop_last=False + ) + + # build gallery loader + galleryset = init_video_dataset( + name, + transform=self.transform_te, + mode='gallery', + combineall=combineall, + verbose=False, + root=root, + split_id=split_id, + seq_len=seq_len, + sample_method=sample_method + ) + self.test_loader[name]['gallery'] = torch.utils.data.DataLoader( + galleryset, + batch_size=batch_size_test, + shuffle=False, + num_workers=workers, + pin_memory=self.use_gpu, + drop_last=False + ) + + self.test_dataset[name]['query'] = queryset.query + self.test_dataset[name]['gallery'] = galleryset.gallery + + print('\n') + print(' **************** Summary ****************') + print(' source : {}'.format(self.sources)) + print(' # source datasets : {}'.format(len(self.sources))) + print(' # source ids : {}'.format(self.num_train_pids)) + print(' # source tracklets : {}'.format(len(trainset))) + print(' # source cameras : {}'.format(self.num_train_cams)) + print(' target : {}'.format(self.targets)) + print(' *****************************************') + print('\n') diff --git a/strong_sort/deep/reid/torchreid/data/datasets/__init__.py b/strong_sort/deep/reid/torchreid/data/datasets/__init__.py new file mode 100644 index 0000000..afb02a2 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/data/datasets/__init__.py @@ -0,0 +1,119 @@ +from __future__ import print_function, absolute_import + +from .image import ( + GRID, PRID, CUHK01, CUHK02, CUHK03, MSMT17, CUHKSYSU, VIPeR, SenseReID, + Market1501, DukeMTMCreID, University1652, iLIDS +) +from .video import PRID2011, Mars, DukeMTMCVidReID, iLIDSVID +from .dataset import Dataset, ImageDataset, VideoDataset + +__image_datasets = { + 'market1501': Market1501, + 'cuhk03': CUHK03, + 'dukemtmcreid': DukeMTMCreID, + 'msmt17': MSMT17, + 'viper': VIPeR, + 'grid': GRID, + 'cuhk01': CUHK01, + 'ilids': iLIDS, + 'sensereid': SenseReID, + 'prid': PRID, + 'cuhk02': CUHK02, + 'university1652': University1652, + 'cuhksysu': CUHKSYSU +} + +__video_datasets = { + 'mars': Mars, + 'ilidsvid': iLIDSVID, + 'prid2011': PRID2011, + 'dukemtmcvidreid': DukeMTMCVidReID +} + + +def init_image_dataset(name, **kwargs): + """Initializes an image dataset.""" + avai_datasets = list(__image_datasets.keys()) + if name not in avai_datasets: + raise ValueError( + 'Invalid dataset name. Received "{}", ' + 'but expected to be one of {}'.format(name, avai_datasets) + ) + return __image_datasets[name](**kwargs) + + +def init_video_dataset(name, **kwargs): + """Initializes a video dataset.""" + avai_datasets = list(__video_datasets.keys()) + if name not in avai_datasets: + raise ValueError( + 'Invalid dataset name. Received "{}", ' + 'but expected to be one of {}'.format(name, avai_datasets) + ) + return __video_datasets[name](**kwargs) + + +def register_image_dataset(name, dataset): + """Registers a new image dataset. + + Args: + name (str): key corresponding to the new dataset. + dataset (Dataset): the new dataset class. + + Examples:: + + import torchreid + import NewDataset + torchreid.data.register_image_dataset('new_dataset', NewDataset) + # single dataset case + datamanager = torchreid.data.ImageDataManager( + root='reid-data', + sources='new_dataset' + ) + # multiple dataset case + datamanager = torchreid.data.ImageDataManager( + root='reid-data', + sources=['new_dataset', 'dukemtmcreid'] + ) + """ + global __image_datasets + curr_datasets = list(__image_datasets.keys()) + if name in curr_datasets: + raise ValueError( + 'The given name already exists, please choose ' + 'another name excluding {}'.format(curr_datasets) + ) + __image_datasets[name] = dataset + + +def register_video_dataset(name, dataset): + """Registers a new video dataset. + + Args: + name (str): key corresponding to the new dataset. + dataset (Dataset): the new dataset class. + + Examples:: + + import torchreid + import NewDataset + torchreid.data.register_video_dataset('new_dataset', NewDataset) + # single dataset case + datamanager = torchreid.data.VideoDataManager( + root='reid-data', + sources='new_dataset' + ) + # multiple dataset case + datamanager = torchreid.data.VideoDataManager( + root='reid-data', + sources=['new_dataset', 'ilidsvid'] + ) + """ + global __video_datasets + curr_datasets = list(__video_datasets.keys()) + if name in curr_datasets: + raise ValueError( + 'The given name already exists, please choose ' + 'another name excluding {}'.format(curr_datasets) + ) + __video_datasets[name] = dataset diff --git a/strong_sort/deep/reid/torchreid/data/datasets/dataset.py b/strong_sort/deep/reid/torchreid/data/datasets/dataset.py new file mode 100644 index 0000000..66b1e7a --- /dev/null +++ b/strong_sort/deep/reid/torchreid/data/datasets/dataset.py @@ -0,0 +1,482 @@ +from __future__ import division, print_function, absolute_import +import copy +import numpy as np +import os.path as osp +import tarfile +import zipfile +import torch + +from torchreid.utils import read_image, download_url, mkdir_if_missing + + +class Dataset(object): + """An abstract class representing a Dataset. + + This is the base class for ``ImageDataset`` and ``VideoDataset``. + + Args: + train (list): contains tuples of (img_path(s), pid, camid). + query (list): contains tuples of (img_path(s), pid, camid). + gallery (list): contains tuples of (img_path(s), pid, camid). + transform: transform function. + k_tfm (int): number of times to apply augmentation to an image + independently. If k_tfm > 1, the transform function will be + applied k_tfm times to an image. This variable will only be + useful for training and is currently valid for image datasets only. + mode (str): 'train', 'query' or 'gallery'. + combineall (bool): combines train, query and gallery in a + dataset for training. + verbose (bool): show information. + """ + + # junk_pids contains useless person IDs, e.g. background, + # false detections, distractors. These IDs will be ignored + # when combining all images in a dataset for training, i.e. + # combineall=True + _junk_pids = [] + + # Some datasets are only used for training, like CUHK-SYSU + # In this case, "combineall=True" is not used for them + _train_only = False + + def __init__( + self, + train, + query, + gallery, + transform=None, + k_tfm=1, + mode='train', + combineall=False, + verbose=True, + **kwargs + ): + # extend 3-tuple (img_path(s), pid, camid) to + # 4-tuple (img_path(s), pid, camid, dsetid) by + # adding a dataset indicator "dsetid" + if len(train[0]) == 3: + train = [(*items, 0) for items in train] + if len(query[0]) == 3: + query = [(*items, 0) for items in query] + if len(gallery[0]) == 3: + gallery = [(*items, 0) for items in gallery] + + self.train = train + self.query = query + self.gallery = gallery + self.transform = transform + self.k_tfm = k_tfm + self.mode = mode + self.combineall = combineall + self.verbose = verbose + + self.num_train_pids = self.get_num_pids(self.train) + self.num_train_cams = self.get_num_cams(self.train) + self.num_datasets = self.get_num_datasets(self.train) + + if self.combineall: + self.combine_all() + + if self.mode == 'train': + self.data = self.train + elif self.mode == 'query': + self.data = self.query + elif self.mode == 'gallery': + self.data = self.gallery + else: + raise ValueError( + 'Invalid mode. Got {}, but expected to be ' + 'one of [train | query | gallery]'.format(self.mode) + ) + + if self.verbose: + self.show_summary() + + def __getitem__(self, index): + raise NotImplementedError + + def __len__(self): + return len(self.data) + + def __add__(self, other): + """Adds two datasets together (only the train set).""" + train = copy.deepcopy(self.train) + + for img_path, pid, camid, dsetid in other.train: + pid += self.num_train_pids + camid += self.num_train_cams + dsetid += self.num_datasets + train.append((img_path, pid, camid, dsetid)) + + ################################### + # Note that + # 1. set verbose=False to avoid unnecessary print + # 2. set combineall=False because combineall would have been applied + # if it was True for a specific dataset; setting it to True will + # create new IDs that should have already been included + ################################### + if isinstance(train[0][0], str): + return ImageDataset( + train, + self.query, + self.gallery, + transform=self.transform, + mode=self.mode, + combineall=False, + verbose=False + ) + else: + return VideoDataset( + train, + self.query, + self.gallery, + transform=self.transform, + mode=self.mode, + combineall=False, + verbose=False, + seq_len=self.seq_len, + sample_method=self.sample_method + ) + + def __radd__(self, other): + """Supports sum([dataset1, dataset2, dataset3]).""" + if other == 0: + return self + else: + return self.__add__(other) + + def get_num_pids(self, data): + """Returns the number of training person identities. + + Each tuple in data contains (img_path(s), pid, camid, dsetid). + """ + pids = set() + for items in data: + pid = items[1] + pids.add(pid) + return len(pids) + + def get_num_cams(self, data): + """Returns the number of training cameras. + + Each tuple in data contains (img_path(s), pid, camid, dsetid). + """ + cams = set() + for items in data: + camid = items[2] + cams.add(camid) + return len(cams) + + def get_num_datasets(self, data): + """Returns the number of datasets included. + + Each tuple in data contains (img_path(s), pid, camid, dsetid). + """ + dsets = set() + for items in data: + dsetid = items[3] + dsets.add(dsetid) + return len(dsets) + + def show_summary(self): + """Shows dataset statistics.""" + pass + + def combine_all(self): + """Combines train, query and gallery in a dataset for training.""" + if self._train_only: + return + + combined = copy.deepcopy(self.train) + + # relabel pids in gallery (query shares the same scope) + g_pids = set() + for items in self.gallery: + pid = items[1] + if pid in self._junk_pids: + continue + g_pids.add(pid) + pid2label = {pid: i for i, pid in enumerate(g_pids)} + + def _combine_data(data): + for img_path, pid, camid, dsetid in data: + if pid in self._junk_pids: + continue + pid = pid2label[pid] + self.num_train_pids + combined.append((img_path, pid, camid, dsetid)) + + _combine_data(self.query) + _combine_data(self.gallery) + + self.train = combined + self.num_train_pids = self.get_num_pids(self.train) + + def download_dataset(self, dataset_dir, dataset_url): + """Downloads and extracts dataset. + + Args: + dataset_dir (str): dataset directory. + dataset_url (str): url to download dataset. + """ + if osp.exists(dataset_dir): + return + + if dataset_url is None: + raise RuntimeError( + '{} dataset needs to be manually ' + 'prepared, please follow the ' + 'document to prepare this dataset'.format( + self.__class__.__name__ + ) + ) + + print('Creating directory "{}"'.format(dataset_dir)) + mkdir_if_missing(dataset_dir) + fpath = osp.join(dataset_dir, osp.basename(dataset_url)) + + print( + 'Downloading {} dataset to "{}"'.format( + self.__class__.__name__, dataset_dir + ) + ) + download_url(dataset_url, fpath) + + print('Extracting "{}"'.format(fpath)) + try: + tar = tarfile.open(fpath) + tar.extractall(path=dataset_dir) + tar.close() + except: + zip_ref = zipfile.ZipFile(fpath, 'r') + zip_ref.extractall(dataset_dir) + zip_ref.close() + + print('{} dataset is ready'.format(self.__class__.__name__)) + + def check_before_run(self, required_files): + """Checks if required files exist before going deeper. + + Args: + required_files (str or list): string file name(s). + """ + if isinstance(required_files, str): + required_files = [required_files] + + for fpath in required_files: + if not osp.exists(fpath): + raise RuntimeError('"{}" is not found'.format(fpath)) + + def __repr__(self): + num_train_pids = self.get_num_pids(self.train) + num_train_cams = self.get_num_cams(self.train) + + num_query_pids = self.get_num_pids(self.query) + num_query_cams = self.get_num_cams(self.query) + + num_gallery_pids = self.get_num_pids(self.gallery) + num_gallery_cams = self.get_num_cams(self.gallery) + + msg = ' ----------------------------------------\n' \ + ' subset | # ids | # items | # cameras\n' \ + ' ----------------------------------------\n' \ + ' train | {:5d} | {:7d} | {:9d}\n' \ + ' query | {:5d} | {:7d} | {:9d}\n' \ + ' gallery | {:5d} | {:7d} | {:9d}\n' \ + ' ----------------------------------------\n' \ + ' items: images/tracklets for image/video dataset\n'.format( + num_train_pids, len(self.train), num_train_cams, + num_query_pids, len(self.query), num_query_cams, + num_gallery_pids, len(self.gallery), num_gallery_cams + ) + + return msg + + def _transform_image(self, tfm, k_tfm, img0): + """Transforms a raw image (img0) k_tfm times with + the transform function tfm. + """ + img_list = [] + + for k in range(k_tfm): + img_list.append(tfm(img0)) + + img = img_list + if len(img) == 1: + img = img[0] + + return img + + +class ImageDataset(Dataset): + """A base class representing ImageDataset. + + All other image datasets should subclass it. + + ``__getitem__`` returns an image given index. + It will return ``img``, ``pid``, ``camid`` and ``img_path`` + where ``img`` has shape (channel, height, width). As a result, + data in each batch has shape (batch_size, channel, height, width). + """ + + def __init__(self, train, query, gallery, **kwargs): + super(ImageDataset, self).__init__(train, query, gallery, **kwargs) + + def __getitem__(self, index): + img_path, pid, camid, dsetid = self.data[index] + img = read_image(img_path) + if self.transform is not None: + img = self._transform_image(self.transform, self.k_tfm, img) + item = { + 'img': img, + 'pid': pid, + 'camid': camid, + 'impath': img_path, + 'dsetid': dsetid + } + return item + + def show_summary(self): + num_train_pids = self.get_num_pids(self.train) + num_train_cams = self.get_num_cams(self.train) + + num_query_pids = self.get_num_pids(self.query) + num_query_cams = self.get_num_cams(self.query) + + num_gallery_pids = self.get_num_pids(self.gallery) + num_gallery_cams = self.get_num_cams(self.gallery) + + print('=> Loaded {}'.format(self.__class__.__name__)) + print(' ----------------------------------------') + print(' subset | # ids | # images | # cameras') + print(' ----------------------------------------') + print( + ' train | {:5d} | {:8d} | {:9d}'.format( + num_train_pids, len(self.train), num_train_cams + ) + ) + print( + ' query | {:5d} | {:8d} | {:9d}'.format( + num_query_pids, len(self.query), num_query_cams + ) + ) + print( + ' gallery | {:5d} | {:8d} | {:9d}'.format( + num_gallery_pids, len(self.gallery), num_gallery_cams + ) + ) + print(' ----------------------------------------') + + +class VideoDataset(Dataset): + """A base class representing VideoDataset. + + All other video datasets should subclass it. + + ``__getitem__`` returns an image given index. + It will return ``imgs``, ``pid`` and ``camid`` + where ``imgs`` has shape (seq_len, channel, height, width). As a result, + data in each batch has shape (batch_size, seq_len, channel, height, width). + """ + + def __init__( + self, + train, + query, + gallery, + seq_len=15, + sample_method='evenly', + **kwargs + ): + super(VideoDataset, self).__init__(train, query, gallery, **kwargs) + self.seq_len = seq_len + self.sample_method = sample_method + + if self.transform is None: + raise RuntimeError('transform must not be None') + + def __getitem__(self, index): + img_paths, pid, camid, dsetid = self.data[index] + num_imgs = len(img_paths) + + if self.sample_method == 'random': + # Randomly samples seq_len images from a tracklet of length num_imgs, + # if num_imgs is smaller than seq_len, then replicates images + indices = np.arange(num_imgs) + replace = False if num_imgs >= self.seq_len else True + indices = np.random.choice( + indices, size=self.seq_len, replace=replace + ) + # sort indices to keep temporal order (comment it to be order-agnostic) + indices = np.sort(indices) + + elif self.sample_method == 'evenly': + # Evenly samples seq_len images from a tracklet + if num_imgs >= self.seq_len: + num_imgs -= num_imgs % self.seq_len + indices = np.arange(0, num_imgs, num_imgs / self.seq_len) + else: + # if num_imgs is smaller than seq_len, simply replicate the last image + # until the seq_len requirement is satisfied + indices = np.arange(0, num_imgs) + num_pads = self.seq_len - num_imgs + indices = np.concatenate( + [ + indices, + np.ones(num_pads).astype(np.int32) * (num_imgs-1) + ] + ) + assert len(indices) == self.seq_len + + elif self.sample_method == 'all': + # Samples all images in a tracklet. batch_size must be set to 1 + indices = np.arange(num_imgs) + + else: + raise ValueError( + 'Unknown sample method: {}'.format(self.sample_method) + ) + + imgs = [] + for index in indices: + img_path = img_paths[int(index)] + img = read_image(img_path) + if self.transform is not None: + img = self.transform(img) + img = img.unsqueeze(0) # img must be torch.Tensor + imgs.append(img) + imgs = torch.cat(imgs, dim=0) + + item = {'img': imgs, 'pid': pid, 'camid': camid, 'dsetid': dsetid} + + return item + + def show_summary(self): + num_train_pids = self.get_num_pids(self.train) + num_train_cams = self.get_num_cams(self.train) + + num_query_pids = self.get_num_pids(self.query) + num_query_cams = self.get_num_cams(self.query) + + num_gallery_pids = self.get_num_pids(self.gallery) + num_gallery_cams = self.get_num_cams(self.gallery) + + print('=> Loaded {}'.format(self.__class__.__name__)) + print(' -------------------------------------------') + print(' subset | # ids | # tracklets | # cameras') + print(' -------------------------------------------') + print( + ' train | {:5d} | {:11d} | {:9d}'.format( + num_train_pids, len(self.train), num_train_cams + ) + ) + print( + ' query | {:5d} | {:11d} | {:9d}'.format( + num_query_pids, len(self.query), num_query_cams + ) + ) + print( + ' gallery | {:5d} | {:11d} | {:9d}'.format( + num_gallery_pids, len(self.gallery), num_gallery_cams + ) + ) + print(' -------------------------------------------') diff --git a/strong_sort/deep/reid/torchreid/data/datasets/image/__init__.py b/strong_sort/deep/reid/torchreid/data/datasets/image/__init__.py new file mode 100644 index 0000000..f2216e9 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/data/datasets/image/__init__.py @@ -0,0 +1,15 @@ +from __future__ import print_function, absolute_import + +from .grid import GRID +from .prid import PRID +from .ilids import iLIDS +from .viper import VIPeR +from .cuhk01 import CUHK01 +from .cuhk02 import CUHK02 +from .cuhk03 import CUHK03 +from .msmt17 import MSMT17 +from .cuhksysu import CUHKSYSU +from .sensereid import SenseReID +from .market1501 import Market1501 +from .dukemtmcreid import DukeMTMCreID +from .university1652 import University1652 diff --git a/strong_sort/deep/reid/torchreid/data/datasets/image/cuhk01.py b/strong_sort/deep/reid/torchreid/data/datasets/image/cuhk01.py new file mode 100644 index 0000000..c4c332e --- /dev/null +++ b/strong_sort/deep/reid/torchreid/data/datasets/image/cuhk01.py @@ -0,0 +1,137 @@ +from __future__ import division, print_function, absolute_import +import glob +import numpy as np +import os.path as osp +import zipfile + +from torchreid.utils import read_json, write_json + +from ..dataset import ImageDataset + + +class CUHK01(ImageDataset): + """CUHK01. + + Reference: + Li et al. Human Reidentification with Transferred Metric Learning. ACCV 2012. + + URL: ``_ + + Dataset statistics: + - identities: 971. + - images: 3884. + - cameras: 4. + + Note: CUHK01 and CUHK02 overlap. + """ + dataset_dir = 'cuhk01' + dataset_url = None + + def __init__(self, root='', split_id=0, **kwargs): + self.root = osp.abspath(osp.expanduser(root)) + self.dataset_dir = osp.join(self.root, self.dataset_dir) + self.download_dataset(self.dataset_dir, self.dataset_url) + + self.zip_path = osp.join(self.dataset_dir, 'CUHK01.zip') + self.campus_dir = osp.join(self.dataset_dir, 'campus') + self.split_path = osp.join(self.dataset_dir, 'splits.json') + + self.extract_file() + + required_files = [self.dataset_dir, self.campus_dir] + self.check_before_run(required_files) + + self.prepare_split() + splits = read_json(self.split_path) + if split_id >= len(splits): + raise ValueError( + 'split_id exceeds range, received {}, but expected between 0 and {}' + .format(split_id, + len(splits) - 1) + ) + split = splits[split_id] + + train = split['train'] + query = split['query'] + gallery = split['gallery'] + + train = [tuple(item) for item in train] + query = [tuple(item) for item in query] + gallery = [tuple(item) for item in gallery] + + super(CUHK01, self).__init__(train, query, gallery, **kwargs) + + def extract_file(self): + if not osp.exists(self.campus_dir): + print('Extracting files') + zip_ref = zipfile.ZipFile(self.zip_path, 'r') + zip_ref.extractall(self.dataset_dir) + zip_ref.close() + + def prepare_split(self): + """ + Image name format: 0001001.png, where first four digits represent identity + and last four digits represent cameras. Camera 1&2 are considered the same + view and camera 3&4 are considered the same view. + """ + if not osp.exists(self.split_path): + print('Creating 10 random splits of train ids and test ids') + img_paths = sorted(glob.glob(osp.join(self.campus_dir, '*.png'))) + img_list = [] + pid_container = set() + for img_path in img_paths: + img_name = osp.basename(img_path) + pid = int(img_name[:4]) - 1 + camid = (int(img_name[4:7]) - 1) // 2 # result is either 0 or 1 + img_list.append((img_path, pid, camid)) + pid_container.add(pid) + + num_pids = len(pid_container) + num_train_pids = num_pids // 2 + + splits = [] + for _ in range(10): + order = np.arange(num_pids) + np.random.shuffle(order) + train_idxs = order[:num_train_pids] + train_idxs = np.sort(train_idxs) + idx2label = { + idx: label + for label, idx in enumerate(train_idxs) + } + + train, test_a, test_b = [], [], [] + for img_path, pid, camid in img_list: + if pid in train_idxs: + train.append((img_path, idx2label[pid], camid)) + else: + if camid == 0: + test_a.append((img_path, pid, camid)) + else: + test_b.append((img_path, pid, camid)) + + # use cameraA as query and cameraB as gallery + split = { + 'train': train, + 'query': test_a, + 'gallery': test_b, + 'num_train_pids': num_train_pids, + 'num_query_pids': num_pids - num_train_pids, + 'num_gallery_pids': num_pids - num_train_pids + } + splits.append(split) + + # use cameraB as query and cameraA as gallery + split = { + 'train': train, + 'query': test_b, + 'gallery': test_a, + 'num_train_pids': num_train_pids, + 'num_query_pids': num_pids - num_train_pids, + 'num_gallery_pids': num_pids - num_train_pids + } + splits.append(split) + + print('Totally {} splits are created'.format(len(splits))) + write_json(splits, self.split_path) + print('Split file saved to {}'.format(self.split_path)) diff --git a/strong_sort/deep/reid/torchreid/data/datasets/image/cuhk02.py b/strong_sort/deep/reid/torchreid/data/datasets/image/cuhk02.py new file mode 100644 index 0000000..dd92588 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/data/datasets/image/cuhk02.py @@ -0,0 +1,97 @@ +from __future__ import division, print_function, absolute_import +import glob +import os.path as osp + +from ..dataset import ImageDataset + + +class CUHK02(ImageDataset): + """CUHK02. + + Reference: + Li and Wang. Locally Aligned Feature Transforms across Views. CVPR 2013. + + URL: ``_ + + Dataset statistics: + - 5 camera view pairs each with two cameras + - 971, 306, 107, 193 and 239 identities from P1 - P5 + - totally 1,816 identities + - image format is png + + Protocol: Use P1 - P4 for training and P5 for evaluation. + + Note: CUHK01 and CUHK02 overlap. + """ + dataset_dir = 'cuhk02' + cam_pairs = ['P1', 'P2', 'P3', 'P4', 'P5'] + test_cam_pair = 'P5' + + def __init__(self, root='', **kwargs): + self.root = osp.abspath(osp.expanduser(root)) + self.dataset_dir = osp.join(self.root, self.dataset_dir, 'Dataset') + + required_files = [self.dataset_dir] + self.check_before_run(required_files) + + train, query, gallery = self.get_data_list() + + super(CUHK02, self).__init__(train, query, gallery, **kwargs) + + def get_data_list(self): + num_train_pids, camid = 0, 0 + train, query, gallery = [], [], [] + + for cam_pair in self.cam_pairs: + cam_pair_dir = osp.join(self.dataset_dir, cam_pair) + + cam1_dir = osp.join(cam_pair_dir, 'cam1') + cam2_dir = osp.join(cam_pair_dir, 'cam2') + + impaths1 = glob.glob(osp.join(cam1_dir, '*.png')) + impaths2 = glob.glob(osp.join(cam2_dir, '*.png')) + + if cam_pair == self.test_cam_pair: + # add images to query + for impath in impaths1: + pid = osp.basename(impath).split('_')[0] + pid = int(pid) + query.append((impath, pid, camid)) + camid += 1 + + # add images to gallery + for impath in impaths2: + pid = osp.basename(impath).split('_')[0] + pid = int(pid) + gallery.append((impath, pid, camid)) + camid += 1 + + else: + pids1 = [ + osp.basename(impath).split('_')[0] for impath in impaths1 + ] + pids2 = [ + osp.basename(impath).split('_')[0] for impath in impaths2 + ] + pids = set(pids1 + pids2) + pid2label = { + pid: label + num_train_pids + for label, pid in enumerate(pids) + } + + # add images to train from cam1 + for impath in impaths1: + pid = osp.basename(impath).split('_')[0] + pid = pid2label[pid] + train.append((impath, pid, camid)) + camid += 1 + + # add images to train from cam2 + for impath in impaths2: + pid = osp.basename(impath).split('_')[0] + pid = pid2label[pid] + train.append((impath, pid, camid)) + camid += 1 + num_train_pids += len(pids) + + return train, query, gallery diff --git a/strong_sort/deep/reid/torchreid/data/datasets/image/cuhk03.py b/strong_sort/deep/reid/torchreid/data/datasets/image/cuhk03.py new file mode 100644 index 0000000..cd27bc2 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/data/datasets/image/cuhk03.py @@ -0,0 +1,307 @@ +from __future__ import division, print_function, absolute_import +import os.path as osp + +from torchreid.utils import read_json, write_json, mkdir_if_missing + +from ..dataset import ImageDataset + + +class CUHK03(ImageDataset): + """CUHK03. + + Reference: + Li et al. DeepReID: Deep Filter Pairing Neural Network for Person Re-identification. CVPR 2014. + + URL: ``_ + + Dataset statistics: + - identities: 1360. + - images: 13164. + - cameras: 6. + - splits: 20 (classic). + """ + dataset_dir = 'cuhk03' + dataset_url = None + + def __init__( + self, + root='', + split_id=0, + cuhk03_labeled=False, + cuhk03_classic_split=False, + **kwargs + ): + self.root = osp.abspath(osp.expanduser(root)) + self.dataset_dir = osp.join(self.root, self.dataset_dir) + self.download_dataset(self.dataset_dir, self.dataset_url) + + self.data_dir = osp.join(self.dataset_dir, 'cuhk03_release') + self.raw_mat_path = osp.join(self.data_dir, 'cuhk-03.mat') + + self.imgs_detected_dir = osp.join(self.dataset_dir, 'images_detected') + self.imgs_labeled_dir = osp.join(self.dataset_dir, 'images_labeled') + + self.split_classic_det_json_path = osp.join( + self.dataset_dir, 'splits_classic_detected.json' + ) + self.split_classic_lab_json_path = osp.join( + self.dataset_dir, 'splits_classic_labeled.json' + ) + + self.split_new_det_json_path = osp.join( + self.dataset_dir, 'splits_new_detected.json' + ) + self.split_new_lab_json_path = osp.join( + self.dataset_dir, 'splits_new_labeled.json' + ) + + self.split_new_det_mat_path = osp.join( + self.dataset_dir, 'cuhk03_new_protocol_config_detected.mat' + ) + self.split_new_lab_mat_path = osp.join( + self.dataset_dir, 'cuhk03_new_protocol_config_labeled.mat' + ) + + required_files = [ + self.dataset_dir, self.data_dir, self.raw_mat_path, + self.split_new_det_mat_path, self.split_new_lab_mat_path + ] + self.check_before_run(required_files) + + self.preprocess_split() + + if cuhk03_labeled: + split_path = self.split_classic_lab_json_path if cuhk03_classic_split else self.split_new_lab_json_path + else: + split_path = self.split_classic_det_json_path if cuhk03_classic_split else self.split_new_det_json_path + + splits = read_json(split_path) + assert split_id < len( + splits + ), 'Condition split_id ({}) < len(splits) ({}) is false'.format( + split_id, len(splits) + ) + split = splits[split_id] + + train = split['train'] + query = split['query'] + gallery = split['gallery'] + + super(CUHK03, self).__init__(train, query, gallery, **kwargs) + + def preprocess_split(self): + # This function is a bit complex and ugly, what it does is + # 1. extract data from cuhk-03.mat and save as png images + # 2. create 20 classic splits (Li et al. CVPR'14) + # 3. create new split (Zhong et al. CVPR'17) + if osp.exists(self.imgs_labeled_dir) \ + and osp.exists(self.imgs_detected_dir) \ + and osp.exists(self.split_classic_det_json_path) \ + and osp.exists(self.split_classic_lab_json_path) \ + and osp.exists(self.split_new_det_json_path) \ + and osp.exists(self.split_new_lab_json_path): + return + + import h5py + import imageio + from scipy.io import loadmat + + mkdir_if_missing(self.imgs_detected_dir) + mkdir_if_missing(self.imgs_labeled_dir) + + print( + 'Extract image data from "{}" and save as png'.format( + self.raw_mat_path + ) + ) + mat = h5py.File(self.raw_mat_path, 'r') + + def _deref(ref): + return mat[ref][:].T + + def _process_images(img_refs, campid, pid, save_dir): + img_paths = [] # Note: some persons only have images for one view + for imgid, img_ref in enumerate(img_refs): + img = _deref(img_ref) + if img.size == 0 or img.ndim < 3: + continue # skip empty cell + # images are saved with the following format, index-1 (ensure uniqueness) + # campid: index of camera pair (1-5) + # pid: index of person in 'campid'-th camera pair + # viewid: index of view, {1, 2} + # imgid: index of image, (1-10) + viewid = 1 if imgid < 5 else 2 + img_name = '{:01d}_{:03d}_{:01d}_{:02d}.png'.format( + campid + 1, pid + 1, viewid, imgid + 1 + ) + img_path = osp.join(save_dir, img_name) + if not osp.isfile(img_path): + imageio.imwrite(img_path, img) + img_paths.append(img_path) + return img_paths + + def _extract_img(image_type): + print('Processing {} images ...'.format(image_type)) + meta_data = [] + imgs_dir = self.imgs_detected_dir if image_type == 'detected' else self.imgs_labeled_dir + for campid, camp_ref in enumerate(mat[image_type][0]): + camp = _deref(camp_ref) + num_pids = camp.shape[0] + for pid in range(num_pids): + img_paths = _process_images( + camp[pid, :], campid, pid, imgs_dir + ) + assert len(img_paths) > 0, \ + 'campid{}-pid{} has no images'.format(campid, pid) + meta_data.append((campid + 1, pid + 1, img_paths)) + print( + '- done camera pair {} with {} identities'.format( + campid + 1, num_pids + ) + ) + return meta_data + + meta_detected = _extract_img('detected') + meta_labeled = _extract_img('labeled') + + def _extract_classic_split(meta_data, test_split): + train, test = [], [] + num_train_pids, num_test_pids = 0, 0 + num_train_imgs, num_test_imgs = 0, 0 + for i, (campid, pid, img_paths) in enumerate(meta_data): + + if [campid, pid] in test_split: + for img_path in img_paths: + camid = int( + osp.basename(img_path).split('_')[2] + ) - 1 # make it 0-based + test.append((img_path, num_test_pids, camid)) + num_test_pids += 1 + num_test_imgs += len(img_paths) + else: + for img_path in img_paths: + camid = int( + osp.basename(img_path).split('_')[2] + ) - 1 # make it 0-based + train.append((img_path, num_train_pids, camid)) + num_train_pids += 1 + num_train_imgs += len(img_paths) + return train, num_train_pids, num_train_imgs, test, num_test_pids, num_test_imgs + + print('Creating classic splits (# = 20) ...') + splits_classic_det, splits_classic_lab = [], [] + for split_ref in mat['testsets'][0]: + test_split = _deref(split_ref).tolist() + + # create split for detected images + train, num_train_pids, num_train_imgs, test, num_test_pids, num_test_imgs = \ + _extract_classic_split(meta_detected, test_split) + splits_classic_det.append( + { + 'train': train, + 'query': test, + 'gallery': test, + 'num_train_pids': num_train_pids, + 'num_train_imgs': num_train_imgs, + 'num_query_pids': num_test_pids, + 'num_query_imgs': num_test_imgs, + 'num_gallery_pids': num_test_pids, + 'num_gallery_imgs': num_test_imgs + } + ) + + # create split for labeled images + train, num_train_pids, num_train_imgs, test, num_test_pids, num_test_imgs = \ + _extract_classic_split(meta_labeled, test_split) + splits_classic_lab.append( + { + 'train': train, + 'query': test, + 'gallery': test, + 'num_train_pids': num_train_pids, + 'num_train_imgs': num_train_imgs, + 'num_query_pids': num_test_pids, + 'num_query_imgs': num_test_imgs, + 'num_gallery_pids': num_test_pids, + 'num_gallery_imgs': num_test_imgs + } + ) + + write_json(splits_classic_det, self.split_classic_det_json_path) + write_json(splits_classic_lab, self.split_classic_lab_json_path) + + def _extract_set(filelist, pids, pid2label, idxs, img_dir, relabel): + tmp_set = [] + unique_pids = set() + for idx in idxs: + img_name = filelist[idx][0] + camid = int(img_name.split('_')[2]) - 1 # make it 0-based + pid = pids[idx] + if relabel: + pid = pid2label[pid] + img_path = osp.join(img_dir, img_name) + tmp_set.append((img_path, int(pid), camid)) + unique_pids.add(pid) + return tmp_set, len(unique_pids), len(idxs) + + def _extract_new_split(split_dict, img_dir): + train_idxs = split_dict['train_idx'].flatten() - 1 # index-0 + pids = split_dict['labels'].flatten() + train_pids = set(pids[train_idxs]) + pid2label = {pid: label for label, pid in enumerate(train_pids)} + query_idxs = split_dict['query_idx'].flatten() - 1 + gallery_idxs = split_dict['gallery_idx'].flatten() - 1 + filelist = split_dict['filelist'].flatten() + train_info = _extract_set( + filelist, pids, pid2label, train_idxs, img_dir, relabel=True + ) + query_info = _extract_set( + filelist, pids, pid2label, query_idxs, img_dir, relabel=False + ) + gallery_info = _extract_set( + filelist, + pids, + pid2label, + gallery_idxs, + img_dir, + relabel=False + ) + return train_info, query_info, gallery_info + + print('Creating new split for detected images (767/700) ...') + train_info, query_info, gallery_info = _extract_new_split( + loadmat(self.split_new_det_mat_path), self.imgs_detected_dir + ) + split = [ + { + 'train': train_info[0], + 'query': query_info[0], + 'gallery': gallery_info[0], + 'num_train_pids': train_info[1], + 'num_train_imgs': train_info[2], + 'num_query_pids': query_info[1], + 'num_query_imgs': query_info[2], + 'num_gallery_pids': gallery_info[1], + 'num_gallery_imgs': gallery_info[2] + } + ] + write_json(split, self.split_new_det_json_path) + + print('Creating new split for labeled images (767/700) ...') + train_info, query_info, gallery_info = _extract_new_split( + loadmat(self.split_new_lab_mat_path), self.imgs_labeled_dir + ) + split = [ + { + 'train': train_info[0], + 'query': query_info[0], + 'gallery': gallery_info[0], + 'num_train_pids': train_info[1], + 'num_train_imgs': train_info[2], + 'num_query_pids': query_info[1], + 'num_query_imgs': query_info[2], + 'num_gallery_pids': gallery_info[1], + 'num_gallery_imgs': gallery_info[2] + } + ] + write_json(split, self.split_new_lab_json_path) diff --git a/strong_sort/deep/reid/torchreid/data/datasets/image/cuhksysu.py b/strong_sort/deep/reid/torchreid/data/datasets/image/cuhksysu.py new file mode 100644 index 0000000..f6c9edd --- /dev/null +++ b/strong_sort/deep/reid/torchreid/data/datasets/image/cuhksysu.py @@ -0,0 +1,60 @@ +from __future__ import division, print_function, absolute_import +import copy +import glob +import os.path as osp + +from ..dataset import ImageDataset + + +class CUHKSYSU(ImageDataset): + """CUHKSYSU. + + This dataset can only be used for model training. + + Reference: + Xiao et al. End-to-end deep learning for person search. + + URL: ``_ + + Dataset statistics: + - identities: 11,934 + - images: 34,574 + """ + _train_only = True + dataset_dir = 'cuhksysu' + + def __init__(self, root='', **kwargs): + self.root = osp.abspath(osp.expanduser(root)) + self.dataset_dir = osp.join(self.root, self.dataset_dir) + self.data_dir = osp.join(self.dataset_dir, 'cropped_images') + + # image name format: p11422_s16929_1.jpg + train = self.process_dir(self.data_dir) + query = [copy.deepcopy(train[0])] + gallery = [copy.deepcopy(train[0])] + + super(CUHKSYSU, self).__init__(train, query, gallery, **kwargs) + + def process_dir(self, dirname): + img_paths = glob.glob(osp.join(dirname, '*.jpg')) + # num_imgs = len(img_paths) + + # get all identities: + pid_container = set() + for img_path in img_paths: + img_name = osp.basename(img_path) + pid = img_name.split('_')[0] + pid_container.add(pid) + pid2label = {pid: label for label, pid in enumerate(pid_container)} + + # num_pids = len(pid_container) + + # extract data + data = [] + for img_path in img_paths: + img_name = osp.basename(img_path) + pid = img_name.split('_')[0] + label = pid2label[pid] + data.append((img_path, label, 0)) # dummy camera id + + return data diff --git a/strong_sort/deep/reid/torchreid/data/datasets/image/dukemtmcreid.py b/strong_sort/deep/reid/torchreid/data/datasets/image/dukemtmcreid.py new file mode 100644 index 0000000..5915da5 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/data/datasets/image/dukemtmcreid.py @@ -0,0 +1,68 @@ +from __future__ import division, print_function, absolute_import +import re +import glob +import os.path as osp + +from ..dataset import ImageDataset + + +class DukeMTMCreID(ImageDataset): + """DukeMTMC-reID. + + Reference: + - Ristani et al. Performance Measures and a Data Set for Multi-Target, Multi-Camera Tracking. ECCVW 2016. + - Zheng et al. Unlabeled Samples Generated by GAN Improve the Person Re-identification Baseline in vitro. ICCV 2017. + + URL: ``_ + + Dataset statistics: + - identities: 1404 (train + query). + - images:16522 (train) + 2228 (query) + 17661 (gallery). + - cameras: 8. + """ + dataset_dir = 'dukemtmc-reid' + dataset_url = 'http://vision.cs.duke.edu/DukeMTMC/data/misc/DukeMTMC-reID.zip' + + def __init__(self, root='', **kwargs): + self.root = osp.abspath(osp.expanduser(root)) + self.dataset_dir = osp.join(self.root, self.dataset_dir) + self.download_dataset(self.dataset_dir, self.dataset_url) + self.train_dir = osp.join( + self.dataset_dir, 'DukeMTMC-reID/bounding_box_train' + ) + self.query_dir = osp.join(self.dataset_dir, 'DukeMTMC-reID/query') + self.gallery_dir = osp.join( + self.dataset_dir, 'DukeMTMC-reID/bounding_box_test' + ) + + required_files = [ + self.dataset_dir, self.train_dir, self.query_dir, self.gallery_dir + ] + self.check_before_run(required_files) + + train = self.process_dir(self.train_dir, relabel=True) + query = self.process_dir(self.query_dir, relabel=False) + gallery = self.process_dir(self.gallery_dir, relabel=False) + + super(DukeMTMCreID, self).__init__(train, query, gallery, **kwargs) + + def process_dir(self, dir_path, relabel=False): + img_paths = glob.glob(osp.join(dir_path, '*.jpg')) + pattern = re.compile(r'([-\d]+)_c(\d)') + + pid_container = set() + for img_path in img_paths: + pid, _ = map(int, pattern.search(img_path).groups()) + pid_container.add(pid) + pid2label = {pid: label for label, pid in enumerate(pid_container)} + + data = [] + for img_path in img_paths: + pid, camid = map(int, pattern.search(img_path).groups()) + assert 1 <= camid <= 8 + camid -= 1 # index starts from 0 + if relabel: + pid = pid2label[pid] + data.append((img_path, pid, camid)) + + return data diff --git a/strong_sort/deep/reid/torchreid/data/datasets/image/grid.py b/strong_sort/deep/reid/torchreid/data/datasets/image/grid.py new file mode 100644 index 0000000..96023d6 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/data/datasets/image/grid.py @@ -0,0 +1,131 @@ +from __future__ import division, print_function, absolute_import +import glob +import os.path as osp +from scipy.io import loadmat + +from torchreid.utils import read_json, write_json + +from ..dataset import ImageDataset + + +class GRID(ImageDataset): + """GRID. + + Reference: + Loy et al. Multi-camera activity correlation analysis. CVPR 2009. + + URL: ``_ + + Dataset statistics: + - identities: 250. + - images: 1275. + - cameras: 8. + """ + dataset_dir = 'grid' + dataset_url = 'http://personal.ie.cuhk.edu.hk/~ccloy/files/datasets/underground_reid.zip' + _junk_pids = [0] + + def __init__(self, root='', split_id=0, **kwargs): + self.root = osp.abspath(osp.expanduser(root)) + self.dataset_dir = osp.join(self.root, self.dataset_dir) + self.download_dataset(self.dataset_dir, self.dataset_url) + + self.probe_path = osp.join( + self.dataset_dir, 'underground_reid', 'probe' + ) + self.gallery_path = osp.join( + self.dataset_dir, 'underground_reid', 'gallery' + ) + self.split_mat_path = osp.join( + self.dataset_dir, 'underground_reid', 'features_and_partitions.mat' + ) + self.split_path = osp.join(self.dataset_dir, 'splits.json') + + required_files = [ + self.dataset_dir, self.probe_path, self.gallery_path, + self.split_mat_path + ] + self.check_before_run(required_files) + + self.prepare_split() + splits = read_json(self.split_path) + if split_id >= len(splits): + raise ValueError( + 'split_id exceeds range, received {}, ' + 'but expected between 0 and {}'.format( + split_id, + len(splits) - 1 + ) + ) + split = splits[split_id] + + train = split['train'] + query = split['query'] + gallery = split['gallery'] + + train = [tuple(item) for item in train] + query = [tuple(item) for item in query] + gallery = [tuple(item) for item in gallery] + + super(GRID, self).__init__(train, query, gallery, **kwargs) + + def prepare_split(self): + if not osp.exists(self.split_path): + print('Creating 10 random splits') + split_mat = loadmat(self.split_mat_path) + trainIdxAll = split_mat['trainIdxAll'][0] # length = 10 + probe_img_paths = sorted( + glob.glob(osp.join(self.probe_path, '*.jpeg')) + ) + gallery_img_paths = sorted( + glob.glob(osp.join(self.gallery_path, '*.jpeg')) + ) + + splits = [] + for split_idx in range(10): + train_idxs = trainIdxAll[split_idx][0][0][2][0].tolist() + assert len(train_idxs) == 125 + idx2label = { + idx: label + for label, idx in enumerate(train_idxs) + } + + train, query, gallery = [], [], [] + + # processing probe folder + for img_path in probe_img_paths: + img_name = osp.basename(img_path) + img_idx = int(img_name.split('_')[0]) + camid = int( + img_name.split('_')[1] + ) - 1 # index starts from 0 + if img_idx in train_idxs: + train.append((img_path, idx2label[img_idx], camid)) + else: + query.append((img_path, img_idx, camid)) + + # process gallery folder + for img_path in gallery_img_paths: + img_name = osp.basename(img_path) + img_idx = int(img_name.split('_')[0]) + camid = int( + img_name.split('_')[1] + ) - 1 # index starts from 0 + if img_idx in train_idxs: + train.append((img_path, idx2label[img_idx], camid)) + else: + gallery.append((img_path, img_idx, camid)) + + split = { + 'train': train, + 'query': query, + 'gallery': gallery, + 'num_train_pids': 125, + 'num_query_pids': 125, + 'num_gallery_pids': 900 + } + splits.append(split) + + print('Totally {} splits are created'.format(len(splits))) + write_json(splits, self.split_path) + print('Split file saved to {}'.format(self.split_path)) diff --git a/strong_sort/deep/reid/torchreid/data/datasets/image/ilids.py b/strong_sort/deep/reid/torchreid/data/datasets/image/ilids.py new file mode 100644 index 0000000..42971b0 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/data/datasets/image/ilids.py @@ -0,0 +1,135 @@ +from __future__ import division, print_function, absolute_import +import copy +import glob +import random +import os.path as osp +from collections import defaultdict + +from torchreid.utils import read_json, write_json + +from ..dataset import ImageDataset + + +class iLIDS(ImageDataset): + """QMUL-iLIDS. + + Reference: + Zheng et al. Associating Groups of People. BMVC 2009. + + Dataset statistics: + - identities: 119. + - images: 476. + - cameras: 8 (not explicitly provided). + """ + dataset_dir = 'ilids' + dataset_url = 'http://www.eecs.qmul.ac.uk/~jason/data/i-LIDS_Pedestrian.tgz' + + def __init__(self, root='', split_id=0, **kwargs): + self.root = osp.abspath(osp.expanduser(root)) + self.dataset_dir = osp.join(self.root, self.dataset_dir) + self.download_dataset(self.dataset_dir, self.dataset_url) + + self.data_dir = osp.join(self.dataset_dir, 'i-LIDS_Pedestrian/Persons') + self.split_path = osp.join(self.dataset_dir, 'splits.json') + + required_files = [self.dataset_dir, self.data_dir] + self.check_before_run(required_files) + + self.prepare_split() + splits = read_json(self.split_path) + if split_id >= len(splits): + raise ValueError( + 'split_id exceeds range, received {}, but ' + 'expected between 0 and {}'.format(split_id, + len(splits) - 1) + ) + split = splits[split_id] + + train, query, gallery = self.process_split(split) + + super(iLIDS, self).__init__(train, query, gallery, **kwargs) + + def prepare_split(self): + if not osp.exists(self.split_path): + print('Creating splits ...') + + paths = glob.glob(osp.join(self.data_dir, '*.jpg')) + img_names = [osp.basename(path) for path in paths] + num_imgs = len(img_names) + assert num_imgs == 476, 'There should be 476 images, but ' \ + 'got {}, please check the data'.format(num_imgs) + + # store image names + # image naming format: + # the first four digits denote the person ID + # the last four digits denote the sequence index + pid_dict = defaultdict(list) + for img_name in img_names: + pid = int(img_name[:4]) + pid_dict[pid].append(img_name) + pids = list(pid_dict.keys()) + num_pids = len(pids) + assert num_pids == 119, 'There should be 119 identities, ' \ + 'but got {}, please check the data'.format(num_pids) + + num_train_pids = int(num_pids * 0.5) + + splits = [] + for _ in range(10): + # randomly choose num_train_pids train IDs and the rest for test IDs + pids_copy = copy.deepcopy(pids) + random.shuffle(pids_copy) + train_pids = pids_copy[:num_train_pids] + test_pids = pids_copy[num_train_pids:] + + train = [] + query = [] + gallery = [] + + # for train IDs, all images are used in the train set. + for pid in train_pids: + img_names = pid_dict[pid] + train.extend(img_names) + + # for each test ID, randomly choose two images, one for + # query and the other one for gallery. + for pid in test_pids: + img_names = pid_dict[pid] + samples = random.sample(img_names, 2) + query.append(samples[0]) + gallery.append(samples[1]) + + split = {'train': train, 'query': query, 'gallery': gallery} + splits.append(split) + + print('Totally {} splits are created'.format(len(splits))) + write_json(splits, self.split_path) + print('Split file is saved to {}'.format(self.split_path)) + + def get_pid2label(self, img_names): + pid_container = set() + for img_name in img_names: + pid = int(img_name[:4]) + pid_container.add(pid) + pid2label = {pid: label for label, pid in enumerate(pid_container)} + return pid2label + + def parse_img_names(self, img_names, pid2label=None): + data = [] + + for img_name in img_names: + pid = int(img_name[:4]) + if pid2label is not None: + pid = pid2label[pid] + camid = int(img_name[4:7]) - 1 # 0-based + img_path = osp.join(self.data_dir, img_name) + data.append((img_path, pid, camid)) + + return data + + def process_split(self, split): + train_pid2label = self.get_pid2label(split['train']) + train = self.parse_img_names(split['train'], train_pid2label) + query = self.parse_img_names(split['query']) + gallery = self.parse_img_names(split['gallery']) + return train, query, gallery diff --git a/strong_sort/deep/reid/torchreid/data/datasets/image/market1501.py b/strong_sort/deep/reid/torchreid/data/datasets/image/market1501.py new file mode 100644 index 0000000..7d138d1 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/data/datasets/image/market1501.py @@ -0,0 +1,88 @@ +from __future__ import division, print_function, absolute_import +import re +import glob +import os.path as osp +import warnings + +from ..dataset import ImageDataset + + +class Market1501(ImageDataset): + """Market1501. + + Reference: + Zheng et al. Scalable Person Re-identification: A Benchmark. ICCV 2015. + + URL: ``_ + + Dataset statistics: + - identities: 1501 (+1 for background). + - images: 12936 (train) + 3368 (query) + 15913 (gallery). + """ + _junk_pids = [0, -1] + dataset_dir = 'market1501' + dataset_url = 'http://188.138.127.15:81/Datasets/Market-1501-v15.09.15.zip' + + def __init__(self, root='', market1501_500k=False, **kwargs): + self.root = osp.abspath(osp.expanduser(root)) + self.dataset_dir = osp.join(self.root, self.dataset_dir) + self.download_dataset(self.dataset_dir, self.dataset_url) + + # allow alternative directory structure + self.data_dir = self.dataset_dir + data_dir = osp.join(self.data_dir, 'Market-1501-v15.09.15') + if osp.isdir(data_dir): + self.data_dir = data_dir + else: + warnings.warn( + 'The current data structure is deprecated. Please ' + 'put data folders such as "bounding_box_train" under ' + '"Market-1501-v15.09.15".' + ) + + self.train_dir = osp.join(self.data_dir, 'bounding_box_train') + self.query_dir = osp.join(self.data_dir, 'query') + self.gallery_dir = osp.join(self.data_dir, 'bounding_box_test') + self.extra_gallery_dir = osp.join(self.data_dir, 'images') + self.market1501_500k = market1501_500k + + required_files = [ + self.data_dir, self.train_dir, self.query_dir, self.gallery_dir + ] + if self.market1501_500k: + required_files.append(self.extra_gallery_dir) + self.check_before_run(required_files) + + train = self.process_dir(self.train_dir, relabel=True) + query = self.process_dir(self.query_dir, relabel=False) + gallery = self.process_dir(self.gallery_dir, relabel=False) + if self.market1501_500k: + gallery += self.process_dir(self.extra_gallery_dir, relabel=False) + + super(Market1501, self).__init__(train, query, gallery, **kwargs) + + def process_dir(self, dir_path, relabel=False): + img_paths = glob.glob(osp.join(dir_path, '*.jpg')) + pattern = re.compile(r'([-\d]+)_c(\d)') + + pid_container = set() + for img_path in img_paths: + pid, _ = map(int, pattern.search(img_path).groups()) + if pid == -1: + continue # junk images are just ignored + pid_container.add(pid) + pid2label = {pid: label for label, pid in enumerate(pid_container)} + + data = [] + for img_path in img_paths: + pid, camid = map(int, pattern.search(img_path).groups()) + if pid == -1: + continue # junk images are just ignored + assert 0 <= pid <= 1501 # pid == 0 means background + assert 1 <= camid <= 6 + camid -= 1 # index starts from 0 + if relabel: + pid = pid2label[pid] + data.append((img_path, pid, camid)) + + return data diff --git a/strong_sort/deep/reid/torchreid/data/datasets/image/msmt17.py b/strong_sort/deep/reid/torchreid/data/datasets/image/msmt17.py new file mode 100644 index 0000000..c4741e6 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/data/datasets/image/msmt17.py @@ -0,0 +1,98 @@ +from __future__ import division, print_function, absolute_import +import os.path as osp + +from ..dataset import ImageDataset + +# Log +# 22.01.2019 +# - add v2 +# - v1 and v2 differ in dir names +# - note that faces in v2 are blurred +TRAIN_DIR_KEY = 'train_dir' +TEST_DIR_KEY = 'test_dir' +VERSION_DICT = { + 'MSMT17_V1': { + TRAIN_DIR_KEY: 'train', + TEST_DIR_KEY: 'test', + }, + 'MSMT17_V2': { + TRAIN_DIR_KEY: 'mask_train_v2', + TEST_DIR_KEY: 'mask_test_v2', + } +} + + +class MSMT17(ImageDataset): + """MSMT17. + + Reference: + Wei et al. Person Transfer GAN to Bridge Domain Gap for Person Re-Identification. CVPR 2018. + + URL: ``_ + + Dataset statistics: + - identities: 4101. + - images: 32621 (train) + 11659 (query) + 82161 (gallery). + - cameras: 15. + """ + dataset_dir = 'msmt17' + dataset_url = None + + def __init__(self, root='', **kwargs): + self.root = osp.abspath(osp.expanduser(root)) + self.dataset_dir = osp.join(self.root, self.dataset_dir) + self.download_dataset(self.dataset_dir, self.dataset_url) + + has_main_dir = False + for main_dir in VERSION_DICT: + if osp.exists(osp.join(self.dataset_dir, main_dir)): + train_dir = VERSION_DICT[main_dir][TRAIN_DIR_KEY] + test_dir = VERSION_DICT[main_dir][TEST_DIR_KEY] + has_main_dir = True + break + assert has_main_dir, 'Dataset folder not found' + + self.train_dir = osp.join(self.dataset_dir, main_dir, train_dir) + self.test_dir = osp.join(self.dataset_dir, main_dir, test_dir) + self.list_train_path = osp.join( + self.dataset_dir, main_dir, 'list_train.txt' + ) + self.list_val_path = osp.join( + self.dataset_dir, main_dir, 'list_val.txt' + ) + self.list_query_path = osp.join( + self.dataset_dir, main_dir, 'list_query.txt' + ) + self.list_gallery_path = osp.join( + self.dataset_dir, main_dir, 'list_gallery.txt' + ) + + required_files = [self.dataset_dir, self.train_dir, self.test_dir] + self.check_before_run(required_files) + + train = self.process_dir(self.train_dir, self.list_train_path) + val = self.process_dir(self.train_dir, self.list_val_path) + query = self.process_dir(self.test_dir, self.list_query_path) + gallery = self.process_dir(self.test_dir, self.list_gallery_path) + + # Note: to fairly compare with published methods on the conventional ReID setting, + # do not add val images to the training set. + if 'combineall' in kwargs and kwargs['combineall']: + train += val + + super(MSMT17, self).__init__(train, query, gallery, **kwargs) + + def process_dir(self, dir_path, list_path): + with open(list_path, 'r') as txt: + lines = txt.readlines() + + data = [] + + for img_idx, img_info in enumerate(lines): + img_path, pid = img_info.split(' ') + pid = int(pid) # no need to relabel + camid = int(img_path.split('_')[2]) - 1 # index starts from 0 + img_path = osp.join(dir_path, img_path) + data.append((img_path, pid, camid)) + + return data diff --git a/strong_sort/deep/reid/torchreid/data/datasets/image/prid.py b/strong_sort/deep/reid/torchreid/data/datasets/image/prid.py new file mode 100644 index 0000000..d6d6c20 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/data/datasets/image/prid.py @@ -0,0 +1,107 @@ +from __future__ import division, print_function, absolute_import +import random +import os.path as osp + +from torchreid.utils import read_json, write_json + +from ..dataset import ImageDataset + + +class PRID(ImageDataset): + """PRID (single-shot version of prid-2011) + + Reference: + Hirzer et al. Person Re-Identification by Descriptive and Discriminative + Classification. SCIA 2011. + + URL: ``_ + + Dataset statistics: + - Two views. + - View A captures 385 identities. + - View B captures 749 identities. + - 200 identities appear in both views (index starts from 1 to 200). + """ + dataset_dir = 'prid2011' + dataset_url = None + _junk_pids = list(range(201, 750)) + + def __init__(self, root='', split_id=0, **kwargs): + self.root = osp.abspath(osp.expanduser(root)) + self.dataset_dir = osp.join(self.root, self.dataset_dir) + self.download_dataset(self.dataset_dir, self.dataset_url) + + self.cam_a_dir = osp.join( + self.dataset_dir, 'prid_2011', 'single_shot', 'cam_a' + ) + self.cam_b_dir = osp.join( + self.dataset_dir, 'prid_2011', 'single_shot', 'cam_b' + ) + self.split_path = osp.join(self.dataset_dir, 'splits_single_shot.json') + + required_files = [self.dataset_dir, self.cam_a_dir, self.cam_b_dir] + self.check_before_run(required_files) + + self.prepare_split() + splits = read_json(self.split_path) + if split_id >= len(splits): + raise ValueError( + 'split_id exceeds range, received {}, but expected between 0 and {}' + .format(split_id, + len(splits) - 1) + ) + split = splits[split_id] + + train, query, gallery = self.process_split(split) + + super(PRID, self).__init__(train, query, gallery, **kwargs) + + def prepare_split(self): + if not osp.exists(self.split_path): + print('Creating splits ...') + + splits = [] + for _ in range(10): + # randomly sample 100 IDs for train and use the rest 100 IDs for test + # (note: there are only 200 IDs appearing in both views) + pids = [i for i in range(1, 201)] + train_pids = random.sample(pids, 100) + train_pids.sort() + test_pids = [i for i in pids if i not in train_pids] + split = {'train': train_pids, 'test': test_pids} + splits.append(split) + + print('Totally {} splits are created'.format(len(splits))) + write_json(splits, self.split_path) + print('Split file is saved to {}'.format(self.split_path)) + + def process_split(self, split): + train_pids = split['train'] + test_pids = split['test'] + + train_pid2label = {pid: label for label, pid in enumerate(train_pids)} + + # train + train = [] + for pid in train_pids: + img_name = 'person_' + str(pid).zfill(4) + '.png' + pid = train_pid2label[pid] + img_a_path = osp.join(self.cam_a_dir, img_name) + train.append((img_a_path, pid, 0)) + img_b_path = osp.join(self.cam_b_dir, img_name) + train.append((img_b_path, pid, 1)) + + # query and gallery + query, gallery = [], [] + for pid in test_pids: + img_name = 'person_' + str(pid).zfill(4) + '.png' + img_a_path = osp.join(self.cam_a_dir, img_name) + query.append((img_a_path, pid, 0)) + img_b_path = osp.join(self.cam_b_dir, img_name) + gallery.append((img_b_path, pid, 1)) + for pid in range(201, 750): + img_name = 'person_' + str(pid).zfill(4) + '.png' + img_b_path = osp.join(self.cam_b_dir, img_name) + gallery.append((img_b_path, pid, 1)) + + return train, query, gallery diff --git a/strong_sort/deep/reid/torchreid/data/datasets/image/sensereid.py b/strong_sort/deep/reid/torchreid/data/datasets/image/sensereid.py new file mode 100644 index 0000000..7cf5f32 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/data/datasets/image/sensereid.py @@ -0,0 +1,70 @@ +from __future__ import division, print_function, absolute_import +import copy +import glob +import os.path as osp + +from ..dataset import ImageDataset + + +class SenseReID(ImageDataset): + """SenseReID. + + This dataset is used for test purpose only. + + Reference: + Zhao et al. Spindle Net: Person Re-identification with Human Body + Region Guided Feature Decomposition and Fusion. CVPR 2017. + + URL: ``_ + + Dataset statistics: + - query: 522 ids, 1040 images. + - gallery: 1717 ids, 3388 images. + """ + dataset_dir = 'sensereid' + dataset_url = None + + def __init__(self, root='', **kwargs): + self.root = osp.abspath(osp.expanduser(root)) + self.dataset_dir = osp.join(self.root, self.dataset_dir) + self.download_dataset(self.dataset_dir, self.dataset_url) + + self.query_dir = osp.join(self.dataset_dir, 'SenseReID', 'test_probe') + self.gallery_dir = osp.join( + self.dataset_dir, 'SenseReID', 'test_gallery' + ) + + required_files = [self.dataset_dir, self.query_dir, self.gallery_dir] + self.check_before_run(required_files) + + query = self.process_dir(self.query_dir) + gallery = self.process_dir(self.gallery_dir) + + # relabel + g_pids = set() + for _, pid, _ in gallery: + g_pids.add(pid) + pid2label = {pid: i for i, pid in enumerate(g_pids)} + + query = [ + (img_path, pid2label[pid], camid) for img_path, pid, camid in query + ] + gallery = [ + (img_path, pid2label[pid], camid) + for img_path, pid, camid in gallery + ] + train = copy.deepcopy(query) + copy.deepcopy(gallery) # dummy variable + + super(SenseReID, self).__init__(train, query, gallery, **kwargs) + + def process_dir(self, dir_path): + img_paths = glob.glob(osp.join(dir_path, '*.jpg')) + data = [] + + for img_path in img_paths: + img_name = osp.splitext(osp.basename(img_path))[0] + pid, camid = img_name.split('_') + pid, camid = int(pid), int(camid) + data.append((img_path, pid, camid)) + + return data diff --git a/strong_sort/deep/reid/torchreid/data/datasets/image/university1652.py b/strong_sort/deep/reid/torchreid/data/datasets/image/university1652.py new file mode 100644 index 0000000..ce1e386 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/data/datasets/image/university1652.py @@ -0,0 +1,110 @@ +from __future__ import division, print_function, absolute_import +import os +import glob +import os.path as osp +import gdown + +from ..dataset import ImageDataset + + +class University1652(ImageDataset): + """University-1652. + + Reference: + - Zheng et al. University-1652: A Multi-view Multi-source Benchmark for Drone-based Geo-localization. ACM MM 2020. + + URL: ``_ + OneDrive: + https://studentutsedu-my.sharepoint.com/:u:/g/personal/12639605_student_uts_edu_au/Ecrz6xK-PcdCjFdpNb0T0s8B_9J5ynaUy3q63_XumjJyrA?e=z4hpcz + [Backup] GoogleDrive: + https://drive.google.com/file/d/1iVnP4gjw-iHXa0KerZQ1IfIO0i1jADsR/view?usp=sharing + [Backup] Baidu Yun: + https://pan.baidu.com/s/1H_wBnWwikKbaBY1pMPjoqQ password: hrqp + + Dataset statistics: + - buildings: 1652 (train + query). + - The dataset split is as follows: + | Split | #imgs | #buildings | #universities| + | -------- | ----- | ----| ----| + | Training | 50,218 | 701 | 33 | + | Query_drone | 37,855 | 701 | 39 | + | Query_satellite | 701 | 701 | 39| + | Query_ground | 2,579 | 701 | 39| + | Gallery_drone | 51,355 | 951 | 39| + | Gallery_satellite | 951 | 951 | 39| + | Gallery_ground | 2,921 | 793 | 39| + - cameras: None. + + datamanager = torchreid.data.ImageDataManager( + root='reid-data', + sources='university1652', + targets='university1652', + height=256, + width=256, + batch_size_train=32, + batch_size_test=100, + transforms=['random_flip', 'random_crop'] + ) + """ + dataset_dir = 'university1652' + dataset_url = 'https://drive.google.com/uc?id=1iVnP4gjw-iHXa0KerZQ1IfIO0i1jADsR' + + def __init__(self, root='', **kwargs): + self.root = osp.abspath(osp.expanduser(root)) + self.dataset_dir = osp.join(self.root, self.dataset_dir) + print(self.dataset_dir) + if not os.path.isdir(self.dataset_dir): + os.mkdir(self.dataset_dir) + gdown.download( + self.dataset_url, self.dataset_dir + 'data.zip', quiet=False + ) + os.system('unzip %s' % (self.dataset_dir + 'data.zip')) + self.train_dir = osp.join( + self.dataset_dir, 'University-Release/train/' + ) + self.query_dir = osp.join( + self.dataset_dir, 'University-Release/test/query_drone' + ) + self.gallery_dir = osp.join( + self.dataset_dir, 'University-Release/test/gallery_satellite' + ) + + required_files = [ + self.dataset_dir, self.train_dir, self.query_dir, self.gallery_dir + ] + self.check_before_run(required_files) + + self.fake_camid = 0 + train = self.process_dir(self.train_dir, relabel=True, train=True) + query = self.process_dir(self.query_dir, relabel=False) + gallery = self.process_dir(self.gallery_dir, relabel=False) + + super(University1652, self).__init__(train, query, gallery, **kwargs) + + def process_dir(self, dir_path, relabel=False, train=False): + IMG_EXTENSIONS = ( + '.jpg', '.jpeg', '.png', '.ppm', '.bmp', '.pgm', '.tif', '.tiff', + '.webp' + ) + if train: + img_paths = glob.glob(osp.join(dir_path, '*/*/*')) + else: + img_paths = glob.glob(osp.join(dir_path, '*/*')) + pid_container = set() + for img_path in img_paths: + if not img_path.lower().endswith(IMG_EXTENSIONS): + continue + pid = int(os.path.basename(os.path.dirname(img_path))) + pid_container.add(pid) + pid2label = {pid: label for label, pid in enumerate(pid_container)} + data = [] + # no camera for university + for img_path in img_paths: + if not img_path.lower().endswith(IMG_EXTENSIONS): + continue + pid = int(os.path.basename(os.path.dirname(img_path))) + if relabel: + pid = pid2label[pid] + data.append((img_path, pid, self.fake_camid)) + self.fake_camid += 1 + return data diff --git a/strong_sort/deep/reid/torchreid/data/datasets/image/viper.py b/strong_sort/deep/reid/torchreid/data/datasets/image/viper.py new file mode 100644 index 0000000..161dd99 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/data/datasets/image/viper.py @@ -0,0 +1,128 @@ +from __future__ import division, print_function, absolute_import +import glob +import numpy as np +import os.path as osp + +from torchreid.utils import read_json, write_json + +from ..dataset import ImageDataset + + +class VIPeR(ImageDataset): + """VIPeR. + + Reference: + Gray et al. Evaluating appearance models for recognition, reacquisition, and tracking. PETS 2007. + + URL: ``_ + + Dataset statistics: + - identities: 632. + - images: 632 x 2 = 1264. + - cameras: 2. + """ + dataset_dir = 'viper' + dataset_url = 'http://users.soe.ucsc.edu/~manduchi/VIPeR.v1.0.zip' + + def __init__(self, root='', split_id=0, **kwargs): + self.root = osp.abspath(osp.expanduser(root)) + self.dataset_dir = osp.join(self.root, self.dataset_dir) + self.download_dataset(self.dataset_dir, self.dataset_url) + + self.cam_a_dir = osp.join(self.dataset_dir, 'VIPeR', 'cam_a') + self.cam_b_dir = osp.join(self.dataset_dir, 'VIPeR', 'cam_b') + self.split_path = osp.join(self.dataset_dir, 'splits.json') + + required_files = [self.dataset_dir, self.cam_a_dir, self.cam_b_dir] + self.check_before_run(required_files) + + self.prepare_split() + splits = read_json(self.split_path) + if split_id >= len(splits): + raise ValueError( + 'split_id exceeds range, received {}, ' + 'but expected between 0 and {}'.format( + split_id, + len(splits) - 1 + ) + ) + split = splits[split_id] + + train = split['train'] + query = split['query'] # query and gallery share the same images + gallery = split['gallery'] + + train = [tuple(item) for item in train] + query = [tuple(item) for item in query] + gallery = [tuple(item) for item in gallery] + + super(VIPeR, self).__init__(train, query, gallery, **kwargs) + + def prepare_split(self): + if not osp.exists(self.split_path): + print('Creating 10 random splits of train ids and test ids') + + cam_a_imgs = sorted(glob.glob(osp.join(self.cam_a_dir, '*.bmp'))) + cam_b_imgs = sorted(glob.glob(osp.join(self.cam_b_dir, '*.bmp'))) + assert len(cam_a_imgs) == len(cam_b_imgs) + num_pids = len(cam_a_imgs) + print('Number of identities: {}'.format(num_pids)) + num_train_pids = num_pids // 2 + """ + In total, there will be 20 splits because each random split creates two + sub-splits, one using cameraA as query and cameraB as gallery + while the other using cameraB as query and cameraA as gallery. + Therefore, results should be averaged over 20 splits (split_id=0~19). + + In practice, a model trained on split_id=0 can be applied to split_id=0&1 + as split_id=0&1 share the same training data (so on and so forth). + """ + splits = [] + for _ in range(10): + order = np.arange(num_pids) + np.random.shuffle(order) + train_idxs = order[:num_train_pids] + test_idxs = order[num_train_pids:] + assert not bool(set(train_idxs) & set(test_idxs)), \ + 'Error: train and test overlap' + + train = [] + for pid, idx in enumerate(train_idxs): + cam_a_img = cam_a_imgs[idx] + cam_b_img = cam_b_imgs[idx] + train.append((cam_a_img, pid, 0)) + train.append((cam_b_img, pid, 1)) + + test_a = [] + test_b = [] + for pid, idx in enumerate(test_idxs): + cam_a_img = cam_a_imgs[idx] + cam_b_img = cam_b_imgs[idx] + test_a.append((cam_a_img, pid, 0)) + test_b.append((cam_b_img, pid, 1)) + + # use cameraA as query and cameraB as gallery + split = { + 'train': train, + 'query': test_a, + 'gallery': test_b, + 'num_train_pids': num_train_pids, + 'num_query_pids': num_pids - num_train_pids, + 'num_gallery_pids': num_pids - num_train_pids + } + splits.append(split) + + # use cameraB as query and cameraA as gallery + split = { + 'train': train, + 'query': test_b, + 'gallery': test_a, + 'num_train_pids': num_train_pids, + 'num_query_pids': num_pids - num_train_pids, + 'num_gallery_pids': num_pids - num_train_pids + } + splits.append(split) + + print('Totally {} splits are created'.format(len(splits))) + write_json(splits, self.split_path) + print('Split file saved to {}'.format(self.split_path)) diff --git a/strong_sort/deep/reid/torchreid/data/datasets/video/__init__.py b/strong_sort/deep/reid/torchreid/data/datasets/video/__init__.py new file mode 100644 index 0000000..f4e75d3 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/data/datasets/video/__init__.py @@ -0,0 +1,6 @@ +from __future__ import print_function, absolute_import + +from .mars import Mars +from .ilidsvid import iLIDSVID +from .prid2011 import PRID2011 +from .dukemtmcvidreid import DukeMTMCVidReID diff --git a/strong_sort/deep/reid/torchreid/data/datasets/video/dukemtmcvidreid.py b/strong_sort/deep/reid/torchreid/data/datasets/video/dukemtmcvidreid.py new file mode 100644 index 0000000..4b4c82f --- /dev/null +++ b/strong_sort/deep/reid/torchreid/data/datasets/video/dukemtmcvidreid.py @@ -0,0 +1,128 @@ +from __future__ import division, print_function, absolute_import +import glob +import os.path as osp +import warnings + +from torchreid.utils import read_json, write_json + +from ..dataset import VideoDataset + + +class DukeMTMCVidReID(VideoDataset): + """DukeMTMCVidReID. + + Reference: + - Ristani et al. Performance Measures and a Data Set for Multi-Target, + Multi-Camera Tracking. ECCVW 2016. + - Wu et al. Exploit the Unknown Gradually: One-Shot Video-Based Person + Re-Identification by Stepwise Learning. CVPR 2018. + + URL: ``_ + + Dataset statistics: + - identities: 702 (train) + 702 (test). + - tracklets: 2196 (train) + 2636 (test). + """ + dataset_dir = 'dukemtmc-vidreid' + dataset_url = 'http://vision.cs.duke.edu/DukeMTMC/data/misc/DukeMTMC-VideoReID.zip' + + def __init__(self, root='', min_seq_len=0, **kwargs): + self.root = osp.abspath(osp.expanduser(root)) + self.dataset_dir = osp.join(self.root, self.dataset_dir) + self.download_dataset(self.dataset_dir, self.dataset_url) + + self.train_dir = osp.join(self.dataset_dir, 'DukeMTMC-VideoReID/train') + self.query_dir = osp.join(self.dataset_dir, 'DukeMTMC-VideoReID/query') + self.gallery_dir = osp.join( + self.dataset_dir, 'DukeMTMC-VideoReID/gallery' + ) + self.split_train_json_path = osp.join( + self.dataset_dir, 'split_train.json' + ) + self.split_query_json_path = osp.join( + self.dataset_dir, 'split_query.json' + ) + self.split_gallery_json_path = osp.join( + self.dataset_dir, 'split_gallery.json' + ) + self.min_seq_len = min_seq_len + + required_files = [ + self.dataset_dir, self.train_dir, self.query_dir, self.gallery_dir + ] + self.check_before_run(required_files) + + train = self.process_dir( + self.train_dir, self.split_train_json_path, relabel=True + ) + query = self.process_dir( + self.query_dir, self.split_query_json_path, relabel=False + ) + gallery = self.process_dir( + self.gallery_dir, self.split_gallery_json_path, relabel=False + ) + + super(DukeMTMCVidReID, self).__init__(train, query, gallery, **kwargs) + + def process_dir(self, dir_path, json_path, relabel): + if osp.exists(json_path): + split = read_json(json_path) + return split['tracklets'] + + print('=> Generating split json file (** this might take a while **)') + pdirs = glob.glob(osp.join(dir_path, '*')) # avoid .DS_Store + print( + 'Processing "{}" with {} person identities'.format( + dir_path, len(pdirs) + ) + ) + + pid_container = set() + for pdir in pdirs: + pid = int(osp.basename(pdir)) + pid_container.add(pid) + pid2label = {pid: label for label, pid in enumerate(pid_container)} + + tracklets = [] + for pdir in pdirs: + pid = int(osp.basename(pdir)) + if relabel: + pid = pid2label[pid] + tdirs = glob.glob(osp.join(pdir, '*')) + for tdir in tdirs: + raw_img_paths = glob.glob(osp.join(tdir, '*.jpg')) + num_imgs = len(raw_img_paths) + + if num_imgs < self.min_seq_len: + continue + + img_paths = [] + for img_idx in range(num_imgs): + # some tracklet starts from 0002 instead of 0001 + img_idx_name = 'F' + str(img_idx + 1).zfill(4) + res = glob.glob( + osp.join(tdir, '*' + img_idx_name + '*.jpg') + ) + if len(res) == 0: + warnings.warn( + 'Index name {} in {} is missing, skip'.format( + img_idx_name, tdir + ) + ) + continue + img_paths.append(res[0]) + img_name = osp.basename(img_paths[0]) + if img_name.find('_') == -1: + # old naming format: 0001C6F0099X30823.jpg + camid = int(img_name[5]) - 1 + else: + # new naming format: 0001_C6_F0099_X30823.jpg + camid = int(img_name[6]) - 1 + img_paths = tuple(img_paths) + tracklets.append((img_paths, pid, camid)) + + print('Saving split to {}'.format(json_path)) + split_dict = {'tracklets': tracklets} + write_json(split_dict, json_path) + + return tracklets diff --git a/strong_sort/deep/reid/torchreid/data/datasets/video/ilidsvid.py b/strong_sort/deep/reid/torchreid/data/datasets/video/ilidsvid.py new file mode 100644 index 0000000..c3ac1bb --- /dev/null +++ b/strong_sort/deep/reid/torchreid/data/datasets/video/ilidsvid.py @@ -0,0 +1,143 @@ +from __future__ import division, print_function, absolute_import +import glob +import os.path as osp +from scipy.io import loadmat + +from torchreid.utils import read_json, write_json + +from ..dataset import VideoDataset + + +class iLIDSVID(VideoDataset): + """iLIDS-VID. + + Reference: + Wang et al. Person Re-Identification by Video Ranking. ECCV 2014. + + URL: ``_ + + Dataset statistics: + - identities: 300. + - tracklets: 600. + - cameras: 2. + """ + dataset_dir = 'ilids-vid' + dataset_url = 'http://www.eecs.qmul.ac.uk/~xiatian/iLIDS-VID/iLIDS-VID.tar' + + def __init__(self, root='', split_id=0, **kwargs): + self.root = osp.abspath(osp.expanduser(root)) + self.dataset_dir = osp.join(self.root, self.dataset_dir) + self.download_dataset(self.dataset_dir, self.dataset_url) + + self.data_dir = osp.join(self.dataset_dir, 'i-LIDS-VID') + self.split_dir = osp.join(self.dataset_dir, 'train-test people splits') + self.split_mat_path = osp.join( + self.split_dir, 'train_test_splits_ilidsvid.mat' + ) + self.split_path = osp.join(self.dataset_dir, 'splits.json') + self.cam_1_path = osp.join( + self.dataset_dir, 'i-LIDS-VID/sequences/cam1' + ) + self.cam_2_path = osp.join( + self.dataset_dir, 'i-LIDS-VID/sequences/cam2' + ) + + required_files = [self.dataset_dir, self.data_dir, self.split_dir] + self.check_before_run(required_files) + + self.prepare_split() + splits = read_json(self.split_path) + if split_id >= len(splits): + raise ValueError( + 'split_id exceeds range, received {}, but expected between 0 and {}' + .format(split_id, + len(splits) - 1) + ) + split = splits[split_id] + train_dirs, test_dirs = split['train'], split['test'] + + train = self.process_data(train_dirs, cam1=True, cam2=True) + query = self.process_data(test_dirs, cam1=True, cam2=False) + gallery = self.process_data(test_dirs, cam1=False, cam2=True) + + super(iLIDSVID, self).__init__(train, query, gallery, **kwargs) + + def prepare_split(self): + if not osp.exists(self.split_path): + print('Creating splits ...') + mat_split_data = loadmat(self.split_mat_path)['ls_set'] + + num_splits = mat_split_data.shape[0] + num_total_ids = mat_split_data.shape[1] + assert num_splits == 10 + assert num_total_ids == 300 + num_ids_each = num_total_ids // 2 + + # pids in mat_split_data are indices, so we need to transform them + # to real pids + person_cam1_dirs = sorted( + glob.glob(osp.join(self.cam_1_path, '*')) + ) + person_cam2_dirs = sorted( + glob.glob(osp.join(self.cam_2_path, '*')) + ) + + person_cam1_dirs = [ + osp.basename(item) for item in person_cam1_dirs + ] + person_cam2_dirs = [ + osp.basename(item) for item in person_cam2_dirs + ] + + # make sure persons in one camera view can be found in the other camera view + assert set(person_cam1_dirs) == set(person_cam2_dirs) + + splits = [] + for i_split in range(num_splits): + # first 50% for testing and the remaining for training, following Wang et al. ECCV'14. + train_idxs = sorted( + list(mat_split_data[i_split, num_ids_each:]) + ) + test_idxs = sorted( + list(mat_split_data[i_split, :num_ids_each]) + ) + + train_idxs = [int(i) - 1 for i in train_idxs] + test_idxs = [int(i) - 1 for i in test_idxs] + + # transform pids to person dir names + train_dirs = [person_cam1_dirs[i] for i in train_idxs] + test_dirs = [person_cam1_dirs[i] for i in test_idxs] + + split = {'train': train_dirs, 'test': test_dirs} + splits.append(split) + + print( + 'Totally {} splits are created, following Wang et al. ECCV\'14' + .format(len(splits)) + ) + print('Split file is saved to {}'.format(self.split_path)) + write_json(splits, self.split_path) + + def process_data(self, dirnames, cam1=True, cam2=True): + tracklets = [] + dirname2pid = {dirname: i for i, dirname in enumerate(dirnames)} + + for dirname in dirnames: + if cam1: + person_dir = osp.join(self.cam_1_path, dirname) + img_names = glob.glob(osp.join(person_dir, '*.png')) + assert len(img_names) > 0 + img_names = tuple(img_names) + pid = dirname2pid[dirname] + tracklets.append((img_names, pid, 0)) + + if cam2: + person_dir = osp.join(self.cam_2_path, dirname) + img_names = glob.glob(osp.join(person_dir, '*.png')) + assert len(img_names) > 0 + img_names = tuple(img_names) + pid = dirname2pid[dirname] + tracklets.append((img_names, pid, 1)) + + return tracklets diff --git a/strong_sort/deep/reid/torchreid/data/datasets/video/mars.py b/strong_sort/deep/reid/torchreid/data/datasets/video/mars.py new file mode 100644 index 0000000..4128e1c --- /dev/null +++ b/strong_sort/deep/reid/torchreid/data/datasets/video/mars.py @@ -0,0 +1,133 @@ +from __future__ import division, print_function, absolute_import +import os.path as osp +import warnings +from scipy.io import loadmat + +from ..dataset import VideoDataset + + +class Mars(VideoDataset): + """MARS. + + Reference: + Zheng et al. MARS: A Video Benchmark for Large-Scale Person Re-identification. ECCV 2016. + + URL: ``_ + + Dataset statistics: + - identities: 1261. + - tracklets: 8298 (train) + 1980 (query) + 9330 (gallery). + - cameras: 6. + """ + dataset_dir = 'mars' + dataset_url = None + + def __init__(self, root='', **kwargs): + self.root = osp.abspath(osp.expanduser(root)) + self.dataset_dir = osp.join(self.root, self.dataset_dir) + self.download_dataset(self.dataset_dir, self.dataset_url) + + self.train_name_path = osp.join( + self.dataset_dir, 'info/train_name.txt' + ) + self.test_name_path = osp.join(self.dataset_dir, 'info/test_name.txt') + self.track_train_info_path = osp.join( + self.dataset_dir, 'info/tracks_train_info.mat' + ) + self.track_test_info_path = osp.join( + self.dataset_dir, 'info/tracks_test_info.mat' + ) + self.query_IDX_path = osp.join(self.dataset_dir, 'info/query_IDX.mat') + + required_files = [ + self.dataset_dir, self.train_name_path, self.test_name_path, + self.track_train_info_path, self.track_test_info_path, + self.query_IDX_path + ] + self.check_before_run(required_files) + + train_names = self.get_names(self.train_name_path) + test_names = self.get_names(self.test_name_path) + track_train = loadmat(self.track_train_info_path + )['track_train_info'] # numpy.ndarray (8298, 4) + track_test = loadmat(self.track_test_info_path + )['track_test_info'] # numpy.ndarray (12180, 4) + query_IDX = loadmat(self.query_IDX_path + )['query_IDX'].squeeze() # numpy.ndarray (1980,) + query_IDX -= 1 # index from 0 + track_query = track_test[query_IDX, :] + gallery_IDX = [ + i for i in range(track_test.shape[0]) if i not in query_IDX + ] + track_gallery = track_test[gallery_IDX, :] + + train = self.process_data( + train_names, track_train, home_dir='bbox_train', relabel=True + ) + query = self.process_data( + test_names, track_query, home_dir='bbox_test', relabel=False + ) + gallery = self.process_data( + test_names, track_gallery, home_dir='bbox_test', relabel=False + ) + + super(Mars, self).__init__(train, query, gallery, **kwargs) + + def get_names(self, fpath): + names = [] + with open(fpath, 'r') as f: + for line in f: + new_line = line.rstrip() + names.append(new_line) + return names + + def process_data( + self, names, meta_data, home_dir=None, relabel=False, min_seq_len=0 + ): + assert home_dir in ['bbox_train', 'bbox_test'] + num_tracklets = meta_data.shape[0] + pid_list = list(set(meta_data[:, 2].tolist())) + + if relabel: + pid2label = {pid: label for label, pid in enumerate(pid_list)} + tracklets = [] + + for tracklet_idx in range(num_tracklets): + data = meta_data[tracklet_idx, ...] + start_index, end_index, pid, camid = data + if pid == -1: + continue # junk images are just ignored + assert 1 <= camid <= 6 + if relabel: + pid = pid2label[pid] + camid -= 1 # index starts from 0 + img_names = names[start_index - 1:end_index] + + # make sure image names correspond to the same person + pnames = [img_name[:4] for img_name in img_names] + assert len( + set(pnames) + ) == 1, 'Error: a single tracklet contains different person images' + + # make sure all images are captured under the same camera + camnames = [img_name[5] for img_name in img_names] + assert len( + set(camnames) + ) == 1, 'Error: images are captured under different cameras!' + + # append image names with directory information + img_paths = [ + osp.join(self.dataset_dir, home_dir, img_name[:4], img_name) + for img_name in img_names + ] + if len(img_paths) >= min_seq_len: + img_paths = tuple(img_paths) + tracklets.append((img_paths, pid, camid)) + + return tracklets + + def combine_all(self): + warnings.warn( + 'Some query IDs do not appear in gallery. Therefore, combineall ' + 'does not make any difference to Mars' + ) diff --git a/strong_sort/deep/reid/torchreid/data/datasets/video/prid2011.py b/strong_sort/deep/reid/torchreid/data/datasets/video/prid2011.py new file mode 100644 index 0000000..3af2e4d --- /dev/null +++ b/strong_sort/deep/reid/torchreid/data/datasets/video/prid2011.py @@ -0,0 +1,80 @@ +from __future__ import division, print_function, absolute_import +import glob +import os.path as osp + +from torchreid.utils import read_json + +from ..dataset import VideoDataset + + +class PRID2011(VideoDataset): + """PRID2011. + + Reference: + Hirzer et al. Person Re-Identification by Descriptive and + Discriminative Classification. SCIA 2011. + + URL: ``_ + + Dataset statistics: + - identities: 200. + - tracklets: 400. + - cameras: 2. + """ + dataset_dir = 'prid2011' + dataset_url = None + + def __init__(self, root='', split_id=0, **kwargs): + self.root = osp.abspath(osp.expanduser(root)) + self.dataset_dir = osp.join(self.root, self.dataset_dir) + self.download_dataset(self.dataset_dir, self.dataset_url) + + self.split_path = osp.join(self.dataset_dir, 'splits_prid2011.json') + self.cam_a_dir = osp.join( + self.dataset_dir, 'prid_2011', 'multi_shot', 'cam_a' + ) + self.cam_b_dir = osp.join( + self.dataset_dir, 'prid_2011', 'multi_shot', 'cam_b' + ) + + required_files = [self.dataset_dir, self.cam_a_dir, self.cam_b_dir] + self.check_before_run(required_files) + + splits = read_json(self.split_path) + if split_id >= len(splits): + raise ValueError( + 'split_id exceeds range, received {}, but expected between 0 and {}' + .format(split_id, + len(splits) - 1) + ) + split = splits[split_id] + train_dirs, test_dirs = split['train'], split['test'] + + train = self.process_dir(train_dirs, cam1=True, cam2=True) + query = self.process_dir(test_dirs, cam1=True, cam2=False) + gallery = self.process_dir(test_dirs, cam1=False, cam2=True) + + super(PRID2011, self).__init__(train, query, gallery, **kwargs) + + def process_dir(self, dirnames, cam1=True, cam2=True): + tracklets = [] + dirname2pid = {dirname: i for i, dirname in enumerate(dirnames)} + + for dirname in dirnames: + if cam1: + person_dir = osp.join(self.cam_a_dir, dirname) + img_names = glob.glob(osp.join(person_dir, '*.png')) + assert len(img_names) > 0 + img_names = tuple(img_names) + pid = dirname2pid[dirname] + tracklets.append((img_names, pid, 0)) + + if cam2: + person_dir = osp.join(self.cam_b_dir, dirname) + img_names = glob.glob(osp.join(person_dir, '*.png')) + assert len(img_names) > 0 + img_names = tuple(img_names) + pid = dirname2pid[dirname] + tracklets.append((img_names, pid, 1)) + + return tracklets diff --git a/strong_sort/deep/reid/torchreid/data/sampler.py b/strong_sort/deep/reid/torchreid/data/sampler.py new file mode 100644 index 0000000..f69b3e0 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/data/sampler.py @@ -0,0 +1,245 @@ +from __future__ import division, absolute_import +import copy +import numpy as np +import random +from collections import defaultdict +from torch.utils.data.sampler import Sampler, RandomSampler, SequentialSampler + +AVAI_SAMPLERS = [ + 'RandomIdentitySampler', 'SequentialSampler', 'RandomSampler', + 'RandomDomainSampler', 'RandomDatasetSampler' +] + + +class RandomIdentitySampler(Sampler): + """Randomly samples N identities each with K instances. + + Args: + data_source (list): contains tuples of (img_path(s), pid, camid, dsetid). + batch_size (int): batch size. + num_instances (int): number of instances per identity in a batch. + """ + + def __init__(self, data_source, batch_size, num_instances): + if batch_size < num_instances: + raise ValueError( + 'batch_size={} must be no less ' + 'than num_instances={}'.format(batch_size, num_instances) + ) + + self.data_source = data_source + self.batch_size = batch_size + self.num_instances = num_instances + self.num_pids_per_batch = self.batch_size // self.num_instances + self.index_dic = defaultdict(list) + for index, items in enumerate(data_source): + pid = items[1] + self.index_dic[pid].append(index) + self.pids = list(self.index_dic.keys()) + assert len(self.pids) >= self.num_pids_per_batch + + # estimate number of examples in an epoch + # TODO: improve precision + self.length = 0 + for pid in self.pids: + idxs = self.index_dic[pid] + num = len(idxs) + if num < self.num_instances: + num = self.num_instances + self.length += num - num % self.num_instances + + def __iter__(self): + batch_idxs_dict = defaultdict(list) + + for pid in self.pids: + idxs = copy.deepcopy(self.index_dic[pid]) + if len(idxs) < self.num_instances: + idxs = np.random.choice( + idxs, size=self.num_instances, replace=True + ) + random.shuffle(idxs) + batch_idxs = [] + for idx in idxs: + batch_idxs.append(idx) + if len(batch_idxs) == self.num_instances: + batch_idxs_dict[pid].append(batch_idxs) + batch_idxs = [] + + avai_pids = copy.deepcopy(self.pids) + final_idxs = [] + + while len(avai_pids) >= self.num_pids_per_batch: + selected_pids = random.sample(avai_pids, self.num_pids_per_batch) + for pid in selected_pids: + batch_idxs = batch_idxs_dict[pid].pop(0) + final_idxs.extend(batch_idxs) + if len(batch_idxs_dict[pid]) == 0: + avai_pids.remove(pid) + + return iter(final_idxs) + + def __len__(self): + return self.length + + +class RandomDomainSampler(Sampler): + """Random domain sampler. + + We consider each camera as a visual domain. + + How does the sampling work: + 1. Randomly sample N cameras (based on the "camid" label). + 2. From each camera, randomly sample K images. + + Args: + data_source (list): contains tuples of (img_path(s), pid, camid, dsetid). + batch_size (int): batch size. + n_domain (int): number of cameras to sample in a batch. + """ + + def __init__(self, data_source, batch_size, n_domain): + self.data_source = data_source + + # Keep track of image indices for each domain + self.domain_dict = defaultdict(list) + for i, items in enumerate(data_source): + camid = items[2] + self.domain_dict[camid].append(i) + self.domains = list(self.domain_dict.keys()) + + # Make sure each domain can be assigned an equal number of images + if n_domain is None or n_domain <= 0: + n_domain = len(self.domains) + assert batch_size % n_domain == 0 + self.n_img_per_domain = batch_size // n_domain + + self.batch_size = batch_size + self.n_domain = n_domain + self.length = len(list(self.__iter__())) + + def __iter__(self): + domain_dict = copy.deepcopy(self.domain_dict) + final_idxs = [] + stop_sampling = False + + while not stop_sampling: + selected_domains = random.sample(self.domains, self.n_domain) + + for domain in selected_domains: + idxs = domain_dict[domain] + selected_idxs = random.sample(idxs, self.n_img_per_domain) + final_idxs.extend(selected_idxs) + + for idx in selected_idxs: + domain_dict[domain].remove(idx) + + remaining = len(domain_dict[domain]) + if remaining < self.n_img_per_domain: + stop_sampling = True + + return iter(final_idxs) + + def __len__(self): + return self.length + + +class RandomDatasetSampler(Sampler): + """Random dataset sampler. + + How does the sampling work: + 1. Randomly sample N datasets (based on the "dsetid" label). + 2. From each dataset, randomly sample K images. + + Args: + data_source (list): contains tuples of (img_path(s), pid, camid, dsetid). + batch_size (int): batch size. + n_dataset (int): number of datasets to sample in a batch. + """ + + def __init__(self, data_source, batch_size, n_dataset): + self.data_source = data_source + + # Keep track of image indices for each dataset + self.dataset_dict = defaultdict(list) + for i, items in enumerate(data_source): + dsetid = items[3] + self.dataset_dict[dsetid].append(i) + self.datasets = list(self.dataset_dict.keys()) + + # Make sure each dataset can be assigned an equal number of images + if n_dataset is None or n_dataset <= 0: + n_dataset = len(self.datasets) + assert batch_size % n_dataset == 0 + self.n_img_per_dset = batch_size // n_dataset + + self.batch_size = batch_size + self.n_dataset = n_dataset + self.length = len(list(self.__iter__())) + + def __iter__(self): + dataset_dict = copy.deepcopy(self.dataset_dict) + final_idxs = [] + stop_sampling = False + + while not stop_sampling: + selected_datasets = random.sample(self.datasets, self.n_dataset) + + for dset in selected_datasets: + idxs = dataset_dict[dset] + selected_idxs = random.sample(idxs, self.n_img_per_dset) + final_idxs.extend(selected_idxs) + + for idx in selected_idxs: + dataset_dict[dset].remove(idx) + + remaining = len(dataset_dict[dset]) + if remaining < self.n_img_per_dset: + stop_sampling = True + + return iter(final_idxs) + + def __len__(self): + return self.length + + +def build_train_sampler( + data_source, + train_sampler, + batch_size=32, + num_instances=4, + num_cams=1, + num_datasets=1, + **kwargs +): + """Builds a training sampler. + + Args: + data_source (list): contains tuples of (img_path(s), pid, camid). + train_sampler (str): sampler name (default: ``RandomSampler``). + batch_size (int, optional): batch size. Default is 32. + num_instances (int, optional): number of instances per identity in a + batch (when using ``RandomIdentitySampler``). Default is 4. + num_cams (int, optional): number of cameras to sample in a batch (when using + ``RandomDomainSampler``). Default is 1. + num_datasets (int, optional): number of datasets to sample in a batch (when + using ``RandomDatasetSampler``). Default is 1. + """ + assert train_sampler in AVAI_SAMPLERS, \ + 'train_sampler must be one of {}, but got {}'.format(AVAI_SAMPLERS, train_sampler) + + if train_sampler == 'RandomIdentitySampler': + sampler = RandomIdentitySampler(data_source, batch_size, num_instances) + + elif train_sampler == 'RandomDomainSampler': + sampler = RandomDomainSampler(data_source, batch_size, num_cams) + + elif train_sampler == 'RandomDatasetSampler': + sampler = RandomDatasetSampler(data_source, batch_size, num_datasets) + + elif train_sampler == 'SequentialSampler': + sampler = SequentialSampler(data_source) + + elif train_sampler == 'RandomSampler': + sampler = RandomSampler(data_source) + + return sampler diff --git a/strong_sort/deep/reid/torchreid/data/transforms.py b/strong_sort/deep/reid/torchreid/data/transforms.py new file mode 100644 index 0000000..0c09ca0 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/data/transforms.py @@ -0,0 +1,326 @@ +from __future__ import division, print_function, absolute_import +import math +import random +from collections import deque +import torch +from PIL import Image +from torchvision.transforms import ( + Resize, Compose, ToTensor, Normalize, ColorJitter, RandomHorizontalFlip +) + + +class Random2DTranslation(object): + """Randomly translates the input image with a probability. + + Specifically, given a predefined shape (height, width), the input is first + resized with a factor of 1.125, leading to (height*1.125, width*1.125), then + a random crop is performed. Such operation is done with a probability. + + Args: + height (int): target image height. + width (int): target image width. + p (float, optional): probability that this operation takes place. + Default is 0.5. + interpolation (int, optional): desired interpolation. Default is + ``PIL.Image.BILINEAR`` + """ + + def __init__(self, height, width, p=0.5, interpolation=Image.BILINEAR): + self.height = height + self.width = width + self.p = p + self.interpolation = interpolation + + def __call__(self, img): + if random.uniform(0, 1) > self.p: + return img.resize((self.width, self.height), self.interpolation) + + new_width, new_height = int(round(self.width * 1.125) + ), int(round(self.height * 1.125)) + resized_img = img.resize((new_width, new_height), self.interpolation) + x_maxrange = new_width - self.width + y_maxrange = new_height - self.height + x1 = int(round(random.uniform(0, x_maxrange))) + y1 = int(round(random.uniform(0, y_maxrange))) + croped_img = resized_img.crop( + (x1, y1, x1 + self.width, y1 + self.height) + ) + return croped_img + + +class RandomErasing(object): + """Randomly erases an image patch. + + Origin: ``_ + + Reference: + Zhong et al. Random Erasing Data Augmentation. + + Args: + probability (float, optional): probability that this operation takes place. + Default is 0.5. + sl (float, optional): min erasing area. + sh (float, optional): max erasing area. + r1 (float, optional): min aspect ratio. + mean (list, optional): erasing value. + """ + + def __init__( + self, + probability=0.5, + sl=0.02, + sh=0.4, + r1=0.3, + mean=[0.4914, 0.4822, 0.4465] + ): + self.probability = probability + self.mean = mean + self.sl = sl + self.sh = sh + self.r1 = r1 + + def __call__(self, img): + if random.uniform(0, 1) > self.probability: + return img + + for attempt in range(100): + area = img.size()[1] * img.size()[2] + + target_area = random.uniform(self.sl, self.sh) * area + aspect_ratio = random.uniform(self.r1, 1 / self.r1) + + h = int(round(math.sqrt(target_area * aspect_ratio))) + w = int(round(math.sqrt(target_area / aspect_ratio))) + + if w < img.size()[2] and h < img.size()[1]: + x1 = random.randint(0, img.size()[1] - h) + y1 = random.randint(0, img.size()[2] - w) + if img.size()[0] == 3: + img[0, x1:x1 + h, y1:y1 + w] = self.mean[0] + img[1, x1:x1 + h, y1:y1 + w] = self.mean[1] + img[2, x1:x1 + h, y1:y1 + w] = self.mean[2] + else: + img[0, x1:x1 + h, y1:y1 + w] = self.mean[0] + return img + + return img + + +class ColorAugmentation(object): + """Randomly alters the intensities of RGB channels. + + Reference: + Krizhevsky et al. ImageNet Classification with Deep ConvolutionalNeural + Networks. NIPS 2012. + + Args: + p (float, optional): probability that this operation takes place. + Default is 0.5. + """ + + def __init__(self, p=0.5): + self.p = p + self.eig_vec = torch.Tensor( + [ + [0.4009, 0.7192, -0.5675], + [-0.8140, -0.0045, -0.5808], + [0.4203, -0.6948, -0.5836], + ] + ) + self.eig_val = torch.Tensor([[0.2175, 0.0188, 0.0045]]) + + def _check_input(self, tensor): + assert tensor.dim() == 3 and tensor.size(0) == 3 + + def __call__(self, tensor): + if random.uniform(0, 1) > self.p: + return tensor + alpha = torch.normal(mean=torch.zeros_like(self.eig_val)) * 0.1 + quatity = torch.mm(self.eig_val * alpha, self.eig_vec) + tensor = tensor + quatity.view(3, 1, 1) + return tensor + + +class RandomPatch(object): + """Random patch data augmentation. + + There is a patch pool that stores randomly extracted pathces from person images. + + For each input image, RandomPatch + 1) extracts a random patch and stores the patch in the patch pool; + 2) randomly selects a patch from the patch pool and pastes it on the + input (at random position) to simulate occlusion. + + Reference: + - Zhou et al. Omni-Scale Feature Learning for Person Re-Identification. ICCV, 2019. + - Zhou et al. Learning Generalisable Omni-Scale Representations + for Person Re-Identification. TPAMI, 2021. + """ + + def __init__( + self, + prob_happen=0.5, + pool_capacity=50000, + min_sample_size=100, + patch_min_area=0.01, + patch_max_area=0.5, + patch_min_ratio=0.1, + prob_rotate=0.5, + prob_flip_leftright=0.5, + ): + self.prob_happen = prob_happen + + self.patch_min_area = patch_min_area + self.patch_max_area = patch_max_area + self.patch_min_ratio = patch_min_ratio + + self.prob_rotate = prob_rotate + self.prob_flip_leftright = prob_flip_leftright + + self.patchpool = deque(maxlen=pool_capacity) + self.min_sample_size = min_sample_size + + def generate_wh(self, W, H): + area = W * H + for attempt in range(100): + target_area = random.uniform( + self.patch_min_area, self.patch_max_area + ) * area + aspect_ratio = random.uniform( + self.patch_min_ratio, 1. / self.patch_min_ratio + ) + h = int(round(math.sqrt(target_area * aspect_ratio))) + w = int(round(math.sqrt(target_area / aspect_ratio))) + if w < W and h < H: + return w, h + return None, None + + def transform_patch(self, patch): + if random.uniform(0, 1) > self.prob_flip_leftright: + patch = patch.transpose(Image.FLIP_LEFT_RIGHT) + if random.uniform(0, 1) > self.prob_rotate: + patch = patch.rotate(random.randint(-10, 10)) + return patch + + def __call__(self, img): + W, H = img.size # original image size + + # collect new patch + w, h = self.generate_wh(W, H) + if w is not None and h is not None: + x1 = random.randint(0, W - w) + y1 = random.randint(0, H - h) + new_patch = img.crop((x1, y1, x1 + w, y1 + h)) + self.patchpool.append(new_patch) + + if len(self.patchpool) < self.min_sample_size: + return img + + if random.uniform(0, 1) > self.prob_happen: + return img + + # paste a randomly selected patch on a random position + patch = random.sample(self.patchpool, 1)[0] + patchW, patchH = patch.size + x1 = random.randint(0, W - patchW) + y1 = random.randint(0, H - patchH) + patch = self.transform_patch(patch) + img.paste(patch, (x1, y1)) + + return img + + +def build_transforms( + height, + width, + transforms='random_flip', + norm_mean=[0.485, 0.456, 0.406], + norm_std=[0.229, 0.224, 0.225], + **kwargs +): + """Builds train and test transform functions. + + Args: + height (int): target image height. + width (int): target image width. + transforms (str or list of str, optional): transformations applied to model training. + Default is 'random_flip'. + norm_mean (list or None, optional): normalization mean values. Default is ImageNet means. + norm_std (list or None, optional): normalization standard deviation values. Default is + ImageNet standard deviation values. + """ + if transforms is None: + transforms = [] + + if isinstance(transforms, str): + transforms = [transforms] + + if not isinstance(transforms, list): + raise ValueError( + 'transforms must be a list of strings, but found to be {}'.format( + type(transforms) + ) + ) + + if len(transforms) > 0: + transforms = [t.lower() for t in transforms] + + if norm_mean is None or norm_std is None: + norm_mean = [0.485, 0.456, 0.406] # imagenet mean + norm_std = [0.229, 0.224, 0.225] # imagenet std + normalize = Normalize(mean=norm_mean, std=norm_std) + + print('Building train transforms ...') + transform_tr = [] + + print('+ resize to {}x{}'.format(height, width)) + transform_tr += [Resize((height, width))] + + if 'random_flip' in transforms: + print('+ random flip') + transform_tr += [RandomHorizontalFlip()] + + if 'random_crop' in transforms: + print( + '+ random crop (enlarge to {}x{} and ' + 'crop {}x{})'.format( + int(round(height * 1.125)), int(round(width * 1.125)), height, + width + ) + ) + transform_tr += [Random2DTranslation(height, width)] + + if 'random_patch' in transforms: + print('+ random patch') + transform_tr += [RandomPatch()] + + if 'color_jitter' in transforms: + print('+ color jitter') + transform_tr += [ + ColorJitter(brightness=0.2, contrast=0.15, saturation=0, hue=0) + ] + + print('+ to torch tensor of range [0, 1]') + transform_tr += [ToTensor()] + + print('+ normalization (mean={}, std={})'.format(norm_mean, norm_std)) + transform_tr += [normalize] + + if 'random_erase' in transforms: + print('+ random erase') + transform_tr += [RandomErasing(mean=norm_mean)] + + transform_tr = Compose(transform_tr) + + print('Building test transforms ...') + print('+ resize to {}x{}'.format(height, width)) + print('+ to torch tensor of range [0, 1]') + print('+ normalization (mean={}, std={})'.format(norm_mean, norm_std)) + + transform_te = Compose([ + Resize((height, width)), + ToTensor(), + normalize, + ]) + + return transform_tr, transform_te diff --git a/strong_sort/deep/reid/torchreid/engine/__init__.py b/strong_sort/deep/reid/torchreid/engine/__init__.py new file mode 100644 index 0000000..a39cc7f --- /dev/null +++ b/strong_sort/deep/reid/torchreid/engine/__init__.py @@ -0,0 +1,5 @@ +from __future__ import print_function, absolute_import + +from .image import ImageSoftmaxEngine, ImageTripletEngine +from .video import VideoSoftmaxEngine, VideoTripletEngine +from .engine import Engine diff --git a/strong_sort/deep/reid/torchreid/engine/engine.py b/strong_sort/deep/reid/torchreid/engine/engine.py new file mode 100644 index 0000000..5fe3e25 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/engine/engine.py @@ -0,0 +1,478 @@ +from __future__ import division, print_function, absolute_import +import time +import numpy as np +import os.path as osp +import datetime +from collections import OrderedDict +import torch +from torch.nn import functional as F +from torch.utils.tensorboard import SummaryWriter + +from torchreid import metrics +from torchreid.utils import ( + MetricMeter, AverageMeter, re_ranking, open_all_layers, save_checkpoint, + open_specified_layers, visualize_ranked_results +) +from torchreid.losses import DeepSupervision + + +class Engine(object): + r"""A generic base Engine class for both image- and video-reid. + + Args: + datamanager (DataManager): an instance of ``torchreid.data.ImageDataManager`` + or ``torchreid.data.VideoDataManager``. + use_gpu (bool, optional): use gpu. Default is True. + """ + + def __init__(self, datamanager, use_gpu=True): + self.datamanager = datamanager + self.train_loader = self.datamanager.train_loader + self.test_loader = self.datamanager.test_loader + self.use_gpu = (torch.cuda.is_available() and use_gpu) + self.writer = None + self.epoch = 0 + + self.model = None + self.optimizer = None + self.scheduler = None + + self._models = OrderedDict() + self._optims = OrderedDict() + self._scheds = OrderedDict() + + def register_model(self, name='model', model=None, optim=None, sched=None): + if self.__dict__.get('_models') is None: + raise AttributeError( + 'Cannot assign model before super().__init__() call' + ) + + if self.__dict__.get('_optims') is None: + raise AttributeError( + 'Cannot assign optim before super().__init__() call' + ) + + if self.__dict__.get('_scheds') is None: + raise AttributeError( + 'Cannot assign sched before super().__init__() call' + ) + + self._models[name] = model + self._optims[name] = optim + self._scheds[name] = sched + + def get_model_names(self, names=None): + names_real = list(self._models.keys()) + if names is not None: + if not isinstance(names, list): + names = [names] + for name in names: + assert name in names_real + return names + else: + return names_real + + def save_model(self, epoch, rank1, save_dir, is_best=False): + names = self.get_model_names() + + for name in names: + save_checkpoint( + { + 'state_dict': self._models[name].state_dict(), + 'epoch': epoch + 1, + 'rank1': rank1, + 'optimizer': self._optims[name].state_dict(), + 'scheduler': self._scheds[name].state_dict() + }, + osp.join(save_dir, name), + is_best=is_best + ) + + def set_model_mode(self, mode='train', names=None): + assert mode in ['train', 'eval', 'test'] + names = self.get_model_names(names) + + for name in names: + if mode == 'train': + self._models[name].train() + else: + self._models[name].eval() + + def get_current_lr(self, names=None): + names = self.get_model_names(names) + name = names[0] + return self._optims[name].param_groups[-1]['lr'] + + def update_lr(self, names=None): + names = self.get_model_names(names) + + for name in names: + if self._scheds[name] is not None: + self._scheds[name].step() + + def run( + self, + save_dir='log', + max_epoch=0, + start_epoch=0, + print_freq=10, + fixbase_epoch=0, + open_layers=None, + start_eval=0, + eval_freq=-1, + test_only=False, + dist_metric='euclidean', + normalize_feature=False, + visrank=False, + visrank_topk=10, + use_metric_cuhk03=False, + ranks=[1, 5, 10, 20], + rerank=False + ): + r"""A unified pipeline for training and evaluating a model. + + Args: + save_dir (str): directory to save model. + max_epoch (int): maximum epoch. + start_epoch (int, optional): starting epoch. Default is 0. + print_freq (int, optional): print_frequency. Default is 10. + fixbase_epoch (int, optional): number of epochs to train ``open_layers`` (new layers) + while keeping base layers frozen. Default is 0. ``fixbase_epoch`` is counted + in ``max_epoch``. + open_layers (str or list, optional): layers (attribute names) open for training. + start_eval (int, optional): from which epoch to start evaluation. Default is 0. + eval_freq (int, optional): evaluation frequency. Default is -1 (meaning evaluation + is only performed at the end of training). + test_only (bool, optional): if True, only runs evaluation on test datasets. + Default is False. + dist_metric (str, optional): distance metric used to compute distance matrix + between query and gallery. Default is "euclidean". + normalize_feature (bool, optional): performs L2 normalization on feature vectors before + computing feature distance. Default is False. + visrank (bool, optional): visualizes ranked results. Default is False. It is recommended to + enable ``visrank`` when ``test_only`` is True. The ranked images will be saved to + "save_dir/visrank_dataset", e.g. "save_dir/visrank_market1501". + visrank_topk (int, optional): top-k ranked images to be visualized. Default is 10. + use_metric_cuhk03 (bool, optional): use single-gallery-shot setting for cuhk03. + Default is False. This should be enabled when using cuhk03 classic split. + ranks (list, optional): cmc ranks to be computed. Default is [1, 5, 10, 20]. + rerank (bool, optional): uses person re-ranking (by Zhong et al. CVPR'17). + Default is False. This is only enabled when test_only=True. + """ + + if visrank and not test_only: + raise ValueError( + 'visrank can be set to True only if test_only=True' + ) + + if test_only: + self.test( + dist_metric=dist_metric, + normalize_feature=normalize_feature, + visrank=visrank, + visrank_topk=visrank_topk, + save_dir=save_dir, + use_metric_cuhk03=use_metric_cuhk03, + ranks=ranks, + rerank=rerank + ) + return + + if self.writer is None: + self.writer = SummaryWriter(log_dir=save_dir) + + time_start = time.time() + self.start_epoch = start_epoch + self.max_epoch = max_epoch + print('=> Start training') + + for self.epoch in range(self.start_epoch, self.max_epoch): + self.train( + print_freq=print_freq, + fixbase_epoch=fixbase_epoch, + open_layers=open_layers + ) + + if (self.epoch + 1) >= start_eval \ + and eval_freq > 0 \ + and (self.epoch+1) % eval_freq == 0 \ + and (self.epoch + 1) != self.max_epoch: + rank1 = self.test( + dist_metric=dist_metric, + normalize_feature=normalize_feature, + visrank=visrank, + visrank_topk=visrank_topk, + save_dir=save_dir, + use_metric_cuhk03=use_metric_cuhk03, + ranks=ranks + ) + self.save_model(self.epoch, rank1, save_dir) + + if self.max_epoch > 0: + print('=> Final test') + rank1 = self.test( + dist_metric=dist_metric, + normalize_feature=normalize_feature, + visrank=visrank, + visrank_topk=visrank_topk, + save_dir=save_dir, + use_metric_cuhk03=use_metric_cuhk03, + ranks=ranks + ) + self.save_model(self.epoch, rank1, save_dir) + + elapsed = round(time.time() - time_start) + elapsed = str(datetime.timedelta(seconds=elapsed)) + print('Elapsed {}'.format(elapsed)) + if self.writer is not None: + self.writer.close() + + def train(self, print_freq=10, fixbase_epoch=0, open_layers=None): + losses = MetricMeter() + batch_time = AverageMeter() + data_time = AverageMeter() + + self.set_model_mode('train') + + self.two_stepped_transfer_learning( + self.epoch, fixbase_epoch, open_layers + ) + + self.num_batches = len(self.train_loader) + end = time.time() + for self.batch_idx, data in enumerate(self.train_loader): + data_time.update(time.time() - end) + loss_summary = self.forward_backward(data) + batch_time.update(time.time() - end) + losses.update(loss_summary) + + if (self.batch_idx + 1) % print_freq == 0: + nb_this_epoch = self.num_batches - (self.batch_idx + 1) + nb_future_epochs = ( + self.max_epoch - (self.epoch + 1) + ) * self.num_batches + eta_seconds = batch_time.avg * (nb_this_epoch+nb_future_epochs) + eta_str = str(datetime.timedelta(seconds=int(eta_seconds))) + print( + 'epoch: [{0}/{1}][{2}/{3}]\t' + 'time {batch_time.val:.3f} ({batch_time.avg:.3f})\t' + 'data {data_time.val:.3f} ({data_time.avg:.3f})\t' + 'eta {eta}\t' + '{losses}\t' + 'lr {lr:.6f}'.format( + self.epoch + 1, + self.max_epoch, + self.batch_idx + 1, + self.num_batches, + batch_time=batch_time, + data_time=data_time, + eta=eta_str, + losses=losses, + lr=self.get_current_lr() + ) + ) + + if self.writer is not None: + n_iter = self.epoch * self.num_batches + self.batch_idx + self.writer.add_scalar('Train/time', batch_time.avg, n_iter) + self.writer.add_scalar('Train/data', data_time.avg, n_iter) + for name, meter in losses.meters.items(): + self.writer.add_scalar('Train/' + name, meter.avg, n_iter) + self.writer.add_scalar( + 'Train/lr', self.get_current_lr(), n_iter + ) + + end = time.time() + + self.update_lr() + + def forward_backward(self, data): + raise NotImplementedError + + def test( + self, + dist_metric='euclidean', + normalize_feature=False, + visrank=False, + visrank_topk=10, + save_dir='', + use_metric_cuhk03=False, + ranks=[1, 5, 10, 20], + rerank=False + ): + r"""Tests model on target datasets. + + .. note:: + + This function has been called in ``run()``. + + .. note:: + + The test pipeline implemented in this function suits both image- and + video-reid. In general, a subclass of Engine only needs to re-implement + ``extract_features()`` and ``parse_data_for_eval()`` (most of the time), + but not a must. Please refer to the source code for more details. + """ + self.set_model_mode('eval') + targets = list(self.test_loader.keys()) + + for name in targets: + domain = 'source' if name in self.datamanager.sources else 'target' + print('##### Evaluating {} ({}) #####'.format(name, domain)) + query_loader = self.test_loader[name]['query'] + gallery_loader = self.test_loader[name]['gallery'] + rank1, mAP = self._evaluate( + dataset_name=name, + query_loader=query_loader, + gallery_loader=gallery_loader, + dist_metric=dist_metric, + normalize_feature=normalize_feature, + visrank=visrank, + visrank_topk=visrank_topk, + save_dir=save_dir, + use_metric_cuhk03=use_metric_cuhk03, + ranks=ranks, + rerank=rerank + ) + + if self.writer is not None: + self.writer.add_scalar(f'Test/{name}/rank1', rank1, self.epoch) + self.writer.add_scalar(f'Test/{name}/mAP', mAP, self.epoch) + + return rank1 + + @torch.no_grad() + def _evaluate( + self, + dataset_name='', + query_loader=None, + gallery_loader=None, + dist_metric='euclidean', + normalize_feature=False, + visrank=False, + visrank_topk=10, + save_dir='', + use_metric_cuhk03=False, + ranks=[1, 5, 10, 20], + rerank=False + ): + batch_time = AverageMeter() + + def _feature_extraction(data_loader): + f_, pids_, camids_ = [], [], [] + for batch_idx, data in enumerate(data_loader): + imgs, pids, camids = self.parse_data_for_eval(data) + if self.use_gpu: + imgs = imgs.cuda() + end = time.time() + features = self.extract_features(imgs) + batch_time.update(time.time() - end) + features = features.cpu().clone() + f_.append(features) + pids_.extend(pids) + camids_.extend(camids) + f_ = torch.cat(f_, 0) + pids_ = np.asarray(pids_) + camids_ = np.asarray(camids_) + return f_, pids_, camids_ + + print('Extracting features from query set ...') + qf, q_pids, q_camids = _feature_extraction(query_loader) + print('Done, obtained {}-by-{} matrix'.format(qf.size(0), qf.size(1))) + + print('Extracting features from gallery set ...') + gf, g_pids, g_camids = _feature_extraction(gallery_loader) + print('Done, obtained {}-by-{} matrix'.format(gf.size(0), gf.size(1))) + + print('Speed: {:.4f} sec/batch'.format(batch_time.avg)) + + if normalize_feature: + print('Normalzing features with L2 norm ...') + qf = F.normalize(qf, p=2, dim=1) + gf = F.normalize(gf, p=2, dim=1) + + print( + 'Computing distance matrix with metric={} ...'.format(dist_metric) + ) + distmat = metrics.compute_distance_matrix(qf, gf, dist_metric) + distmat = distmat.numpy() + + if rerank: + print('Applying person re-ranking ...') + distmat_qq = metrics.compute_distance_matrix(qf, qf, dist_metric) + distmat_gg = metrics.compute_distance_matrix(gf, gf, dist_metric) + distmat = re_ranking(distmat, distmat_qq, distmat_gg) + + print('Computing CMC and mAP ...') + cmc, mAP = metrics.evaluate_rank( + distmat, + q_pids, + g_pids, + q_camids, + g_camids, + use_metric_cuhk03=use_metric_cuhk03 + ) + + print('** Results **') + print('mAP: {:.1%}'.format(mAP)) + print('CMC curve') + for r in ranks: + print('Rank-{:<3}: {:.1%}'.format(r, cmc[r - 1])) + + if visrank: + visualize_ranked_results( + distmat, + self.datamanager.fetch_test_loaders(dataset_name), + self.datamanager.data_type, + width=self.datamanager.width, + height=self.datamanager.height, + save_dir=osp.join(save_dir, 'visrank_' + dataset_name), + topk=visrank_topk + ) + + return cmc[0], mAP + + def compute_loss(self, criterion, outputs, targets): + if isinstance(outputs, (tuple, list)): + loss = DeepSupervision(criterion, outputs, targets) + else: + loss = criterion(outputs, targets) + return loss + + def extract_features(self, input): + return self.model(input) + + def parse_data_for_train(self, data): + imgs = data['img'] + pids = data['pid'] + return imgs, pids + + def parse_data_for_eval(self, data): + imgs = data['img'] + pids = data['pid'] + camids = data['camid'] + return imgs, pids, camids + + def two_stepped_transfer_learning( + self, epoch, fixbase_epoch, open_layers, model=None + ): + """Two-stepped transfer learning. + + The idea is to freeze base layers for a certain number of epochs + and then open all layers for training. + + Reference: https://arxiv.org/abs/1611.05244 + """ + model = self.model if model is None else model + if model is None: + return + + if (epoch + 1) <= fixbase_epoch and open_layers is not None: + print( + '* Only train {} (epoch: {}/{})'.format( + open_layers, epoch + 1, fixbase_epoch + ) + ) + open_specified_layers(model, open_layers) + else: + open_all_layers(model) diff --git a/strong_sort/deep/reid/torchreid/engine/image/__init__.py b/strong_sort/deep/reid/torchreid/engine/image/__init__.py new file mode 100644 index 0000000..08d313a --- /dev/null +++ b/strong_sort/deep/reid/torchreid/engine/image/__init__.py @@ -0,0 +1,4 @@ +from __future__ import absolute_import + +from .softmax import ImageSoftmaxEngine +from .triplet import ImageTripletEngine diff --git a/strong_sort/deep/reid/torchreid/engine/image/softmax.py b/strong_sort/deep/reid/torchreid/engine/image/softmax.py new file mode 100644 index 0000000..5785d4f --- /dev/null +++ b/strong_sort/deep/reid/torchreid/engine/image/softmax.py @@ -0,0 +1,97 @@ +from __future__ import division, print_function, absolute_import + +from torchreid import metrics +from torchreid.losses import CrossEntropyLoss + +from ..engine import Engine + + +class ImageSoftmaxEngine(Engine): + r"""Softmax-loss engine for image-reid. + + Args: + datamanager (DataManager): an instance of ``torchreid.data.ImageDataManager`` + or ``torchreid.data.VideoDataManager``. + model (nn.Module): model instance. + optimizer (Optimizer): an Optimizer. + scheduler (LRScheduler, optional): if None, no learning rate decay will be performed. + use_gpu (bool, optional): use gpu. Default is True. + label_smooth (bool, optional): use label smoothing regularizer. Default is True. + + Examples:: + + import torchreid + datamanager = torchreid.data.ImageDataManager( + root='path/to/reid-data', + sources='market1501', + height=256, + width=128, + combineall=False, + batch_size=32 + ) + model = torchreid.models.build_model( + name='resnet50', + num_classes=datamanager.num_train_pids, + loss='softmax' + ) + model = model.cuda() + optimizer = torchreid.optim.build_optimizer( + model, optim='adam', lr=0.0003 + ) + scheduler = torchreid.optim.build_lr_scheduler( + optimizer, + lr_scheduler='single_step', + stepsize=20 + ) + engine = torchreid.engine.ImageSoftmaxEngine( + datamanager, model, optimizer, scheduler=scheduler + ) + engine.run( + max_epoch=60, + save_dir='log/resnet50-softmax-market1501', + print_freq=10 + ) + """ + + def __init__( + self, + datamanager, + model, + optimizer, + scheduler=None, + use_gpu=True, + label_smooth=True + ): + super(ImageSoftmaxEngine, self).__init__(datamanager, use_gpu) + + self.model = model + self.optimizer = optimizer + self.scheduler = scheduler + self.register_model('model', model, optimizer, scheduler) + + self.criterion = CrossEntropyLoss( + num_classes=self.datamanager.num_train_pids, + use_gpu=self.use_gpu, + label_smooth=label_smooth + ) + + def forward_backward(self, data): + imgs, pids = self.parse_data_for_train(data) + + if self.use_gpu: + imgs = imgs.cuda() + pids = pids.cuda() + + outputs = self.model(imgs) + loss = self.compute_loss(self.criterion, outputs, pids) + + self.optimizer.zero_grad() + loss.backward() + self.optimizer.step() + + loss_summary = { + 'loss': loss.item(), + 'acc': metrics.accuracy(outputs, pids)[0].item() + } + + return loss_summary diff --git a/strong_sort/deep/reid/torchreid/engine/image/triplet.py b/strong_sort/deep/reid/torchreid/engine/image/triplet.py new file mode 100644 index 0000000..cd15cfb --- /dev/null +++ b/strong_sort/deep/reid/torchreid/engine/image/triplet.py @@ -0,0 +1,122 @@ +from __future__ import division, print_function, absolute_import + +from torchreid import metrics +from torchreid.losses import TripletLoss, CrossEntropyLoss + +from ..engine import Engine + + +class ImageTripletEngine(Engine): + r"""Triplet-loss engine for image-reid. + + Args: + datamanager (DataManager): an instance of ``torchreid.data.ImageDataManager`` + or ``torchreid.data.VideoDataManager``. + model (nn.Module): model instance. + optimizer (Optimizer): an Optimizer. + margin (float, optional): margin for triplet loss. Default is 0.3. + weight_t (float, optional): weight for triplet loss. Default is 1. + weight_x (float, optional): weight for softmax loss. Default is 1. + scheduler (LRScheduler, optional): if None, no learning rate decay will be performed. + use_gpu (bool, optional): use gpu. Default is True. + label_smooth (bool, optional): use label smoothing regularizer. Default is True. + + Examples:: + + import torchreid + datamanager = torchreid.data.ImageDataManager( + root='path/to/reid-data', + sources='market1501', + height=256, + width=128, + combineall=False, + batch_size=32, + num_instances=4, + train_sampler='RandomIdentitySampler' # this is important + ) + model = torchreid.models.build_model( + name='resnet50', + num_classes=datamanager.num_train_pids, + loss='triplet' + ) + model = model.cuda() + optimizer = torchreid.optim.build_optimizer( + model, optim='adam', lr=0.0003 + ) + scheduler = torchreid.optim.build_lr_scheduler( + optimizer, + lr_scheduler='single_step', + stepsize=20 + ) + engine = torchreid.engine.ImageTripletEngine( + datamanager, model, optimizer, margin=0.3, + weight_t=0.7, weight_x=1, scheduler=scheduler + ) + engine.run( + max_epoch=60, + save_dir='log/resnet50-triplet-market1501', + print_freq=10 + ) + """ + + def __init__( + self, + datamanager, + model, + optimizer, + margin=0.3, + weight_t=1, + weight_x=1, + scheduler=None, + use_gpu=True, + label_smooth=True + ): + super(ImageTripletEngine, self).__init__(datamanager, use_gpu) + + self.model = model + self.optimizer = optimizer + self.scheduler = scheduler + self.register_model('model', model, optimizer, scheduler) + + assert weight_t >= 0 and weight_x >= 0 + assert weight_t + weight_x > 0 + self.weight_t = weight_t + self.weight_x = weight_x + + self.criterion_t = TripletLoss(margin=margin) + self.criterion_x = CrossEntropyLoss( + num_classes=self.datamanager.num_train_pids, + use_gpu=self.use_gpu, + label_smooth=label_smooth + ) + + def forward_backward(self, data): + imgs, pids = self.parse_data_for_train(data) + + if self.use_gpu: + imgs = imgs.cuda() + pids = pids.cuda() + + outputs, features = self.model(imgs) + + loss = 0 + loss_summary = {} + + if self.weight_t > 0: + loss_t = self.compute_loss(self.criterion_t, features, pids) + loss += self.weight_t * loss_t + loss_summary['loss_t'] = loss_t.item() + + if self.weight_x > 0: + loss_x = self.compute_loss(self.criterion_x, outputs, pids) + loss += self.weight_x * loss_x + loss_summary['loss_x'] = loss_x.item() + loss_summary['acc'] = metrics.accuracy(outputs, pids)[0].item() + + assert loss_summary + + self.optimizer.zero_grad() + loss.backward() + self.optimizer.step() + + return loss_summary diff --git a/strong_sort/deep/reid/torchreid/engine/video/__init__.py b/strong_sort/deep/reid/torchreid/engine/video/__init__.py new file mode 100644 index 0000000..b818bf4 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/engine/video/__init__.py @@ -0,0 +1,4 @@ +from __future__ import absolute_import + +from .softmax import VideoSoftmaxEngine +from .triplet import VideoTripletEngine diff --git a/strong_sort/deep/reid/torchreid/engine/video/softmax.py b/strong_sort/deep/reid/torchreid/engine/video/softmax.py new file mode 100644 index 0000000..fe92feb --- /dev/null +++ b/strong_sort/deep/reid/torchreid/engine/video/softmax.py @@ -0,0 +1,109 @@ +from __future__ import division, print_function, absolute_import +import torch + +from torchreid.engine.image import ImageSoftmaxEngine + + +class VideoSoftmaxEngine(ImageSoftmaxEngine): + """Softmax-loss engine for video-reid. + + Args: + datamanager (DataManager): an instance of ``torchreid.data.ImageDataManager`` + or ``torchreid.data.VideoDataManager``. + model (nn.Module): model instance. + optimizer (Optimizer): an Optimizer. + scheduler (LRScheduler, optional): if None, no learning rate decay will be performed. + use_gpu (bool, optional): use gpu. Default is True. + label_smooth (bool, optional): use label smoothing regularizer. Default is True. + pooling_method (str, optional): how to pool features for a tracklet. + Default is "avg" (average). Choices are ["avg", "max"]. + + Examples:: + + import torch + import torchreid + # Each batch contains batch_size*seq_len images + datamanager = torchreid.data.VideoDataManager( + root='path/to/reid-data', + sources='mars', + height=256, + width=128, + combineall=False, + batch_size=8, # number of tracklets + seq_len=15 # number of images in each tracklet + ) + model = torchreid.models.build_model( + name='resnet50', + num_classes=datamanager.num_train_pids, + loss='softmax' + ) + model = model.cuda() + optimizer = torchreid.optim.build_optimizer( + model, optim='adam', lr=0.0003 + ) + scheduler = torchreid.optim.build_lr_scheduler( + optimizer, + lr_scheduler='single_step', + stepsize=20 + ) + engine = torchreid.engine.VideoSoftmaxEngine( + datamanager, model, optimizer, scheduler=scheduler, + pooling_method='avg' + ) + engine.run( + max_epoch=60, + save_dir='log/resnet50-softmax-mars', + print_freq=10 + ) + """ + + def __init__( + self, + datamanager, + model, + optimizer, + scheduler=None, + use_gpu=True, + label_smooth=True, + pooling_method='avg' + ): + super(VideoSoftmaxEngine, self).__init__( + datamanager, + model, + optimizer, + scheduler=scheduler, + use_gpu=use_gpu, + label_smooth=label_smooth + ) + self.pooling_method = pooling_method + + def parse_data_for_train(self, data): + imgs = data['img'] + pids = data['pid'] + if imgs.dim() == 5: + # b: batch size + # s: sqeuence length + # c: channel depth + # h: height + # w: width + b, s, c, h, w = imgs.size() + imgs = imgs.view(b * s, c, h, w) + pids = pids.view(b, 1).expand(b, s) + pids = pids.contiguous().view(b * s) + return imgs, pids + + def extract_features(self, input): + # b: batch size + # s: sqeuence length + # c: channel depth + # h: height + # w: width + b, s, c, h, w = input.size() + input = input.view(b * s, c, h, w) + features = self.model(input) + features = features.view(b, s, -1) + if self.pooling_method == 'avg': + features = torch.mean(features, 1) + else: + features = torch.max(features, 1)[0] + return features diff --git a/strong_sort/deep/reid/torchreid/engine/video/triplet.py b/strong_sort/deep/reid/torchreid/engine/video/triplet.py new file mode 100644 index 0000000..b2778db --- /dev/null +++ b/strong_sort/deep/reid/torchreid/engine/video/triplet.py @@ -0,0 +1,122 @@ +from __future__ import division, print_function, absolute_import +import torch + +from torchreid.engine.image import ImageTripletEngine + + +class VideoTripletEngine(ImageTripletEngine): + """Triplet-loss engine for video-reid. + + Args: + datamanager (DataManager): an instance of ``torchreid.data.ImageDataManager`` + or ``torchreid.data.VideoDataManager``. + model (nn.Module): model instance. + optimizer (Optimizer): an Optimizer. + margin (float, optional): margin for triplet loss. Default is 0.3. + weight_t (float, optional): weight for triplet loss. Default is 1. + weight_x (float, optional): weight for softmax loss. Default is 1. + scheduler (LRScheduler, optional): if None, no learning rate decay will be performed. + use_gpu (bool, optional): use gpu. Default is True. + label_smooth (bool, optional): use label smoothing regularizer. Default is True. + pooling_method (str, optional): how to pool features for a tracklet. + Default is "avg" (average). Choices are ["avg", "max"]. + + Examples:: + + import torch + import torchreid + # Each batch contains batch_size*seq_len images + # Each identity is sampled with num_instances tracklets + datamanager = torchreid.data.VideoDataManager( + root='path/to/reid-data', + sources='mars', + height=256, + width=128, + combineall=False, + num_instances=4, + train_sampler='RandomIdentitySampler' + batch_size=8, # number of tracklets + seq_len=15 # number of images in each tracklet + ) + model = torchreid.models.build_model( + name='resnet50', + num_classes=datamanager.num_train_pids, + loss='triplet' + ) + model = model.cuda() + optimizer = torchreid.optim.build_optimizer( + model, optim='adam', lr=0.0003 + ) + scheduler = torchreid.optim.build_lr_scheduler( + optimizer, + lr_scheduler='single_step', + stepsize=20 + ) + engine = torchreid.engine.VideoTripletEngine( + datamanager, model, optimizer, margin=0.3, + weight_t=0.7, weight_x=1, scheduler=scheduler, + pooling_method='avg' + ) + engine.run( + max_epoch=60, + save_dir='log/resnet50-triplet-mars', + print_freq=10 + ) + """ + + def __init__( + self, + datamanager, + model, + optimizer, + margin=0.3, + weight_t=1, + weight_x=1, + scheduler=None, + use_gpu=True, + label_smooth=True, + pooling_method='avg' + ): + super(VideoTripletEngine, self).__init__( + datamanager, + model, + optimizer, + margin=margin, + weight_t=weight_t, + weight_x=weight_x, + scheduler=scheduler, + use_gpu=use_gpu, + label_smooth=label_smooth + ) + self.pooling_method = pooling_method + + def parse_data_for_train(self, data): + imgs = data['img'] + pids = data['pid'] + if imgs.dim() == 5: + # b: batch size + # s: sqeuence length + # c: channel depth + # h: height + # w: width + b, s, c, h, w = imgs.size() + imgs = imgs.view(b * s, c, h, w) + pids = pids.view(b, 1).expand(b, s) + pids = pids.contiguous().view(b * s) + return imgs, pids + + def extract_features(self, input): + # b: batch size + # s: sqeuence length + # c: channel depth + # h: height + # w: width + b, s, c, h, w = input.size() + input = input.view(b * s, c, h, w) + features = self.model(input) + features = features.view(b, s, -1) + if self.pooling_method == 'avg': + features = torch.mean(features, 1) + else: + features = torch.max(features, 1)[0] + return features diff --git a/strong_sort/deep/reid/torchreid/losses/__init__.py b/strong_sort/deep/reid/torchreid/losses/__init__.py new file mode 100644 index 0000000..1625749 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/losses/__init__.py @@ -0,0 +1,21 @@ +from __future__ import division, print_function, absolute_import + +from .cross_entropy_loss import CrossEntropyLoss +from .hard_mine_triplet_loss import TripletLoss + + +def DeepSupervision(criterion, xs, y): + """DeepSupervision + + Applies criterion to each element in a list. + + Args: + criterion: loss function + xs: tuple of inputs + y: ground truth + """ + loss = 0. + for x in xs: + loss += criterion(x, y) + loss /= len(xs) + return loss diff --git a/strong_sort/deep/reid/torchreid/losses/cross_entropy_loss.py b/strong_sort/deep/reid/torchreid/losses/cross_entropy_loss.py new file mode 100644 index 0000000..4cfa5d4 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/losses/cross_entropy_loss.py @@ -0,0 +1,50 @@ +from __future__ import division, absolute_import +import torch +import torch.nn as nn + + +class CrossEntropyLoss(nn.Module): + r"""Cross entropy loss with label smoothing regularizer. + + Reference: + Szegedy et al. Rethinking the Inception Architecture for Computer Vision. CVPR 2016. + + With label smoothing, the label :math:`y` for a class is computed by + + .. math:: + \begin{equation} + (1 - \eps) \times y + \frac{\eps}{K}, + \end{equation} + + where :math:`K` denotes the number of classes and :math:`\eps` is a weight. When + :math:`\eps = 0`, the loss function reduces to the normal cross entropy. + + Args: + num_classes (int): number of classes. + eps (float, optional): weight. Default is 0.1. + use_gpu (bool, optional): whether to use gpu devices. Default is True. + label_smooth (bool, optional): whether to apply label smoothing. Default is True. + """ + + def __init__(self, num_classes, eps=0.1, use_gpu=True, label_smooth=True): + super(CrossEntropyLoss, self).__init__() + self.num_classes = num_classes + self.eps = eps if label_smooth else 0 + self.use_gpu = use_gpu + self.logsoftmax = nn.LogSoftmax(dim=1) + + def forward(self, inputs, targets): + """ + Args: + inputs (torch.Tensor): prediction matrix (before softmax) with + shape (batch_size, num_classes). + targets (torch.LongTensor): ground truth labels with shape (batch_size). + Each position contains the label index. + """ + log_probs = self.logsoftmax(inputs) + zeros = torch.zeros(log_probs.size()) + targets = zeros.scatter_(1, targets.unsqueeze(1).data.cpu(), 1) + if self.use_gpu: + targets = targets.cuda() + targets = (1 - self.eps) * targets + self.eps / self.num_classes + return (-targets * log_probs).mean(0).sum() diff --git a/strong_sort/deep/reid/torchreid/losses/hard_mine_triplet_loss.py b/strong_sort/deep/reid/torchreid/losses/hard_mine_triplet_loss.py new file mode 100644 index 0000000..ef9019b --- /dev/null +++ b/strong_sort/deep/reid/torchreid/losses/hard_mine_triplet_loss.py @@ -0,0 +1,48 @@ +from __future__ import division, absolute_import +import torch +import torch.nn as nn + + +class TripletLoss(nn.Module): + """Triplet loss with hard positive/negative mining. + + Reference: + Hermans et al. In Defense of the Triplet Loss for Person Re-Identification. arXiv:1703.07737. + + Imported from ``_. + + Args: + margin (float, optional): margin for triplet. Default is 0.3. + """ + + def __init__(self, margin=0.3): + super(TripletLoss, self).__init__() + self.margin = margin + self.ranking_loss = nn.MarginRankingLoss(margin=margin) + + def forward(self, inputs, targets): + """ + Args: + inputs (torch.Tensor): feature matrix with shape (batch_size, feat_dim). + targets (torch.LongTensor): ground truth labels with shape (num_classes). + """ + n = inputs.size(0) + + # Compute pairwise distance, replace by the official when merged + dist = torch.pow(inputs, 2).sum(dim=1, keepdim=True).expand(n, n) + dist = dist + dist.t() + dist.addmm_(inputs, inputs.t(), beta=1, alpha=-2) + dist = dist.clamp(min=1e-12).sqrt() # for numerical stability + + # For each anchor, find the hardest positive and negative + mask = targets.expand(n, n).eq(targets.expand(n, n).t()) + dist_ap, dist_an = [], [] + for i in range(n): + dist_ap.append(dist[i][mask[i]].max().unsqueeze(0)) + dist_an.append(dist[i][mask[i] == 0].min().unsqueeze(0)) + dist_ap = torch.cat(dist_ap) + dist_an = torch.cat(dist_an) + + # Compute ranking hinge loss + y = torch.ones_like(dist_an) + return self.ranking_loss(dist_an, dist_ap, y) diff --git a/strong_sort/deep/reid/torchreid/metrics/__init__.py b/strong_sort/deep/reid/torchreid/metrics/__init__.py new file mode 100644 index 0000000..5159e1e --- /dev/null +++ b/strong_sort/deep/reid/torchreid/metrics/__init__.py @@ -0,0 +1,5 @@ +from __future__ import absolute_import + +from .rank import evaluate_rank +from .accuracy import accuracy +from .distance import compute_distance_matrix diff --git a/strong_sort/deep/reid/torchreid/metrics/accuracy.py b/strong_sort/deep/reid/torchreid/metrics/accuracy.py new file mode 100644 index 0000000..3161f7b --- /dev/null +++ b/strong_sort/deep/reid/torchreid/metrics/accuracy.py @@ -0,0 +1,37 @@ +from __future__ import division, print_function, absolute_import + + +def accuracy(output, target, topk=(1, )): + """Computes the accuracy over the k top predictions for + the specified values of k. + + Args: + output (torch.Tensor): prediction matrix with shape (batch_size, num_classes). + target (torch.LongTensor): ground truth labels with shape (batch_size). + topk (tuple, optional): accuracy at top-k will be computed. For example, + topk=(1, 5) means accuracy at top-1 and top-5 will be computed. + + Returns: + list: accuracy at top-k. + + Examples:: + >>> from torchreid import metrics + >>> metrics.accuracy(output, target) + """ + maxk = max(topk) + batch_size = target.size(0) + + if isinstance(output, (tuple, list)): + output = output[0] + + _, pred = output.topk(maxk, 1, True, True) + pred = pred.t() + correct = pred.eq(target.view(1, -1).expand_as(pred)) + + res = [] + for k in topk: + correct_k = correct[:k].view(-1).float().sum(0, keepdim=True) + acc = correct_k.mul_(100.0 / batch_size) + res.append(acc) + + return res diff --git a/strong_sort/deep/reid/torchreid/metrics/distance.py b/strong_sort/deep/reid/torchreid/metrics/distance.py new file mode 100644 index 0000000..f4fb383 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/metrics/distance.py @@ -0,0 +1,80 @@ +from __future__ import division, print_function, absolute_import +import torch +from torch.nn import functional as F + + +def compute_distance_matrix(input1, input2, metric='euclidean'): + """A wrapper function for computing distance matrix. + + Args: + input1 (torch.Tensor): 2-D feature matrix. + input2 (torch.Tensor): 2-D feature matrix. + metric (str, optional): "euclidean" or "cosine". + Default is "euclidean". + + Returns: + torch.Tensor: distance matrix. + + Examples:: + >>> from torchreid import metrics + >>> input1 = torch.rand(10, 2048) + >>> input2 = torch.rand(100, 2048) + >>> distmat = metrics.compute_distance_matrix(input1, input2) + >>> distmat.size() # (10, 100) + """ + # check input + assert isinstance(input1, torch.Tensor) + assert isinstance(input2, torch.Tensor) + assert input1.dim() == 2, 'Expected 2-D tensor, but got {}-D'.format( + input1.dim() + ) + assert input2.dim() == 2, 'Expected 2-D tensor, but got {}-D'.format( + input2.dim() + ) + assert input1.size(1) == input2.size(1) + + if metric == 'euclidean': + distmat = euclidean_squared_distance(input1, input2) + elif metric == 'cosine': + distmat = cosine_distance(input1, input2) + else: + raise ValueError( + 'Unknown distance metric: {}. ' + 'Please choose either "euclidean" or "cosine"'.format(metric) + ) + + return distmat + + +def euclidean_squared_distance(input1, input2): + """Computes euclidean squared distance. + + Args: + input1 (torch.Tensor): 2-D feature matrix. + input2 (torch.Tensor): 2-D feature matrix. + + Returns: + torch.Tensor: distance matrix. + """ + m, n = input1.size(0), input2.size(0) + mat1 = torch.pow(input1, 2).sum(dim=1, keepdim=True).expand(m, n) + mat2 = torch.pow(input2, 2).sum(dim=1, keepdim=True).expand(n, m).t() + distmat = mat1 + mat2 + distmat.addmm_(input1, input2.t(), beta=1, alpha=-2) + return distmat + + +def cosine_distance(input1, input2): + """Computes cosine distance. + + Args: + input1 (torch.Tensor): 2-D feature matrix. + input2 (torch.Tensor): 2-D feature matrix. + + Returns: + torch.Tensor: distance matrix. + """ + input1_normed = F.normalize(input1, p=2, dim=1) + input2_normed = F.normalize(input2, p=2, dim=1) + distmat = 1 - torch.mm(input1_normed, input2_normed.t()) + return distmat diff --git a/strong_sort/deep/reid/torchreid/metrics/rank.py b/strong_sort/deep/reid/torchreid/metrics/rank.py new file mode 100644 index 0000000..bf6205b --- /dev/null +++ b/strong_sort/deep/reid/torchreid/metrics/rank.py @@ -0,0 +1,207 @@ +from __future__ import division, print_function, absolute_import +import numpy as np +import warnings +from collections import defaultdict + +try: + from torchreid.metrics.rank_cylib.rank_cy import evaluate_cy + IS_CYTHON_AVAI = True +except ImportError: + IS_CYTHON_AVAI = False + warnings.warn( + 'Cython evaluation (very fast so highly recommended) is ' + 'unavailable, now use python evaluation.' + ) + + +def eval_cuhk03(distmat, q_pids, g_pids, q_camids, g_camids, max_rank): + """Evaluation with cuhk03 metric + Key: one image for each gallery identity is randomly sampled for each query identity. + Random sampling is performed num_repeats times. + """ + num_repeats = 10 + num_q, num_g = distmat.shape + + if num_g < max_rank: + max_rank = num_g + print( + 'Note: number of gallery samples is quite small, got {}'. + format(num_g) + ) + + indices = np.argsort(distmat, axis=1) + matches = (g_pids[indices] == q_pids[:, np.newaxis]).astype(np.int32) + + # compute cmc curve for each query + all_cmc = [] + all_AP = [] + num_valid_q = 0. # number of valid query + + for q_idx in range(num_q): + # get query pid and camid + q_pid = q_pids[q_idx] + q_camid = q_camids[q_idx] + + # remove gallery samples that have the same pid and camid with query + order = indices[q_idx] + remove = (g_pids[order] == q_pid) & (g_camids[order] == q_camid) + keep = np.invert(remove) + + # compute cmc curve + raw_cmc = matches[q_idx][ + keep] # binary vector, positions with value 1 are correct matches + if not np.any(raw_cmc): + # this condition is true when query identity does not appear in gallery + continue + + kept_g_pids = g_pids[order][keep] + g_pids_dict = defaultdict(list) + for idx, pid in enumerate(kept_g_pids): + g_pids_dict[pid].append(idx) + + cmc = 0. + for repeat_idx in range(num_repeats): + mask = np.zeros(len(raw_cmc), dtype=np.bool) + for _, idxs in g_pids_dict.items(): + # randomly sample one image for each gallery person + rnd_idx = np.random.choice(idxs) + mask[rnd_idx] = True + masked_raw_cmc = raw_cmc[mask] + _cmc = masked_raw_cmc.cumsum() + _cmc[_cmc > 1] = 1 + cmc += _cmc[:max_rank].astype(np.float32) + + cmc /= num_repeats + all_cmc.append(cmc) + # compute AP + num_rel = raw_cmc.sum() + tmp_cmc = raw_cmc.cumsum() + tmp_cmc = [x / (i+1.) for i, x in enumerate(tmp_cmc)] + tmp_cmc = np.asarray(tmp_cmc) * raw_cmc + AP = tmp_cmc.sum() / num_rel + all_AP.append(AP) + num_valid_q += 1. + + assert num_valid_q > 0, 'Error: all query identities do not appear in gallery' + + all_cmc = np.asarray(all_cmc).astype(np.float32) + all_cmc = all_cmc.sum(0) / num_valid_q + mAP = np.mean(all_AP) + + return all_cmc, mAP + + +def eval_market1501(distmat, q_pids, g_pids, q_camids, g_camids, max_rank): + """Evaluation with market1501 metric + Key: for each query identity, its gallery images from the same camera view are discarded. + """ + num_q, num_g = distmat.shape + + if num_g < max_rank: + max_rank = num_g + print( + 'Note: number of gallery samples is quite small, got {}'. + format(num_g) + ) + + indices = np.argsort(distmat, axis=1) + matches = (g_pids[indices] == q_pids[:, np.newaxis]).astype(np.int32) + + # compute cmc curve for each query + all_cmc = [] + all_AP = [] + num_valid_q = 0. # number of valid query + + for q_idx in range(num_q): + # get query pid and camid + q_pid = q_pids[q_idx] + q_camid = q_camids[q_idx] + + # remove gallery samples that have the same pid and camid with query + order = indices[q_idx] + remove = (g_pids[order] == q_pid) & (g_camids[order] == q_camid) + keep = np.invert(remove) + + # compute cmc curve + raw_cmc = matches[q_idx][ + keep] # binary vector, positions with value 1 are correct matches + if not np.any(raw_cmc): + # this condition is true when query identity does not appear in gallery + continue + + cmc = raw_cmc.cumsum() + cmc[cmc > 1] = 1 + + all_cmc.append(cmc[:max_rank]) + num_valid_q += 1. + + # compute average precision + # reference: https://en.wikipedia.org/wiki/Evaluation_measures_(information_retrieval)#Average_precision + num_rel = raw_cmc.sum() + tmp_cmc = raw_cmc.cumsum() + tmp_cmc = [x / (i+1.) for i, x in enumerate(tmp_cmc)] + tmp_cmc = np.asarray(tmp_cmc) * raw_cmc + AP = tmp_cmc.sum() / num_rel + all_AP.append(AP) + + assert num_valid_q > 0, 'Error: all query identities do not appear in gallery' + + all_cmc = np.asarray(all_cmc).astype(np.float32) + all_cmc = all_cmc.sum(0) / num_valid_q + mAP = np.mean(all_AP) + + return all_cmc, mAP + + +def evaluate_py( + distmat, q_pids, g_pids, q_camids, g_camids, max_rank, use_metric_cuhk03 +): + if use_metric_cuhk03: + return eval_cuhk03( + distmat, q_pids, g_pids, q_camids, g_camids, max_rank + ) + else: + return eval_market1501( + distmat, q_pids, g_pids, q_camids, g_camids, max_rank + ) + + +def evaluate_rank( + distmat, + q_pids, + g_pids, + q_camids, + g_camids, + max_rank=50, + use_metric_cuhk03=False, + use_cython=True +): + """Evaluates CMC rank. + + Args: + distmat (numpy.ndarray): distance matrix of shape (num_query, num_gallery). + q_pids (numpy.ndarray): 1-D array containing person identities + of each query instance. + g_pids (numpy.ndarray): 1-D array containing person identities + of each gallery instance. + q_camids (numpy.ndarray): 1-D array containing camera views under + which each query instance is captured. + g_camids (numpy.ndarray): 1-D array containing camera views under + which each gallery instance is captured. + max_rank (int, optional): maximum CMC rank to be computed. Default is 50. + use_metric_cuhk03 (bool, optional): use single-gallery-shot setting for cuhk03. + Default is False. This should be enabled when using cuhk03 classic split. + use_cython (bool, optional): use cython code for evaluation. Default is True. + This is highly recommended as the cython code can speed up the cmc computation + by more than 10x. This requires Cython to be installed. + """ + if use_cython and IS_CYTHON_AVAI: + return evaluate_cy( + distmat, q_pids, g_pids, q_camids, g_camids, max_rank, + use_metric_cuhk03 + ) + else: + return evaluate_py( + distmat, q_pids, g_pids, q_camids, g_camids, max_rank, + use_metric_cuhk03 + ) diff --git a/strong_sort/deep/reid/torchreid/metrics/rank_cylib/Makefile b/strong_sort/deep/reid/torchreid/metrics/rank_cylib/Makefile new file mode 100644 index 0000000..d49e655 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/metrics/rank_cylib/Makefile @@ -0,0 +1,6 @@ +all: + $(PYTHON) setup.py build_ext --inplace + rm -rf build +clean: + rm -rf build + rm -f rank_cy.c *.so \ No newline at end of file diff --git a/strong_sort/deep/reid/torchreid/metrics/rank_cylib/__init__.py b/strong_sort/deep/reid/torchreid/metrics/rank_cylib/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/strong_sort/deep/reid/torchreid/metrics/rank_cylib/rank_cy.pyx b/strong_sort/deep/reid/torchreid/metrics/rank_cylib/rank_cy.pyx new file mode 100644 index 0000000..b4a8690 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/metrics/rank_cylib/rank_cy.pyx @@ -0,0 +1,251 @@ +# cython: boundscheck=False, wraparound=False, nonecheck=False, cdivision=True + +from __future__ import print_function +import numpy as np +from libc.stdint cimport int64_t, uint64_t + +import cython + +cimport numpy as np + +import random +from collections import defaultdict + +""" +Compiler directives: +https://github.com/cython/cython/wiki/enhancements-compilerdirectives + +Cython tutorial: +https://cython.readthedocs.io/en/latest/src/userguide/numpy_tutorial.html + +Credit to https://github.com/luzai +""" + + +# Main interface +cpdef evaluate_cy(distmat, q_pids, g_pids, q_camids, g_camids, max_rank, use_metric_cuhk03=False): + distmat = np.asarray(distmat, dtype=np.float32) + q_pids = np.asarray(q_pids, dtype=np.int64) + g_pids = np.asarray(g_pids, dtype=np.int64) + q_camids = np.asarray(q_camids, dtype=np.int64) + g_camids = np.asarray(g_camids, dtype=np.int64) + if use_metric_cuhk03: + return eval_cuhk03_cy(distmat, q_pids, g_pids, q_camids, g_camids, max_rank) + return eval_market1501_cy(distmat, q_pids, g_pids, q_camids, g_camids, max_rank) + + +cpdef eval_cuhk03_cy(float[:,:] distmat, int64_t[:] q_pids, int64_t[:]g_pids, + int64_t[:]q_camids, int64_t[:]g_camids, int64_t max_rank): + + cdef int64_t num_q = distmat.shape[0] + cdef int64_t num_g = distmat.shape[1] + + if num_g < max_rank: + max_rank = num_g + print('Note: number of gallery samples is quite small, got {}'.format(num_g)) + + cdef: + int64_t num_repeats = 10 + int64_t[:,:] indices = np.argsort(distmat, axis=1) + int64_t[:,:] matches = (np.asarray(g_pids)[np.asarray(indices)] == np.asarray(q_pids)[:, np.newaxis]).astype(np.int64) + + float[:,:] all_cmc = np.zeros((num_q, max_rank), dtype=np.float32) + float[:] all_AP = np.zeros(num_q, dtype=np.float32) + float num_valid_q = 0. # number of valid query + + int64_t q_idx, q_pid, q_camid, g_idx + int64_t[:] order = np.zeros(num_g, dtype=np.int64) + int64_t keep + + float[:] raw_cmc = np.zeros(num_g, dtype=np.float32) # binary vector, positions with value 1 are correct matches + float[:] masked_raw_cmc = np.zeros(num_g, dtype=np.float32) + float[:] cmc, masked_cmc + int64_t num_g_real, num_g_real_masked, rank_idx, rnd_idx + uint64_t meet_condition + float AP + int64_t[:] kept_g_pids, mask + + float num_rel + float[:] tmp_cmc = np.zeros(num_g, dtype=np.float32) + float tmp_cmc_sum + + for q_idx in range(num_q): + # get query pid and camid + q_pid = q_pids[q_idx] + q_camid = q_camids[q_idx] + + # remove gallery samples that have the same pid and camid with query + for g_idx in range(num_g): + order[g_idx] = indices[q_idx, g_idx] + num_g_real = 0 + meet_condition = 0 + kept_g_pids = np.zeros(num_g, dtype=np.int64) + + for g_idx in range(num_g): + if (g_pids[order[g_idx]] != q_pid) or (g_camids[order[g_idx]] != q_camid): + raw_cmc[num_g_real] = matches[q_idx][g_idx] + kept_g_pids[num_g_real] = g_pids[order[g_idx]] + num_g_real += 1 + if matches[q_idx][g_idx] > 1e-31: + meet_condition = 1 + + if not meet_condition: + # this condition is true when query identity does not appear in gallery + continue + + # cuhk03-specific setting + g_pids_dict = defaultdict(list) # overhead! + for g_idx in range(num_g_real): + g_pids_dict[kept_g_pids[g_idx]].append(g_idx) + + cmc = np.zeros(max_rank, dtype=np.float32) + for _ in range(num_repeats): + mask = np.zeros(num_g_real, dtype=np.int64) + + for _, idxs in g_pids_dict.items(): + # randomly sample one image for each gallery person + rnd_idx = np.random.choice(idxs) + #rnd_idx = idxs[0] # use deterministic for debugging + mask[rnd_idx] = 1 + + num_g_real_masked = 0 + for g_idx in range(num_g_real): + if mask[g_idx] == 1: + masked_raw_cmc[num_g_real_masked] = raw_cmc[g_idx] + num_g_real_masked += 1 + + masked_cmc = np.zeros(num_g, dtype=np.float32) + function_cumsum(masked_raw_cmc, masked_cmc, num_g_real_masked) + for g_idx in range(num_g_real_masked): + if masked_cmc[g_idx] > 1: + masked_cmc[g_idx] = 1 + + for rank_idx in range(max_rank): + cmc[rank_idx] += masked_cmc[rank_idx] / num_repeats + + for rank_idx in range(max_rank): + all_cmc[q_idx, rank_idx] = cmc[rank_idx] + # compute average precision + # reference: https://en.wikipedia.org/wiki/Evaluation_measures_(information_retrieval)#Average_precision + function_cumsum(raw_cmc, tmp_cmc, num_g_real) + num_rel = 0 + tmp_cmc_sum = 0 + for g_idx in range(num_g_real): + tmp_cmc_sum += (tmp_cmc[g_idx] / (g_idx + 1.)) * raw_cmc[g_idx] + num_rel += raw_cmc[g_idx] + all_AP[q_idx] = tmp_cmc_sum / num_rel + num_valid_q += 1. + + assert num_valid_q > 0, 'Error: all query identities do not appear in gallery' + + # compute averaged cmc + cdef float[:] avg_cmc = np.zeros(max_rank, dtype=np.float32) + for rank_idx in range(max_rank): + for q_idx in range(num_q): + avg_cmc[rank_idx] += all_cmc[q_idx, rank_idx] + avg_cmc[rank_idx] /= num_valid_q + + cdef float mAP = 0 + for q_idx in range(num_q): + mAP += all_AP[q_idx] + mAP /= num_valid_q + + return np.asarray(avg_cmc).astype(np.float32), mAP + + +cpdef eval_market1501_cy(float[:,:] distmat, int64_t[:] q_pids, int64_t[:]g_pids, + int64_t[:]q_camids, int64_t[:]g_camids, int64_t max_rank): + + cdef int64_t num_q = distmat.shape[0] + cdef int64_t num_g = distmat.shape[1] + + if num_g < max_rank: + max_rank = num_g + print('Note: number of gallery samples is quite small, got {}'.format(num_g)) + + cdef: + int64_t[:,:] indices = np.argsort(distmat, axis=1) + int64_t[:,:] matches = (np.asarray(g_pids)[np.asarray(indices)] == np.asarray(q_pids)[:, np.newaxis]).astype(np.int64) + + float[:,:] all_cmc = np.zeros((num_q, max_rank), dtype=np.float32) + float[:] all_AP = np.zeros(num_q, dtype=np.float32) + float num_valid_q = 0. # number of valid query + + int64_t q_idx, q_pid, q_camid, g_idx + int64_t[:] order = np.zeros(num_g, dtype=np.int64) + int64_t keep + + float[:] raw_cmc = np.zeros(num_g, dtype=np.float32) # binary vector, positions with value 1 are correct matches + float[:] cmc = np.zeros(num_g, dtype=np.float32) + int64_t num_g_real, rank_idx + uint64_t meet_condition + + float num_rel + float[:] tmp_cmc = np.zeros(num_g, dtype=np.float32) + float tmp_cmc_sum + + for q_idx in range(num_q): + # get query pid and camid + q_pid = q_pids[q_idx] + q_camid = q_camids[q_idx] + + # remove gallery samples that have the same pid and camid with query + for g_idx in range(num_g): + order[g_idx] = indices[q_idx, g_idx] + num_g_real = 0 + meet_condition = 0 + + for g_idx in range(num_g): + if (g_pids[order[g_idx]] != q_pid) or (g_camids[order[g_idx]] != q_camid): + raw_cmc[num_g_real] = matches[q_idx][g_idx] + num_g_real += 1 + if matches[q_idx][g_idx] > 1e-31: + meet_condition = 1 + + if not meet_condition: + # this condition is true when query identity does not appear in gallery + continue + + # compute cmc + function_cumsum(raw_cmc, cmc, num_g_real) + for g_idx in range(num_g_real): + if cmc[g_idx] > 1: + cmc[g_idx] = 1 + + for rank_idx in range(max_rank): + all_cmc[q_idx, rank_idx] = cmc[rank_idx] + num_valid_q += 1. + + # compute average precision + # reference: https://en.wikipedia.org/wiki/Evaluation_measures_(information_retrieval)#Average_precision + function_cumsum(raw_cmc, tmp_cmc, num_g_real) + num_rel = 0 + tmp_cmc_sum = 0 + for g_idx in range(num_g_real): + tmp_cmc_sum += (tmp_cmc[g_idx] / (g_idx + 1.)) * raw_cmc[g_idx] + num_rel += raw_cmc[g_idx] + all_AP[q_idx] = tmp_cmc_sum / num_rel + + assert num_valid_q > 0, 'Error: all query identities do not appear in gallery' + + # compute averaged cmc + cdef float[:] avg_cmc = np.zeros(max_rank, dtype=np.float32) + for rank_idx in range(max_rank): + for q_idx in range(num_q): + avg_cmc[rank_idx] += all_cmc[q_idx, rank_idx] + avg_cmc[rank_idx] /= num_valid_q + + cdef float mAP = 0 + for q_idx in range(num_q): + mAP += all_AP[q_idx] + mAP /= num_valid_q + + return np.asarray(avg_cmc).astype(np.float32), mAP + + +# Compute the cumulative sum +cdef void function_cumsum(cython.numeric[:] src, cython.numeric[:] dst, int64_t n): + cdef int64_t i + dst[0] = src[0] + for i in range(1, n): + dst[i] = src[i] + dst[i - 1] \ No newline at end of file diff --git a/strong_sort/deep/reid/torchreid/metrics/rank_cylib/setup.py b/strong_sort/deep/reid/torchreid/metrics/rank_cylib/setup.py new file mode 100644 index 0000000..ce2aeb7 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/metrics/rank_cylib/setup.py @@ -0,0 +1,26 @@ +import numpy as np +from distutils.core import setup +from distutils.extension import Extension +from Cython.Build import cythonize + + +def numpy_include(): + try: + numpy_include = np.get_include() + except AttributeError: + numpy_include = np.get_numpy_include() + return numpy_include + + +ext_modules = [ + Extension( + 'rank_cy', + ['rank_cy.pyx'], + include_dirs=[numpy_include()], + ) +] + +setup( + name='Cython-based reid evaluation code', + ext_modules=cythonize(ext_modules) +) diff --git a/strong_sort/deep/reid/torchreid/metrics/rank_cylib/test_cython.py b/strong_sort/deep/reid/torchreid/metrics/rank_cylib/test_cython.py new file mode 100644 index 0000000..5d1175d --- /dev/null +++ b/strong_sort/deep/reid/torchreid/metrics/rank_cylib/test_cython.py @@ -0,0 +1,83 @@ +from __future__ import print_function +import sys +import numpy as np +import timeit +import os.path as osp + +from torchreid import metrics + +sys.path.insert(0, osp.dirname(osp.abspath(__file__)) + '/../../..') +""" +Test the speed of cython-based evaluation code. The speed improvements +can be much bigger when using the real reid data, which contains a larger +amount of query and gallery images. + +Note: you might encounter the following error: + 'AssertionError: Error: all query identities do not appear in gallery'. +This is normal because the inputs are random numbers. Just try again. +""" + +print('*** Compare running time ***') + +setup = ''' +import sys +import os.path as osp +import numpy as np +sys.path.insert(0, osp.dirname(osp.abspath(__file__)) + '/../../..') +from torchreid import metrics +num_q = 30 +num_g = 300 +max_rank = 5 +distmat = np.random.rand(num_q, num_g) * 20 +q_pids = np.random.randint(0, num_q, size=num_q) +g_pids = np.random.randint(0, num_g, size=num_g) +q_camids = np.random.randint(0, 5, size=num_q) +g_camids = np.random.randint(0, 5, size=num_g) +''' + +print('=> Using market1501\'s metric') +pytime = timeit.timeit( + 'metrics.evaluate_rank(distmat, q_pids, g_pids, q_camids, g_camids, max_rank, use_cython=False)', + setup=setup, + number=20 +) +cytime = timeit.timeit( + 'metrics.evaluate_rank(distmat, q_pids, g_pids, q_camids, g_camids, max_rank, use_cython=True)', + setup=setup, + number=20 +) +print('Python time: {} s'.format(pytime)) +print('Cython time: {} s'.format(cytime)) +print('Cython is {} times faster than python\n'.format(pytime / cytime)) + +print('=> Using cuhk03\'s metric') +pytime = timeit.timeit( + 'metrics.evaluate_rank(distmat, q_pids, g_pids, q_camids, g_camids, max_rank, use_metric_cuhk03=True, use_cython=False)', + setup=setup, + number=20 +) +cytime = timeit.timeit( + 'metrics.evaluate_rank(distmat, q_pids, g_pids, q_camids, g_camids, max_rank, use_metric_cuhk03=True, use_cython=True)', + setup=setup, + number=20 +) +print('Python time: {} s'.format(pytime)) +print('Cython time: {} s'.format(cytime)) +print('Cython is {} times faster than python\n'.format(pytime / cytime)) +""" +print("=> Check precision") + +num_q = 30 +num_g = 300 +max_rank = 5 +distmat = np.random.rand(num_q, num_g) * 20 +q_pids = np.random.randint(0, num_q, size=num_q) +g_pids = np.random.randint(0, num_g, size=num_g) +q_camids = np.random.randint(0, 5, size=num_q) +g_camids = np.random.randint(0, 5, size=num_g) + +cmc, mAP = evaluate(distmat, q_pids, g_pids, q_camids, g_camids, max_rank, use_cython=False) +print("Python:\nmAP = {} \ncmc = {}\n".format(mAP, cmc)) +cmc, mAP = evaluate(distmat, q_pids, g_pids, q_camids, g_camids, max_rank, use_cython=True) +print("Cython:\nmAP = {} \ncmc = {}\n".format(mAP, cmc)) +""" diff --git a/strong_sort/deep/reid/torchreid/models/__init__.py b/strong_sort/deep/reid/torchreid/models/__init__.py new file mode 100644 index 0000000..3c60ba6 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/models/__init__.py @@ -0,0 +1,122 @@ +from __future__ import absolute_import +import torch + +from .pcb import * +from .mlfn import * +from .hacnn import * +from .osnet import * +from .senet import * +from .mudeep import * +from .nasnet import * +from .resnet import * +from .densenet import * +from .xception import * +from .osnet_ain import * +from .resnetmid import * +from .shufflenet import * +from .squeezenet import * +from .inceptionv4 import * +from .mobilenetv2 import * +from .resnet_ibn_a import * +from .resnet_ibn_b import * +from .shufflenetv2 import * +from .inceptionresnetv2 import * + +__model_factory = { + # image classification models + 'resnet18': resnet18, + 'resnet34': resnet34, + 'resnet50': resnet50, + 'resnet101': resnet101, + 'resnet152': resnet152, + 'resnext50_32x4d': resnext50_32x4d, + 'resnext101_32x8d': resnext101_32x8d, + 'resnet50_fc512': resnet50_fc512, + 'se_resnet50': se_resnet50, + 'se_resnet50_fc512': se_resnet50_fc512, + 'se_resnet101': se_resnet101, + 'se_resnext50_32x4d': se_resnext50_32x4d, + 'se_resnext101_32x4d': se_resnext101_32x4d, + 'densenet121': densenet121, + 'densenet169': densenet169, + 'densenet201': densenet201, + 'densenet161': densenet161, + 'densenet121_fc512': densenet121_fc512, + 'inceptionresnetv2': inceptionresnetv2, + 'inceptionv4': inceptionv4, + 'xception': xception, + 'resnet50_ibn_a': resnet50_ibn_a, + 'resnet50_ibn_b': resnet50_ibn_b, + # lightweight models + 'nasnsetmobile': nasnetamobile, + 'mobilenetv2_x1_0': mobilenetv2_x1_0, + 'mobilenetv2_x1_4': mobilenetv2_x1_4, + 'shufflenet': shufflenet, + 'squeezenet1_0': squeezenet1_0, + 'squeezenet1_0_fc512': squeezenet1_0_fc512, + 'squeezenet1_1': squeezenet1_1, + 'shufflenet_v2_x0_5': shufflenet_v2_x0_5, + 'shufflenet_v2_x1_0': shufflenet_v2_x1_0, + 'shufflenet_v2_x1_5': shufflenet_v2_x1_5, + 'shufflenet_v2_x2_0': shufflenet_v2_x2_0, + # reid-specific models + 'mudeep': MuDeep, + 'resnet50mid': resnet50mid, + 'hacnn': HACNN, + 'pcb_p6': pcb_p6, + 'pcb_p4': pcb_p4, + 'mlfn': mlfn, + 'osnet_x1_0': osnet_x1_0, + 'osnet_x0_75': osnet_x0_75, + 'osnet_x0_5': osnet_x0_5, + 'osnet_x0_25': osnet_x0_25, + 'osnet_ibn_x1_0': osnet_ibn_x1_0, + 'osnet_ain_x1_0': osnet_ain_x1_0, + 'osnet_ain_x0_75': osnet_ain_x0_75, + 'osnet_ain_x0_5': osnet_ain_x0_5, + 'osnet_ain_x0_25': osnet_ain_x0_25 +} + + +def show_avai_models(): + """Displays available models. + + Examples:: + >>> from torchreid import models + >>> models.show_avai_models() + """ + print(list(__model_factory.keys())) + + +def build_model( + name, num_classes, loss='softmax', pretrained=True, use_gpu=True +): + """A function wrapper for building a model. + + Args: + name (str): model name. + num_classes (int): number of training identities. + loss (str, optional): loss function to optimize the model. Currently + supports "softmax" and "triplet". Default is "softmax". + pretrained (bool, optional): whether to load ImageNet-pretrained weights. + Default is True. + use_gpu (bool, optional): whether to use gpu. Default is True. + + Returns: + nn.Module + + Examples:: + >>> from torchreid import models + >>> model = models.build_model('resnet50', 751, loss='softmax') + """ + avai_models = list(__model_factory.keys()) + if name not in avai_models: + raise KeyError( + 'Unknown model: {}. Must be one of {}'.format(name, avai_models) + ) + return __model_factory[name]( + num_classes=num_classes, + loss=loss, + pretrained=pretrained, + use_gpu=use_gpu + ) diff --git a/strong_sort/deep/reid/torchreid/models/densenet.py b/strong_sort/deep/reid/torchreid/models/densenet.py new file mode 100644 index 0000000..a1d9b7e --- /dev/null +++ b/strong_sort/deep/reid/torchreid/models/densenet.py @@ -0,0 +1,380 @@ +""" +Code source: https://github.com/pytorch/vision +""" +from __future__ import division, absolute_import +import re +from collections import OrderedDict +import torch +import torch.nn as nn +from torch.nn import functional as F +from torch.utils import model_zoo + +__all__ = [ + 'densenet121', 'densenet169', 'densenet201', 'densenet161', + 'densenet121_fc512' +] + +model_urls = { + 'densenet121': + 'https://download.pytorch.org/models/densenet121-a639ec97.pth', + 'densenet169': + 'https://download.pytorch.org/models/densenet169-b2777c0a.pth', + 'densenet201': + 'https://download.pytorch.org/models/densenet201-c1103571.pth', + 'densenet161': + 'https://download.pytorch.org/models/densenet161-8d451a50.pth', +} + + +class _DenseLayer(nn.Sequential): + + def __init__(self, num_input_features, growth_rate, bn_size, drop_rate): + super(_DenseLayer, self).__init__() + self.add_module('norm1', nn.BatchNorm2d(num_input_features)), + self.add_module('relu1', nn.ReLU(inplace=True)), + self.add_module( + 'conv1', + nn.Conv2d( + num_input_features, + bn_size * growth_rate, + kernel_size=1, + stride=1, + bias=False + ) + ), + self.add_module('norm2', nn.BatchNorm2d(bn_size * growth_rate)), + self.add_module('relu2', nn.ReLU(inplace=True)), + self.add_module( + 'conv2', + nn.Conv2d( + bn_size * growth_rate, + growth_rate, + kernel_size=3, + stride=1, + padding=1, + bias=False + ) + ), + self.drop_rate = drop_rate + + def forward(self, x): + new_features = super(_DenseLayer, self).forward(x) + if self.drop_rate > 0: + new_features = F.dropout( + new_features, p=self.drop_rate, training=self.training + ) + return torch.cat([x, new_features], 1) + + +class _DenseBlock(nn.Sequential): + + def __init__( + self, num_layers, num_input_features, bn_size, growth_rate, drop_rate + ): + super(_DenseBlock, self).__init__() + for i in range(num_layers): + layer = _DenseLayer( + num_input_features + i*growth_rate, growth_rate, bn_size, + drop_rate + ) + self.add_module('denselayer%d' % (i+1), layer) + + +class _Transition(nn.Sequential): + + def __init__(self, num_input_features, num_output_features): + super(_Transition, self).__init__() + self.add_module('norm', nn.BatchNorm2d(num_input_features)) + self.add_module('relu', nn.ReLU(inplace=True)) + self.add_module( + 'conv', + nn.Conv2d( + num_input_features, + num_output_features, + kernel_size=1, + stride=1, + bias=False + ) + ) + self.add_module('pool', nn.AvgPool2d(kernel_size=2, stride=2)) + + +class DenseNet(nn.Module): + """Densely connected network. + + Reference: + Huang et al. Densely Connected Convolutional Networks. CVPR 2017. + + Public keys: + - ``densenet121``: DenseNet121. + - ``densenet169``: DenseNet169. + - ``densenet201``: DenseNet201. + - ``densenet161``: DenseNet161. + - ``densenet121_fc512``: DenseNet121 + FC. + """ + + def __init__( + self, + num_classes, + loss, + growth_rate=32, + block_config=(6, 12, 24, 16), + num_init_features=64, + bn_size=4, + drop_rate=0, + fc_dims=None, + dropout_p=None, + **kwargs + ): + + super(DenseNet, self).__init__() + self.loss = loss + + # First convolution + self.features = nn.Sequential( + OrderedDict( + [ + ( + 'conv0', + nn.Conv2d( + 3, + num_init_features, + kernel_size=7, + stride=2, + padding=3, + bias=False + ) + ), + ('norm0', nn.BatchNorm2d(num_init_features)), + ('relu0', nn.ReLU(inplace=True)), + ( + 'pool0', + nn.MaxPool2d(kernel_size=3, stride=2, padding=1) + ), + ] + ) + ) + + # Each denseblock + num_features = num_init_features + for i, num_layers in enumerate(block_config): + block = _DenseBlock( + num_layers=num_layers, + num_input_features=num_features, + bn_size=bn_size, + growth_rate=growth_rate, + drop_rate=drop_rate + ) + self.features.add_module('denseblock%d' % (i+1), block) + num_features = num_features + num_layers*growth_rate + if i != len(block_config) - 1: + trans = _Transition( + num_input_features=num_features, + num_output_features=num_features // 2 + ) + self.features.add_module('transition%d' % (i+1), trans) + num_features = num_features // 2 + + # Final batch norm + self.features.add_module('norm5', nn.BatchNorm2d(num_features)) + + self.global_avgpool = nn.AdaptiveAvgPool2d(1) + self.feature_dim = num_features + self.fc = self._construct_fc_layer(fc_dims, num_features, dropout_p) + + # Linear layer + self.classifier = nn.Linear(self.feature_dim, num_classes) + + self._init_params() + + def _construct_fc_layer(self, fc_dims, input_dim, dropout_p=None): + """Constructs fully connected layer. + + Args: + fc_dims (list or tuple): dimensions of fc layers, if None, no fc layers are constructed + input_dim (int): input dimension + dropout_p (float): dropout probability, if None, dropout is unused + """ + if fc_dims is None: + self.feature_dim = input_dim + return None + + assert isinstance( + fc_dims, (list, tuple) + ), 'fc_dims must be either list or tuple, but got {}'.format( + type(fc_dims) + ) + + layers = [] + for dim in fc_dims: + layers.append(nn.Linear(input_dim, dim)) + layers.append(nn.BatchNorm1d(dim)) + layers.append(nn.ReLU(inplace=True)) + if dropout_p is not None: + layers.append(nn.Dropout(p=dropout_p)) + input_dim = dim + + self.feature_dim = fc_dims[-1] + + return nn.Sequential(*layers) + + def _init_params(self): + for m in self.modules(): + if isinstance(m, nn.Conv2d): + nn.init.kaiming_normal_( + m.weight, mode='fan_out', nonlinearity='relu' + ) + if m.bias is not None: + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.BatchNorm2d): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.BatchNorm1d): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.Linear): + nn.init.normal_(m.weight, 0, 0.01) + if m.bias is not None: + nn.init.constant_(m.bias, 0) + + def forward(self, x): + f = self.features(x) + f = F.relu(f, inplace=True) + v = self.global_avgpool(f) + v = v.view(v.size(0), -1) + + if self.fc is not None: + v = self.fc(v) + + if not self.training: + return v + + y = self.classifier(v) + + if self.loss == 'softmax': + return y + elif self.loss == 'triplet': + return y, v + else: + raise KeyError('Unsupported loss: {}'.format(self.loss)) + + +def init_pretrained_weights(model, model_url): + """Initializes model with pretrained weights. + + Layers that don't match with pretrained layers in name or size are kept unchanged. + """ + pretrain_dict = model_zoo.load_url(model_url) + + # '.'s are no longer allowed in module names, but pervious _DenseLayer + # has keys 'norm.1', 'relu.1', 'conv.1', 'norm.2', 'relu.2', 'conv.2'. + # They are also in the checkpoints in model_urls. This pattern is used + # to find such keys. + pattern = re.compile( + r'^(.*denselayer\d+\.(?:norm|relu|conv))\.((?:[12])\.(?:weight|bias|running_mean|running_var))$' + ) + for key in list(pretrain_dict.keys()): + res = pattern.match(key) + if res: + new_key = res.group(1) + res.group(2) + pretrain_dict[new_key] = pretrain_dict[key] + del pretrain_dict[key] + + model_dict = model.state_dict() + pretrain_dict = { + k: v + for k, v in pretrain_dict.items() + if k in model_dict and model_dict[k].size() == v.size() + } + model_dict.update(pretrain_dict) + model.load_state_dict(model_dict) + + +""" +Dense network configurations: +-- +densenet121: num_init_features=64, growth_rate=32, block_config=(6, 12, 24, 16) +densenet169: num_init_features=64, growth_rate=32, block_config=(6, 12, 32, 32) +densenet201: num_init_features=64, growth_rate=32, block_config=(6, 12, 48, 32) +densenet161: num_init_features=96, growth_rate=48, block_config=(6, 12, 36, 24) +""" + + +def densenet121(num_classes, loss='softmax', pretrained=True, **kwargs): + model = DenseNet( + num_classes=num_classes, + loss=loss, + num_init_features=64, + growth_rate=32, + block_config=(6, 12, 24, 16), + fc_dims=None, + dropout_p=None, + **kwargs + ) + if pretrained: + init_pretrained_weights(model, model_urls['densenet121']) + return model + + +def densenet169(num_classes, loss='softmax', pretrained=True, **kwargs): + model = DenseNet( + num_classes=num_classes, + loss=loss, + num_init_features=64, + growth_rate=32, + block_config=(6, 12, 32, 32), + fc_dims=None, + dropout_p=None, + **kwargs + ) + if pretrained: + init_pretrained_weights(model, model_urls['densenet169']) + return model + + +def densenet201(num_classes, loss='softmax', pretrained=True, **kwargs): + model = DenseNet( + num_classes=num_classes, + loss=loss, + num_init_features=64, + growth_rate=32, + block_config=(6, 12, 48, 32), + fc_dims=None, + dropout_p=None, + **kwargs + ) + if pretrained: + init_pretrained_weights(model, model_urls['densenet201']) + return model + + +def densenet161(num_classes, loss='softmax', pretrained=True, **kwargs): + model = DenseNet( + num_classes=num_classes, + loss=loss, + num_init_features=96, + growth_rate=48, + block_config=(6, 12, 36, 24), + fc_dims=None, + dropout_p=None, + **kwargs + ) + if pretrained: + init_pretrained_weights(model, model_urls['densenet161']) + return model + + +def densenet121_fc512(num_classes, loss='softmax', pretrained=True, **kwargs): + model = DenseNet( + num_classes=num_classes, + loss=loss, + num_init_features=64, + growth_rate=32, + block_config=(6, 12, 24, 16), + fc_dims=[512], + dropout_p=None, + **kwargs + ) + if pretrained: + init_pretrained_weights(model, model_urls['densenet121']) + return model diff --git a/strong_sort/deep/reid/torchreid/models/hacnn.py b/strong_sort/deep/reid/torchreid/models/hacnn.py new file mode 100644 index 0000000..f21cc82 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/models/hacnn.py @@ -0,0 +1,414 @@ +from __future__ import division, absolute_import +import torch +from torch import nn +from torch.nn import functional as F + +__all__ = ['HACNN'] + + +class ConvBlock(nn.Module): + """Basic convolutional block. + + convolution + batch normalization + relu. + + Args: + in_c (int): number of input channels. + out_c (int): number of output channels. + k (int or tuple): kernel size. + s (int or tuple): stride. + p (int or tuple): padding. + """ + + def __init__(self, in_c, out_c, k, s=1, p=0): + super(ConvBlock, self).__init__() + self.conv = nn.Conv2d(in_c, out_c, k, stride=s, padding=p) + self.bn = nn.BatchNorm2d(out_c) + + def forward(self, x): + return F.relu(self.bn(self.conv(x))) + + +class InceptionA(nn.Module): + + def __init__(self, in_channels, out_channels): + super(InceptionA, self).__init__() + mid_channels = out_channels // 4 + + self.stream1 = nn.Sequential( + ConvBlock(in_channels, mid_channels, 1), + ConvBlock(mid_channels, mid_channels, 3, p=1), + ) + self.stream2 = nn.Sequential( + ConvBlock(in_channels, mid_channels, 1), + ConvBlock(mid_channels, mid_channels, 3, p=1), + ) + self.stream3 = nn.Sequential( + ConvBlock(in_channels, mid_channels, 1), + ConvBlock(mid_channels, mid_channels, 3, p=1), + ) + self.stream4 = nn.Sequential( + nn.AvgPool2d(3, stride=1, padding=1), + ConvBlock(in_channels, mid_channels, 1), + ) + + def forward(self, x): + s1 = self.stream1(x) + s2 = self.stream2(x) + s3 = self.stream3(x) + s4 = self.stream4(x) + y = torch.cat([s1, s2, s3, s4], dim=1) + return y + + +class InceptionB(nn.Module): + + def __init__(self, in_channels, out_channels): + super(InceptionB, self).__init__() + mid_channels = out_channels // 4 + + self.stream1 = nn.Sequential( + ConvBlock(in_channels, mid_channels, 1), + ConvBlock(mid_channels, mid_channels, 3, s=2, p=1), + ) + self.stream2 = nn.Sequential( + ConvBlock(in_channels, mid_channels, 1), + ConvBlock(mid_channels, mid_channels, 3, p=1), + ConvBlock(mid_channels, mid_channels, 3, s=2, p=1), + ) + self.stream3 = nn.Sequential( + nn.MaxPool2d(3, stride=2, padding=1), + ConvBlock(in_channels, mid_channels * 2, 1), + ) + + def forward(self, x): + s1 = self.stream1(x) + s2 = self.stream2(x) + s3 = self.stream3(x) + y = torch.cat([s1, s2, s3], dim=1) + return y + + +class SpatialAttn(nn.Module): + """Spatial Attention (Sec. 3.1.I.1)""" + + def __init__(self): + super(SpatialAttn, self).__init__() + self.conv1 = ConvBlock(1, 1, 3, s=2, p=1) + self.conv2 = ConvBlock(1, 1, 1) + + def forward(self, x): + # global cross-channel averaging + x = x.mean(1, keepdim=True) + # 3-by-3 conv + x = self.conv1(x) + # bilinear resizing + x = F.upsample( + x, (x.size(2) * 2, x.size(3) * 2), + mode='bilinear', + align_corners=True + ) + # scaling conv + x = self.conv2(x) + return x + + +class ChannelAttn(nn.Module): + """Channel Attention (Sec. 3.1.I.2)""" + + def __init__(self, in_channels, reduction_rate=16): + super(ChannelAttn, self).__init__() + assert in_channels % reduction_rate == 0 + self.conv1 = ConvBlock(in_channels, in_channels // reduction_rate, 1) + self.conv2 = ConvBlock(in_channels // reduction_rate, in_channels, 1) + + def forward(self, x): + # squeeze operation (global average pooling) + x = F.avg_pool2d(x, x.size()[2:]) + # excitation operation (2 conv layers) + x = self.conv1(x) + x = self.conv2(x) + return x + + +class SoftAttn(nn.Module): + """Soft Attention (Sec. 3.1.I) + + Aim: Spatial Attention + Channel Attention + + Output: attention maps with shape identical to input. + """ + + def __init__(self, in_channels): + super(SoftAttn, self).__init__() + self.spatial_attn = SpatialAttn() + self.channel_attn = ChannelAttn(in_channels) + self.conv = ConvBlock(in_channels, in_channels, 1) + + def forward(self, x): + y_spatial = self.spatial_attn(x) + y_channel = self.channel_attn(x) + y = y_spatial * y_channel + y = torch.sigmoid(self.conv(y)) + return y + + +class HardAttn(nn.Module): + """Hard Attention (Sec. 3.1.II)""" + + def __init__(self, in_channels): + super(HardAttn, self).__init__() + self.fc = nn.Linear(in_channels, 4 * 2) + self.init_params() + + def init_params(self): + self.fc.weight.data.zero_() + self.fc.bias.data.copy_( + torch.tensor( + [0, -0.75, 0, -0.25, 0, 0.25, 0, 0.75], dtype=torch.float + ) + ) + + def forward(self, x): + # squeeze operation (global average pooling) + x = F.avg_pool2d(x, x.size()[2:]).view(x.size(0), x.size(1)) + # predict transformation parameters + theta = torch.tanh(self.fc(x)) + theta = theta.view(-1, 4, 2) + return theta + + +class HarmAttn(nn.Module): + """Harmonious Attention (Sec. 3.1)""" + + def __init__(self, in_channels): + super(HarmAttn, self).__init__() + self.soft_attn = SoftAttn(in_channels) + self.hard_attn = HardAttn(in_channels) + + def forward(self, x): + y_soft_attn = self.soft_attn(x) + theta = self.hard_attn(x) + return y_soft_attn, theta + + +class HACNN(nn.Module): + """Harmonious Attention Convolutional Neural Network. + + Reference: + Li et al. Harmonious Attention Network for Person Re-identification. CVPR 2018. + + Public keys: + - ``hacnn``: HACNN. + """ + + # Args: + # num_classes (int): number of classes to predict + # nchannels (list): number of channels AFTER concatenation + # feat_dim (int): feature dimension for a single stream + # learn_region (bool): whether to learn region features (i.e. local branch) + + def __init__( + self, + num_classes, + loss='softmax', + nchannels=[128, 256, 384], + feat_dim=512, + learn_region=True, + use_gpu=True, + **kwargs + ): + super(HACNN, self).__init__() + self.loss = loss + self.learn_region = learn_region + self.use_gpu = use_gpu + + self.conv = ConvBlock(3, 32, 3, s=2, p=1) + + # Construct Inception + HarmAttn blocks + # ============== Block 1 ============== + self.inception1 = nn.Sequential( + InceptionA(32, nchannels[0]), + InceptionB(nchannels[0], nchannels[0]), + ) + self.ha1 = HarmAttn(nchannels[0]) + + # ============== Block 2 ============== + self.inception2 = nn.Sequential( + InceptionA(nchannels[0], nchannels[1]), + InceptionB(nchannels[1], nchannels[1]), + ) + self.ha2 = HarmAttn(nchannels[1]) + + # ============== Block 3 ============== + self.inception3 = nn.Sequential( + InceptionA(nchannels[1], nchannels[2]), + InceptionB(nchannels[2], nchannels[2]), + ) + self.ha3 = HarmAttn(nchannels[2]) + + self.fc_global = nn.Sequential( + nn.Linear(nchannels[2], feat_dim), + nn.BatchNorm1d(feat_dim), + nn.ReLU(), + ) + self.classifier_global = nn.Linear(feat_dim, num_classes) + + if self.learn_region: + self.init_scale_factors() + self.local_conv1 = InceptionB(32, nchannels[0]) + self.local_conv2 = InceptionB(nchannels[0], nchannels[1]) + self.local_conv3 = InceptionB(nchannels[1], nchannels[2]) + self.fc_local = nn.Sequential( + nn.Linear(nchannels[2] * 4, feat_dim), + nn.BatchNorm1d(feat_dim), + nn.ReLU(), + ) + self.classifier_local = nn.Linear(feat_dim, num_classes) + self.feat_dim = feat_dim * 2 + else: + self.feat_dim = feat_dim + + def init_scale_factors(self): + # initialize scale factors (s_w, s_h) for four regions + self.scale_factors = [] + self.scale_factors.append( + torch.tensor([[1, 0], [0, 0.25]], dtype=torch.float) + ) + self.scale_factors.append( + torch.tensor([[1, 0], [0, 0.25]], dtype=torch.float) + ) + self.scale_factors.append( + torch.tensor([[1, 0], [0, 0.25]], dtype=torch.float) + ) + self.scale_factors.append( + torch.tensor([[1, 0], [0, 0.25]], dtype=torch.float) + ) + + def stn(self, x, theta): + """Performs spatial transform + + x: (batch, channel, height, width) + theta: (batch, 2, 3) + """ + grid = F.affine_grid(theta, x.size()) + x = F.grid_sample(x, grid) + return x + + def transform_theta(self, theta_i, region_idx): + """Transforms theta to include (s_w, s_h), resulting in (batch, 2, 3)""" + scale_factors = self.scale_factors[region_idx] + theta = torch.zeros(theta_i.size(0), 2, 3) + theta[:, :, :2] = scale_factors + theta[:, :, -1] = theta_i + if self.use_gpu: + theta = theta.cuda() + return theta + + def forward(self, x): + assert x.size(2) == 160 and x.size(3) == 64, \ + 'Input size does not match, expected (160, 64) but got ({}, {})'.format(x.size(2), x.size(3)) + x = self.conv(x) + + # ============== Block 1 ============== + # global branch + x1 = self.inception1(x) + x1_attn, x1_theta = self.ha1(x1) + x1_out = x1 * x1_attn + # local branch + if self.learn_region: + x1_local_list = [] + for region_idx in range(4): + x1_theta_i = x1_theta[:, region_idx, :] + x1_theta_i = self.transform_theta(x1_theta_i, region_idx) + x1_trans_i = self.stn(x, x1_theta_i) + x1_trans_i = F.upsample( + x1_trans_i, (24, 28), mode='bilinear', align_corners=True + ) + x1_local_i = self.local_conv1(x1_trans_i) + x1_local_list.append(x1_local_i) + + # ============== Block 2 ============== + # Block 2 + # global branch + x2 = self.inception2(x1_out) + x2_attn, x2_theta = self.ha2(x2) + x2_out = x2 * x2_attn + # local branch + if self.learn_region: + x2_local_list = [] + for region_idx in range(4): + x2_theta_i = x2_theta[:, region_idx, :] + x2_theta_i = self.transform_theta(x2_theta_i, region_idx) + x2_trans_i = self.stn(x1_out, x2_theta_i) + x2_trans_i = F.upsample( + x2_trans_i, (12, 14), mode='bilinear', align_corners=True + ) + x2_local_i = x2_trans_i + x1_local_list[region_idx] + x2_local_i = self.local_conv2(x2_local_i) + x2_local_list.append(x2_local_i) + + # ============== Block 3 ============== + # Block 3 + # global branch + x3 = self.inception3(x2_out) + x3_attn, x3_theta = self.ha3(x3) + x3_out = x3 * x3_attn + # local branch + if self.learn_region: + x3_local_list = [] + for region_idx in range(4): + x3_theta_i = x3_theta[:, region_idx, :] + x3_theta_i = self.transform_theta(x3_theta_i, region_idx) + x3_trans_i = self.stn(x2_out, x3_theta_i) + x3_trans_i = F.upsample( + x3_trans_i, (6, 7), mode='bilinear', align_corners=True + ) + x3_local_i = x3_trans_i + x2_local_list[region_idx] + x3_local_i = self.local_conv3(x3_local_i) + x3_local_list.append(x3_local_i) + + # ============== Feature generation ============== + # global branch + x_global = F.avg_pool2d(x3_out, + x3_out.size()[2:] + ).view(x3_out.size(0), x3_out.size(1)) + x_global = self.fc_global(x_global) + # local branch + if self.learn_region: + x_local_list = [] + for region_idx in range(4): + x_local_i = x3_local_list[region_idx] + x_local_i = F.avg_pool2d(x_local_i, + x_local_i.size()[2:] + ).view(x_local_i.size(0), -1) + x_local_list.append(x_local_i) + x_local = torch.cat(x_local_list, 1) + x_local = self.fc_local(x_local) + + if not self.training: + # l2 normalization before concatenation + if self.learn_region: + x_global = x_global / x_global.norm(p=2, dim=1, keepdim=True) + x_local = x_local / x_local.norm(p=2, dim=1, keepdim=True) + return torch.cat([x_global, x_local], 1) + else: + return x_global + + prelogits_global = self.classifier_global(x_global) + if self.learn_region: + prelogits_local = self.classifier_local(x_local) + + if self.loss == 'softmax': + if self.learn_region: + return (prelogits_global, prelogits_local) + else: + return prelogits_global + + elif self.loss == 'triplet': + if self.learn_region: + return (prelogits_global, prelogits_local), (x_global, x_local) + else: + return prelogits_global, x_global + + else: + raise KeyError("Unsupported loss: {}".format(self.loss)) diff --git a/strong_sort/deep/reid/torchreid/models/inceptionresnetv2.py b/strong_sort/deep/reid/torchreid/models/inceptionresnetv2.py new file mode 100644 index 0000000..03e4034 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/models/inceptionresnetv2.py @@ -0,0 +1,361 @@ +""" +Code imported from https://github.com/Cadene/pretrained-models.pytorch +""" +from __future__ import division, absolute_import +import torch +import torch.nn as nn +import torch.utils.model_zoo as model_zoo + +__all__ = ['inceptionresnetv2'] + +pretrained_settings = { + 'inceptionresnetv2': { + 'imagenet': { + 'url': + 'http://data.lip6.fr/cadene/pretrainedmodels/inceptionresnetv2-520b38e4.pth', + 'input_space': 'RGB', + 'input_size': [3, 299, 299], + 'input_range': [0, 1], + 'mean': [0.5, 0.5, 0.5], + 'std': [0.5, 0.5, 0.5], + 'num_classes': 1000 + }, + 'imagenet+background': { + 'url': + 'http://data.lip6.fr/cadene/pretrainedmodels/inceptionresnetv2-520b38e4.pth', + 'input_space': 'RGB', + 'input_size': [3, 299, 299], + 'input_range': [0, 1], + 'mean': [0.5, 0.5, 0.5], + 'std': [0.5, 0.5, 0.5], + 'num_classes': 1001 + } + } +} + + +class BasicConv2d(nn.Module): + + def __init__(self, in_planes, out_planes, kernel_size, stride, padding=0): + super(BasicConv2d, self).__init__() + self.conv = nn.Conv2d( + in_planes, + out_planes, + kernel_size=kernel_size, + stride=stride, + padding=padding, + bias=False + ) # verify bias false + self.bn = nn.BatchNorm2d( + out_planes, + eps=0.001, # value found in tensorflow + momentum=0.1, # default pytorch value + affine=True + ) + self.relu = nn.ReLU(inplace=False) + + def forward(self, x): + x = self.conv(x) + x = self.bn(x) + x = self.relu(x) + return x + + +class Mixed_5b(nn.Module): + + def __init__(self): + super(Mixed_5b, self).__init__() + + self.branch0 = BasicConv2d(192, 96, kernel_size=1, stride=1) + + self.branch1 = nn.Sequential( + BasicConv2d(192, 48, kernel_size=1, stride=1), + BasicConv2d(48, 64, kernel_size=5, stride=1, padding=2) + ) + + self.branch2 = nn.Sequential( + BasicConv2d(192, 64, kernel_size=1, stride=1), + BasicConv2d(64, 96, kernel_size=3, stride=1, padding=1), + BasicConv2d(96, 96, kernel_size=3, stride=1, padding=1) + ) + + self.branch3 = nn.Sequential( + nn.AvgPool2d(3, stride=1, padding=1, count_include_pad=False), + BasicConv2d(192, 64, kernel_size=1, stride=1) + ) + + def forward(self, x): + x0 = self.branch0(x) + x1 = self.branch1(x) + x2 = self.branch2(x) + x3 = self.branch3(x) + out = torch.cat((x0, x1, x2, x3), 1) + return out + + +class Block35(nn.Module): + + def __init__(self, scale=1.0): + super(Block35, self).__init__() + + self.scale = scale + + self.branch0 = BasicConv2d(320, 32, kernel_size=1, stride=1) + + self.branch1 = nn.Sequential( + BasicConv2d(320, 32, kernel_size=1, stride=1), + BasicConv2d(32, 32, kernel_size=3, stride=1, padding=1) + ) + + self.branch2 = nn.Sequential( + BasicConv2d(320, 32, kernel_size=1, stride=1), + BasicConv2d(32, 48, kernel_size=3, stride=1, padding=1), + BasicConv2d(48, 64, kernel_size=3, stride=1, padding=1) + ) + + self.conv2d = nn.Conv2d(128, 320, kernel_size=1, stride=1) + self.relu = nn.ReLU(inplace=False) + + def forward(self, x): + x0 = self.branch0(x) + x1 = self.branch1(x) + x2 = self.branch2(x) + out = torch.cat((x0, x1, x2), 1) + out = self.conv2d(out) + out = out * self.scale + x + out = self.relu(out) + return out + + +class Mixed_6a(nn.Module): + + def __init__(self): + super(Mixed_6a, self).__init__() + + self.branch0 = BasicConv2d(320, 384, kernel_size=3, stride=2) + + self.branch1 = nn.Sequential( + BasicConv2d(320, 256, kernel_size=1, stride=1), + BasicConv2d(256, 256, kernel_size=3, stride=1, padding=1), + BasicConv2d(256, 384, kernel_size=3, stride=2) + ) + + self.branch2 = nn.MaxPool2d(3, stride=2) + + def forward(self, x): + x0 = self.branch0(x) + x1 = self.branch1(x) + x2 = self.branch2(x) + out = torch.cat((x0, x1, x2), 1) + return out + + +class Block17(nn.Module): + + def __init__(self, scale=1.0): + super(Block17, self).__init__() + + self.scale = scale + + self.branch0 = BasicConv2d(1088, 192, kernel_size=1, stride=1) + + self.branch1 = nn.Sequential( + BasicConv2d(1088, 128, kernel_size=1, stride=1), + BasicConv2d( + 128, 160, kernel_size=(1, 7), stride=1, padding=(0, 3) + ), + BasicConv2d( + 160, 192, kernel_size=(7, 1), stride=1, padding=(3, 0) + ) + ) + + self.conv2d = nn.Conv2d(384, 1088, kernel_size=1, stride=1) + self.relu = nn.ReLU(inplace=False) + + def forward(self, x): + x0 = self.branch0(x) + x1 = self.branch1(x) + out = torch.cat((x0, x1), 1) + out = self.conv2d(out) + out = out * self.scale + x + out = self.relu(out) + return out + + +class Mixed_7a(nn.Module): + + def __init__(self): + super(Mixed_7a, self).__init__() + + self.branch0 = nn.Sequential( + BasicConv2d(1088, 256, kernel_size=1, stride=1), + BasicConv2d(256, 384, kernel_size=3, stride=2) + ) + + self.branch1 = nn.Sequential( + BasicConv2d(1088, 256, kernel_size=1, stride=1), + BasicConv2d(256, 288, kernel_size=3, stride=2) + ) + + self.branch2 = nn.Sequential( + BasicConv2d(1088, 256, kernel_size=1, stride=1), + BasicConv2d(256, 288, kernel_size=3, stride=1, padding=1), + BasicConv2d(288, 320, kernel_size=3, stride=2) + ) + + self.branch3 = nn.MaxPool2d(3, stride=2) + + def forward(self, x): + x0 = self.branch0(x) + x1 = self.branch1(x) + x2 = self.branch2(x) + x3 = self.branch3(x) + out = torch.cat((x0, x1, x2, x3), 1) + return out + + +class Block8(nn.Module): + + def __init__(self, scale=1.0, noReLU=False): + super(Block8, self).__init__() + + self.scale = scale + self.noReLU = noReLU + + self.branch0 = BasicConv2d(2080, 192, kernel_size=1, stride=1) + + self.branch1 = nn.Sequential( + BasicConv2d(2080, 192, kernel_size=1, stride=1), + BasicConv2d( + 192, 224, kernel_size=(1, 3), stride=1, padding=(0, 1) + ), + BasicConv2d( + 224, 256, kernel_size=(3, 1), stride=1, padding=(1, 0) + ) + ) + + self.conv2d = nn.Conv2d(448, 2080, kernel_size=1, stride=1) + if not self.noReLU: + self.relu = nn.ReLU(inplace=False) + + def forward(self, x): + x0 = self.branch0(x) + x1 = self.branch1(x) + out = torch.cat((x0, x1), 1) + out = self.conv2d(out) + out = out * self.scale + x + if not self.noReLU: + out = self.relu(out) + return out + + +# ---------------- +# Model Definition +# ---------------- +class InceptionResNetV2(nn.Module): + """Inception-ResNet-V2. + + Reference: + Szegedy et al. Inception-v4, Inception-ResNet and the Impact of Residual + Connections on Learning. AAAI 2017. + + Public keys: + - ``inceptionresnetv2``: Inception-ResNet-V2. + """ + + def __init__(self, num_classes, loss='softmax', **kwargs): + super(InceptionResNetV2, self).__init__() + self.loss = loss + + # Modules + self.conv2d_1a = BasicConv2d(3, 32, kernel_size=3, stride=2) + self.conv2d_2a = BasicConv2d(32, 32, kernel_size=3, stride=1) + self.conv2d_2b = BasicConv2d( + 32, 64, kernel_size=3, stride=1, padding=1 + ) + self.maxpool_3a = nn.MaxPool2d(3, stride=2) + self.conv2d_3b = BasicConv2d(64, 80, kernel_size=1, stride=1) + self.conv2d_4a = BasicConv2d(80, 192, kernel_size=3, stride=1) + self.maxpool_5a = nn.MaxPool2d(3, stride=2) + self.mixed_5b = Mixed_5b() + self.repeat = nn.Sequential( + Block35(scale=0.17), Block35(scale=0.17), Block35(scale=0.17), + Block35(scale=0.17), Block35(scale=0.17), Block35(scale=0.17), + Block35(scale=0.17), Block35(scale=0.17), Block35(scale=0.17), + Block35(scale=0.17) + ) + self.mixed_6a = Mixed_6a() + self.repeat_1 = nn.Sequential( + Block17(scale=0.10), Block17(scale=0.10), Block17(scale=0.10), + Block17(scale=0.10), Block17(scale=0.10), Block17(scale=0.10), + Block17(scale=0.10), Block17(scale=0.10), Block17(scale=0.10), + Block17(scale=0.10), Block17(scale=0.10), Block17(scale=0.10), + Block17(scale=0.10), Block17(scale=0.10), Block17(scale=0.10), + Block17(scale=0.10), Block17(scale=0.10), Block17(scale=0.10), + Block17(scale=0.10), Block17(scale=0.10) + ) + self.mixed_7a = Mixed_7a() + self.repeat_2 = nn.Sequential( + Block8(scale=0.20), Block8(scale=0.20), Block8(scale=0.20), + Block8(scale=0.20), Block8(scale=0.20), Block8(scale=0.20), + Block8(scale=0.20), Block8(scale=0.20), Block8(scale=0.20) + ) + + self.block8 = Block8(noReLU=True) + self.conv2d_7b = BasicConv2d(2080, 1536, kernel_size=1, stride=1) + self.global_avgpool = nn.AdaptiveAvgPool2d(1) + self.classifier = nn.Linear(1536, num_classes) + + def load_imagenet_weights(self): + settings = pretrained_settings['inceptionresnetv2']['imagenet'] + pretrain_dict = model_zoo.load_url(settings['url']) + model_dict = self.state_dict() + pretrain_dict = { + k: v + for k, v in pretrain_dict.items() + if k in model_dict and model_dict[k].size() == v.size() + } + model_dict.update(pretrain_dict) + self.load_state_dict(model_dict) + + def featuremaps(self, x): + x = self.conv2d_1a(x) + x = self.conv2d_2a(x) + x = self.conv2d_2b(x) + x = self.maxpool_3a(x) + x = self.conv2d_3b(x) + x = self.conv2d_4a(x) + x = self.maxpool_5a(x) + x = self.mixed_5b(x) + x = self.repeat(x) + x = self.mixed_6a(x) + x = self.repeat_1(x) + x = self.mixed_7a(x) + x = self.repeat_2(x) + x = self.block8(x) + x = self.conv2d_7b(x) + return x + + def forward(self, x): + f = self.featuremaps(x) + v = self.global_avgpool(f) + v = v.view(v.size(0), -1) + + if not self.training: + return v + + y = self.classifier(v) + + if self.loss == 'softmax': + return y + elif self.loss == 'triplet': + return y, v + else: + raise KeyError('Unsupported loss: {}'.format(self.loss)) + + +def inceptionresnetv2(num_classes, loss='softmax', pretrained=True, **kwargs): + model = InceptionResNetV2(num_classes=num_classes, loss=loss, **kwargs) + if pretrained: + model.load_imagenet_weights() + return model diff --git a/strong_sort/deep/reid/torchreid/models/inceptionv4.py b/strong_sort/deep/reid/torchreid/models/inceptionv4.py new file mode 100644 index 0000000..b14916f --- /dev/null +++ b/strong_sort/deep/reid/torchreid/models/inceptionv4.py @@ -0,0 +1,381 @@ +from __future__ import division, absolute_import +import torch +import torch.nn as nn +import torch.utils.model_zoo as model_zoo + +__all__ = ['inceptionv4'] +""" +Code imported from https://github.com/Cadene/pretrained-models.pytorch +""" + +pretrained_settings = { + 'inceptionv4': { + 'imagenet': { + 'url': + 'http://data.lip6.fr/cadene/pretrainedmodels/inceptionv4-8e4777a0.pth', + 'input_space': 'RGB', + 'input_size': [3, 299, 299], + 'input_range': [0, 1], + 'mean': [0.5, 0.5, 0.5], + 'std': [0.5, 0.5, 0.5], + 'num_classes': 1000 + }, + 'imagenet+background': { + 'url': + 'http://data.lip6.fr/cadene/pretrainedmodels/inceptionv4-8e4777a0.pth', + 'input_space': 'RGB', + 'input_size': [3, 299, 299], + 'input_range': [0, 1], + 'mean': [0.5, 0.5, 0.5], + 'std': [0.5, 0.5, 0.5], + 'num_classes': 1001 + } + } +} + + +class BasicConv2d(nn.Module): + + def __init__(self, in_planes, out_planes, kernel_size, stride, padding=0): + super(BasicConv2d, self).__init__() + self.conv = nn.Conv2d( + in_planes, + out_planes, + kernel_size=kernel_size, + stride=stride, + padding=padding, + bias=False + ) # verify bias false + self.bn = nn.BatchNorm2d( + out_planes, + eps=0.001, # value found in tensorflow + momentum=0.1, # default pytorch value + affine=True + ) + self.relu = nn.ReLU(inplace=True) + + def forward(self, x): + x = self.conv(x) + x = self.bn(x) + x = self.relu(x) + return x + + +class Mixed_3a(nn.Module): + + def __init__(self): + super(Mixed_3a, self).__init__() + self.maxpool = nn.MaxPool2d(3, stride=2) + self.conv = BasicConv2d(64, 96, kernel_size=3, stride=2) + + def forward(self, x): + x0 = self.maxpool(x) + x1 = self.conv(x) + out = torch.cat((x0, x1), 1) + return out + + +class Mixed_4a(nn.Module): + + def __init__(self): + super(Mixed_4a, self).__init__() + + self.branch0 = nn.Sequential( + BasicConv2d(160, 64, kernel_size=1, stride=1), + BasicConv2d(64, 96, kernel_size=3, stride=1) + ) + + self.branch1 = nn.Sequential( + BasicConv2d(160, 64, kernel_size=1, stride=1), + BasicConv2d(64, 64, kernel_size=(1, 7), stride=1, padding=(0, 3)), + BasicConv2d(64, 64, kernel_size=(7, 1), stride=1, padding=(3, 0)), + BasicConv2d(64, 96, kernel_size=(3, 3), stride=1) + ) + + def forward(self, x): + x0 = self.branch0(x) + x1 = self.branch1(x) + out = torch.cat((x0, x1), 1) + return out + + +class Mixed_5a(nn.Module): + + def __init__(self): + super(Mixed_5a, self).__init__() + self.conv = BasicConv2d(192, 192, kernel_size=3, stride=2) + self.maxpool = nn.MaxPool2d(3, stride=2) + + def forward(self, x): + x0 = self.conv(x) + x1 = self.maxpool(x) + out = torch.cat((x0, x1), 1) + return out + + +class Inception_A(nn.Module): + + def __init__(self): + super(Inception_A, self).__init__() + self.branch0 = BasicConv2d(384, 96, kernel_size=1, stride=1) + + self.branch1 = nn.Sequential( + BasicConv2d(384, 64, kernel_size=1, stride=1), + BasicConv2d(64, 96, kernel_size=3, stride=1, padding=1) + ) + + self.branch2 = nn.Sequential( + BasicConv2d(384, 64, kernel_size=1, stride=1), + BasicConv2d(64, 96, kernel_size=3, stride=1, padding=1), + BasicConv2d(96, 96, kernel_size=3, stride=1, padding=1) + ) + + self.branch3 = nn.Sequential( + nn.AvgPool2d(3, stride=1, padding=1, count_include_pad=False), + BasicConv2d(384, 96, kernel_size=1, stride=1) + ) + + def forward(self, x): + x0 = self.branch0(x) + x1 = self.branch1(x) + x2 = self.branch2(x) + x3 = self.branch3(x) + out = torch.cat((x0, x1, x2, x3), 1) + return out + + +class Reduction_A(nn.Module): + + def __init__(self): + super(Reduction_A, self).__init__() + self.branch0 = BasicConv2d(384, 384, kernel_size=3, stride=2) + + self.branch1 = nn.Sequential( + BasicConv2d(384, 192, kernel_size=1, stride=1), + BasicConv2d(192, 224, kernel_size=3, stride=1, padding=1), + BasicConv2d(224, 256, kernel_size=3, stride=2) + ) + + self.branch2 = nn.MaxPool2d(3, stride=2) + + def forward(self, x): + x0 = self.branch0(x) + x1 = self.branch1(x) + x2 = self.branch2(x) + out = torch.cat((x0, x1, x2), 1) + return out + + +class Inception_B(nn.Module): + + def __init__(self): + super(Inception_B, self).__init__() + self.branch0 = BasicConv2d(1024, 384, kernel_size=1, stride=1) + + self.branch1 = nn.Sequential( + BasicConv2d(1024, 192, kernel_size=1, stride=1), + BasicConv2d( + 192, 224, kernel_size=(1, 7), stride=1, padding=(0, 3) + ), + BasicConv2d( + 224, 256, kernel_size=(7, 1), stride=1, padding=(3, 0) + ) + ) + + self.branch2 = nn.Sequential( + BasicConv2d(1024, 192, kernel_size=1, stride=1), + BasicConv2d( + 192, 192, kernel_size=(7, 1), stride=1, padding=(3, 0) + ), + BasicConv2d( + 192, 224, kernel_size=(1, 7), stride=1, padding=(0, 3) + ), + BasicConv2d( + 224, 224, kernel_size=(7, 1), stride=1, padding=(3, 0) + ), + BasicConv2d( + 224, 256, kernel_size=(1, 7), stride=1, padding=(0, 3) + ) + ) + + self.branch3 = nn.Sequential( + nn.AvgPool2d(3, stride=1, padding=1, count_include_pad=False), + BasicConv2d(1024, 128, kernel_size=1, stride=1) + ) + + def forward(self, x): + x0 = self.branch0(x) + x1 = self.branch1(x) + x2 = self.branch2(x) + x3 = self.branch3(x) + out = torch.cat((x0, x1, x2, x3), 1) + return out + + +class Reduction_B(nn.Module): + + def __init__(self): + super(Reduction_B, self).__init__() + + self.branch0 = nn.Sequential( + BasicConv2d(1024, 192, kernel_size=1, stride=1), + BasicConv2d(192, 192, kernel_size=3, stride=2) + ) + + self.branch1 = nn.Sequential( + BasicConv2d(1024, 256, kernel_size=1, stride=1), + BasicConv2d( + 256, 256, kernel_size=(1, 7), stride=1, padding=(0, 3) + ), + BasicConv2d( + 256, 320, kernel_size=(7, 1), stride=1, padding=(3, 0) + ), BasicConv2d(320, 320, kernel_size=3, stride=2) + ) + + self.branch2 = nn.MaxPool2d(3, stride=2) + + def forward(self, x): + x0 = self.branch0(x) + x1 = self.branch1(x) + x2 = self.branch2(x) + out = torch.cat((x0, x1, x2), 1) + return out + + +class Inception_C(nn.Module): + + def __init__(self): + super(Inception_C, self).__init__() + + self.branch0 = BasicConv2d(1536, 256, kernel_size=1, stride=1) + + self.branch1_0 = BasicConv2d(1536, 384, kernel_size=1, stride=1) + self.branch1_1a = BasicConv2d( + 384, 256, kernel_size=(1, 3), stride=1, padding=(0, 1) + ) + self.branch1_1b = BasicConv2d( + 384, 256, kernel_size=(3, 1), stride=1, padding=(1, 0) + ) + + self.branch2_0 = BasicConv2d(1536, 384, kernel_size=1, stride=1) + self.branch2_1 = BasicConv2d( + 384, 448, kernel_size=(3, 1), stride=1, padding=(1, 0) + ) + self.branch2_2 = BasicConv2d( + 448, 512, kernel_size=(1, 3), stride=1, padding=(0, 1) + ) + self.branch2_3a = BasicConv2d( + 512, 256, kernel_size=(1, 3), stride=1, padding=(0, 1) + ) + self.branch2_3b = BasicConv2d( + 512, 256, kernel_size=(3, 1), stride=1, padding=(1, 0) + ) + + self.branch3 = nn.Sequential( + nn.AvgPool2d(3, stride=1, padding=1, count_include_pad=False), + BasicConv2d(1536, 256, kernel_size=1, stride=1) + ) + + def forward(self, x): + x0 = self.branch0(x) + + x1_0 = self.branch1_0(x) + x1_1a = self.branch1_1a(x1_0) + x1_1b = self.branch1_1b(x1_0) + x1 = torch.cat((x1_1a, x1_1b), 1) + + x2_0 = self.branch2_0(x) + x2_1 = self.branch2_1(x2_0) + x2_2 = self.branch2_2(x2_1) + x2_3a = self.branch2_3a(x2_2) + x2_3b = self.branch2_3b(x2_2) + x2 = torch.cat((x2_3a, x2_3b), 1) + + x3 = self.branch3(x) + + out = torch.cat((x0, x1, x2, x3), 1) + return out + + +class InceptionV4(nn.Module): + """Inception-v4. + + Reference: + Szegedy et al. Inception-v4, Inception-ResNet and the Impact of Residual + Connections on Learning. AAAI 2017. + + Public keys: + - ``inceptionv4``: InceptionV4. + """ + + def __init__(self, num_classes, loss, **kwargs): + super(InceptionV4, self).__init__() + self.loss = loss + + self.features = nn.Sequential( + BasicConv2d(3, 32, kernel_size=3, stride=2), + BasicConv2d(32, 32, kernel_size=3, stride=1), + BasicConv2d(32, 64, kernel_size=3, stride=1, padding=1), + Mixed_3a(), + Mixed_4a(), + Mixed_5a(), + Inception_A(), + Inception_A(), + Inception_A(), + Inception_A(), + Reduction_A(), # Mixed_6a + Inception_B(), + Inception_B(), + Inception_B(), + Inception_B(), + Inception_B(), + Inception_B(), + Inception_B(), + Reduction_B(), # Mixed_7a + Inception_C(), + Inception_C(), + Inception_C() + ) + self.global_avgpool = nn.AdaptiveAvgPool2d(1) + self.classifier = nn.Linear(1536, num_classes) + + def forward(self, x): + f = self.features(x) + v = self.global_avgpool(f) + v = v.view(v.size(0), -1) + + if not self.training: + return v + + y = self.classifier(v) + + if self.loss == 'softmax': + return y + elif self.loss == 'triplet': + return y, v + else: + raise KeyError('Unsupported loss: {}'.format(self.loss)) + + +def init_pretrained_weights(model, model_url): + """Initializes model with pretrained weights. + + Layers that don't match with pretrained layers in name or size are kept unchanged. + """ + pretrain_dict = model_zoo.load_url(model_url) + model_dict = model.state_dict() + pretrain_dict = { + k: v + for k, v in pretrain_dict.items() + if k in model_dict and model_dict[k].size() == v.size() + } + model_dict.update(pretrain_dict) + model.load_state_dict(model_dict) + + +def inceptionv4(num_classes, loss='softmax', pretrained=True, **kwargs): + model = InceptionV4(num_classes, loss, **kwargs) + if pretrained: + model_url = pretrained_settings['inceptionv4']['imagenet']['url'] + init_pretrained_weights(model, model_url) + return model diff --git a/strong_sort/deep/reid/torchreid/models/mlfn.py b/strong_sort/deep/reid/torchreid/models/mlfn.py new file mode 100644 index 0000000..ac7e126 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/models/mlfn.py @@ -0,0 +1,269 @@ +from __future__ import division, absolute_import +import torch +import torch.utils.model_zoo as model_zoo +from torch import nn +from torch.nn import functional as F + +__all__ = ['mlfn'] + +model_urls = { + # training epoch = 5, top1 = 51.6 + 'imagenet': + 'https://mega.nz/#!YHxAhaxC!yu9E6zWl0x5zscSouTdbZu8gdFFytDdl-RAdD2DEfpk', +} + + +class MLFNBlock(nn.Module): + + def __init__( + self, in_channels, out_channels, stride, fsm_channels, groups=32 + ): + super(MLFNBlock, self).__init__() + self.groups = groups + mid_channels = out_channels // 2 + + # Factor Modules + self.fm_conv1 = nn.Conv2d(in_channels, mid_channels, 1, bias=False) + self.fm_bn1 = nn.BatchNorm2d(mid_channels) + self.fm_conv2 = nn.Conv2d( + mid_channels, + mid_channels, + 3, + stride=stride, + padding=1, + bias=False, + groups=self.groups + ) + self.fm_bn2 = nn.BatchNorm2d(mid_channels) + self.fm_conv3 = nn.Conv2d(mid_channels, out_channels, 1, bias=False) + self.fm_bn3 = nn.BatchNorm2d(out_channels) + + # Factor Selection Module + self.fsm = nn.Sequential( + nn.AdaptiveAvgPool2d(1), + nn.Conv2d(in_channels, fsm_channels[0], 1), + nn.BatchNorm2d(fsm_channels[0]), + nn.ReLU(inplace=True), + nn.Conv2d(fsm_channels[0], fsm_channels[1], 1), + nn.BatchNorm2d(fsm_channels[1]), + nn.ReLU(inplace=True), + nn.Conv2d(fsm_channels[1], self.groups, 1), + nn.BatchNorm2d(self.groups), + nn.Sigmoid(), + ) + + self.downsample = None + if in_channels != out_channels or stride > 1: + self.downsample = nn.Sequential( + nn.Conv2d( + in_channels, out_channels, 1, stride=stride, bias=False + ), + nn.BatchNorm2d(out_channels), + ) + + def forward(self, x): + residual = x + s = self.fsm(x) + + # reduce dimension + x = self.fm_conv1(x) + x = self.fm_bn1(x) + x = F.relu(x, inplace=True) + + # group convolution + x = self.fm_conv2(x) + x = self.fm_bn2(x) + x = F.relu(x, inplace=True) + + # factor selection + b, c = x.size(0), x.size(1) + n = c // self.groups + ss = s.repeat(1, n, 1, 1) # from (b, g, 1, 1) to (b, g*n=c, 1, 1) + ss = ss.view(b, n, self.groups, 1, 1) + ss = ss.permute(0, 2, 1, 3, 4).contiguous() + ss = ss.view(b, c, 1, 1) + x = ss * x + + # recover dimension + x = self.fm_conv3(x) + x = self.fm_bn3(x) + x = F.relu(x, inplace=True) + + if self.downsample is not None: + residual = self.downsample(residual) + + return F.relu(residual + x, inplace=True), s + + +class MLFN(nn.Module): + """Multi-Level Factorisation Net. + + Reference: + Chang et al. Multi-Level Factorisation Net for + Person Re-Identification. CVPR 2018. + + Public keys: + - ``mlfn``: MLFN (Multi-Level Factorisation Net). + """ + + def __init__( + self, + num_classes, + loss='softmax', + groups=32, + channels=[64, 256, 512, 1024, 2048], + embed_dim=1024, + **kwargs + ): + super(MLFN, self).__init__() + self.loss = loss + self.groups = groups + + # first convolutional layer + self.conv1 = nn.Conv2d(3, channels[0], 7, stride=2, padding=3) + self.bn1 = nn.BatchNorm2d(channels[0]) + self.maxpool = nn.MaxPool2d(3, stride=2, padding=1) + + # main body + self.feature = nn.ModuleList( + [ + # layer 1-3 + MLFNBlock(channels[0], channels[1], 1, [128, 64], self.groups), + MLFNBlock(channels[1], channels[1], 1, [128, 64], self.groups), + MLFNBlock(channels[1], channels[1], 1, [128, 64], self.groups), + # layer 4-7 + MLFNBlock( + channels[1], channels[2], 2, [256, 128], self.groups + ), + MLFNBlock( + channels[2], channels[2], 1, [256, 128], self.groups + ), + MLFNBlock( + channels[2], channels[2], 1, [256, 128], self.groups + ), + MLFNBlock( + channels[2], channels[2], 1, [256, 128], self.groups + ), + # layer 8-13 + MLFNBlock( + channels[2], channels[3], 2, [512, 128], self.groups + ), + MLFNBlock( + channels[3], channels[3], 1, [512, 128], self.groups + ), + MLFNBlock( + channels[3], channels[3], 1, [512, 128], self.groups + ), + MLFNBlock( + channels[3], channels[3], 1, [512, 128], self.groups + ), + MLFNBlock( + channels[3], channels[3], 1, [512, 128], self.groups + ), + MLFNBlock( + channels[3], channels[3], 1, [512, 128], self.groups + ), + # layer 14-16 + MLFNBlock( + channels[3], channels[4], 2, [512, 128], self.groups + ), + MLFNBlock( + channels[4], channels[4], 1, [512, 128], self.groups + ), + MLFNBlock( + channels[4], channels[4], 1, [512, 128], self.groups + ), + ] + ) + self.global_avgpool = nn.AdaptiveAvgPool2d(1) + + # projection functions + self.fc_x = nn.Sequential( + nn.Conv2d(channels[4], embed_dim, 1, bias=False), + nn.BatchNorm2d(embed_dim), + nn.ReLU(inplace=True), + ) + self.fc_s = nn.Sequential( + nn.Conv2d(self.groups * 16, embed_dim, 1, bias=False), + nn.BatchNorm2d(embed_dim), + nn.ReLU(inplace=True), + ) + + self.classifier = nn.Linear(embed_dim, num_classes) + + self.init_params() + + def init_params(self): + for m in self.modules(): + if isinstance(m, nn.Conv2d): + nn.init.kaiming_normal_( + m.weight, mode='fan_out', nonlinearity='relu' + ) + if m.bias is not None: + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.BatchNorm2d): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.Linear): + nn.init.normal_(m.weight, 0, 0.01) + if m.bias is not None: + nn.init.constant_(m.bias, 0) + + def forward(self, x): + x = self.conv1(x) + x = self.bn1(x) + x = F.relu(x, inplace=True) + x = self.maxpool(x) + + s_hat = [] + for block in self.feature: + x, s = block(x) + s_hat.append(s) + s_hat = torch.cat(s_hat, 1) + + x = self.global_avgpool(x) + x = self.fc_x(x) + s_hat = self.fc_s(s_hat) + + v = (x+s_hat) * 0.5 + v = v.view(v.size(0), -1) + + if not self.training: + return v + + y = self.classifier(v) + + if self.loss == 'softmax': + return y + elif self.loss == 'triplet': + return y, v + else: + raise KeyError('Unsupported loss: {}'.format(self.loss)) + + +def init_pretrained_weights(model, model_url): + """Initializes model with pretrained weights. + + Layers that don't match with pretrained layers in name or size are kept unchanged. + """ + pretrain_dict = model_zoo.load_url(model_url) + model_dict = model.state_dict() + pretrain_dict = { + k: v + for k, v in pretrain_dict.items() + if k in model_dict and model_dict[k].size() == v.size() + } + model_dict.update(pretrain_dict) + model.load_state_dict(model_dict) + + +def mlfn(num_classes, loss='softmax', pretrained=True, **kwargs): + model = MLFN(num_classes, loss, **kwargs) + if pretrained: + # init_pretrained_weights(model, model_urls['imagenet']) + import warnings + warnings.warn( + 'The imagenet pretrained weights need to be manually downloaded from {}' + .format(model_urls['imagenet']) + ) + return model diff --git a/strong_sort/deep/reid/torchreid/models/mobilenetv2.py b/strong_sort/deep/reid/torchreid/models/mobilenetv2.py new file mode 100644 index 0000000..c451ef8 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/models/mobilenetv2.py @@ -0,0 +1,274 @@ +from __future__ import division, absolute_import +import torch.utils.model_zoo as model_zoo +from torch import nn +from torch.nn import functional as F + +__all__ = ['mobilenetv2_x1_0', 'mobilenetv2_x1_4'] + +model_urls = { + # 1.0: top-1 71.3 + 'mobilenetv2_x1_0': + 'https://mega.nz/#!NKp2wAIA!1NH1pbNzY_M2hVk_hdsxNM1NUOWvvGPHhaNr-fASF6c', + # 1.4: top-1 73.9 + 'mobilenetv2_x1_4': + 'https://mega.nz/#!RGhgEIwS!xN2s2ZdyqI6vQ3EwgmRXLEW3khr9tpXg96G9SUJugGk', +} + + +class ConvBlock(nn.Module): + """Basic convolutional block. + + convolution (bias discarded) + batch normalization + relu6. + + Args: + in_c (int): number of input channels. + out_c (int): number of output channels. + k (int or tuple): kernel size. + s (int or tuple): stride. + p (int or tuple): padding. + g (int): number of blocked connections from input channels + to output channels (default: 1). + """ + + def __init__(self, in_c, out_c, k, s=1, p=0, g=1): + super(ConvBlock, self).__init__() + self.conv = nn.Conv2d( + in_c, out_c, k, stride=s, padding=p, bias=False, groups=g + ) + self.bn = nn.BatchNorm2d(out_c) + + def forward(self, x): + return F.relu6(self.bn(self.conv(x))) + + +class Bottleneck(nn.Module): + + def __init__(self, in_channels, out_channels, expansion_factor, stride=1): + super(Bottleneck, self).__init__() + mid_channels = in_channels * expansion_factor + self.use_residual = stride == 1 and in_channels == out_channels + self.conv1 = ConvBlock(in_channels, mid_channels, 1) + self.dwconv2 = ConvBlock( + mid_channels, mid_channels, 3, stride, 1, g=mid_channels + ) + self.conv3 = nn.Sequential( + nn.Conv2d(mid_channels, out_channels, 1, bias=False), + nn.BatchNorm2d(out_channels), + ) + + def forward(self, x): + m = self.conv1(x) + m = self.dwconv2(m) + m = self.conv3(m) + if self.use_residual: + return x + m + else: + return m + + +class MobileNetV2(nn.Module): + """MobileNetV2. + + Reference: + Sandler et al. MobileNetV2: Inverted Residuals and + Linear Bottlenecks. CVPR 2018. + + Public keys: + - ``mobilenetv2_x1_0``: MobileNetV2 x1.0. + - ``mobilenetv2_x1_4``: MobileNetV2 x1.4. + """ + + def __init__( + self, + num_classes, + width_mult=1, + loss='softmax', + fc_dims=None, + dropout_p=None, + **kwargs + ): + super(MobileNetV2, self).__init__() + self.loss = loss + self.in_channels = int(32 * width_mult) + self.feature_dim = int(1280 * width_mult) if width_mult > 1 else 1280 + + # construct layers + self.conv1 = ConvBlock(3, self.in_channels, 3, s=2, p=1) + self.conv2 = self._make_layer( + Bottleneck, 1, int(16 * width_mult), 1, 1 + ) + self.conv3 = self._make_layer( + Bottleneck, 6, int(24 * width_mult), 2, 2 + ) + self.conv4 = self._make_layer( + Bottleneck, 6, int(32 * width_mult), 3, 2 + ) + self.conv5 = self._make_layer( + Bottleneck, 6, int(64 * width_mult), 4, 2 + ) + self.conv6 = self._make_layer( + Bottleneck, 6, int(96 * width_mult), 3, 1 + ) + self.conv7 = self._make_layer( + Bottleneck, 6, int(160 * width_mult), 3, 2 + ) + self.conv8 = self._make_layer( + Bottleneck, 6, int(320 * width_mult), 1, 1 + ) + self.conv9 = ConvBlock(self.in_channels, self.feature_dim, 1) + + self.global_avgpool = nn.AdaptiveAvgPool2d(1) + self.fc = self._construct_fc_layer( + fc_dims, self.feature_dim, dropout_p + ) + self.classifier = nn.Linear(self.feature_dim, num_classes) + + self._init_params() + + def _make_layer(self, block, t, c, n, s): + # t: expansion factor + # c: output channels + # n: number of blocks + # s: stride for first layer + layers = [] + layers.append(block(self.in_channels, c, t, s)) + self.in_channels = c + for i in range(1, n): + layers.append(block(self.in_channels, c, t)) + return nn.Sequential(*layers) + + def _construct_fc_layer(self, fc_dims, input_dim, dropout_p=None): + """Constructs fully connected layer. + + Args: + fc_dims (list or tuple): dimensions of fc layers, if None, no fc layers are constructed + input_dim (int): input dimension + dropout_p (float): dropout probability, if None, dropout is unused + """ + if fc_dims is None: + self.feature_dim = input_dim + return None + + assert isinstance( + fc_dims, (list, tuple) + ), 'fc_dims must be either list or tuple, but got {}'.format( + type(fc_dims) + ) + + layers = [] + for dim in fc_dims: + layers.append(nn.Linear(input_dim, dim)) + layers.append(nn.BatchNorm1d(dim)) + layers.append(nn.ReLU(inplace=True)) + if dropout_p is not None: + layers.append(nn.Dropout(p=dropout_p)) + input_dim = dim + + self.feature_dim = fc_dims[-1] + + return nn.Sequential(*layers) + + def _init_params(self): + for m in self.modules(): + if isinstance(m, nn.Conv2d): + nn.init.kaiming_normal_( + m.weight, mode='fan_out', nonlinearity='relu' + ) + if m.bias is not None: + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.BatchNorm2d): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.BatchNorm1d): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.Linear): + nn.init.normal_(m.weight, 0, 0.01) + if m.bias is not None: + nn.init.constant_(m.bias, 0) + + def featuremaps(self, x): + x = self.conv1(x) + x = self.conv2(x) + x = self.conv3(x) + x = self.conv4(x) + x = self.conv5(x) + x = self.conv6(x) + x = self.conv7(x) + x = self.conv8(x) + x = self.conv9(x) + return x + + def forward(self, x): + f = self.featuremaps(x) + v = self.global_avgpool(f) + v = v.view(v.size(0), -1) + + if self.fc is not None: + v = self.fc(v) + + if not self.training: + return v + + y = self.classifier(v) + + if self.loss == 'softmax': + return y + elif self.loss == 'triplet': + return y, v + else: + raise KeyError("Unsupported loss: {}".format(self.loss)) + + +def init_pretrained_weights(model, model_url): + """Initializes model with pretrained weights. + + Layers that don't match with pretrained layers in name or size are kept unchanged. + """ + pretrain_dict = model_zoo.load_url(model_url) + model_dict = model.state_dict() + pretrain_dict = { + k: v + for k, v in pretrain_dict.items() + if k in model_dict and model_dict[k].size() == v.size() + } + model_dict.update(pretrain_dict) + model.load_state_dict(model_dict) + + +def mobilenetv2_x1_0(num_classes, loss, pretrained=True, **kwargs): + model = MobileNetV2( + num_classes, + loss=loss, + width_mult=1, + fc_dims=None, + dropout_p=None, + **kwargs + ) + if pretrained: + # init_pretrained_weights(model, model_urls['mobilenetv2_x1_0']) + import warnings + warnings.warn( + 'The imagenet pretrained weights need to be manually downloaded from {}' + .format(model_urls['mobilenetv2_x1_0']) + ) + return model + + +def mobilenetv2_x1_4(num_classes, loss, pretrained=True, **kwargs): + model = MobileNetV2( + num_classes, + loss=loss, + width_mult=1.4, + fc_dims=None, + dropout_p=None, + **kwargs + ) + if pretrained: + # init_pretrained_weights(model, model_urls['mobilenetv2_x1_4']) + import warnings + warnings.warn( + 'The imagenet pretrained weights need to be manually downloaded from {}' + .format(model_urls['mobilenetv2_x1_4']) + ) + return model diff --git a/strong_sort/deep/reid/torchreid/models/mudeep.py b/strong_sort/deep/reid/torchreid/models/mudeep.py new file mode 100644 index 0000000..ddbca67 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/models/mudeep.py @@ -0,0 +1,206 @@ +from __future__ import division, absolute_import +import torch +from torch import nn +from torch.nn import functional as F + +__all__ = ['MuDeep'] + + +class ConvBlock(nn.Module): + """Basic convolutional block. + + convolution + batch normalization + relu. + + Args: + in_c (int): number of input channels. + out_c (int): number of output channels. + k (int or tuple): kernel size. + s (int or tuple): stride. + p (int or tuple): padding. + """ + + def __init__(self, in_c, out_c, k, s, p): + super(ConvBlock, self).__init__() + self.conv = nn.Conv2d(in_c, out_c, k, stride=s, padding=p) + self.bn = nn.BatchNorm2d(out_c) + + def forward(self, x): + return F.relu(self.bn(self.conv(x))) + + +class ConvLayers(nn.Module): + """Preprocessing layers.""" + + def __init__(self): + super(ConvLayers, self).__init__() + self.conv1 = ConvBlock(3, 48, k=3, s=1, p=1) + self.conv2 = ConvBlock(48, 96, k=3, s=1, p=1) + self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) + + def forward(self, x): + x = self.conv1(x) + x = self.conv2(x) + x = self.maxpool(x) + return x + + +class MultiScaleA(nn.Module): + """Multi-scale stream layer A (Sec.3.1)""" + + def __init__(self): + super(MultiScaleA, self).__init__() + self.stream1 = nn.Sequential( + ConvBlock(96, 96, k=1, s=1, p=0), + ConvBlock(96, 24, k=3, s=1, p=1), + ) + self.stream2 = nn.Sequential( + nn.AvgPool2d(kernel_size=3, stride=1, padding=1), + ConvBlock(96, 24, k=1, s=1, p=0), + ) + self.stream3 = ConvBlock(96, 24, k=1, s=1, p=0) + self.stream4 = nn.Sequential( + ConvBlock(96, 16, k=1, s=1, p=0), + ConvBlock(16, 24, k=3, s=1, p=1), + ConvBlock(24, 24, k=3, s=1, p=1), + ) + + def forward(self, x): + s1 = self.stream1(x) + s2 = self.stream2(x) + s3 = self.stream3(x) + s4 = self.stream4(x) + y = torch.cat([s1, s2, s3, s4], dim=1) + return y + + +class Reduction(nn.Module): + """Reduction layer (Sec.3.1)""" + + def __init__(self): + super(Reduction, self).__init__() + self.stream1 = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) + self.stream2 = ConvBlock(96, 96, k=3, s=2, p=1) + self.stream3 = nn.Sequential( + ConvBlock(96, 48, k=1, s=1, p=0), + ConvBlock(48, 56, k=3, s=1, p=1), + ConvBlock(56, 64, k=3, s=2, p=1), + ) + + def forward(self, x): + s1 = self.stream1(x) + s2 = self.stream2(x) + s3 = self.stream3(x) + y = torch.cat([s1, s2, s3], dim=1) + return y + + +class MultiScaleB(nn.Module): + """Multi-scale stream layer B (Sec.3.1)""" + + def __init__(self): + super(MultiScaleB, self).__init__() + self.stream1 = nn.Sequential( + nn.AvgPool2d(kernel_size=3, stride=1, padding=1), + ConvBlock(256, 256, k=1, s=1, p=0), + ) + self.stream2 = nn.Sequential( + ConvBlock(256, 64, k=1, s=1, p=0), + ConvBlock(64, 128, k=(1, 3), s=1, p=(0, 1)), + ConvBlock(128, 256, k=(3, 1), s=1, p=(1, 0)), + ) + self.stream3 = ConvBlock(256, 256, k=1, s=1, p=0) + self.stream4 = nn.Sequential( + ConvBlock(256, 64, k=1, s=1, p=0), + ConvBlock(64, 64, k=(1, 3), s=1, p=(0, 1)), + ConvBlock(64, 128, k=(3, 1), s=1, p=(1, 0)), + ConvBlock(128, 128, k=(1, 3), s=1, p=(0, 1)), + ConvBlock(128, 256, k=(3, 1), s=1, p=(1, 0)), + ) + + def forward(self, x): + s1 = self.stream1(x) + s2 = self.stream2(x) + s3 = self.stream3(x) + s4 = self.stream4(x) + return s1, s2, s3, s4 + + +class Fusion(nn.Module): + """Saliency-based learning fusion layer (Sec.3.2)""" + + def __init__(self): + super(Fusion, self).__init__() + self.a1 = nn.Parameter(torch.rand(1, 256, 1, 1)) + self.a2 = nn.Parameter(torch.rand(1, 256, 1, 1)) + self.a3 = nn.Parameter(torch.rand(1, 256, 1, 1)) + self.a4 = nn.Parameter(torch.rand(1, 256, 1, 1)) + + # We add an average pooling layer to reduce the spatial dimension + # of feature maps, which differs from the original paper. + self.avgpool = nn.AvgPool2d(kernel_size=4, stride=4, padding=0) + + def forward(self, x1, x2, x3, x4): + s1 = self.a1.expand_as(x1) * x1 + s2 = self.a2.expand_as(x2) * x2 + s3 = self.a3.expand_as(x3) * x3 + s4 = self.a4.expand_as(x4) * x4 + y = self.avgpool(s1 + s2 + s3 + s4) + return y + + +class MuDeep(nn.Module): + """Multiscale deep neural network. + + Reference: + Qian et al. Multi-scale Deep Learning Architectures + for Person Re-identification. ICCV 2017. + + Public keys: + - ``mudeep``: Multiscale deep neural network. + """ + + def __init__(self, num_classes, loss='softmax', **kwargs): + super(MuDeep, self).__init__() + self.loss = loss + + self.block1 = ConvLayers() + self.block2 = MultiScaleA() + self.block3 = Reduction() + self.block4 = MultiScaleB() + self.block5 = Fusion() + + # Due to this fully connected layer, input image has to be fixed + # in shape, i.e. (3, 256, 128), such that the last convolutional feature + # maps are of shape (256, 16, 8). If input shape is changed, + # the input dimension of this layer has to be changed accordingly. + self.fc = nn.Sequential( + nn.Linear(256 * 16 * 8, 4096), + nn.BatchNorm1d(4096), + nn.ReLU(), + ) + self.classifier = nn.Linear(4096, num_classes) + self.feat_dim = 4096 + + def featuremaps(self, x): + x = self.block1(x) + x = self.block2(x) + x = self.block3(x) + x = self.block4(x) + x = self.block5(*x) + return x + + def forward(self, x): + x = self.featuremaps(x) + x = x.view(x.size(0), -1) + x = self.fc(x) + y = self.classifier(x) + + if not self.training: + return x + + if self.loss == 'softmax': + return y + elif self.loss == 'triplet': + return y, x + else: + raise KeyError('Unsupported loss: {}'.format(self.loss)) diff --git a/strong_sort/deep/reid/torchreid/models/nasnet.py b/strong_sort/deep/reid/torchreid/models/nasnet.py new file mode 100644 index 0000000..b1f31de --- /dev/null +++ b/strong_sort/deep/reid/torchreid/models/nasnet.py @@ -0,0 +1,1131 @@ +from __future__ import division, absolute_import +import torch +import torch.nn as nn +import torch.nn.functional as F +import torch.utils.model_zoo as model_zoo + +__all__ = ['nasnetamobile'] +""" +NASNet Mobile +Thanks to Anastasiia (https://github.com/DagnyT) for the great help, support and motivation! + + +------------------------------------------------------------------------------------ + Architecture | Top-1 Acc | Top-5 Acc | Multiply-Adds | Params (M) +------------------------------------------------------------------------------------ +| NASNet-A (4 @ 1056) | 74.08% | 91.74% | 564 M | 5.3 | +------------------------------------------------------------------------------------ +# References: + - [Learning Transferable Architectures for Scalable Image Recognition] + (https://arxiv.org/abs/1707.07012) +""" +""" +Code imported from https://github.com/Cadene/pretrained-models.pytorch +""" + +pretrained_settings = { + 'nasnetamobile': { + 'imagenet': { + # 'url': 'https://github.com/veronikayurchuk/pretrained-models.pytorch/releases/download/v1.0/nasnetmobile-7e03cead.pth.tar', + 'url': + 'http://data.lip6.fr/cadene/pretrainedmodels/nasnetamobile-7e03cead.pth', + 'input_space': 'RGB', + 'input_size': [3, 224, 224], # resize 256 + 'input_range': [0, 1], + 'mean': [0.5, 0.5, 0.5], + 'std': [0.5, 0.5, 0.5], + 'num_classes': 1000 + }, + # 'imagenet+background': { + # # 'url': 'http://data.lip6.fr/cadene/pretrainedmodels/nasnetalarge-a1897284.pth', + # 'input_space': 'RGB', + # 'input_size': [3, 224, 224], # resize 256 + # 'input_range': [0, 1], + # 'mean': [0.5, 0.5, 0.5], + # 'std': [0.5, 0.5, 0.5], + # 'num_classes': 1001 + # } + } +} + + +class MaxPoolPad(nn.Module): + + def __init__(self): + super(MaxPoolPad, self).__init__() + self.pad = nn.ZeroPad2d((1, 0, 1, 0)) + self.pool = nn.MaxPool2d(3, stride=2, padding=1) + + def forward(self, x): + x = self.pad(x) + x = self.pool(x) + x = x[:, :, 1:, 1:].contiguous() + return x + + +class AvgPoolPad(nn.Module): + + def __init__(self, stride=2, padding=1): + super(AvgPoolPad, self).__init__() + self.pad = nn.ZeroPad2d((1, 0, 1, 0)) + self.pool = nn.AvgPool2d( + 3, stride=stride, padding=padding, count_include_pad=False + ) + + def forward(self, x): + x = self.pad(x) + x = self.pool(x) + x = x[:, :, 1:, 1:].contiguous() + return x + + +class SeparableConv2d(nn.Module): + + def __init__( + self, + in_channels, + out_channels, + dw_kernel, + dw_stride, + dw_padding, + bias=False + ): + super(SeparableConv2d, self).__init__() + self.depthwise_conv2d = nn.Conv2d( + in_channels, + in_channels, + dw_kernel, + stride=dw_stride, + padding=dw_padding, + bias=bias, + groups=in_channels + ) + self.pointwise_conv2d = nn.Conv2d( + in_channels, out_channels, 1, stride=1, bias=bias + ) + + def forward(self, x): + x = self.depthwise_conv2d(x) + x = self.pointwise_conv2d(x) + return x + + +class BranchSeparables(nn.Module): + + def __init__( + self, + in_channels, + out_channels, + kernel_size, + stride, + padding, + name=None, + bias=False + ): + super(BranchSeparables, self).__init__() + self.relu = nn.ReLU() + self.separable_1 = SeparableConv2d( + in_channels, in_channels, kernel_size, stride, padding, bias=bias + ) + self.bn_sep_1 = nn.BatchNorm2d( + in_channels, eps=0.001, momentum=0.1, affine=True + ) + self.relu1 = nn.ReLU() + self.separable_2 = SeparableConv2d( + in_channels, out_channels, kernel_size, 1, padding, bias=bias + ) + self.bn_sep_2 = nn.BatchNorm2d( + out_channels, eps=0.001, momentum=0.1, affine=True + ) + self.name = name + + def forward(self, x): + x = self.relu(x) + if self.name == 'specific': + x = nn.ZeroPad2d((1, 0, 1, 0))(x) + x = self.separable_1(x) + if self.name == 'specific': + x = x[:, :, 1:, 1:].contiguous() + + x = self.bn_sep_1(x) + x = self.relu1(x) + x = self.separable_2(x) + x = self.bn_sep_2(x) + return x + + +class BranchSeparablesStem(nn.Module): + + def __init__( + self, + in_channels, + out_channels, + kernel_size, + stride, + padding, + bias=False + ): + super(BranchSeparablesStem, self).__init__() + self.relu = nn.ReLU() + self.separable_1 = SeparableConv2d( + in_channels, out_channels, kernel_size, stride, padding, bias=bias + ) + self.bn_sep_1 = nn.BatchNorm2d( + out_channels, eps=0.001, momentum=0.1, affine=True + ) + self.relu1 = nn.ReLU() + self.separable_2 = SeparableConv2d( + out_channels, out_channels, kernel_size, 1, padding, bias=bias + ) + self.bn_sep_2 = nn.BatchNorm2d( + out_channels, eps=0.001, momentum=0.1, affine=True + ) + + def forward(self, x): + x = self.relu(x) + x = self.separable_1(x) + x = self.bn_sep_1(x) + x = self.relu1(x) + x = self.separable_2(x) + x = self.bn_sep_2(x) + return x + + +class BranchSeparablesReduction(BranchSeparables): + + def __init__( + self, + in_channels, + out_channels, + kernel_size, + stride, + padding, + z_padding=1, + bias=False + ): + BranchSeparables.__init__( + self, in_channels, out_channels, kernel_size, stride, padding, bias + ) + self.padding = nn.ZeroPad2d((z_padding, 0, z_padding, 0)) + + def forward(self, x): + x = self.relu(x) + x = self.padding(x) + x = self.separable_1(x) + x = x[:, :, 1:, 1:].contiguous() + x = self.bn_sep_1(x) + x = self.relu1(x) + x = self.separable_2(x) + x = self.bn_sep_2(x) + return x + + +class CellStem0(nn.Module): + + def __init__(self, stem_filters, num_filters=42): + super(CellStem0, self).__init__() + self.num_filters = num_filters + self.stem_filters = stem_filters + self.conv_1x1 = nn.Sequential() + self.conv_1x1.add_module('relu', nn.ReLU()) + self.conv_1x1.add_module( + 'conv', + nn.Conv2d( + self.stem_filters, self.num_filters, 1, stride=1, bias=False + ) + ) + self.conv_1x1.add_module( + 'bn', + nn.BatchNorm2d( + self.num_filters, eps=0.001, momentum=0.1, affine=True + ) + ) + + self.comb_iter_0_left = BranchSeparables( + self.num_filters, self.num_filters, 5, 2, 2 + ) + self.comb_iter_0_right = BranchSeparablesStem( + self.stem_filters, self.num_filters, 7, 2, 3, bias=False + ) + + self.comb_iter_1_left = nn.MaxPool2d(3, stride=2, padding=1) + self.comb_iter_1_right = BranchSeparablesStem( + self.stem_filters, self.num_filters, 7, 2, 3, bias=False + ) + + self.comb_iter_2_left = nn.AvgPool2d( + 3, stride=2, padding=1, count_include_pad=False + ) + self.comb_iter_2_right = BranchSeparablesStem( + self.stem_filters, self.num_filters, 5, 2, 2, bias=False + ) + + self.comb_iter_3_right = nn.AvgPool2d( + 3, stride=1, padding=1, count_include_pad=False + ) + + self.comb_iter_4_left = BranchSeparables( + self.num_filters, self.num_filters, 3, 1, 1, bias=False + ) + self.comb_iter_4_right = nn.MaxPool2d(3, stride=2, padding=1) + + def forward(self, x): + x1 = self.conv_1x1(x) + + x_comb_iter_0_left = self.comb_iter_0_left(x1) + x_comb_iter_0_right = self.comb_iter_0_right(x) + x_comb_iter_0 = x_comb_iter_0_left + x_comb_iter_0_right + + x_comb_iter_1_left = self.comb_iter_1_left(x1) + x_comb_iter_1_right = self.comb_iter_1_right(x) + x_comb_iter_1 = x_comb_iter_1_left + x_comb_iter_1_right + + x_comb_iter_2_left = self.comb_iter_2_left(x1) + x_comb_iter_2_right = self.comb_iter_2_right(x) + x_comb_iter_2 = x_comb_iter_2_left + x_comb_iter_2_right + + x_comb_iter_3_right = self.comb_iter_3_right(x_comb_iter_0) + x_comb_iter_3 = x_comb_iter_3_right + x_comb_iter_1 + + x_comb_iter_4_left = self.comb_iter_4_left(x_comb_iter_0) + x_comb_iter_4_right = self.comb_iter_4_right(x1) + x_comb_iter_4 = x_comb_iter_4_left + x_comb_iter_4_right + + x_out = torch.cat( + [x_comb_iter_1, x_comb_iter_2, x_comb_iter_3, x_comb_iter_4], 1 + ) + return x_out + + +class CellStem1(nn.Module): + + def __init__(self, stem_filters, num_filters): + super(CellStem1, self).__init__() + self.num_filters = num_filters + self.stem_filters = stem_filters + self.conv_1x1 = nn.Sequential() + self.conv_1x1.add_module('relu', nn.ReLU()) + self.conv_1x1.add_module( + 'conv', + nn.Conv2d( + 2 * self.num_filters, + self.num_filters, + 1, + stride=1, + bias=False + ) + ) + self.conv_1x1.add_module( + 'bn', + nn.BatchNorm2d( + self.num_filters, eps=0.001, momentum=0.1, affine=True + ) + ) + + self.relu = nn.ReLU() + self.path_1 = nn.Sequential() + self.path_1.add_module( + 'avgpool', nn.AvgPool2d(1, stride=2, count_include_pad=False) + ) + self.path_1.add_module( + 'conv', + nn.Conv2d( + self.stem_filters, + self.num_filters // 2, + 1, + stride=1, + bias=False + ) + ) + self.path_2 = nn.ModuleList() + self.path_2.add_module('pad', nn.ZeroPad2d((0, 1, 0, 1))) + self.path_2.add_module( + 'avgpool', nn.AvgPool2d(1, stride=2, count_include_pad=False) + ) + self.path_2.add_module( + 'conv', + nn.Conv2d( + self.stem_filters, + self.num_filters // 2, + 1, + stride=1, + bias=False + ) + ) + + self.final_path_bn = nn.BatchNorm2d( + self.num_filters, eps=0.001, momentum=0.1, affine=True + ) + + self.comb_iter_0_left = BranchSeparables( + self.num_filters, + self.num_filters, + 5, + 2, + 2, + name='specific', + bias=False + ) + self.comb_iter_0_right = BranchSeparables( + self.num_filters, + self.num_filters, + 7, + 2, + 3, + name='specific', + bias=False + ) + + # self.comb_iter_1_left = nn.MaxPool2d(3, stride=2, padding=1) + self.comb_iter_1_left = MaxPoolPad() + self.comb_iter_1_right = BranchSeparables( + self.num_filters, + self.num_filters, + 7, + 2, + 3, + name='specific', + bias=False + ) + + # self.comb_iter_2_left = nn.AvgPool2d(3, stride=2, padding=1, count_include_pad=False) + self.comb_iter_2_left = AvgPoolPad() + self.comb_iter_2_right = BranchSeparables( + self.num_filters, + self.num_filters, + 5, + 2, + 2, + name='specific', + bias=False + ) + + self.comb_iter_3_right = nn.AvgPool2d( + 3, stride=1, padding=1, count_include_pad=False + ) + + self.comb_iter_4_left = BranchSeparables( + self.num_filters, + self.num_filters, + 3, + 1, + 1, + name='specific', + bias=False + ) + # self.comb_iter_4_right = nn.MaxPool2d(3, stride=2, padding=1) + self.comb_iter_4_right = MaxPoolPad() + + def forward(self, x_conv0, x_stem_0): + x_left = self.conv_1x1(x_stem_0) + + x_relu = self.relu(x_conv0) + # path 1 + x_path1 = self.path_1(x_relu) + # path 2 + x_path2 = self.path_2.pad(x_relu) + x_path2 = x_path2[:, :, 1:, 1:] + x_path2 = self.path_2.avgpool(x_path2) + x_path2 = self.path_2.conv(x_path2) + # final path + x_right = self.final_path_bn(torch.cat([x_path1, x_path2], 1)) + + x_comb_iter_0_left = self.comb_iter_0_left(x_left) + x_comb_iter_0_right = self.comb_iter_0_right(x_right) + x_comb_iter_0 = x_comb_iter_0_left + x_comb_iter_0_right + + x_comb_iter_1_left = self.comb_iter_1_left(x_left) + x_comb_iter_1_right = self.comb_iter_1_right(x_right) + x_comb_iter_1 = x_comb_iter_1_left + x_comb_iter_1_right + + x_comb_iter_2_left = self.comb_iter_2_left(x_left) + x_comb_iter_2_right = self.comb_iter_2_right(x_right) + x_comb_iter_2 = x_comb_iter_2_left + x_comb_iter_2_right + + x_comb_iter_3_right = self.comb_iter_3_right(x_comb_iter_0) + x_comb_iter_3 = x_comb_iter_3_right + x_comb_iter_1 + + x_comb_iter_4_left = self.comb_iter_4_left(x_comb_iter_0) + x_comb_iter_4_right = self.comb_iter_4_right(x_left) + x_comb_iter_4 = x_comb_iter_4_left + x_comb_iter_4_right + + x_out = torch.cat( + [x_comb_iter_1, x_comb_iter_2, x_comb_iter_3, x_comb_iter_4], 1 + ) + return x_out + + +class FirstCell(nn.Module): + + def __init__( + self, in_channels_left, out_channels_left, in_channels_right, + out_channels_right + ): + super(FirstCell, self).__init__() + self.conv_1x1 = nn.Sequential() + self.conv_1x1.add_module('relu', nn.ReLU()) + self.conv_1x1.add_module( + 'conv', + nn.Conv2d( + in_channels_right, out_channels_right, 1, stride=1, bias=False + ) + ) + self.conv_1x1.add_module( + 'bn', + nn.BatchNorm2d( + out_channels_right, eps=0.001, momentum=0.1, affine=True + ) + ) + + self.relu = nn.ReLU() + self.path_1 = nn.Sequential() + self.path_1.add_module( + 'avgpool', nn.AvgPool2d(1, stride=2, count_include_pad=False) + ) + self.path_1.add_module( + 'conv', + nn.Conv2d( + in_channels_left, out_channels_left, 1, stride=1, bias=False + ) + ) + self.path_2 = nn.ModuleList() + self.path_2.add_module('pad', nn.ZeroPad2d((0, 1, 0, 1))) + self.path_2.add_module( + 'avgpool', nn.AvgPool2d(1, stride=2, count_include_pad=False) + ) + self.path_2.add_module( + 'conv', + nn.Conv2d( + in_channels_left, out_channels_left, 1, stride=1, bias=False + ) + ) + + self.final_path_bn = nn.BatchNorm2d( + out_channels_left * 2, eps=0.001, momentum=0.1, affine=True + ) + + self.comb_iter_0_left = BranchSeparables( + out_channels_right, out_channels_right, 5, 1, 2, bias=False + ) + self.comb_iter_0_right = BranchSeparables( + out_channels_right, out_channels_right, 3, 1, 1, bias=False + ) + + self.comb_iter_1_left = BranchSeparables( + out_channels_right, out_channels_right, 5, 1, 2, bias=False + ) + self.comb_iter_1_right = BranchSeparables( + out_channels_right, out_channels_right, 3, 1, 1, bias=False + ) + + self.comb_iter_2_left = nn.AvgPool2d( + 3, stride=1, padding=1, count_include_pad=False + ) + + self.comb_iter_3_left = nn.AvgPool2d( + 3, stride=1, padding=1, count_include_pad=False + ) + self.comb_iter_3_right = nn.AvgPool2d( + 3, stride=1, padding=1, count_include_pad=False + ) + + self.comb_iter_4_left = BranchSeparables( + out_channels_right, out_channels_right, 3, 1, 1, bias=False + ) + + def forward(self, x, x_prev): + x_relu = self.relu(x_prev) + # path 1 + x_path1 = self.path_1(x_relu) + # path 2 + x_path2 = self.path_2.pad(x_relu) + x_path2 = x_path2[:, :, 1:, 1:] + x_path2 = self.path_2.avgpool(x_path2) + x_path2 = self.path_2.conv(x_path2) + # final path + x_left = self.final_path_bn(torch.cat([x_path1, x_path2], 1)) + + x_right = self.conv_1x1(x) + + x_comb_iter_0_left = self.comb_iter_0_left(x_right) + x_comb_iter_0_right = self.comb_iter_0_right(x_left) + x_comb_iter_0 = x_comb_iter_0_left + x_comb_iter_0_right + + x_comb_iter_1_left = self.comb_iter_1_left(x_left) + x_comb_iter_1_right = self.comb_iter_1_right(x_left) + x_comb_iter_1 = x_comb_iter_1_left + x_comb_iter_1_right + + x_comb_iter_2_left = self.comb_iter_2_left(x_right) + x_comb_iter_2 = x_comb_iter_2_left + x_left + + x_comb_iter_3_left = self.comb_iter_3_left(x_left) + x_comb_iter_3_right = self.comb_iter_3_right(x_left) + x_comb_iter_3 = x_comb_iter_3_left + x_comb_iter_3_right + + x_comb_iter_4_left = self.comb_iter_4_left(x_right) + x_comb_iter_4 = x_comb_iter_4_left + x_right + + x_out = torch.cat( + [ + x_left, x_comb_iter_0, x_comb_iter_1, x_comb_iter_2, + x_comb_iter_3, x_comb_iter_4 + ], 1 + ) + return x_out + + +class NormalCell(nn.Module): + + def __init__( + self, in_channels_left, out_channels_left, in_channels_right, + out_channels_right + ): + super(NormalCell, self).__init__() + self.conv_prev_1x1 = nn.Sequential() + self.conv_prev_1x1.add_module('relu', nn.ReLU()) + self.conv_prev_1x1.add_module( + 'conv', + nn.Conv2d( + in_channels_left, out_channels_left, 1, stride=1, bias=False + ) + ) + self.conv_prev_1x1.add_module( + 'bn', + nn.BatchNorm2d( + out_channels_left, eps=0.001, momentum=0.1, affine=True + ) + ) + + self.conv_1x1 = nn.Sequential() + self.conv_1x1.add_module('relu', nn.ReLU()) + self.conv_1x1.add_module( + 'conv', + nn.Conv2d( + in_channels_right, out_channels_right, 1, stride=1, bias=False + ) + ) + self.conv_1x1.add_module( + 'bn', + nn.BatchNorm2d( + out_channels_right, eps=0.001, momentum=0.1, affine=True + ) + ) + + self.comb_iter_0_left = BranchSeparables( + out_channels_right, out_channels_right, 5, 1, 2, bias=False + ) + self.comb_iter_0_right = BranchSeparables( + out_channels_left, out_channels_left, 3, 1, 1, bias=False + ) + + self.comb_iter_1_left = BranchSeparables( + out_channels_left, out_channels_left, 5, 1, 2, bias=False + ) + self.comb_iter_1_right = BranchSeparables( + out_channels_left, out_channels_left, 3, 1, 1, bias=False + ) + + self.comb_iter_2_left = nn.AvgPool2d( + 3, stride=1, padding=1, count_include_pad=False + ) + + self.comb_iter_3_left = nn.AvgPool2d( + 3, stride=1, padding=1, count_include_pad=False + ) + self.comb_iter_3_right = nn.AvgPool2d( + 3, stride=1, padding=1, count_include_pad=False + ) + + self.comb_iter_4_left = BranchSeparables( + out_channels_right, out_channels_right, 3, 1, 1, bias=False + ) + + def forward(self, x, x_prev): + x_left = self.conv_prev_1x1(x_prev) + x_right = self.conv_1x1(x) + + x_comb_iter_0_left = self.comb_iter_0_left(x_right) + x_comb_iter_0_right = self.comb_iter_0_right(x_left) + x_comb_iter_0 = x_comb_iter_0_left + x_comb_iter_0_right + + x_comb_iter_1_left = self.comb_iter_1_left(x_left) + x_comb_iter_1_right = self.comb_iter_1_right(x_left) + x_comb_iter_1 = x_comb_iter_1_left + x_comb_iter_1_right + + x_comb_iter_2_left = self.comb_iter_2_left(x_right) + x_comb_iter_2 = x_comb_iter_2_left + x_left + + x_comb_iter_3_left = self.comb_iter_3_left(x_left) + x_comb_iter_3_right = self.comb_iter_3_right(x_left) + x_comb_iter_3 = x_comb_iter_3_left + x_comb_iter_3_right + + x_comb_iter_4_left = self.comb_iter_4_left(x_right) + x_comb_iter_4 = x_comb_iter_4_left + x_right + + x_out = torch.cat( + [ + x_left, x_comb_iter_0, x_comb_iter_1, x_comb_iter_2, + x_comb_iter_3, x_comb_iter_4 + ], 1 + ) + return x_out + + +class ReductionCell0(nn.Module): + + def __init__( + self, in_channels_left, out_channels_left, in_channels_right, + out_channels_right + ): + super(ReductionCell0, self).__init__() + self.conv_prev_1x1 = nn.Sequential() + self.conv_prev_1x1.add_module('relu', nn.ReLU()) + self.conv_prev_1x1.add_module( + 'conv', + nn.Conv2d( + in_channels_left, out_channels_left, 1, stride=1, bias=False + ) + ) + self.conv_prev_1x1.add_module( + 'bn', + nn.BatchNorm2d( + out_channels_left, eps=0.001, momentum=0.1, affine=True + ) + ) + + self.conv_1x1 = nn.Sequential() + self.conv_1x1.add_module('relu', nn.ReLU()) + self.conv_1x1.add_module( + 'conv', + nn.Conv2d( + in_channels_right, out_channels_right, 1, stride=1, bias=False + ) + ) + self.conv_1x1.add_module( + 'bn', + nn.BatchNorm2d( + out_channels_right, eps=0.001, momentum=0.1, affine=True + ) + ) + + self.comb_iter_0_left = BranchSeparablesReduction( + out_channels_right, out_channels_right, 5, 2, 2, bias=False + ) + self.comb_iter_0_right = BranchSeparablesReduction( + out_channels_right, out_channels_right, 7, 2, 3, bias=False + ) + + self.comb_iter_1_left = MaxPoolPad() + self.comb_iter_1_right = BranchSeparablesReduction( + out_channels_right, out_channels_right, 7, 2, 3, bias=False + ) + + self.comb_iter_2_left = AvgPoolPad() + self.comb_iter_2_right = BranchSeparablesReduction( + out_channels_right, out_channels_right, 5, 2, 2, bias=False + ) + + self.comb_iter_3_right = nn.AvgPool2d( + 3, stride=1, padding=1, count_include_pad=False + ) + + self.comb_iter_4_left = BranchSeparablesReduction( + out_channels_right, out_channels_right, 3, 1, 1, bias=False + ) + self.comb_iter_4_right = MaxPoolPad() + + def forward(self, x, x_prev): + x_left = self.conv_prev_1x1(x_prev) + x_right = self.conv_1x1(x) + + x_comb_iter_0_left = self.comb_iter_0_left(x_right) + x_comb_iter_0_right = self.comb_iter_0_right(x_left) + x_comb_iter_0 = x_comb_iter_0_left + x_comb_iter_0_right + + x_comb_iter_1_left = self.comb_iter_1_left(x_right) + x_comb_iter_1_right = self.comb_iter_1_right(x_left) + x_comb_iter_1 = x_comb_iter_1_left + x_comb_iter_1_right + + x_comb_iter_2_left = self.comb_iter_2_left(x_right) + x_comb_iter_2_right = self.comb_iter_2_right(x_left) + x_comb_iter_2 = x_comb_iter_2_left + x_comb_iter_2_right + + x_comb_iter_3_right = self.comb_iter_3_right(x_comb_iter_0) + x_comb_iter_3 = x_comb_iter_3_right + x_comb_iter_1 + + x_comb_iter_4_left = self.comb_iter_4_left(x_comb_iter_0) + x_comb_iter_4_right = self.comb_iter_4_right(x_right) + x_comb_iter_4 = x_comb_iter_4_left + x_comb_iter_4_right + + x_out = torch.cat( + [x_comb_iter_1, x_comb_iter_2, x_comb_iter_3, x_comb_iter_4], 1 + ) + return x_out + + +class ReductionCell1(nn.Module): + + def __init__( + self, in_channels_left, out_channels_left, in_channels_right, + out_channels_right + ): + super(ReductionCell1, self).__init__() + self.conv_prev_1x1 = nn.Sequential() + self.conv_prev_1x1.add_module('relu', nn.ReLU()) + self.conv_prev_1x1.add_module( + 'conv', + nn.Conv2d( + in_channels_left, out_channels_left, 1, stride=1, bias=False + ) + ) + self.conv_prev_1x1.add_module( + 'bn', + nn.BatchNorm2d( + out_channels_left, eps=0.001, momentum=0.1, affine=True + ) + ) + + self.conv_1x1 = nn.Sequential() + self.conv_1x1.add_module('relu', nn.ReLU()) + self.conv_1x1.add_module( + 'conv', + nn.Conv2d( + in_channels_right, out_channels_right, 1, stride=1, bias=False + ) + ) + self.conv_1x1.add_module( + 'bn', + nn.BatchNorm2d( + out_channels_right, eps=0.001, momentum=0.1, affine=True + ) + ) + + self.comb_iter_0_left = BranchSeparables( + out_channels_right, + out_channels_right, + 5, + 2, + 2, + name='specific', + bias=False + ) + self.comb_iter_0_right = BranchSeparables( + out_channels_right, + out_channels_right, + 7, + 2, + 3, + name='specific', + bias=False + ) + + # self.comb_iter_1_left = nn.MaxPool2d(3, stride=2, padding=1) + self.comb_iter_1_left = MaxPoolPad() + self.comb_iter_1_right = BranchSeparables( + out_channels_right, + out_channels_right, + 7, + 2, + 3, + name='specific', + bias=False + ) + + # self.comb_iter_2_left = nn.AvgPool2d(3, stride=2, padding=1, count_include_pad=False) + self.comb_iter_2_left = AvgPoolPad() + self.comb_iter_2_right = BranchSeparables( + out_channels_right, + out_channels_right, + 5, + 2, + 2, + name='specific', + bias=False + ) + + self.comb_iter_3_right = nn.AvgPool2d( + 3, stride=1, padding=1, count_include_pad=False + ) + + self.comb_iter_4_left = BranchSeparables( + out_channels_right, + out_channels_right, + 3, + 1, + 1, + name='specific', + bias=False + ) + # self.comb_iter_4_right = nn.MaxPool2d(3, stride=2, padding=1) + self.comb_iter_4_right = MaxPoolPad() + + def forward(self, x, x_prev): + x_left = self.conv_prev_1x1(x_prev) + x_right = self.conv_1x1(x) + + x_comb_iter_0_left = self.comb_iter_0_left(x_right) + x_comb_iter_0_right = self.comb_iter_0_right(x_left) + x_comb_iter_0 = x_comb_iter_0_left + x_comb_iter_0_right + + x_comb_iter_1_left = self.comb_iter_1_left(x_right) + x_comb_iter_1_right = self.comb_iter_1_right(x_left) + x_comb_iter_1 = x_comb_iter_1_left + x_comb_iter_1_right + + x_comb_iter_2_left = self.comb_iter_2_left(x_right) + x_comb_iter_2_right = self.comb_iter_2_right(x_left) + x_comb_iter_2 = x_comb_iter_2_left + x_comb_iter_2_right + + x_comb_iter_3_right = self.comb_iter_3_right(x_comb_iter_0) + x_comb_iter_3 = x_comb_iter_3_right + x_comb_iter_1 + + x_comb_iter_4_left = self.comb_iter_4_left(x_comb_iter_0) + x_comb_iter_4_right = self.comb_iter_4_right(x_right) + x_comb_iter_4 = x_comb_iter_4_left + x_comb_iter_4_right + + x_out = torch.cat( + [x_comb_iter_1, x_comb_iter_2, x_comb_iter_3, x_comb_iter_4], 1 + ) + return x_out + + +class NASNetAMobile(nn.Module): + """Neural Architecture Search (NAS). + + Reference: + Zoph et al. Learning Transferable Architectures + for Scalable Image Recognition. CVPR 2018. + + Public keys: + - ``nasnetamobile``: NASNet-A Mobile. + """ + + def __init__( + self, + num_classes, + loss, + stem_filters=32, + penultimate_filters=1056, + filters_multiplier=2, + **kwargs + ): + super(NASNetAMobile, self).__init__() + self.stem_filters = stem_filters + self.penultimate_filters = penultimate_filters + self.filters_multiplier = filters_multiplier + self.loss = loss + + filters = self.penultimate_filters // 24 + # 24 is default value for the architecture + + self.conv0 = nn.Sequential() + self.conv0.add_module( + 'conv', + nn.Conv2d( + in_channels=3, + out_channels=self.stem_filters, + kernel_size=3, + padding=0, + stride=2, + bias=False + ) + ) + self.conv0.add_module( + 'bn', + nn.BatchNorm2d( + self.stem_filters, eps=0.001, momentum=0.1, affine=True + ) + ) + + self.cell_stem_0 = CellStem0( + self.stem_filters, num_filters=filters // (filters_multiplier**2) + ) + self.cell_stem_1 = CellStem1( + self.stem_filters, num_filters=filters // filters_multiplier + ) + + self.cell_0 = FirstCell( + in_channels_left=filters, + out_channels_left=filters // 2, # 1, 0.5 + in_channels_right=2 * filters, + out_channels_right=filters + ) # 2, 1 + self.cell_1 = NormalCell( + in_channels_left=2 * filters, + out_channels_left=filters, # 2, 1 + in_channels_right=6 * filters, + out_channels_right=filters + ) # 6, 1 + self.cell_2 = NormalCell( + in_channels_left=6 * filters, + out_channels_left=filters, # 6, 1 + in_channels_right=6 * filters, + out_channels_right=filters + ) # 6, 1 + self.cell_3 = NormalCell( + in_channels_left=6 * filters, + out_channels_left=filters, # 6, 1 + in_channels_right=6 * filters, + out_channels_right=filters + ) # 6, 1 + + self.reduction_cell_0 = ReductionCell0( + in_channels_left=6 * filters, + out_channels_left=2 * filters, # 6, 2 + in_channels_right=6 * filters, + out_channels_right=2 * filters + ) # 6, 2 + + self.cell_6 = FirstCell( + in_channels_left=6 * filters, + out_channels_left=filters, # 6, 1 + in_channels_right=8 * filters, + out_channels_right=2 * filters + ) # 8, 2 + self.cell_7 = NormalCell( + in_channels_left=8 * filters, + out_channels_left=2 * filters, # 8, 2 + in_channels_right=12 * filters, + out_channels_right=2 * filters + ) # 12, 2 + self.cell_8 = NormalCell( + in_channels_left=12 * filters, + out_channels_left=2 * filters, # 12, 2 + in_channels_right=12 * filters, + out_channels_right=2 * filters + ) # 12, 2 + self.cell_9 = NormalCell( + in_channels_left=12 * filters, + out_channels_left=2 * filters, # 12, 2 + in_channels_right=12 * filters, + out_channels_right=2 * filters + ) # 12, 2 + + self.reduction_cell_1 = ReductionCell1( + in_channels_left=12 * filters, + out_channels_left=4 * filters, # 12, 4 + in_channels_right=12 * filters, + out_channels_right=4 * filters + ) # 12, 4 + + self.cell_12 = FirstCell( + in_channels_left=12 * filters, + out_channels_left=2 * filters, # 12, 2 + in_channels_right=16 * filters, + out_channels_right=4 * filters + ) # 16, 4 + self.cell_13 = NormalCell( + in_channels_left=16 * filters, + out_channels_left=4 * filters, # 16, 4 + in_channels_right=24 * filters, + out_channels_right=4 * filters + ) # 24, 4 + self.cell_14 = NormalCell( + in_channels_left=24 * filters, + out_channels_left=4 * filters, # 24, 4 + in_channels_right=24 * filters, + out_channels_right=4 * filters + ) # 24, 4 + self.cell_15 = NormalCell( + in_channels_left=24 * filters, + out_channels_left=4 * filters, # 24, 4 + in_channels_right=24 * filters, + out_channels_right=4 * filters + ) # 24, 4 + + self.relu = nn.ReLU() + self.dropout = nn.Dropout() + self.classifier = nn.Linear(24 * filters, num_classes) + + self._init_params() + + def _init_params(self): + for m in self.modules(): + if isinstance(m, nn.Conv2d): + nn.init.kaiming_normal_( + m.weight, mode='fan_out', nonlinearity='relu' + ) + if m.bias is not None: + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.BatchNorm2d): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.BatchNorm1d): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.Linear): + nn.init.normal_(m.weight, 0, 0.01) + if m.bias is not None: + nn.init.constant_(m.bias, 0) + + def features(self, input): + x_conv0 = self.conv0(input) + x_stem_0 = self.cell_stem_0(x_conv0) + x_stem_1 = self.cell_stem_1(x_conv0, x_stem_0) + + x_cell_0 = self.cell_0(x_stem_1, x_stem_0) + x_cell_1 = self.cell_1(x_cell_0, x_stem_1) + x_cell_2 = self.cell_2(x_cell_1, x_cell_0) + x_cell_3 = self.cell_3(x_cell_2, x_cell_1) + + x_reduction_cell_0 = self.reduction_cell_0(x_cell_3, x_cell_2) + + x_cell_6 = self.cell_6(x_reduction_cell_0, x_cell_3) + x_cell_7 = self.cell_7(x_cell_6, x_reduction_cell_0) + x_cell_8 = self.cell_8(x_cell_7, x_cell_6) + x_cell_9 = self.cell_9(x_cell_8, x_cell_7) + + x_reduction_cell_1 = self.reduction_cell_1(x_cell_9, x_cell_8) + + x_cell_12 = self.cell_12(x_reduction_cell_1, x_cell_9) + x_cell_13 = self.cell_13(x_cell_12, x_reduction_cell_1) + x_cell_14 = self.cell_14(x_cell_13, x_cell_12) + x_cell_15 = self.cell_15(x_cell_14, x_cell_13) + + x_cell_15 = self.relu(x_cell_15) + x_cell_15 = F.avg_pool2d( + x_cell_15, + x_cell_15.size()[2:] + ) # global average pool + x_cell_15 = x_cell_15.view(x_cell_15.size(0), -1) + x_cell_15 = self.dropout(x_cell_15) + + return x_cell_15 + + def forward(self, input): + v = self.features(input) + + if not self.training: + return v + + y = self.classifier(v) + + if self.loss == 'softmax': + return y + elif self.loss == 'triplet': + return y, v + else: + raise KeyError('Unsupported loss: {}'.format(self.loss)) + + +def init_pretrained_weights(model, model_url): + """Initializes model with pretrained weights. + + Layers that don't match with pretrained layers in name or size are kept unchanged. + """ + pretrain_dict = model_zoo.load_url(model_url) + model_dict = model.state_dict() + pretrain_dict = { + k: v + for k, v in pretrain_dict.items() + if k in model_dict and model_dict[k].size() == v.size() + } + model_dict.update(pretrain_dict) + model.load_state_dict(model_dict) + + +def nasnetamobile(num_classes, loss='softmax', pretrained=True, **kwargs): + model = NASNetAMobile(num_classes, loss, **kwargs) + if pretrained: + model_url = pretrained_settings['nasnetamobile']['imagenet']['url'] + init_pretrained_weights(model, model_url) + return model diff --git a/strong_sort/deep/reid/torchreid/models/osnet.py b/strong_sort/deep/reid/torchreid/models/osnet.py new file mode 100644 index 0000000..b331b28 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/models/osnet.py @@ -0,0 +1,661 @@ +from __future__ import division, absolute_import +from copy import deepcopy +import warnings +import torch +from torch import nn +from torch.nn import functional as F +from torch.nn.modules import adaptive + +__all__ = [ + 'osnet_x1_0', 'osnet_x0_75', 'osnet_x0_5', 'osnet_x0_25', 'osnet_ibn_x1_0' +] + +pretrained_urls = { + 'osnet_x1_0': + 'https://drive.google.com/uc?id=1LaG1EJpHrxdAxKnSCJ_i0u-nbxSAeiFY', + 'osnet_x0_75': + 'https://drive.google.com/uc?id=1uwA9fElHOk3ZogwbeY5GkLI6QPTX70Hq', + 'osnet_x0_5': + 'https://drive.google.com/uc?id=16DGLbZukvVYgINws8u8deSaOqjybZ83i', + 'osnet_x0_25': + 'https://drive.google.com/uc?id=1rb8UN5ZzPKRc_xvtHlyDh-cSz88YX9hs', + 'osnet_ibn_x1_0': + 'https://drive.google.com/uc?id=1sr90V6irlYYDd4_4ISU2iruoRG8J__6l' +} + + +########## +# Basic layers +########## +class ConvLayer(nn.Module): + """Convolution layer (conv + bn + relu).""" + + def __init__( + self, + in_channels, + out_channels, + kernel_size, + stride=1, + padding=0, + groups=1, + IN=False + ): + super(ConvLayer, self).__init__() + self.conv = nn.Conv2d( + in_channels, + out_channels, + kernel_size, + stride=stride, + padding=padding, + bias=False, + groups=groups + ) + if IN: + self.bn = nn.InstanceNorm2d(out_channels, affine=True) + else: + self.bn = nn.BatchNorm2d(out_channels) + self.relu = nn.ReLU(inplace=True) + + def forward(self, x): + x = self.conv(x) + x = self.bn(x) + x = self.relu(x) + return x + + +class Conv1x1(nn.Module): + """1x1 convolution + bn + relu.""" + + def __init__(self, in_channels, out_channels, stride=1, groups=1): + super(Conv1x1, self).__init__() + self.conv = nn.Conv2d( + in_channels, + out_channels, + 1, + stride=stride, + padding=0, + bias=False, + groups=groups + ) + self.bn = nn.BatchNorm2d(out_channels) + self.relu = nn.ReLU(inplace=True) + + def forward(self, x): + x = self.conv(x) + x = self.bn(x) + x = self.relu(x) + return x + + +class Conv1x1Linear(nn.Module): + """1x1 convolution + bn (w/o non-linearity).""" + + def __init__(self, in_channels, out_channels, stride=1): + super(Conv1x1Linear, self).__init__() + self.conv = nn.Conv2d( + in_channels, out_channels, 1, stride=stride, padding=0, bias=False + ) + self.bn = nn.BatchNorm2d(out_channels) + + def forward(self, x): + x = self.conv(x) + x = self.bn(x) + return x + + +class Conv3x3(nn.Module): + """3x3 convolution + bn + relu.""" + + def __init__(self, in_channels, out_channels, stride=1, groups=1): + super(Conv3x3, self).__init__() + self.conv = nn.Conv2d( + in_channels, + out_channels, + 3, + stride=stride, + padding=1, + bias=False, + groups=groups + ) + self.bn = nn.BatchNorm2d(out_channels) + self.relu = nn.ReLU(inplace=True) + + def forward(self, x): + x = self.conv(x) + x = self.bn(x) + x = self.relu(x) + return x + + +class LightConv3x3(nn.Module): + """Lightweight 3x3 convolution. + + 1x1 (linear) + dw 3x3 (nonlinear). + """ + + def __init__(self, in_channels, out_channels): + super(LightConv3x3, self).__init__() + self.conv1 = nn.Conv2d( + in_channels, out_channels, 1, stride=1, padding=0, bias=False + ) + self.conv2 = nn.Conv2d( + out_channels, + out_channels, + 3, + stride=1, + padding=1, + bias=False, + groups=out_channels + ) + self.bn = nn.BatchNorm2d(out_channels) + self.relu = nn.ReLU(inplace=True) + + def forward(self, x): + x = self.conv1(x) + x = self.conv2(x) + x = self.bn(x) + x = self.relu(x) + return x + + +########## +# Building blocks for omni-scale feature learning +########## +class ChannelGate(nn.Module): + """A mini-network that generates channel-wise gates conditioned on input tensor.""" + + def __init__( + self, + in_channels, + num_gates=None, + return_gates=False, + gate_activation='sigmoid', + reduction=16, + layer_norm=False + ): + super(ChannelGate, self).__init__() + if num_gates is None: + num_gates = in_channels + self.return_gates = return_gates + self.global_avgpool = nn.AdaptiveAvgPool2d(1) + self.fc1 = nn.Conv2d( + in_channels, + in_channels // reduction, + kernel_size=1, + bias=True, + padding=0 + ) + self.norm1 = None + if layer_norm: + self.norm1 = nn.LayerNorm((in_channels // reduction, 1, 1)) + self.relu = nn.ReLU(inplace=True) + self.fc2 = nn.Conv2d( + in_channels // reduction, + num_gates, + kernel_size=1, + bias=True, + padding=0 + ) + if gate_activation == 'sigmoid': + self.gate_activation = nn.Sigmoid() + elif gate_activation == 'relu': + self.gate_activation = nn.ReLU(inplace=True) + elif gate_activation == 'linear': + self.gate_activation = None + else: + raise RuntimeError( + "Unknown gate activation: {}".format(gate_activation) + ) + + def forward(self, x): + input = x + x = self.global_avgpool(x) + x = self.fc1(x) + if self.norm1 is not None: + x = self.norm1(x) + x = self.relu(x) + x = self.fc2(x) + if self.gate_activation is not None: + x = self.gate_activation(x) + if self.return_gates: + return x + return input * x + + +class OSBlock(nn.Module): + """Omni-scale feature learning block.""" + + def __init__( + self, + in_channels, + out_channels, + IN=False, + bottleneck_reduction=4, + **kwargs + ): + super(OSBlock, self).__init__() + mid_channels = out_channels // bottleneck_reduction + self.conv1 = Conv1x1(in_channels, mid_channels) + self.conv2a = LightConv3x3(mid_channels, mid_channels) + self.conv2b = nn.Sequential( + LightConv3x3(mid_channels, mid_channels), + LightConv3x3(mid_channels, mid_channels), + ) + self.conv2c = nn.Sequential( + LightConv3x3(mid_channels, mid_channels), + LightConv3x3(mid_channels, mid_channels), + LightConv3x3(mid_channels, mid_channels), + ) + self.conv2d = nn.Sequential( + LightConv3x3(mid_channels, mid_channels), + LightConv3x3(mid_channels, mid_channels), + LightConv3x3(mid_channels, mid_channels), + LightConv3x3(mid_channels, mid_channels), + ) + self.gate = ChannelGate(mid_channels) + self.conv3 = Conv1x1Linear(mid_channels, out_channels) + self.downsample = None + if in_channels != out_channels: + self.downsample = Conv1x1Linear(in_channels, out_channels) + self.IN = None + if IN: + self.IN = nn.InstanceNorm2d(out_channels, affine=True) + + def forward(self, x): + identity = x + x1 = self.conv1(x) + x2a = self.conv2a(x1) + x2b = self.conv2b(x1) + x2c = self.conv2c(x1) + x2d = self.conv2d(x1) + x2 = self.gate(x2a) + self.gate(x2b) + self.gate(x2c) + self.gate(x2d) + x3 = self.conv3(x2) + if self.downsample is not None: + identity = self.downsample(identity) + out = x3 + identity + if self.IN is not None: + out = self.IN(out) + return F.relu(out) + + +########## +# Network architecture +########## +class OSNet(nn.Module): + """Omni-Scale Network. + + Reference: + - Zhou et al. Omni-Scale Feature Learning for Person Re-Identification. ICCV, 2019. + - Zhou et al. Learning Generalisable Omni-Scale Representations + for Person Re-Identification. TPAMI, 2021. + """ + + def __init__( + self, + num_classes, + blocks, + layers, + channels, + feature_dim=512, + loss='softmax', + IN=False, + **kwargs + ): + super(OSNet, self).__init__() + num_blocks = len(blocks) + assert num_blocks == len(layers) + assert num_blocks == len(channels) - 1 + self.loss = loss + self.feature_dim = feature_dim + + # convolutional backbone + self.conv1 = ConvLayer(3, channels[0], 7, stride=2, padding=3, IN=IN) + self.maxpool = nn.MaxPool2d(3, stride=2, padding=1) + self.conv2 = self._make_layer( + blocks[0], + layers[0], + channels[0], + channels[1], + reduce_spatial_size=True, + IN=IN + ) + self.conv3 = self._make_layer( + blocks[1], + layers[1], + channels[1], + channels[2], + reduce_spatial_size=True + ) + self.conv4 = self._make_layer( + blocks[2], + layers[2], + channels[2], + channels[3], + reduce_spatial_size=False + ) + self.conv5 = Conv1x1(channels[3], channels[3]) + self.global_avgpool = nn.AdaptiveAvgPool2d(1) + # fully connected layer + self.fc = self._construct_fc_layer( + self.feature_dim, channels[3], dropout_p=None + ) + # identity classification layer + self.classifier = nn.Linear(self.feature_dim, num_classes) + + self._init_params() + + def _make_layer( + self, + block, + layer, + in_channels, + out_channels, + reduce_spatial_size, + IN=False + ): + layers = [] + + layers.append(block(in_channels, out_channels, IN=IN)) + for i in range(1, layer): + layers.append(block(out_channels, out_channels, IN=IN)) + + if reduce_spatial_size: + layers.append( + nn.Sequential( + Conv1x1(out_channels, out_channels), + nn.AvgPool2d(2, stride=2) + ) + ) + + return nn.Sequential(*layers) + + def _construct_fc_layer(self, fc_dims, input_dim, dropout_p=None): + if fc_dims is None or fc_dims < 0: + self.feature_dim = input_dim + return None + + if isinstance(fc_dims, int): + fc_dims = [fc_dims] + + layers = [] + for dim in fc_dims: + layers.append(nn.Linear(input_dim, dim)) + layers.append(nn.BatchNorm1d(dim)) + layers.append(nn.ReLU(inplace=True)) + if dropout_p is not None: + layers.append(nn.Dropout(p=dropout_p)) + input_dim = dim + + self.feature_dim = fc_dims[-1] + + return nn.Sequential(*layers) + + def _init_params(self): + for m in self.modules(): + if isinstance(m, nn.Conv2d): + nn.init.kaiming_normal_( + m.weight, mode='fan_out', nonlinearity='relu' + ) + if m.bias is not None: + nn.init.constant_(m.bias, 0) + + elif isinstance(m, nn.BatchNorm2d): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + + elif isinstance(m, nn.BatchNorm1d): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + + elif isinstance(m, nn.Linear): + nn.init.normal_(m.weight, 0, 0.01) + if m.bias is not None: + nn.init.constant_(m.bias, 0) + + def featuremaps(self, x): + x = self.conv1(x) + x = self.maxpool(x) + x = self.conv2(x) + x = self.conv3(x) + x = self.conv4(x) + x = self.conv5(x) + return x + + def forward(self, x, return_featuremaps=False): + x = self.featuremaps(x) + if return_featuremaps: + return x + v = self.global_avgpool(x) + v = v.view(v.size(0), -1) + if self.fc is not None: + v = self.fc(v) + if not self.training: + return v + y = self.classifier(v) + if self.loss == 'softmax': + return y + elif self.loss == 'triplet': + return y, v + else: + raise KeyError("Unsupported loss: {}".format(self.loss)) + + def layer_extractor(self, x, layer, return_featuremaps=False): + if layer == 'conv1': + x = self.conv1(x) + return x + elif layer == 'maxpool': + x = self.conv1(x) + x = self.maxpool(x) + return x + elif layer == 'conv2': + x = self.conv1(x) + x = self.maxpool(x) + x = self.conv2(x) + return x + elif layer == 'conv3': + x = self.conv1(x) + x = self.maxpool(x) + x = self.conv2(x) + x = self.conv3(x) + return x + elif layer == 'conv4': + x = self.conv1(x) + x = self.maxpool(x) + x = self.conv2(x) + x = self.conv3(x) + x = self.conv4(x) + return x + elif layer == 'conv5': + x = self.conv1(x) + x = self.maxpool(x) + x = self.conv2(x) + x = self.conv3(x) + x = self.conv4(x) + x = self.conv5(x) + return x + elif layer == 'global_avgpool': + x = self.conv1(x) + x = self.maxpool(x) + x = self.conv2(x) + x = self.conv3(x) + x = self.conv4(x) + x = self.conv5(x) + x = self.global_avgpool(x) + return x + elif layer == 'fc': + x = self.conv1(x) + x = self.maxpool(x) + x = self.conv2(x) + x = self.conv3(x) + x = self.conv4(x) + y = x.clone().detach() + x = self.conv5(x) + x = self.global_avgpool(x) + x = x.view(x.size(0), -1) + x = self.fc(x) + return x, y + """ + adaptive_pool = torch.nn.AdaptiveAvgPool2d((16,16)) + out_conv1 = adaptive_pool(out_conv1) + return {'out_conv1': out_conv1, 'out_maxpool': out_maxpool, 'out_conv2': out_conv2, + 'out_conv3': out_conv3, 'out_conv4': out_conv4, 'out_featuremap': out_featuremap, + 'out_globalavg': out_globalavg, 'out_fc': out_fc} + """ +def init_pretrained_weights(model, key=''): + """Initializes model with pretrained weights. + + Layers that don't match with pretrained layers in name or size are kept unchanged. + """ + import os + import errno + import gdown + from collections import OrderedDict + + def _get_torch_home(): + ENV_TORCH_HOME = 'TORCH_HOME' + ENV_XDG_CACHE_HOME = 'XDG_CACHE_HOME' + DEFAULT_CACHE_DIR = '~/.cache' + torch_home = os.path.expanduser( + os.getenv( + ENV_TORCH_HOME, + os.path.join( + os.getenv(ENV_XDG_CACHE_HOME, DEFAULT_CACHE_DIR), 'torch' + ) + ) + ) + return torch_home + + torch_home = _get_torch_home() + model_dir = os.path.join(torch_home, 'checkpoints') + try: + os.makedirs(model_dir) + except OSError as e: + if e.errno == errno.EEXIST: + # Directory already exists, ignore. + pass + else: + # Unexpected OSError, re-raise. + raise + filename = key + '_imagenet.pth' + cached_file = os.path.join(model_dir, filename) + + if not os.path.exists(cached_file): + gdown.download(pretrained_urls[key], cached_file, quiet=False) + + state_dict = torch.load(cached_file) + model_dict = model.state_dict() + new_state_dict = OrderedDict() + matched_layers, discarded_layers = [], [] + + for k, v in state_dict.items(): + if k.startswith('module.'): + k = k[7:] # discard module. + + if k in model_dict and model_dict[k].size() == v.size(): + new_state_dict[k] = v + matched_layers.append(k) + else: + discarded_layers.append(k) + + model_dict.update(new_state_dict) + model.load_state_dict(model_dict) + + if len(matched_layers) == 0: + warnings.warn( + 'The pretrained weights from "{}" cannot be loaded, ' + 'please check the key names manually ' + '(** ignored and continue **)'.format(cached_file) + ) + else: + print( + 'Successfully loaded imagenet pretrained weights from "{}"'. + format(cached_file) + ) + if len(discarded_layers) > 0: + print( + '** The following layers are discarded ' + 'due to unmatched keys or layer size: {}'. + format(discarded_layers) + ) + + +########## +# Instantiation +########## +def osnet_x1_0(num_classes=1000, pretrained=True, loss='softmax', **kwargs): + # standard size (width x1.0) + model = OSNet( + num_classes, + blocks=[OSBlock, OSBlock, OSBlock], + layers=[2, 2, 2], + channels=[64, 256, 384, 512], + loss=loss, + **kwargs + ) + if pretrained: + init_pretrained_weights(model, key='osnet_x1_0') + return model + + +def osnet_x0_75(num_classes=1000, pretrained=True, loss='softmax', **kwargs): + # medium size (width x0.75) + model = OSNet( + num_classes, + blocks=[OSBlock, OSBlock, OSBlock], + layers=[2, 2, 2], + channels=[48, 192, 288, 384], + loss=loss, + **kwargs + ) + if pretrained: + init_pretrained_weights(model, key='osnet_x0_75') + return model + + +def osnet_x0_5(num_classes=1000, pretrained=True, loss='softmax', **kwargs): + # tiny size (width x0.5) + model = OSNet( + num_classes, + blocks=[OSBlock, OSBlock, OSBlock], + layers=[2, 2, 2], + channels=[32, 128, 192, 256], + loss=loss, + **kwargs + ) + if pretrained: + init_pretrained_weights(model, key='osnet_x0_5') + return model + + +def osnet_x0_25(num_classes=1000, pretrained=True, loss='softmax', **kwargs): + # very tiny size (width x0.25) + model = OSNet( + num_classes, + blocks=[OSBlock, OSBlock, OSBlock], + layers=[2, 2, 2], + channels=[16, 64, 96, 128], + loss=loss, + **kwargs + ) + if pretrained: + init_pretrained_weights(model, key='osnet_x0_25') + return model + + +def osnet_ibn_x1_0( + num_classes=1000, pretrained=True, loss='softmax', **kwargs +): + # standard size (width x1.0) + IBN layer + # Ref: Pan et al. Two at Once: Enhancing Learning and Generalization Capacities via IBN-Net. ECCV, 2018. + model = OSNet( + num_classes, + blocks=[OSBlock, OSBlock, OSBlock], + layers=[2, 2, 2], + channels=[64, 256, 384, 512], + loss=loss, + IN=True, + **kwargs + ) + if pretrained: + init_pretrained_weights(model, key='osnet_ibn_x1_0') + return model diff --git a/strong_sort/deep/reid/torchreid/models/osnet_ain.py b/strong_sort/deep/reid/torchreid/models/osnet_ain.py new file mode 100644 index 0000000..e532a6c --- /dev/null +++ b/strong_sort/deep/reid/torchreid/models/osnet_ain.py @@ -0,0 +1,629 @@ +from __future__ import division, absolute_import +import warnings +import torch +from torch import nn +from torch.nn import functional as F + +__all__ = [ + 'osnet_ain_x1_0', 'osnet_ain_x0_75', 'osnet_ain_x0_5', 'osnet_ain_x0_25' +] + +pretrained_urls = { + 'osnet_ain_x1_0': + 'https://drive.google.com/uc?id=1-CaioD9NaqbHK_kzSMW8VE4_3KcsRjEo', + 'osnet_ain_x0_75': + 'https://drive.google.com/uc?id=1apy0hpsMypqstfencdH-jKIUEFOW4xoM', + 'osnet_ain_x0_5': + 'https://drive.google.com/uc?id=1KusKvEYyKGDTUBVRxRiz55G31wkihB6l', + 'osnet_ain_x0_25': + 'https://drive.google.com/uc?id=1SxQt2AvmEcgWNhaRb2xC4rP6ZwVDP0Wt' +} + + +########## +# Basic layers +########## +class ConvLayer(nn.Module): + """Convolution layer (conv + bn + relu).""" + + def __init__( + self, + in_channels, + out_channels, + kernel_size, + stride=1, + padding=0, + groups=1, + IN=False + ): + super(ConvLayer, self).__init__() + self.conv = nn.Conv2d( + in_channels, + out_channels, + kernel_size, + stride=stride, + padding=padding, + bias=False, + groups=groups + ) + if IN: + self.bn = nn.InstanceNorm2d(out_channels, affine=True) + else: + self.bn = nn.BatchNorm2d(out_channels) + self.relu = nn.ReLU() + + def forward(self, x): + x = self.conv(x) + x = self.bn(x) + return self.relu(x) + + +class Conv1x1(nn.Module): + """1x1 convolution + bn + relu.""" + + def __init__(self, in_channels, out_channels, stride=1, groups=1): + super(Conv1x1, self).__init__() + self.conv = nn.Conv2d( + in_channels, + out_channels, + 1, + stride=stride, + padding=0, + bias=False, + groups=groups + ) + self.bn = nn.BatchNorm2d(out_channels) + self.relu = nn.ReLU() + + def forward(self, x): + x = self.conv(x) + x = self.bn(x) + return self.relu(x) + + +class Conv1x1Linear(nn.Module): + """1x1 convolution + bn (w/o non-linearity).""" + + def __init__(self, in_channels, out_channels, stride=1, bn=True): + super(Conv1x1Linear, self).__init__() + self.conv = nn.Conv2d( + in_channels, out_channels, 1, stride=stride, padding=0, bias=False + ) + self.bn = None + if bn: + self.bn = nn.BatchNorm2d(out_channels) + + def forward(self, x): + x = self.conv(x) + if self.bn is not None: + x = self.bn(x) + return x + + +class Conv3x3(nn.Module): + """3x3 convolution + bn + relu.""" + + def __init__(self, in_channels, out_channels, stride=1, groups=1): + super(Conv3x3, self).__init__() + self.conv = nn.Conv2d( + in_channels, + out_channels, + 3, + stride=stride, + padding=1, + bias=False, + groups=groups + ) + self.bn = nn.BatchNorm2d(out_channels) + self.relu = nn.ReLU() + + def forward(self, x): + x = self.conv(x) + x = self.bn(x) + return self.relu(x) + + +class LightConv3x3(nn.Module): + """Lightweight 3x3 convolution. + + 1x1 (linear) + dw 3x3 (nonlinear). + """ + + def __init__(self, in_channels, out_channels): + super(LightConv3x3, self).__init__() + self.conv1 = nn.Conv2d( + in_channels, out_channels, 1, stride=1, padding=0, bias=False + ) + self.conv2 = nn.Conv2d( + out_channels, + out_channels, + 3, + stride=1, + padding=1, + bias=False, + groups=out_channels + ) + self.bn = nn.BatchNorm2d(out_channels) + self.relu = nn.ReLU() + + def forward(self, x): + x = self.conv1(x) + x = self.conv2(x) + x = self.bn(x) + return self.relu(x) + + +class LightConvStream(nn.Module): + """Lightweight convolution stream.""" + + def __init__(self, in_channels, out_channels, depth): + super(LightConvStream, self).__init__() + assert depth >= 1, 'depth must be equal to or larger than 1, but got {}'.format( + depth + ) + layers = [] + layers += [LightConv3x3(in_channels, out_channels)] + for i in range(depth - 1): + layers += [LightConv3x3(out_channels, out_channels)] + self.layers = nn.Sequential(*layers) + + def forward(self, x): + return self.layers(x) + + +########## +# Building blocks for omni-scale feature learning +########## +class ChannelGate(nn.Module): + """A mini-network that generates channel-wise gates conditioned on input tensor.""" + + def __init__( + self, + in_channels, + num_gates=None, + return_gates=False, + gate_activation='sigmoid', + reduction=16, + layer_norm=False + ): + super(ChannelGate, self).__init__() + if num_gates is None: + num_gates = in_channels + self.return_gates = return_gates + self.global_avgpool = nn.AdaptiveAvgPool2d(1) + self.fc1 = nn.Conv2d( + in_channels, + in_channels // reduction, + kernel_size=1, + bias=True, + padding=0 + ) + self.norm1 = None + if layer_norm: + self.norm1 = nn.LayerNorm((in_channels // reduction, 1, 1)) + self.relu = nn.ReLU() + self.fc2 = nn.Conv2d( + in_channels // reduction, + num_gates, + kernel_size=1, + bias=True, + padding=0 + ) + if gate_activation == 'sigmoid': + self.gate_activation = nn.Sigmoid() + elif gate_activation == 'relu': + self.gate_activation = nn.ReLU() + elif gate_activation == 'linear': + self.gate_activation = None + else: + raise RuntimeError( + "Unknown gate activation: {}".format(gate_activation) + ) + + def forward(self, x): + input = x + x = self.global_avgpool(x) + x = self.fc1(x) + if self.norm1 is not None: + x = self.norm1(x) + x = self.relu(x) + x = self.fc2(x) + if self.gate_activation is not None: + x = self.gate_activation(x) + if self.return_gates: + return x + return input * x + + +class OSBlock(nn.Module): + """Omni-scale feature learning block.""" + + def __init__(self, in_channels, out_channels, reduction=4, T=4, **kwargs): + super(OSBlock, self).__init__() + assert T >= 1 + assert out_channels >= reduction and out_channels % reduction == 0 + mid_channels = out_channels // reduction + + self.conv1 = Conv1x1(in_channels, mid_channels) + self.conv2 = nn.ModuleList() + for t in range(1, T + 1): + self.conv2 += [LightConvStream(mid_channels, mid_channels, t)] + self.gate = ChannelGate(mid_channels) + self.conv3 = Conv1x1Linear(mid_channels, out_channels) + self.downsample = None + if in_channels != out_channels: + self.downsample = Conv1x1Linear(in_channels, out_channels) + + def forward(self, x): + identity = x + x1 = self.conv1(x) + x2 = 0 + for conv2_t in self.conv2: + x2_t = conv2_t(x1) + x2 = x2 + self.gate(x2_t) + x3 = self.conv3(x2) + if self.downsample is not None: + identity = self.downsample(identity) + out = x3 + identity + return F.relu(out) + + +class OSBlockINin(nn.Module): + """Omni-scale feature learning block with instance normalization.""" + + def __init__(self, in_channels, out_channels, reduction=4, T=4, **kwargs): + super(OSBlockINin, self).__init__() + assert T >= 1 + assert out_channels >= reduction and out_channels % reduction == 0 + mid_channels = out_channels // reduction + + self.conv1 = Conv1x1(in_channels, mid_channels) + self.conv2 = nn.ModuleList() + for t in range(1, T + 1): + self.conv2 += [LightConvStream(mid_channels, mid_channels, t)] + self.gate = ChannelGate(mid_channels) + self.conv3 = Conv1x1Linear(mid_channels, out_channels, bn=False) + self.downsample = None + if in_channels != out_channels: + self.downsample = Conv1x1Linear(in_channels, out_channels) + self.IN = nn.InstanceNorm2d(out_channels, affine=True) + + def forward(self, x): + identity = x + x1 = self.conv1(x) + x2 = 0 + for conv2_t in self.conv2: + x2_t = conv2_t(x1) + x2 = x2 + self.gate(x2_t) + x3 = self.conv3(x2) + x3 = self.IN(x3) # IN inside residual + if self.downsample is not None: + identity = self.downsample(identity) + out = x3 + identity + return F.relu(out) + + +########## +# Network architecture +########## +class OSNet(nn.Module): + """Omni-Scale Network. + + Reference: + - Zhou et al. Omni-Scale Feature Learning for Person Re-Identification. ICCV, 2019. + - Zhou et al. Learning Generalisable Omni-Scale Representations + for Person Re-Identification. TPAMI, 2021. + """ + + def __init__( + self, + num_classes, + blocks, + layers, + channels, + feature_dim=512, + loss='softmax', + conv1_IN=False, + **kwargs + ): + super(OSNet, self).__init__() + num_blocks = len(blocks) + assert num_blocks == len(layers) + assert num_blocks == len(channels) - 1 + self.loss = loss + self.feature_dim = feature_dim + + # convolutional backbone + self.conv1 = ConvLayer( + 3, channels[0], 7, stride=2, padding=3, IN=conv1_IN + ) + self.maxpool = nn.MaxPool2d(3, stride=2, padding=1) + self.conv2 = self._make_layer( + blocks[0], layers[0], channels[0], channels[1] + ) + self.pool2 = nn.Sequential( + Conv1x1(channels[1], channels[1]), nn.AvgPool2d(2, stride=2) + ) + self.conv3 = self._make_layer( + blocks[1], layers[1], channels[1], channels[2] + ) + self.pool3 = nn.Sequential( + Conv1x1(channels[2], channels[2]), nn.AvgPool2d(2, stride=2) + ) + self.conv4 = self._make_layer( + blocks[2], layers[2], channels[2], channels[3] + ) + self.conv5 = Conv1x1(channels[3], channels[3]) + self.global_avgpool = nn.AdaptiveAvgPool2d(1) + # fully connected layer + self.fc = self._construct_fc_layer( + self.feature_dim, channels[3], dropout_p=None + ) + # identity classification layer + self.classifier = nn.Linear(self.feature_dim, num_classes) + + self._init_params() + + def _make_layer(self, blocks, layer, in_channels, out_channels): + layers = [] + layers += [blocks[0](in_channels, out_channels)] + for i in range(1, len(blocks)): + layers += [blocks[i](out_channels, out_channels)] + return nn.Sequential(*layers) + + def _construct_fc_layer(self, fc_dims, input_dim, dropout_p=None): + if fc_dims is None or fc_dims < 0: + self.feature_dim = input_dim + return None + + if isinstance(fc_dims, int): + fc_dims = [fc_dims] + + layers = [] + for dim in fc_dims: + layers.append(nn.Linear(input_dim, dim)) + layers.append(nn.BatchNorm1d(dim)) + layers.append(nn.ReLU()) + if dropout_p is not None: + layers.append(nn.Dropout(p=dropout_p)) + input_dim = dim + + self.feature_dim = fc_dims[-1] + + return nn.Sequential(*layers) + + def _init_params(self): + for m in self.modules(): + if isinstance(m, nn.Conv2d): + nn.init.kaiming_normal_( + m.weight, mode='fan_out', nonlinearity='relu' + ) + if m.bias is not None: + nn.init.constant_(m.bias, 0) + + elif isinstance(m, nn.BatchNorm2d): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + + elif isinstance(m, nn.BatchNorm1d): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + + elif isinstance(m, nn.InstanceNorm2d): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + + elif isinstance(m, nn.Linear): + nn.init.normal_(m.weight, 0, 0.01) + if m.bias is not None: + nn.init.constant_(m.bias, 0) + + def featuremaps(self, x): + x = self.conv1(x) + x = self.maxpool(x) + x = self.conv2(x) + x = self.pool2(x) + x = self.conv3(x) + x = self.pool3(x) + x = self.conv4(x) + x = self.conv5(x) + return x + + def out_layers_extractor(self, x): + if True: + out_conv1 = self.conv1(x) + out_maxpool = self.maxpool(out_conv1) + out_conv2 = self.conv2(out_maxpool) + return out_conv2 + else: + out_conv1 = self.conv1(x) + out_maxpool = self.maxpool(out_conv1) + out_conv2 = self.conv2(out_maxpool) + x = self.pool2(out_conv2) + out_conv3 = self.conv3(x) + x = self.pool3(out_conv3) + x = self.conv4(x) + out_featuremap = self.conv5(x) + out_globalavg = self.global_avgpool(out_featuremap) + v = out_globalavg.view(out_globalavg.size(0), -1) + #out_fc = self.fc(v) + return out_conv1, out_maxpool, out_conv2, out_conv3, out_featuremap, out_globalavg + + def forward(self, x, return_featuremaps=False): + x = self.featuremaps(x) + if return_featuremaps: + return x + v = self.global_avgpool(x) + v = v.view(v.size(0), -1) + if self.fc is not None: + v = self.fc(v) + if not self.training: + return v + y = self.classifier(v) + if self.loss == 'softmax': + return y + elif self.loss == 'triplet': + return y, v + else: + raise KeyError("Unsupported loss: {}".format(self.loss)) + + +def init_pretrained_weights(model, key=''): + """Initializes model with pretrained weights. + + Layers that don't match with pretrained layers in name or size are kept unchanged. + """ + import os + import errno + import gdown + from collections import OrderedDict + + def _get_torch_home(): + ENV_TORCH_HOME = 'TORCH_HOME' + ENV_XDG_CACHE_HOME = 'XDG_CACHE_HOME' + DEFAULT_CACHE_DIR = '~/.cache' + torch_home = os.path.expanduser( + os.getenv( + ENV_TORCH_HOME, + os.path.join( + os.getenv(ENV_XDG_CACHE_HOME, DEFAULT_CACHE_DIR), 'torch' + ) + ) + ) + return torch_home + + torch_home = _get_torch_home() + model_dir = os.path.join(torch_home, 'checkpoints') + try: + os.makedirs(model_dir) + except OSError as e: + if e.errno == errno.EEXIST: + # Directory already exists, ignore. + pass + else: + # Unexpected OSError, re-raise. + raise + filename = key + '_imagenet.pth' + cached_file = os.path.join(model_dir, filename) + + if not os.path.exists(cached_file): + gdown.download(pretrained_urls[key], cached_file, quiet=False) + + state_dict = torch.load(cached_file) + model_dict = model.state_dict() + new_state_dict = OrderedDict() + matched_layers, discarded_layers = [], [] + + for k, v in state_dict.items(): + if k.startswith('module.'): + k = k[7:] # discard module. + + if k in model_dict and model_dict[k].size() == v.size(): + new_state_dict[k] = v + matched_layers.append(k) + else: + discarded_layers.append(k) + + model_dict.update(new_state_dict) + model.load_state_dict(model_dict) + + if len(matched_layers) == 0: + warnings.warn( + 'The pretrained weights from "{}" cannot be loaded, ' + 'please check the key names manually ' + '(** ignored and continue **)'.format(cached_file) + ) + else: + print( + 'Successfully loaded imagenet pretrained weights from "{}"'. + format(cached_file) + ) + if len(discarded_layers) > 0: + print( + '** The following layers are discarded ' + 'due to unmatched keys or layer size: {}'. + format(discarded_layers) + ) + + +########## +# Instantiation +########## +def osnet_ain_x1_0( + num_classes=1000, pretrained=True, loss='softmax', **kwargs +): + model = OSNet( + num_classes, + blocks=[ + [OSBlockINin, OSBlockINin], [OSBlock, OSBlockINin], + [OSBlockINin, OSBlock] + ], + layers=[2, 2, 2], + channels=[64, 256, 384, 512], + loss=loss, + conv1_IN=True, + **kwargs + ) + if pretrained: + init_pretrained_weights(model, key='osnet_ain_x1_0') + return model + + +def osnet_ain_x0_75( + num_classes=1000, pretrained=True, loss='softmax', **kwargs +): + model = OSNet( + num_classes, + blocks=[ + [OSBlockINin, OSBlockINin], [OSBlock, OSBlockINin], + [OSBlockINin, OSBlock] + ], + layers=[2, 2, 2], + channels=[48, 192, 288, 384], + loss=loss, + conv1_IN=True, + **kwargs + ) + if pretrained: + init_pretrained_weights(model, key='osnet_ain_x0_75') + return model + + +def osnet_ain_x0_5( + num_classes=1000, pretrained=True, loss='softmax', **kwargs +): + model = OSNet( + num_classes, + blocks=[ + [OSBlockINin, OSBlockINin], [OSBlock, OSBlockINin], + [OSBlockINin, OSBlock] + ], + layers=[2, 2, 2], + channels=[32, 128, 192, 256], + loss=loss, + conv1_IN=True, + **kwargs + ) + if pretrained: + init_pretrained_weights(model, key='osnet_ain_x0_5') + return model + + +def osnet_ain_x0_25( + num_classes=1000, pretrained=True, loss='softmax', **kwargs +): + model = OSNet( + num_classes, + blocks=[ + [OSBlockINin, OSBlockINin], [OSBlock, OSBlockINin], + [OSBlockINin, OSBlock] + ], + layers=[2, 2, 2], + channels=[16, 64, 96, 128], + loss=loss, + conv1_IN=True, + **kwargs + ) + if pretrained: + init_pretrained_weights(model, key='osnet_ain_x0_25') + return model diff --git a/strong_sort/deep/reid/torchreid/models/pcb.py b/strong_sort/deep/reid/torchreid/models/pcb.py new file mode 100644 index 0000000..92c7414 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/models/pcb.py @@ -0,0 +1,314 @@ +from __future__ import division, absolute_import +import torch.utils.model_zoo as model_zoo +from torch import nn +from torch.nn import functional as F + +__all__ = ['pcb_p6', 'pcb_p4'] + +model_urls = { + 'resnet18': 'https://download.pytorch.org/models/resnet18-5c106cde.pth', + 'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth', + 'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth', + 'resnet101': 'https://download.pytorch.org/models/resnet101-5d3b4d8f.pth', + 'resnet152': 'https://download.pytorch.org/models/resnet152-b121ed2d.pth', +} + + +def conv3x3(in_planes, out_planes, stride=1): + """3x3 convolution with padding""" + return nn.Conv2d( + in_planes, + out_planes, + kernel_size=3, + stride=stride, + padding=1, + bias=False + ) + + +class BasicBlock(nn.Module): + expansion = 1 + + def __init__(self, inplanes, planes, stride=1, downsample=None): + super(BasicBlock, self).__init__() + self.conv1 = conv3x3(inplanes, planes, stride) + self.bn1 = nn.BatchNorm2d(planes) + self.relu = nn.ReLU(inplace=True) + self.conv2 = conv3x3(planes, planes) + self.bn2 = nn.BatchNorm2d(planes) + self.downsample = downsample + self.stride = stride + + def forward(self, x): + residual = x + + out = self.conv1(x) + out = self.bn1(out) + out = self.relu(out) + + out = self.conv2(out) + out = self.bn2(out) + + if self.downsample is not None: + residual = self.downsample(x) + + out += residual + out = self.relu(out) + + return out + + +class Bottleneck(nn.Module): + expansion = 4 + + def __init__(self, inplanes, planes, stride=1, downsample=None): + super(Bottleneck, self).__init__() + self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False) + self.bn1 = nn.BatchNorm2d(planes) + self.conv2 = nn.Conv2d( + planes, + planes, + kernel_size=3, + stride=stride, + padding=1, + bias=False + ) + self.bn2 = nn.BatchNorm2d(planes) + self.conv3 = nn.Conv2d( + planes, planes * self.expansion, kernel_size=1, bias=False + ) + self.bn3 = nn.BatchNorm2d(planes * self.expansion) + self.relu = nn.ReLU(inplace=True) + self.downsample = downsample + self.stride = stride + + def forward(self, x): + residual = x + + out = self.conv1(x) + out = self.bn1(out) + out = self.relu(out) + + out = self.conv2(out) + out = self.bn2(out) + out = self.relu(out) + + out = self.conv3(out) + out = self.bn3(out) + + if self.downsample is not None: + residual = self.downsample(x) + + out += residual + out = self.relu(out) + + return out + + +class DimReduceLayer(nn.Module): + + def __init__(self, in_channels, out_channels, nonlinear): + super(DimReduceLayer, self).__init__() + layers = [] + layers.append( + nn.Conv2d( + in_channels, out_channels, 1, stride=1, padding=0, bias=False + ) + ) + layers.append(nn.BatchNorm2d(out_channels)) + + if nonlinear == 'relu': + layers.append(nn.ReLU(inplace=True)) + elif nonlinear == 'leakyrelu': + layers.append(nn.LeakyReLU(0.1)) + + self.layers = nn.Sequential(*layers) + + def forward(self, x): + return self.layers(x) + + +class PCB(nn.Module): + """Part-based Convolutional Baseline. + + Reference: + Sun et al. Beyond Part Models: Person Retrieval with Refined + Part Pooling (and A Strong Convolutional Baseline). ECCV 2018. + + Public keys: + - ``pcb_p4``: PCB with 4-part strips. + - ``pcb_p6``: PCB with 6-part strips. + """ + + def __init__( + self, + num_classes, + loss, + block, + layers, + parts=6, + reduced_dim=256, + nonlinear='relu', + **kwargs + ): + self.inplanes = 64 + super(PCB, self).__init__() + self.loss = loss + self.parts = parts + self.feature_dim = 512 * block.expansion + + # backbone network + self.conv1 = nn.Conv2d( + 3, 64, kernel_size=7, stride=2, padding=3, bias=False + ) + self.bn1 = nn.BatchNorm2d(64) + self.relu = nn.ReLU(inplace=True) + self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) + self.layer1 = self._make_layer(block, 64, layers[0]) + self.layer2 = self._make_layer(block, 128, layers[1], stride=2) + self.layer3 = self._make_layer(block, 256, layers[2], stride=2) + self.layer4 = self._make_layer(block, 512, layers[3], stride=1) + + # pcb layers + self.parts_avgpool = nn.AdaptiveAvgPool2d((self.parts, 1)) + self.dropout = nn.Dropout(p=0.5) + self.conv5 = DimReduceLayer( + 512 * block.expansion, reduced_dim, nonlinear=nonlinear + ) + self.feature_dim = reduced_dim + self.classifier = nn.ModuleList( + [ + nn.Linear(self.feature_dim, num_classes) + for _ in range(self.parts) + ] + ) + + self._init_params() + + def _make_layer(self, block, planes, blocks, stride=1): + downsample = None + if stride != 1 or self.inplanes != planes * block.expansion: + downsample = nn.Sequential( + nn.Conv2d( + self.inplanes, + planes * block.expansion, + kernel_size=1, + stride=stride, + bias=False + ), + nn.BatchNorm2d(planes * block.expansion), + ) + + layers = [] + layers.append(block(self.inplanes, planes, stride, downsample)) + self.inplanes = planes * block.expansion + for i in range(1, blocks): + layers.append(block(self.inplanes, planes)) + + return nn.Sequential(*layers) + + def _init_params(self): + for m in self.modules(): + if isinstance(m, nn.Conv2d): + nn.init.kaiming_normal_( + m.weight, mode='fan_out', nonlinearity='relu' + ) + if m.bias is not None: + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.BatchNorm2d): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.BatchNorm1d): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.Linear): + nn.init.normal_(m.weight, 0, 0.01) + if m.bias is not None: + nn.init.constant_(m.bias, 0) + + def featuremaps(self, x): + x = self.conv1(x) + x = self.bn1(x) + x = self.relu(x) + x = self.maxpool(x) + x = self.layer1(x) + x = self.layer2(x) + x = self.layer3(x) + x = self.layer4(x) + return x + + def forward(self, x): + f = self.featuremaps(x) + v_g = self.parts_avgpool(f) + + if not self.training: + v_g = F.normalize(v_g, p=2, dim=1) + return v_g.view(v_g.size(0), -1) + + v_g = self.dropout(v_g) + v_h = self.conv5(v_g) + + y = [] + for i in range(self.parts): + v_h_i = v_h[:, :, i, :] + v_h_i = v_h_i.view(v_h_i.size(0), -1) + y_i = self.classifier[i](v_h_i) + y.append(y_i) + + if self.loss == 'softmax': + return y + elif self.loss == 'triplet': + v_g = F.normalize(v_g, p=2, dim=1) + return y, v_g.view(v_g.size(0), -1) + else: + raise KeyError('Unsupported loss: {}'.format(self.loss)) + + +def init_pretrained_weights(model, model_url): + """Initializes model with pretrained weights. + + Layers that don't match with pretrained layers in name or size are kept unchanged. + """ + pretrain_dict = model_zoo.load_url(model_url) + model_dict = model.state_dict() + pretrain_dict = { + k: v + for k, v in pretrain_dict.items() + if k in model_dict and model_dict[k].size() == v.size() + } + model_dict.update(pretrain_dict) + model.load_state_dict(model_dict) + + +def pcb_p6(num_classes, loss='softmax', pretrained=True, **kwargs): + model = PCB( + num_classes=num_classes, + loss=loss, + block=Bottleneck, + layers=[3, 4, 6, 3], + last_stride=1, + parts=6, + reduced_dim=256, + nonlinear='relu', + **kwargs + ) + if pretrained: + init_pretrained_weights(model, model_urls['resnet50']) + return model + + +def pcb_p4(num_classes, loss='softmax', pretrained=True, **kwargs): + model = PCB( + num_classes=num_classes, + loss=loss, + block=Bottleneck, + layers=[3, 4, 6, 3], + last_stride=1, + parts=4, + reduced_dim=256, + nonlinear='relu', + **kwargs + ) + if pretrained: + init_pretrained_weights(model, model_urls['resnet50']) + return model diff --git a/strong_sort/deep/reid/torchreid/models/resnet.py b/strong_sort/deep/reid/torchreid/models/resnet.py new file mode 100644 index 0000000..d5e23e5 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/models/resnet.py @@ -0,0 +1,521 @@ +""" +Code source: https://github.com/pytorch/vision +""" +from __future__ import division, absolute_import +import torch.utils.model_zoo as model_zoo +from torch import nn + +__all__ = [ + 'resnet18', 'resnet34', 'resnet50', 'resnet101', 'resnet152', + 'resnext50_32x4d', 'resnext101_32x8d', 'resnet50_fc512' +] + +model_urls = { + 'resnet18': + 'https://download.pytorch.org/models/resnet18-5c106cde.pth', + 'resnet34': + 'https://download.pytorch.org/models/resnet34-333f7ec4.pth', + 'resnet50': + 'https://download.pytorch.org/models/resnet50-19c8e357.pth', + 'resnet101': + 'https://download.pytorch.org/models/resnet101-5d3b4d8f.pth', + 'resnet152': + 'https://download.pytorch.org/models/resnet152-b121ed2d.pth', + 'resnext50_32x4d': + 'https://download.pytorch.org/models/resnext50_32x4d-7cdf4587.pth', + 'resnext101_32x8d': + 'https://download.pytorch.org/models/resnext101_32x8d-8ba56ff5.pth', +} + + +def conv3x3(in_planes, out_planes, stride=1, groups=1, dilation=1): + """3x3 convolution with padding""" + return nn.Conv2d( + in_planes, + out_planes, + kernel_size=3, + stride=stride, + padding=dilation, + groups=groups, + bias=False, + dilation=dilation + ) + + +def conv1x1(in_planes, out_planes, stride=1): + """1x1 convolution""" + return nn.Conv2d( + in_planes, out_planes, kernel_size=1, stride=stride, bias=False + ) + + +class BasicBlock(nn.Module): + expansion = 1 + + def __init__( + self, + inplanes, + planes, + stride=1, + downsample=None, + groups=1, + base_width=64, + dilation=1, + norm_layer=None + ): + super(BasicBlock, self).__init__() + if norm_layer is None: + norm_layer = nn.BatchNorm2d + if groups != 1 or base_width != 64: + raise ValueError( + 'BasicBlock only supports groups=1 and base_width=64' + ) + if dilation > 1: + raise NotImplementedError( + "Dilation > 1 not supported in BasicBlock" + ) + # Both self.conv1 and self.downsample layers downsample the input when stride != 1 + self.conv1 = conv3x3(inplanes, planes, stride) + self.bn1 = norm_layer(planes) + self.relu = nn.ReLU(inplace=True) + self.conv2 = conv3x3(planes, planes) + self.bn2 = norm_layer(planes) + self.downsample = downsample + self.stride = stride + + def forward(self, x): + identity = x + + out = self.conv1(x) + out = self.bn1(out) + out = self.relu(out) + + out = self.conv2(out) + out = self.bn2(out) + + if self.downsample is not None: + identity = self.downsample(x) + + out += identity + out = self.relu(out) + + return out + + +class Bottleneck(nn.Module): + expansion = 4 + + def __init__( + self, + inplanes, + planes, + stride=1, + downsample=None, + groups=1, + base_width=64, + dilation=1, + norm_layer=None + ): + super(Bottleneck, self).__init__() + if norm_layer is None: + norm_layer = nn.BatchNorm2d + width = int(planes * (base_width/64.)) * groups + # Both self.conv2 and self.downsample layers downsample the input when stride != 1 + self.conv1 = conv1x1(inplanes, width) + self.bn1 = norm_layer(width) + self.conv2 = conv3x3(width, width, stride, groups, dilation) + self.bn2 = norm_layer(width) + self.conv3 = conv1x1(width, planes * self.expansion) + self.bn3 = norm_layer(planes * self.expansion) + self.relu = nn.ReLU(inplace=True) + self.downsample = downsample + self.stride = stride + + def forward(self, x): + identity = x + + out = self.conv1(x) + out = self.bn1(out) + out = self.relu(out) + + out = self.conv2(out) + out = self.bn2(out) + out = self.relu(out) + + out = self.conv3(out) + out = self.bn3(out) + + if self.downsample is not None: + identity = self.downsample(x) + + out += identity + out = self.relu(out) + + return out + + +class ResNet(nn.Module): + """Residual network. + + Reference: + - He et al. Deep Residual Learning for Image Recognition. CVPR 2016. + - Xie et al. Aggregated Residual Transformations for Deep Neural Networks. CVPR 2017. + + Public keys: + - ``resnet18``: ResNet18. + - ``resnet34``: ResNet34. + - ``resnet50``: ResNet50. + - ``resnet101``: ResNet101. + - ``resnet152``: ResNet152. + - ``resnext50_32x4d``: ResNeXt50. + - ``resnext101_32x8d``: ResNeXt101. + - ``resnet50_fc512``: ResNet50 + FC. + """ + + def __init__( + self, + num_classes, + loss, + block, + layers, + zero_init_residual=False, + groups=1, + width_per_group=64, + replace_stride_with_dilation=None, + norm_layer=None, + last_stride=2, + fc_dims=None, + dropout_p=None, + **kwargs + ): + super(ResNet, self).__init__() + if norm_layer is None: + norm_layer = nn.BatchNorm2d + self._norm_layer = norm_layer + self.loss = loss + self.feature_dim = 512 * block.expansion + self.inplanes = 64 + self.dilation = 1 + if replace_stride_with_dilation is None: + # each element in the tuple indicates if we should replace + # the 2x2 stride with a dilated convolution instead + replace_stride_with_dilation = [False, False, False] + if len(replace_stride_with_dilation) != 3: + raise ValueError( + "replace_stride_with_dilation should be None " + "or a 3-element tuple, got {}". + format(replace_stride_with_dilation) + ) + self.groups = groups + self.base_width = width_per_group + self.conv1 = nn.Conv2d( + 3, self.inplanes, kernel_size=7, stride=2, padding=3, bias=False + ) + self.bn1 = norm_layer(self.inplanes) + self.relu = nn.ReLU(inplace=True) + self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) + self.layer1 = self._make_layer(block, 64, layers[0]) + self.layer2 = self._make_layer( + block, + 128, + layers[1], + stride=2, + dilate=replace_stride_with_dilation[0] + ) + self.layer3 = self._make_layer( + block, + 256, + layers[2], + stride=2, + dilate=replace_stride_with_dilation[1] + ) + self.layer4 = self._make_layer( + block, + 512, + layers[3], + stride=last_stride, + dilate=replace_stride_with_dilation[2] + ) + self.global_avgpool = nn.AdaptiveAvgPool2d((1, 1)) + self.fc = self._construct_fc_layer( + fc_dims, 512 * block.expansion, dropout_p + ) + self.classifier = nn.Linear(self.feature_dim, num_classes) + + self._init_params() + + # Zero-initialize the last BN in each residual branch, + # so that the residual branch starts with zeros, and each residual block behaves like an identity. + # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677 + if zero_init_residual: + for m in self.modules(): + if isinstance(m, Bottleneck): + nn.init.constant_(m.bn3.weight, 0) + elif isinstance(m, BasicBlock): + nn.init.constant_(m.bn2.weight, 0) + + def _make_layer(self, block, planes, blocks, stride=1, dilate=False): + norm_layer = self._norm_layer + downsample = None + previous_dilation = self.dilation + if dilate: + self.dilation *= stride + stride = 1 + if stride != 1 or self.inplanes != planes * block.expansion: + downsample = nn.Sequential( + conv1x1(self.inplanes, planes * block.expansion, stride), + norm_layer(planes * block.expansion), + ) + + layers = [] + layers.append( + block( + self.inplanes, planes, stride, downsample, self.groups, + self.base_width, previous_dilation, norm_layer + ) + ) + self.inplanes = planes * block.expansion + for _ in range(1, blocks): + layers.append( + block( + self.inplanes, + planes, + groups=self.groups, + base_width=self.base_width, + dilation=self.dilation, + norm_layer=norm_layer + ) + ) + + return nn.Sequential(*layers) + + def _construct_fc_layer(self, fc_dims, input_dim, dropout_p=None): + """Constructs fully connected layer + + Args: + fc_dims (list or tuple): dimensions of fc layers, if None, no fc layers are constructed + input_dim (int): input dimension + dropout_p (float): dropout probability, if None, dropout is unused + """ + if fc_dims is None: + self.feature_dim = input_dim + return None + + assert isinstance( + fc_dims, (list, tuple) + ), 'fc_dims must be either list or tuple, but got {}'.format( + type(fc_dims) + ) + + layers = [] + for dim in fc_dims: + layers.append(nn.Linear(input_dim, dim)) + layers.append(nn.BatchNorm1d(dim)) + layers.append(nn.ReLU(inplace=True)) + if dropout_p is not None: + layers.append(nn.Dropout(p=dropout_p)) + input_dim = dim + + self.feature_dim = fc_dims[-1] + + return nn.Sequential(*layers) + + def _init_params(self): + for m in self.modules(): + if isinstance(m, nn.Conv2d): + nn.init.kaiming_normal_( + m.weight, mode='fan_out', nonlinearity='relu' + ) + if m.bias is not None: + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.BatchNorm2d): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.BatchNorm1d): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.Linear): + nn.init.normal_(m.weight, 0, 0.01) + if m.bias is not None: + nn.init.constant_(m.bias, 0) + + def featuremaps(self, x): + x = self.conv1(x) + x = self.bn1(x) + x = self.relu(x) + x = self.maxpool(x) + x = self.layer1(x) + x = self.layer2(x) + x = self.layer3(x) + x = self.layer4(x) + return x + + def forward(self, x): + f = self.featuremaps(x) + v = self.global_avgpool(f) + v = v.view(v.size(0), -1) + + if self.fc is not None: + v = self.fc(v) + + y = self.classifier(v) + + return y, v + +def init_pretrained_weights(model, model_url): + """Initializes model with pretrained weights. + + Layers that don't match with pretrained layers in name or size are kept unchanged. + """ + pretrain_dict = model_zoo.load_url(model_url) + model_dict = model.state_dict() + pretrain_dict = { + k: v + for k, v in pretrain_dict.items() + if k in model_dict and model_dict[k].size() == v.size() + } + model_dict.update(pretrain_dict) + model.load_state_dict(model_dict) + + +"""ResNet""" + + +def resnet18(num_classes, loss='softmax', pretrained=True, **kwargs): + model = ResNet( + num_classes=num_classes, + loss=loss, + block=BasicBlock, + layers=[2, 2, 2, 2], + last_stride=2, + fc_dims=None, + dropout_p=None, + **kwargs + ) + if pretrained: + init_pretrained_weights(model, model_urls['resnet18']) + return model + + +def resnet34(num_classes, loss='softmax', pretrained=True, **kwargs): + model = ResNet( + num_classes=num_classes, + loss=loss, + block=BasicBlock, + layers=[3, 4, 6, 3], + last_stride=2, + fc_dims=None, + dropout_p=None, + **kwargs + ) + if pretrained: + init_pretrained_weights(model, model_urls['resnet34']) + return model + + +def resnet50(num_classes, loss='softmax', pretrained=True, **kwargs): + model = ResNet( + num_classes=num_classes, + loss=loss, + block=Bottleneck, + layers=[3, 4, 6, 3], + last_stride=2, + fc_dims=None, + dropout_p=None, + **kwargs + ) + if pretrained: + init_pretrained_weights(model, model_urls['resnet50']) + return model + + +def resnet101(num_classes, loss='softmax', pretrained=True, **kwargs): + model = ResNet( + num_classes=num_classes, + loss=loss, + block=Bottleneck, + layers=[3, 4, 23, 3], + last_stride=2, + fc_dims=None, + dropout_p=None, + **kwargs + ) + if pretrained: + init_pretrained_weights(model, model_urls['resnet101']) + return model + + +def resnet152(num_classes, loss='softmax', pretrained=True, **kwargs): + model = ResNet( + num_classes=num_classes, + loss=loss, + block=Bottleneck, + layers=[3, 8, 36, 3], + last_stride=2, + fc_dims=None, + dropout_p=None, + **kwargs + ) + if pretrained: + init_pretrained_weights(model, model_urls['resnet152']) + return model + + +"""ResNeXt""" + + +def resnext50_32x4d(num_classes, loss='softmax', pretrained=True, **kwargs): + model = ResNet( + num_classes=num_classes, + loss=loss, + block=Bottleneck, + layers=[3, 4, 6, 3], + last_stride=2, + fc_dims=None, + dropout_p=None, + groups=32, + width_per_group=4, + **kwargs + ) + if pretrained: + init_pretrained_weights(model, model_urls['resnext50_32x4d']) + return model + + +def resnext101_32x8d(num_classes, loss='softmax', pretrained=True, **kwargs): + model = ResNet( + num_classes=num_classes, + loss=loss, + block=Bottleneck, + layers=[3, 4, 23, 3], + last_stride=2, + fc_dims=None, + dropout_p=None, + groups=32, + width_per_group=8, + **kwargs + ) + if pretrained: + init_pretrained_weights(model, model_urls['resnext101_32x8d']) + return model + + +""" +ResNet + FC +""" + + +def resnet50_fc512(num_classes, loss='softmax', pretrained=True, **kwargs): + model = ResNet( + num_classes=num_classes, + loss=loss, + block=Bottleneck, + layers=[3, 4, 6, 3], + last_stride=1, + fc_dims=[512], + dropout_p=None, + **kwargs + ) + if pretrained: + init_pretrained_weights(model, model_urls['resnet50']) + return model diff --git a/strong_sort/deep/reid/torchreid/models/resnet_ibn_a.py b/strong_sort/deep/reid/torchreid/models/resnet_ibn_a.py new file mode 100644 index 0000000..d198e7c --- /dev/null +++ b/strong_sort/deep/reid/torchreid/models/resnet_ibn_a.py @@ -0,0 +1,289 @@ +""" +Credit to https://github.com/XingangPan/IBN-Net. +""" +from __future__ import division, absolute_import +import math +import torch +import torch.nn as nn +import torch.utils.model_zoo as model_zoo + +__all__ = ['resnet50_ibn_a'] + +model_urls = { + 'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth', + 'resnet101': 'https://download.pytorch.org/models/resnet101-5d3b4d8f.pth', + 'resnet152': 'https://download.pytorch.org/models/resnet152-b121ed2d.pth', +} + + +def conv3x3(in_planes, out_planes, stride=1): + "3x3 convolution with padding" + return nn.Conv2d( + in_planes, + out_planes, + kernel_size=3, + stride=stride, + padding=1, + bias=False + ) + + +class BasicBlock(nn.Module): + expansion = 1 + + def __init__(self, inplanes, planes, stride=1, downsample=None): + super(BasicBlock, self).__init__() + self.conv1 = conv3x3(inplanes, planes, stride) + self.bn1 = nn.BatchNorm2d(planes) + self.relu = nn.ReLU(inplace=True) + self.conv2 = conv3x3(planes, planes) + self.bn2 = nn.BatchNorm2d(planes) + self.downsample = downsample + self.stride = stride + + def forward(self, x): + residual = x + + out = self.conv1(x) + out = self.bn1(out) + out = self.relu(out) + + out = self.conv2(out) + out = self.bn2(out) + + if self.downsample is not None: + residual = self.downsample(x) + + out += residual + out = self.relu(out) + + return out + + +class IBN(nn.Module): + + def __init__(self, planes): + super(IBN, self).__init__() + half1 = int(planes / 2) + self.half = half1 + half2 = planes - half1 + self.IN = nn.InstanceNorm2d(half1, affine=True) + self.BN = nn.BatchNorm2d(half2) + + def forward(self, x): + split = torch.split(x, self.half, 1) + out1 = self.IN(split[0].contiguous()) + out2 = self.BN(split[1].contiguous()) + out = torch.cat((out1, out2), 1) + return out + + +class Bottleneck(nn.Module): + expansion = 4 + + def __init__(self, inplanes, planes, ibn=False, stride=1, downsample=None): + super(Bottleneck, self).__init__() + self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False) + if ibn: + self.bn1 = IBN(planes) + else: + self.bn1 = nn.BatchNorm2d(planes) + self.conv2 = nn.Conv2d( + planes, + planes, + kernel_size=3, + stride=stride, + padding=1, + bias=False + ) + self.bn2 = nn.BatchNorm2d(planes) + self.conv3 = nn.Conv2d( + planes, planes * self.expansion, kernel_size=1, bias=False + ) + self.bn3 = nn.BatchNorm2d(planes * self.expansion) + self.relu = nn.ReLU(inplace=True) + self.downsample = downsample + self.stride = stride + + def forward(self, x): + residual = x + + out = self.conv1(x) + out = self.bn1(out) + out = self.relu(out) + + out = self.conv2(out) + out = self.bn2(out) + out = self.relu(out) + + out = self.conv3(out) + out = self.bn3(out) + + if self.downsample is not None: + residual = self.downsample(x) + + out += residual + out = self.relu(out) + + return out + + +class ResNet(nn.Module): + """Residual network + IBN layer. + + Reference: + - He et al. Deep Residual Learning for Image Recognition. CVPR 2016. + - Pan et al. Two at Once: Enhancing Learning and Generalization + Capacities via IBN-Net. ECCV 2018. + """ + + def __init__( + self, + block, + layers, + num_classes=1000, + loss='softmax', + fc_dims=None, + dropout_p=None, + **kwargs + ): + scale = 64 + self.inplanes = scale + super(ResNet, self).__init__() + self.loss = loss + self.feature_dim = scale * 8 * block.expansion + + self.conv1 = nn.Conv2d( + 3, scale, kernel_size=7, stride=2, padding=3, bias=False + ) + self.bn1 = nn.BatchNorm2d(scale) + self.relu = nn.ReLU(inplace=True) + self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) + self.layer1 = self._make_layer(block, scale, layers[0]) + self.layer2 = self._make_layer(block, scale * 2, layers[1], stride=2) + self.layer3 = self._make_layer(block, scale * 4, layers[2], stride=2) + self.layer4 = self._make_layer(block, scale * 8, layers[3], stride=2) + self.avgpool = nn.AdaptiveAvgPool2d((1, 1)) + self.fc = self._construct_fc_layer( + fc_dims, scale * 8 * block.expansion, dropout_p + ) + self.classifier = nn.Linear(self.feature_dim, num_classes) + + for m in self.modules(): + if isinstance(m, nn.Conv2d): + n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels + m.weight.data.normal_(0, math.sqrt(2. / n)) + elif isinstance(m, nn.BatchNorm2d): + m.weight.data.fill_(1) + m.bias.data.zero_() + elif isinstance(m, nn.InstanceNorm2d): + m.weight.data.fill_(1) + m.bias.data.zero_() + + def _make_layer(self, block, planes, blocks, stride=1): + downsample = None + if stride != 1 or self.inplanes != planes * block.expansion: + downsample = nn.Sequential( + nn.Conv2d( + self.inplanes, + planes * block.expansion, + kernel_size=1, + stride=stride, + bias=False + ), + nn.BatchNorm2d(planes * block.expansion), + ) + + layers = [] + ibn = True + if planes == 512: + ibn = False + layers.append(block(self.inplanes, planes, ibn, stride, downsample)) + self.inplanes = planes * block.expansion + for i in range(1, blocks): + layers.append(block(self.inplanes, planes, ibn)) + + return nn.Sequential(*layers) + + def _construct_fc_layer(self, fc_dims, input_dim, dropout_p=None): + """Constructs fully connected layer + + Args: + fc_dims (list or tuple): dimensions of fc layers, if None, no fc layers are constructed + input_dim (int): input dimension + dropout_p (float): dropout probability, if None, dropout is unused + """ + if fc_dims is None: + self.feature_dim = input_dim + return None + + assert isinstance( + fc_dims, (list, tuple) + ), 'fc_dims must be either list or tuple, but got {}'.format( + type(fc_dims) + ) + + layers = [] + for dim in fc_dims: + layers.append(nn.Linear(input_dim, dim)) + layers.append(nn.BatchNorm1d(dim)) + layers.append(nn.ReLU(inplace=True)) + if dropout_p is not None: + layers.append(nn.Dropout(p=dropout_p)) + input_dim = dim + + self.feature_dim = fc_dims[-1] + + return nn.Sequential(*layers) + + def featuremaps(self, x): + x = self.conv1(x) + x = self.bn1(x) + x = self.relu(x) + x = self.maxpool(x) + x = self.layer1(x) + x = self.layer2(x) + x = self.layer3(x) + x = self.layer4(x) + return x + + def forward(self, x): + f = self.featuremaps(x) + v = self.avgpool(f) + v = v.view(v.size(0), -1) + if self.fc is not None: + v = self.fc(v) + if not self.training: + return v + y = self.classifier(v) + if self.loss == 'softmax': + return y + elif self.loss == 'triplet': + return y, v + else: + raise KeyError("Unsupported loss: {}".format(self.loss)) + + +def init_pretrained_weights(model, model_url): + """Initializes model with pretrained weights. + + Layers that don't match with pretrained layers in name or size are kept unchanged. + """ + pretrain_dict = model_zoo.load_url(model_url) + model_dict = model.state_dict() + pretrain_dict = { + k: v + for k, v in pretrain_dict.items() + if k in model_dict and model_dict[k].size() == v.size() + } + model_dict.update(pretrain_dict) + model.load_state_dict(model_dict) + + +def resnet50_ibn_a(num_classes, loss='softmax', pretrained=False, **kwargs): + model = ResNet( + Bottleneck, [3, 4, 6, 3], num_classes=num_classes, loss=loss, **kwargs + ) + if pretrained: + init_pretrained_weights(model, model_urls['resnet50']) + return model diff --git a/strong_sort/deep/reid/torchreid/models/resnet_ibn_b.py b/strong_sort/deep/reid/torchreid/models/resnet_ibn_b.py new file mode 100644 index 0000000..9881cc7 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/models/resnet_ibn_b.py @@ -0,0 +1,274 @@ +""" +Credit to https://github.com/XingangPan/IBN-Net. +""" +from __future__ import division, absolute_import +import math +import torch.nn as nn +import torch.utils.model_zoo as model_zoo + +__all__ = ['resnet50_ibn_b'] + +model_urls = { + 'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth', + 'resnet101': 'https://download.pytorch.org/models/resnet101-5d3b4d8f.pth', + 'resnet152': 'https://download.pytorch.org/models/resnet152-b121ed2d.pth', +} + + +def conv3x3(in_planes, out_planes, stride=1): + "3x3 convolution with padding" + return nn.Conv2d( + in_planes, + out_planes, + kernel_size=3, + stride=stride, + padding=1, + bias=False + ) + + +class BasicBlock(nn.Module): + expansion = 1 + + def __init__(self, inplanes, planes, stride=1, downsample=None): + super(BasicBlock, self).__init__() + self.conv1 = conv3x3(inplanes, planes, stride) + self.bn1 = nn.BatchNorm2d(planes) + self.relu = nn.ReLU(inplace=True) + self.conv2 = conv3x3(planes, planes) + self.bn2 = nn.BatchNorm2d(planes) + self.downsample = downsample + self.stride = stride + + def forward(self, x): + residual = x + + out = self.conv1(x) + out = self.bn1(out) + out = self.relu(out) + + out = self.conv2(out) + out = self.bn2(out) + + if self.downsample is not None: + residual = self.downsample(x) + + out += residual + out = self.relu(out) + + return out + + +class Bottleneck(nn.Module): + expansion = 4 + + def __init__(self, inplanes, planes, stride=1, downsample=None, IN=False): + super(Bottleneck, self).__init__() + self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False) + self.bn1 = nn.BatchNorm2d(planes) + self.conv2 = nn.Conv2d( + planes, + planes, + kernel_size=3, + stride=stride, + padding=1, + bias=False + ) + self.bn2 = nn.BatchNorm2d(planes) + self.conv3 = nn.Conv2d( + planes, planes * self.expansion, kernel_size=1, bias=False + ) + self.bn3 = nn.BatchNorm2d(planes * self.expansion) + self.IN = None + if IN: + self.IN = nn.InstanceNorm2d(planes * 4, affine=True) + self.relu = nn.ReLU(inplace=True) + self.downsample = downsample + self.stride = stride + + def forward(self, x): + residual = x + + out = self.conv1(x) + out = self.bn1(out) + out = self.relu(out) + + out = self.conv2(out) + out = self.bn2(out) + out = self.relu(out) + + out = self.conv3(out) + out = self.bn3(out) + + if self.downsample is not None: + residual = self.downsample(x) + + out += residual + if self.IN is not None: + out = self.IN(out) + out = self.relu(out) + + return out + + +class ResNet(nn.Module): + """Residual network + IBN layer. + + Reference: + - He et al. Deep Residual Learning for Image Recognition. CVPR 2016. + - Pan et al. Two at Once: Enhancing Learning and Generalization + Capacities via IBN-Net. ECCV 2018. + """ + + def __init__( + self, + block, + layers, + num_classes=1000, + loss='softmax', + fc_dims=None, + dropout_p=None, + **kwargs + ): + scale = 64 + self.inplanes = scale + super(ResNet, self).__init__() + self.loss = loss + self.feature_dim = scale * 8 * block.expansion + + self.conv1 = nn.Conv2d( + 3, scale, kernel_size=7, stride=2, padding=3, bias=False + ) + self.bn1 = nn.InstanceNorm2d(scale, affine=True) + self.relu = nn.ReLU(inplace=True) + self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) + self.layer1 = self._make_layer( + block, scale, layers[0], stride=1, IN=True + ) + self.layer2 = self._make_layer( + block, scale * 2, layers[1], stride=2, IN=True + ) + self.layer3 = self._make_layer(block, scale * 4, layers[2], stride=2) + self.layer4 = self._make_layer(block, scale * 8, layers[3], stride=2) + self.avgpool = nn.AdaptiveAvgPool2d((1, 1)) + self.fc = self._construct_fc_layer( + fc_dims, scale * 8 * block.expansion, dropout_p + ) + self.classifier = nn.Linear(self.feature_dim, num_classes) + + for m in self.modules(): + if isinstance(m, nn.Conv2d): + n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels + m.weight.data.normal_(0, math.sqrt(2. / n)) + elif isinstance(m, nn.BatchNorm2d): + m.weight.data.fill_(1) + m.bias.data.zero_() + elif isinstance(m, nn.InstanceNorm2d): + m.weight.data.fill_(1) + m.bias.data.zero_() + + def _make_layer(self, block, planes, blocks, stride=1, IN=False): + downsample = None + if stride != 1 or self.inplanes != planes * block.expansion: + downsample = nn.Sequential( + nn.Conv2d( + self.inplanes, + planes * block.expansion, + kernel_size=1, + stride=stride, + bias=False + ), + nn.BatchNorm2d(planes * block.expansion), + ) + + layers = [] + layers.append(block(self.inplanes, planes, stride, downsample)) + self.inplanes = planes * block.expansion + for i in range(1, blocks - 1): + layers.append(block(self.inplanes, planes)) + layers.append(block(self.inplanes, planes, IN=IN)) + + return nn.Sequential(*layers) + + def _construct_fc_layer(self, fc_dims, input_dim, dropout_p=None): + """Constructs fully connected layer + + Args: + fc_dims (list or tuple): dimensions of fc layers, if None, no fc layers are constructed + input_dim (int): input dimension + dropout_p (float): dropout probability, if None, dropout is unused + """ + if fc_dims is None: + self.feature_dim = input_dim + return None + + assert isinstance( + fc_dims, (list, tuple) + ), 'fc_dims must be either list or tuple, but got {}'.format( + type(fc_dims) + ) + + layers = [] + for dim in fc_dims: + layers.append(nn.Linear(input_dim, dim)) + layers.append(nn.BatchNorm1d(dim)) + layers.append(nn.ReLU(inplace=True)) + if dropout_p is not None: + layers.append(nn.Dropout(p=dropout_p)) + input_dim = dim + + self.feature_dim = fc_dims[-1] + + return nn.Sequential(*layers) + + def featuremaps(self, x): + x = self.conv1(x) + x = self.bn1(x) + x = self.relu(x) + x = self.maxpool(x) + x = self.layer1(x) + x = self.layer2(x) + x = self.layer3(x) + x = self.layer4(x) + return x + + def forward(self, x): + f = self.featuremaps(x) + v = self.avgpool(f) + v = v.view(v.size(0), -1) + if self.fc is not None: + v = self.fc(v) + if not self.training: + return v + y = self.classifier(v) + if self.loss == 'softmax': + return y + elif self.loss == 'triplet': + return y, v + else: + raise KeyError("Unsupported loss: {}".format(self.loss)) + + +def init_pretrained_weights(model, model_url): + """Initializes model with pretrained weights. + + Layers that don't match with pretrained layers in name or size are kept unchanged. + """ + pretrain_dict = model_zoo.load_url(model_url) + model_dict = model.state_dict() + pretrain_dict = { + k: v + for k, v in pretrain_dict.items() + if k in model_dict and model_dict[k].size() == v.size() + } + model_dict.update(pretrain_dict) + model.load_state_dict(model_dict) + + +def resnet50_ibn_b(num_classes, loss='softmax', pretrained=False, **kwargs): + model = ResNet( + Bottleneck, [3, 4, 6, 3], num_classes=num_classes, loss=loss, **kwargs + ) + if pretrained: + init_pretrained_weights(model, model_urls['resnet50']) + return model diff --git a/strong_sort/deep/reid/torchreid/models/resnetmid.py b/strong_sort/deep/reid/torchreid/models/resnetmid.py new file mode 100644 index 0000000..017f6c6 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/models/resnetmid.py @@ -0,0 +1,307 @@ +from __future__ import division, absolute_import +import torch +import torch.utils.model_zoo as model_zoo +from torch import nn + +__all__ = ['resnet50mid'] + +model_urls = { + 'resnet18': 'https://download.pytorch.org/models/resnet18-5c106cde.pth', + 'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth', + 'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth', + 'resnet101': 'https://download.pytorch.org/models/resnet101-5d3b4d8f.pth', + 'resnet152': 'https://download.pytorch.org/models/resnet152-b121ed2d.pth', +} + + +def conv3x3(in_planes, out_planes, stride=1): + """3x3 convolution with padding""" + return nn.Conv2d( + in_planes, + out_planes, + kernel_size=3, + stride=stride, + padding=1, + bias=False + ) + + +class BasicBlock(nn.Module): + expansion = 1 + + def __init__(self, inplanes, planes, stride=1, downsample=None): + super(BasicBlock, self).__init__() + self.conv1 = conv3x3(inplanes, planes, stride) + self.bn1 = nn.BatchNorm2d(planes) + self.relu = nn.ReLU(inplace=True) + self.conv2 = conv3x3(planes, planes) + self.bn2 = nn.BatchNorm2d(planes) + self.downsample = downsample + self.stride = stride + + def forward(self, x): + residual = x + + out = self.conv1(x) + out = self.bn1(out) + out = self.relu(out) + + out = self.conv2(out) + out = self.bn2(out) + + if self.downsample is not None: + residual = self.downsample(x) + + out += residual + out = self.relu(out) + + return out + + +class Bottleneck(nn.Module): + expansion = 4 + + def __init__(self, inplanes, planes, stride=1, downsample=None): + super(Bottleneck, self).__init__() + self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False) + self.bn1 = nn.BatchNorm2d(planes) + self.conv2 = nn.Conv2d( + planes, + planes, + kernel_size=3, + stride=stride, + padding=1, + bias=False + ) + self.bn2 = nn.BatchNorm2d(planes) + self.conv3 = nn.Conv2d( + planes, planes * self.expansion, kernel_size=1, bias=False + ) + self.bn3 = nn.BatchNorm2d(planes * self.expansion) + self.relu = nn.ReLU(inplace=True) + self.downsample = downsample + self.stride = stride + + def forward(self, x): + residual = x + + out = self.conv1(x) + out = self.bn1(out) + out = self.relu(out) + + out = self.conv2(out) + out = self.bn2(out) + out = self.relu(out) + + out = self.conv3(out) + out = self.bn3(out) + + if self.downsample is not None: + residual = self.downsample(x) + + out += residual + out = self.relu(out) + + return out + + +class ResNetMid(nn.Module): + """Residual network + mid-level features. + + Reference: + Yu et al. The Devil is in the Middle: Exploiting Mid-level Representations for + Cross-Domain Instance Matching. arXiv:1711.08106. + + Public keys: + - ``resnet50mid``: ResNet50 + mid-level feature fusion. + """ + + def __init__( + self, + num_classes, + loss, + block, + layers, + last_stride=2, + fc_dims=None, + **kwargs + ): + self.inplanes = 64 + super(ResNetMid, self).__init__() + self.loss = loss + self.feature_dim = 512 * block.expansion + + # backbone network + self.conv1 = nn.Conv2d( + 3, 64, kernel_size=7, stride=2, padding=3, bias=False + ) + self.bn1 = nn.BatchNorm2d(64) + self.relu = nn.ReLU(inplace=True) + self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) + self.layer1 = self._make_layer(block, 64, layers[0]) + self.layer2 = self._make_layer(block, 128, layers[1], stride=2) + self.layer3 = self._make_layer(block, 256, layers[2], stride=2) + self.layer4 = self._make_layer( + block, 512, layers[3], stride=last_stride + ) + + self.global_avgpool = nn.AdaptiveAvgPool2d(1) + assert fc_dims is not None + self.fc_fusion = self._construct_fc_layer( + fc_dims, 512 * block.expansion * 2 + ) + self.feature_dim += 512 * block.expansion + self.classifier = nn.Linear(self.feature_dim, num_classes) + + self._init_params() + + def _make_layer(self, block, planes, blocks, stride=1): + downsample = None + if stride != 1 or self.inplanes != planes * block.expansion: + downsample = nn.Sequential( + nn.Conv2d( + self.inplanes, + planes * block.expansion, + kernel_size=1, + stride=stride, + bias=False + ), + nn.BatchNorm2d(planes * block.expansion), + ) + + layers = [] + layers.append(block(self.inplanes, planes, stride, downsample)) + self.inplanes = planes * block.expansion + for i in range(1, blocks): + layers.append(block(self.inplanes, planes)) + + return nn.Sequential(*layers) + + def _construct_fc_layer(self, fc_dims, input_dim, dropout_p=None): + """Constructs fully connected layer + + Args: + fc_dims (list or tuple): dimensions of fc layers, if None, no fc layers are constructed + input_dim (int): input dimension + dropout_p (float): dropout probability, if None, dropout is unused + """ + if fc_dims is None: + self.feature_dim = input_dim + return None + + assert isinstance( + fc_dims, (list, tuple) + ), 'fc_dims must be either list or tuple, but got {}'.format( + type(fc_dims) + ) + + layers = [] + for dim in fc_dims: + layers.append(nn.Linear(input_dim, dim)) + layers.append(nn.BatchNorm1d(dim)) + layers.append(nn.ReLU(inplace=True)) + if dropout_p is not None: + layers.append(nn.Dropout(p=dropout_p)) + input_dim = dim + + self.feature_dim = fc_dims[-1] + + return nn.Sequential(*layers) + + def _init_params(self): + for m in self.modules(): + if isinstance(m, nn.Conv2d): + nn.init.kaiming_normal_( + m.weight, mode='fan_out', nonlinearity='relu' + ) + if m.bias is not None: + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.BatchNorm2d): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.BatchNorm1d): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.Linear): + nn.init.normal_(m.weight, 0, 0.01) + if m.bias is not None: + nn.init.constant_(m.bias, 0) + + def featuremaps(self, x): + x = self.conv1(x) + x = self.bn1(x) + x = self.relu(x) + x = self.maxpool(x) + x = self.layer1(x) + x = self.layer2(x) + x = self.layer3(x) + x4a = self.layer4[0](x) + x4b = self.layer4[1](x4a) + x4c = self.layer4[2](x4b) + return x4a, x4b, x4c + + def forward(self, x): + x4a, x4b, x4c = self.featuremaps(x) + + v4a = self.global_avgpool(x4a) + v4b = self.global_avgpool(x4b) + v4c = self.global_avgpool(x4c) + v4ab = torch.cat([v4a, v4b], 1) + v4ab = v4ab.view(v4ab.size(0), -1) + v4ab = self.fc_fusion(v4ab) + v4c = v4c.view(v4c.size(0), -1) + v = torch.cat([v4ab, v4c], 1) + + if not self.training: + return v + + y = self.classifier(v) + + if self.loss == 'softmax': + return y + elif self.loss == 'triplet': + return y, v + else: + raise KeyError('Unsupported loss: {}'.format(self.loss)) + + +def init_pretrained_weights(model, model_url): + """Initializes model with pretrained weights. + + Layers that don't match with pretrained layers in name or size are kept unchanged. + """ + pretrain_dict = model_zoo.load_url(model_url) + model_dict = model.state_dict() + pretrain_dict = { + k: v + for k, v in pretrain_dict.items() + if k in model_dict and model_dict[k].size() == v.size() + } + model_dict.update(pretrain_dict) + model.load_state_dict(model_dict) + + +""" +Residual network configurations: +-- +resnet18: block=BasicBlock, layers=[2, 2, 2, 2] +resnet34: block=BasicBlock, layers=[3, 4, 6, 3] +resnet50: block=Bottleneck, layers=[3, 4, 6, 3] +resnet101: block=Bottleneck, layers=[3, 4, 23, 3] +resnet152: block=Bottleneck, layers=[3, 8, 36, 3] +""" + + +def resnet50mid(num_classes, loss='softmax', pretrained=True, **kwargs): + model = ResNetMid( + num_classes=num_classes, + loss=loss, + block=Bottleneck, + layers=[3, 4, 6, 3], + last_stride=2, + fc_dims=[1024], + **kwargs + ) + if pretrained: + init_pretrained_weights(model, model_urls['resnet50']) + return model diff --git a/strong_sort/deep/reid/torchreid/models/senet.py b/strong_sort/deep/reid/torchreid/models/senet.py new file mode 100644 index 0000000..baaf9b0 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/models/senet.py @@ -0,0 +1,688 @@ +from __future__ import division, absolute_import +import math +from collections import OrderedDict +import torch.nn as nn +from torch.utils import model_zoo + +__all__ = [ + 'senet154', 'se_resnet50', 'se_resnet101', 'se_resnet152', + 'se_resnext50_32x4d', 'se_resnext101_32x4d', 'se_resnet50_fc512' +] +""" +Code imported from https://github.com/Cadene/pretrained-models.pytorch +""" + +pretrained_settings = { + 'senet154': { + 'imagenet': { + 'url': + 'http://data.lip6.fr/cadene/pretrainedmodels/senet154-c7b49a05.pth', + 'input_space': 'RGB', + 'input_size': [3, 224, 224], + 'input_range': [0, 1], + 'mean': [0.485, 0.456, 0.406], + 'std': [0.229, 0.224, 0.225], + 'num_classes': 1000 + } + }, + 'se_resnet50': { + 'imagenet': { + 'url': + 'http://data.lip6.fr/cadene/pretrainedmodels/se_resnet50-ce0d4300.pth', + 'input_space': 'RGB', + 'input_size': [3, 224, 224], + 'input_range': [0, 1], + 'mean': [0.485, 0.456, 0.406], + 'std': [0.229, 0.224, 0.225], + 'num_classes': 1000 + } + }, + 'se_resnet101': { + 'imagenet': { + 'url': + 'http://data.lip6.fr/cadene/pretrainedmodels/se_resnet101-7e38fcc6.pth', + 'input_space': 'RGB', + 'input_size': [3, 224, 224], + 'input_range': [0, 1], + 'mean': [0.485, 0.456, 0.406], + 'std': [0.229, 0.224, 0.225], + 'num_classes': 1000 + } + }, + 'se_resnet152': { + 'imagenet': { + 'url': + 'http://data.lip6.fr/cadene/pretrainedmodels/se_resnet152-d17c99b7.pth', + 'input_space': 'RGB', + 'input_size': [3, 224, 224], + 'input_range': [0, 1], + 'mean': [0.485, 0.456, 0.406], + 'std': [0.229, 0.224, 0.225], + 'num_classes': 1000 + } + }, + 'se_resnext50_32x4d': { + 'imagenet': { + 'url': + 'http://data.lip6.fr/cadene/pretrainedmodels/se_resnext50_32x4d-a260b3a4.pth', + 'input_space': 'RGB', + 'input_size': [3, 224, 224], + 'input_range': [0, 1], + 'mean': [0.485, 0.456, 0.406], + 'std': [0.229, 0.224, 0.225], + 'num_classes': 1000 + } + }, + 'se_resnext101_32x4d': { + 'imagenet': { + 'url': + 'http://data.lip6.fr/cadene/pretrainedmodels/se_resnext101_32x4d-3b2fe3d8.pth', + 'input_space': 'RGB', + 'input_size': [3, 224, 224], + 'input_range': [0, 1], + 'mean': [0.485, 0.456, 0.406], + 'std': [0.229, 0.224, 0.225], + 'num_classes': 1000 + } + }, +} + + +class SEModule(nn.Module): + + def __init__(self, channels, reduction): + super(SEModule, self).__init__() + self.avg_pool = nn.AdaptiveAvgPool2d(1) + self.fc1 = nn.Conv2d( + channels, channels // reduction, kernel_size=1, padding=0 + ) + self.relu = nn.ReLU(inplace=True) + self.fc2 = nn.Conv2d( + channels // reduction, channels, kernel_size=1, padding=0 + ) + self.sigmoid = nn.Sigmoid() + + def forward(self, x): + module_input = x + x = self.avg_pool(x) + x = self.fc1(x) + x = self.relu(x) + x = self.fc2(x) + x = self.sigmoid(x) + return module_input * x + + +class Bottleneck(nn.Module): + """ + Base class for bottlenecks that implements `forward()` method. + """ + + def forward(self, x): + residual = x + + out = self.conv1(x) + out = self.bn1(out) + out = self.relu(out) + + out = self.conv2(out) + out = self.bn2(out) + out = self.relu(out) + + out = self.conv3(out) + out = self.bn3(out) + + if self.downsample is not None: + residual = self.downsample(x) + + out = self.se_module(out) + residual + out = self.relu(out) + + return out + + +class SEBottleneck(Bottleneck): + """ + Bottleneck for SENet154. + """ + expansion = 4 + + def __init__( + self, inplanes, planes, groups, reduction, stride=1, downsample=None + ): + super(SEBottleneck, self).__init__() + self.conv1 = nn.Conv2d(inplanes, planes * 2, kernel_size=1, bias=False) + self.bn1 = nn.BatchNorm2d(planes * 2) + self.conv2 = nn.Conv2d( + planes * 2, + planes * 4, + kernel_size=3, + stride=stride, + padding=1, + groups=groups, + bias=False + ) + self.bn2 = nn.BatchNorm2d(planes * 4) + self.conv3 = nn.Conv2d( + planes * 4, planes * 4, kernel_size=1, bias=False + ) + self.bn3 = nn.BatchNorm2d(planes * 4) + self.relu = nn.ReLU(inplace=True) + self.se_module = SEModule(planes * 4, reduction=reduction) + self.downsample = downsample + self.stride = stride + + +class SEResNetBottleneck(Bottleneck): + """ + ResNet bottleneck with a Squeeze-and-Excitation module. It follows Caffe + implementation and uses `stride=stride` in `conv1` and not in `conv2` + (the latter is used in the torchvision implementation of ResNet). + """ + expansion = 4 + + def __init__( + self, inplanes, planes, groups, reduction, stride=1, downsample=None + ): + super(SEResNetBottleneck, self).__init__() + self.conv1 = nn.Conv2d( + inplanes, planes, kernel_size=1, bias=False, stride=stride + ) + self.bn1 = nn.BatchNorm2d(planes) + self.conv2 = nn.Conv2d( + planes, + planes, + kernel_size=3, + padding=1, + groups=groups, + bias=False + ) + self.bn2 = nn.BatchNorm2d(planes) + self.conv3 = nn.Conv2d(planes, planes * 4, kernel_size=1, bias=False) + self.bn3 = nn.BatchNorm2d(planes * 4) + self.relu = nn.ReLU(inplace=True) + self.se_module = SEModule(planes * 4, reduction=reduction) + self.downsample = downsample + self.stride = stride + + +class SEResNeXtBottleneck(Bottleneck): + """ResNeXt bottleneck type C with a Squeeze-and-Excitation module""" + expansion = 4 + + def __init__( + self, + inplanes, + planes, + groups, + reduction, + stride=1, + downsample=None, + base_width=4 + ): + super(SEResNeXtBottleneck, self).__init__() + width = int(math.floor(planes * (base_width/64.)) * groups) + self.conv1 = nn.Conv2d( + inplanes, width, kernel_size=1, bias=False, stride=1 + ) + self.bn1 = nn.BatchNorm2d(width) + self.conv2 = nn.Conv2d( + width, + width, + kernel_size=3, + stride=stride, + padding=1, + groups=groups, + bias=False + ) + self.bn2 = nn.BatchNorm2d(width) + self.conv3 = nn.Conv2d(width, planes * 4, kernel_size=1, bias=False) + self.bn3 = nn.BatchNorm2d(planes * 4) + self.relu = nn.ReLU(inplace=True) + self.se_module = SEModule(planes * 4, reduction=reduction) + self.downsample = downsample + self.stride = stride + + +class SENet(nn.Module): + """Squeeze-and-excitation network. + + Reference: + Hu et al. Squeeze-and-Excitation Networks. CVPR 2018. + + Public keys: + - ``senet154``: SENet154. + - ``se_resnet50``: ResNet50 + SE. + - ``se_resnet101``: ResNet101 + SE. + - ``se_resnet152``: ResNet152 + SE. + - ``se_resnext50_32x4d``: ResNeXt50 (groups=32, width=4) + SE. + - ``se_resnext101_32x4d``: ResNeXt101 (groups=32, width=4) + SE. + - ``se_resnet50_fc512``: (ResNet50 + SE) + FC. + """ + + def __init__( + self, + num_classes, + loss, + block, + layers, + groups, + reduction, + dropout_p=0.2, + inplanes=128, + input_3x3=True, + downsample_kernel_size=3, + downsample_padding=1, + last_stride=2, + fc_dims=None, + **kwargs + ): + """ + Parameters + ---------- + block (nn.Module): Bottleneck class. + - For SENet154: SEBottleneck + - For SE-ResNet models: SEResNetBottleneck + - For SE-ResNeXt models: SEResNeXtBottleneck + layers (list of ints): Number of residual blocks for 4 layers of the + network (layer1...layer4). + groups (int): Number of groups for the 3x3 convolution in each + bottleneck block. + - For SENet154: 64 + - For SE-ResNet models: 1 + - For SE-ResNeXt models: 32 + reduction (int): Reduction ratio for Squeeze-and-Excitation modules. + - For all models: 16 + dropout_p (float or None): Drop probability for the Dropout layer. + If `None` the Dropout layer is not used. + - For SENet154: 0.2 + - For SE-ResNet models: None + - For SE-ResNeXt models: None + inplanes (int): Number of input channels for layer1. + - For SENet154: 128 + - For SE-ResNet models: 64 + - For SE-ResNeXt models: 64 + input_3x3 (bool): If `True`, use three 3x3 convolutions instead of + a single 7x7 convolution in layer0. + - For SENet154: True + - For SE-ResNet models: False + - For SE-ResNeXt models: False + downsample_kernel_size (int): Kernel size for downsampling convolutions + in layer2, layer3 and layer4. + - For SENet154: 3 + - For SE-ResNet models: 1 + - For SE-ResNeXt models: 1 + downsample_padding (int): Padding for downsampling convolutions in + layer2, layer3 and layer4. + - For SENet154: 1 + - For SE-ResNet models: 0 + - For SE-ResNeXt models: 0 + num_classes (int): Number of outputs in `classifier` layer. + """ + super(SENet, self).__init__() + self.inplanes = inplanes + self.loss = loss + + if input_3x3: + layer0_modules = [ + ( + 'conv1', + nn.Conv2d(3, 64, 3, stride=2, padding=1, bias=False) + ), + ('bn1', nn.BatchNorm2d(64)), + ('relu1', nn.ReLU(inplace=True)), + ( + 'conv2', + nn.Conv2d(64, 64, 3, stride=1, padding=1, bias=False) + ), + ('bn2', nn.BatchNorm2d(64)), + ('relu2', nn.ReLU(inplace=True)), + ( + 'conv3', + nn.Conv2d( + 64, inplanes, 3, stride=1, padding=1, bias=False + ) + ), + ('bn3', nn.BatchNorm2d(inplanes)), + ('relu3', nn.ReLU(inplace=True)), + ] + else: + layer0_modules = [ + ( + 'conv1', + nn.Conv2d( + 3, + inplanes, + kernel_size=7, + stride=2, + padding=3, + bias=False + ) + ), + ('bn1', nn.BatchNorm2d(inplanes)), + ('relu1', nn.ReLU(inplace=True)), + ] + # To preserve compatibility with Caffe weights `ceil_mode=True` + # is used instead of `padding=1`. + layer0_modules.append( + ('pool', nn.MaxPool2d(3, stride=2, ceil_mode=True)) + ) + self.layer0 = nn.Sequential(OrderedDict(layer0_modules)) + self.layer1 = self._make_layer( + block, + planes=64, + blocks=layers[0], + groups=groups, + reduction=reduction, + downsample_kernel_size=1, + downsample_padding=0 + ) + self.layer2 = self._make_layer( + block, + planes=128, + blocks=layers[1], + stride=2, + groups=groups, + reduction=reduction, + downsample_kernel_size=downsample_kernel_size, + downsample_padding=downsample_padding + ) + self.layer3 = self._make_layer( + block, + planes=256, + blocks=layers[2], + stride=2, + groups=groups, + reduction=reduction, + downsample_kernel_size=downsample_kernel_size, + downsample_padding=downsample_padding + ) + self.layer4 = self._make_layer( + block, + planes=512, + blocks=layers[3], + stride=last_stride, + groups=groups, + reduction=reduction, + downsample_kernel_size=downsample_kernel_size, + downsample_padding=downsample_padding + ) + + self.global_avgpool = nn.AdaptiveAvgPool2d(1) + self.fc = self._construct_fc_layer( + fc_dims, 512 * block.expansion, dropout_p + ) + self.classifier = nn.Linear(self.feature_dim, num_classes) + + def _make_layer( + self, + block, + planes, + blocks, + groups, + reduction, + stride=1, + downsample_kernel_size=1, + downsample_padding=0 + ): + downsample = None + if stride != 1 or self.inplanes != planes * block.expansion: + downsample = nn.Sequential( + nn.Conv2d( + self.inplanes, + planes * block.expansion, + kernel_size=downsample_kernel_size, + stride=stride, + padding=downsample_padding, + bias=False + ), + nn.BatchNorm2d(planes * block.expansion), + ) + + layers = [] + layers.append( + block( + self.inplanes, planes, groups, reduction, stride, downsample + ) + ) + self.inplanes = planes * block.expansion + for i in range(1, blocks): + layers.append(block(self.inplanes, planes, groups, reduction)) + + return nn.Sequential(*layers) + + def _construct_fc_layer(self, fc_dims, input_dim, dropout_p=None): + """ + Construct fully connected layer + + - fc_dims (list or tuple): dimensions of fc layers, if None, + no fc layers are constructed + - input_dim (int): input dimension + - dropout_p (float): dropout probability, if None, dropout is unused + """ + if fc_dims is None: + self.feature_dim = input_dim + return None + + assert isinstance( + fc_dims, (list, tuple) + ), 'fc_dims must be either list or tuple, but got {}'.format( + type(fc_dims) + ) + + layers = [] + for dim in fc_dims: + layers.append(nn.Linear(input_dim, dim)) + layers.append(nn.BatchNorm1d(dim)) + layers.append(nn.ReLU(inplace=True)) + if dropout_p is not None: + layers.append(nn.Dropout(p=dropout_p)) + input_dim = dim + + self.feature_dim = fc_dims[-1] + + return nn.Sequential(*layers) + + def featuremaps(self, x): + x = self.layer0(x) + x = self.layer1(x) + x = self.layer2(x) + x = self.layer3(x) + x = self.layer4(x) + return x + + def forward(self, x): + f = self.featuremaps(x) + v = self.global_avgpool(f) + v = v.view(v.size(0), -1) + + if self.fc is not None: + v = self.fc(v) + + if not self.training: + return v + + y = self.classifier(v) + + if self.loss == 'softmax': + return y + elif self.loss == 'triplet': + return y, v + else: + raise KeyError("Unsupported loss: {}".format(self.loss)) + + +def init_pretrained_weights(model, model_url): + """Initializes model with pretrained weights. + + Layers that don't match with pretrained layers in name or size are kept unchanged. + """ + pretrain_dict = model_zoo.load_url(model_url) + model_dict = model.state_dict() + pretrain_dict = { + k: v + for k, v in pretrain_dict.items() + if k in model_dict and model_dict[k].size() == v.size() + } + model_dict.update(pretrain_dict) + model.load_state_dict(model_dict) + + +def senet154(num_classes, loss='softmax', pretrained=True, **kwargs): + model = SENet( + num_classes=num_classes, + loss=loss, + block=SEBottleneck, + layers=[3, 8, 36, 3], + groups=64, + reduction=16, + dropout_p=0.2, + last_stride=2, + fc_dims=None, + **kwargs + ) + if pretrained: + model_url = pretrained_settings['senet154']['imagenet']['url'] + init_pretrained_weights(model, model_url) + return model + + +def se_resnet50(num_classes, loss='softmax', pretrained=True, **kwargs): + model = SENet( + num_classes=num_classes, + loss=loss, + block=SEResNetBottleneck, + layers=[3, 4, 6, 3], + groups=1, + reduction=16, + dropout_p=None, + inplanes=64, + input_3x3=False, + downsample_kernel_size=1, + downsample_padding=0, + last_stride=2, + fc_dims=None, + **kwargs + ) + if pretrained: + model_url = pretrained_settings['se_resnet50']['imagenet']['url'] + init_pretrained_weights(model, model_url) + return model + + +def se_resnet50_fc512(num_classes, loss='softmax', pretrained=True, **kwargs): + model = SENet( + num_classes=num_classes, + loss=loss, + block=SEResNetBottleneck, + layers=[3, 4, 6, 3], + groups=1, + reduction=16, + dropout_p=None, + inplanes=64, + input_3x3=False, + downsample_kernel_size=1, + downsample_padding=0, + last_stride=1, + fc_dims=[512], + **kwargs + ) + if pretrained: + model_url = pretrained_settings['se_resnet50']['imagenet']['url'] + init_pretrained_weights(model, model_url) + return model + + +def se_resnet101(num_classes, loss='softmax', pretrained=True, **kwargs): + model = SENet( + num_classes=num_classes, + loss=loss, + block=SEResNetBottleneck, + layers=[3, 4, 23, 3], + groups=1, + reduction=16, + dropout_p=None, + inplanes=64, + input_3x3=False, + downsample_kernel_size=1, + downsample_padding=0, + last_stride=2, + fc_dims=None, + **kwargs + ) + if pretrained: + model_url = pretrained_settings['se_resnet101']['imagenet']['url'] + init_pretrained_weights(model, model_url) + return model + + +def se_resnet152(num_classes, loss='softmax', pretrained=True, **kwargs): + model = SENet( + num_classes=num_classes, + loss=loss, + block=SEResNetBottleneck, + layers=[3, 8, 36, 3], + groups=1, + reduction=16, + dropout_p=None, + inplanes=64, + input_3x3=False, + downsample_kernel_size=1, + downsample_padding=0, + last_stride=2, + fc_dims=None, + **kwargs + ) + if pretrained: + model_url = pretrained_settings['se_resnet152']['imagenet']['url'] + init_pretrained_weights(model, model_url) + return model + + +def se_resnext50_32x4d(num_classes, loss='softmax', pretrained=True, **kwargs): + model = SENet( + num_classes=num_classes, + loss=loss, + block=SEResNeXtBottleneck, + layers=[3, 4, 6, 3], + groups=32, + reduction=16, + dropout_p=None, + inplanes=64, + input_3x3=False, + downsample_kernel_size=1, + downsample_padding=0, + last_stride=2, + fc_dims=None, + **kwargs + ) + if pretrained: + model_url = pretrained_settings['se_resnext50_32x4d']['imagenet']['url' + ] + init_pretrained_weights(model, model_url) + return model + + +def se_resnext101_32x4d( + num_classes, loss='softmax', pretrained=True, **kwargs +): + model = SENet( + num_classes=num_classes, + loss=loss, + block=SEResNeXtBottleneck, + layers=[3, 4, 23, 3], + groups=32, + reduction=16, + dropout_p=None, + inplanes=64, + input_3x3=False, + downsample_kernel_size=1, + downsample_padding=0, + last_stride=2, + fc_dims=None, + **kwargs + ) + if pretrained: + model_url = pretrained_settings['se_resnext101_32x4d']['imagenet'][ + 'url'] + init_pretrained_weights(model, model_url) + return model diff --git a/strong_sort/deep/reid/torchreid/models/shufflenet.py b/strong_sort/deep/reid/torchreid/models/shufflenet.py new file mode 100644 index 0000000..bc4d34f --- /dev/null +++ b/strong_sort/deep/reid/torchreid/models/shufflenet.py @@ -0,0 +1,198 @@ +from __future__ import division, absolute_import +import torch +import torch.utils.model_zoo as model_zoo +from torch import nn +from torch.nn import functional as F + +__all__ = ['shufflenet'] + +model_urls = { + # training epoch = 90, top1 = 61.8 + 'imagenet': + 'https://mega.nz/#!RDpUlQCY!tr_5xBEkelzDjveIYBBcGcovNCOrgfiJO9kiidz9fZM', +} + + +class ChannelShuffle(nn.Module): + + def __init__(self, num_groups): + super(ChannelShuffle, self).__init__() + self.g = num_groups + + def forward(self, x): + b, c, h, w = x.size() + n = c // self.g + # reshape + x = x.view(b, self.g, n, h, w) + # transpose + x = x.permute(0, 2, 1, 3, 4).contiguous() + # flatten + x = x.view(b, c, h, w) + return x + + +class Bottleneck(nn.Module): + + def __init__( + self, + in_channels, + out_channels, + stride, + num_groups, + group_conv1x1=True + ): + super(Bottleneck, self).__init__() + assert stride in [1, 2], 'Warning: stride must be either 1 or 2' + self.stride = stride + mid_channels = out_channels // 4 + if stride == 2: + out_channels -= in_channels + # group conv is not applied to first conv1x1 at stage 2 + num_groups_conv1x1 = num_groups if group_conv1x1 else 1 + self.conv1 = nn.Conv2d( + in_channels, + mid_channels, + 1, + groups=num_groups_conv1x1, + bias=False + ) + self.bn1 = nn.BatchNorm2d(mid_channels) + self.shuffle1 = ChannelShuffle(num_groups) + self.conv2 = nn.Conv2d( + mid_channels, + mid_channels, + 3, + stride=stride, + padding=1, + groups=mid_channels, + bias=False + ) + self.bn2 = nn.BatchNorm2d(mid_channels) + self.conv3 = nn.Conv2d( + mid_channels, out_channels, 1, groups=num_groups, bias=False + ) + self.bn3 = nn.BatchNorm2d(out_channels) + if stride == 2: + self.shortcut = nn.AvgPool2d(3, stride=2, padding=1) + + def forward(self, x): + out = F.relu(self.bn1(self.conv1(x))) + out = self.shuffle1(out) + out = self.bn2(self.conv2(out)) + out = self.bn3(self.conv3(out)) + if self.stride == 2: + res = self.shortcut(x) + out = F.relu(torch.cat([res, out], 1)) + else: + out = F.relu(x + out) + return out + + +# configuration of (num_groups: #out_channels) based on Table 1 in the paper +cfg = { + 1: [144, 288, 576], + 2: [200, 400, 800], + 3: [240, 480, 960], + 4: [272, 544, 1088], + 8: [384, 768, 1536], +} + + +class ShuffleNet(nn.Module): + """ShuffleNet. + + Reference: + Zhang et al. ShuffleNet: An Extremely Efficient Convolutional Neural + Network for Mobile Devices. CVPR 2018. + + Public keys: + - ``shufflenet``: ShuffleNet (groups=3). + """ + + def __init__(self, num_classes, loss='softmax', num_groups=3, **kwargs): + super(ShuffleNet, self).__init__() + self.loss = loss + + self.conv1 = nn.Sequential( + nn.Conv2d(3, 24, 3, stride=2, padding=1, bias=False), + nn.BatchNorm2d(24), + nn.ReLU(), + nn.MaxPool2d(3, stride=2, padding=1), + ) + + self.stage2 = nn.Sequential( + Bottleneck( + 24, cfg[num_groups][0], 2, num_groups, group_conv1x1=False + ), + Bottleneck(cfg[num_groups][0], cfg[num_groups][0], 1, num_groups), + Bottleneck(cfg[num_groups][0], cfg[num_groups][0], 1, num_groups), + Bottleneck(cfg[num_groups][0], cfg[num_groups][0], 1, num_groups), + ) + + self.stage3 = nn.Sequential( + Bottleneck(cfg[num_groups][0], cfg[num_groups][1], 2, num_groups), + Bottleneck(cfg[num_groups][1], cfg[num_groups][1], 1, num_groups), + Bottleneck(cfg[num_groups][1], cfg[num_groups][1], 1, num_groups), + Bottleneck(cfg[num_groups][1], cfg[num_groups][1], 1, num_groups), + Bottleneck(cfg[num_groups][1], cfg[num_groups][1], 1, num_groups), + Bottleneck(cfg[num_groups][1], cfg[num_groups][1], 1, num_groups), + Bottleneck(cfg[num_groups][1], cfg[num_groups][1], 1, num_groups), + Bottleneck(cfg[num_groups][1], cfg[num_groups][1], 1, num_groups), + ) + + self.stage4 = nn.Sequential( + Bottleneck(cfg[num_groups][1], cfg[num_groups][2], 2, num_groups), + Bottleneck(cfg[num_groups][2], cfg[num_groups][2], 1, num_groups), + Bottleneck(cfg[num_groups][2], cfg[num_groups][2], 1, num_groups), + Bottleneck(cfg[num_groups][2], cfg[num_groups][2], 1, num_groups), + ) + + self.classifier = nn.Linear(cfg[num_groups][2], num_classes) + self.feat_dim = cfg[num_groups][2] + + def forward(self, x): + x = self.conv1(x) + x = self.stage2(x) + x = self.stage3(x) + x = self.stage4(x) + x = F.avg_pool2d(x, x.size()[2:]).view(x.size(0), -1) + + if not self.training: + return x + + y = self.classifier(x) + + if self.loss == 'softmax': + return y + elif self.loss == 'triplet': + return y, x + else: + raise KeyError('Unsupported loss: {}'.format(self.loss)) + + +def init_pretrained_weights(model, model_url): + """Initializes model with pretrained weights. + + Layers that don't match with pretrained layers in name or size are kept unchanged. + """ + pretrain_dict = model_zoo.load_url(model_url) + model_dict = model.state_dict() + pretrain_dict = { + k: v + for k, v in pretrain_dict.items() + if k in model_dict and model_dict[k].size() == v.size() + } + model_dict.update(pretrain_dict) + model.load_state_dict(model_dict) + + +def shufflenet(num_classes, loss='softmax', pretrained=True, **kwargs): + model = ShuffleNet(num_classes, loss, **kwargs) + if pretrained: + # init_pretrained_weights(model, model_urls['imagenet']) + import warnings + warnings.warn( + 'The imagenet pretrained weights need to be manually downloaded from {}' + .format(model_urls['imagenet']) + ) + return model diff --git a/strong_sort/deep/reid/torchreid/models/shufflenetv2.py b/strong_sort/deep/reid/torchreid/models/shufflenetv2.py new file mode 100644 index 0000000..3ff879e --- /dev/null +++ b/strong_sort/deep/reid/torchreid/models/shufflenetv2.py @@ -0,0 +1,262 @@ +""" +Code source: https://github.com/pytorch/vision +""" +from __future__ import division, absolute_import +import torch +import torch.utils.model_zoo as model_zoo +from torch import nn + +__all__ = [ + 'shufflenet_v2_x0_5', 'shufflenet_v2_x1_0', 'shufflenet_v2_x1_5', + 'shufflenet_v2_x2_0' +] + +model_urls = { + 'shufflenetv2_x0.5': + 'https://download.pytorch.org/models/shufflenetv2_x0.5-f707e7126e.pth', + 'shufflenetv2_x1.0': + 'https://download.pytorch.org/models/shufflenetv2_x1-5666bf0f80.pth', + 'shufflenetv2_x1.5': None, + 'shufflenetv2_x2.0': None, +} + + +def channel_shuffle(x, groups): + batchsize, num_channels, height, width = x.data.size() + channels_per_group = num_channels // groups + + # reshape + x = x.view(batchsize, groups, channels_per_group, height, width) + + x = torch.transpose(x, 1, 2).contiguous() + + # flatten + x = x.view(batchsize, -1, height, width) + + return x + + +class InvertedResidual(nn.Module): + + def __init__(self, inp, oup, stride): + super(InvertedResidual, self).__init__() + + if not (1 <= stride <= 3): + raise ValueError('illegal stride value') + self.stride = stride + + branch_features = oup // 2 + assert (self.stride != 1) or (inp == branch_features << 1) + + if self.stride > 1: + self.branch1 = nn.Sequential( + self.depthwise_conv( + inp, inp, kernel_size=3, stride=self.stride, padding=1 + ), + nn.BatchNorm2d(inp), + nn.Conv2d( + inp, + branch_features, + kernel_size=1, + stride=1, + padding=0, + bias=False + ), + nn.BatchNorm2d(branch_features), + nn.ReLU(inplace=True), + ) + + self.branch2 = nn.Sequential( + nn.Conv2d( + inp if (self.stride > 1) else branch_features, + branch_features, + kernel_size=1, + stride=1, + padding=0, + bias=False + ), + nn.BatchNorm2d(branch_features), + nn.ReLU(inplace=True), + self.depthwise_conv( + branch_features, + branch_features, + kernel_size=3, + stride=self.stride, + padding=1 + ), + nn.BatchNorm2d(branch_features), + nn.Conv2d( + branch_features, + branch_features, + kernel_size=1, + stride=1, + padding=0, + bias=False + ), + nn.BatchNorm2d(branch_features), + nn.ReLU(inplace=True), + ) + + @staticmethod + def depthwise_conv(i, o, kernel_size, stride=1, padding=0, bias=False): + return nn.Conv2d( + i, o, kernel_size, stride, padding, bias=bias, groups=i + ) + + def forward(self, x): + if self.stride == 1: + x1, x2 = x.chunk(2, dim=1) + out = torch.cat((x1, self.branch2(x2)), dim=1) + else: + out = torch.cat((self.branch1(x), self.branch2(x)), dim=1) + + out = channel_shuffle(out, 2) + + return out + + +class ShuffleNetV2(nn.Module): + """ShuffleNetV2. + + Reference: + Ma et al. ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design. ECCV 2018. + + Public keys: + - ``shufflenet_v2_x0_5``: ShuffleNetV2 x0.5. + - ``shufflenet_v2_x1_0``: ShuffleNetV2 x1.0. + - ``shufflenet_v2_x1_5``: ShuffleNetV2 x1.5. + - ``shufflenet_v2_x2_0``: ShuffleNetV2 x2.0. + """ + + def __init__( + self, num_classes, loss, stages_repeats, stages_out_channels, **kwargs + ): + super(ShuffleNetV2, self).__init__() + self.loss = loss + + if len(stages_repeats) != 3: + raise ValueError( + 'expected stages_repeats as list of 3 positive ints' + ) + if len(stages_out_channels) != 5: + raise ValueError( + 'expected stages_out_channels as list of 5 positive ints' + ) + self._stage_out_channels = stages_out_channels + + input_channels = 3 + output_channels = self._stage_out_channels[0] + self.conv1 = nn.Sequential( + nn.Conv2d(input_channels, output_channels, 3, 2, 1, bias=False), + nn.BatchNorm2d(output_channels), + nn.ReLU(inplace=True), + ) + input_channels = output_channels + + self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) + + stage_names = ['stage{}'.format(i) for i in [2, 3, 4]] + for name, repeats, output_channels in zip( + stage_names, stages_repeats, self._stage_out_channels[1:] + ): + seq = [InvertedResidual(input_channels, output_channels, 2)] + for i in range(repeats - 1): + seq.append( + InvertedResidual(output_channels, output_channels, 1) + ) + setattr(self, name, nn.Sequential(*seq)) + input_channels = output_channels + + output_channels = self._stage_out_channels[-1] + self.conv5 = nn.Sequential( + nn.Conv2d(input_channels, output_channels, 1, 1, 0, bias=False), + nn.BatchNorm2d(output_channels), + nn.ReLU(inplace=True), + ) + self.global_avgpool = nn.AdaptiveAvgPool2d((1, 1)) + + self.classifier = nn.Linear(output_channels, num_classes) + + def featuremaps(self, x): + x = self.conv1(x) + x = self.maxpool(x) + x = self.stage2(x) + x = self.stage3(x) + x = self.stage4(x) + x = self.conv5(x) + return x + + def forward(self, x): + f = self.featuremaps(x) + v = self.global_avgpool(f) + v = v.view(v.size(0), -1) + + if not self.training: + return v + + y = self.classifier(v) + + if self.loss == 'softmax': + return y + elif self.loss == 'triplet': + return y, v + else: + raise KeyError("Unsupported loss: {}".format(self.loss)) + + +def init_pretrained_weights(model, model_url): + """Initializes model with pretrained weights. + + Layers that don't match with pretrained layers in name or size are kept unchanged. + """ + if model_url is None: + import warnings + warnings.warn( + 'ImageNet pretrained weights are unavailable for this model' + ) + return + pretrain_dict = model_zoo.load_url(model_url) + model_dict = model.state_dict() + pretrain_dict = { + k: v + for k, v in pretrain_dict.items() + if k in model_dict and model_dict[k].size() == v.size() + } + model_dict.update(pretrain_dict) + model.load_state_dict(model_dict) + + +def shufflenet_v2_x0_5(num_classes, loss='softmax', pretrained=True, **kwargs): + model = ShuffleNetV2( + num_classes, loss, [4, 8, 4], [24, 48, 96, 192, 1024], **kwargs + ) + if pretrained: + init_pretrained_weights(model, model_urls['shufflenetv2_x0.5']) + return model + + +def shufflenet_v2_x1_0(num_classes, loss='softmax', pretrained=True, **kwargs): + model = ShuffleNetV2( + num_classes, loss, [4, 8, 4], [24, 116, 232, 464, 1024], **kwargs + ) + if pretrained: + init_pretrained_weights(model, model_urls['shufflenetv2_x1.0']) + return model + + +def shufflenet_v2_x1_5(num_classes, loss='softmax', pretrained=True, **kwargs): + model = ShuffleNetV2( + num_classes, loss, [4, 8, 4], [24, 176, 352, 704, 1024], **kwargs + ) + if pretrained: + init_pretrained_weights(model, model_urls['shufflenetv2_x1.5']) + return model + + +def shufflenet_v2_x2_0(num_classes, loss='softmax', pretrained=True, **kwargs): + model = ShuffleNetV2( + num_classes, loss, [4, 8, 4], [24, 244, 488, 976, 2048], **kwargs + ) + if pretrained: + init_pretrained_weights(model, model_urls['shufflenetv2_x2.0']) + return model diff --git a/strong_sort/deep/reid/torchreid/models/squeezenet.py b/strong_sort/deep/reid/torchreid/models/squeezenet.py new file mode 100644 index 0000000..83e8dc9 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/models/squeezenet.py @@ -0,0 +1,236 @@ +""" +Code source: https://github.com/pytorch/vision +""" +from __future__ import division, absolute_import +import torch +import torch.nn as nn +import torch.utils.model_zoo as model_zoo + +__all__ = ['squeezenet1_0', 'squeezenet1_1', 'squeezenet1_0_fc512'] + +model_urls = { + 'squeezenet1_0': + 'https://download.pytorch.org/models/squeezenet1_0-a815701f.pth', + 'squeezenet1_1': + 'https://download.pytorch.org/models/squeezenet1_1-f364aa15.pth', +} + + +class Fire(nn.Module): + + def __init__( + self, inplanes, squeeze_planes, expand1x1_planes, expand3x3_planes + ): + super(Fire, self).__init__() + self.inplanes = inplanes + self.squeeze = nn.Conv2d(inplanes, squeeze_planes, kernel_size=1) + self.squeeze_activation = nn.ReLU(inplace=True) + self.expand1x1 = nn.Conv2d( + squeeze_planes, expand1x1_planes, kernel_size=1 + ) + self.expand1x1_activation = nn.ReLU(inplace=True) + self.expand3x3 = nn.Conv2d( + squeeze_planes, expand3x3_planes, kernel_size=3, padding=1 + ) + self.expand3x3_activation = nn.ReLU(inplace=True) + + def forward(self, x): + x = self.squeeze_activation(self.squeeze(x)) + return torch.cat( + [ + self.expand1x1_activation(self.expand1x1(x)), + self.expand3x3_activation(self.expand3x3(x)) + ], 1 + ) + + +class SqueezeNet(nn.Module): + """SqueezeNet. + + Reference: + Iandola et al. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters + and< 0.5 MB model size. arXiv:1602.07360. + + Public keys: + - ``squeezenet1_0``: SqueezeNet (version=1.0). + - ``squeezenet1_1``: SqueezeNet (version=1.1). + - ``squeezenet1_0_fc512``: SqueezeNet (version=1.0) + FC. + """ + + def __init__( + self, + num_classes, + loss, + version=1.0, + fc_dims=None, + dropout_p=None, + **kwargs + ): + super(SqueezeNet, self).__init__() + self.loss = loss + self.feature_dim = 512 + + if version not in [1.0, 1.1]: + raise ValueError( + 'Unsupported SqueezeNet version {version}:' + '1.0 or 1.1 expected'.format(version=version) + ) + + if version == 1.0: + self.features = nn.Sequential( + nn.Conv2d(3, 96, kernel_size=7, stride=2), + nn.ReLU(inplace=True), + nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True), + Fire(96, 16, 64, 64), + Fire(128, 16, 64, 64), + Fire(128, 32, 128, 128), + nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True), + Fire(256, 32, 128, 128), + Fire(256, 48, 192, 192), + Fire(384, 48, 192, 192), + Fire(384, 64, 256, 256), + nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True), + Fire(512, 64, 256, 256), + ) + else: + self.features = nn.Sequential( + nn.Conv2d(3, 64, kernel_size=3, stride=2), + nn.ReLU(inplace=True), + nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True), + Fire(64, 16, 64, 64), + Fire(128, 16, 64, 64), + nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True), + Fire(128, 32, 128, 128), + Fire(256, 32, 128, 128), + nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True), + Fire(256, 48, 192, 192), + Fire(384, 48, 192, 192), + Fire(384, 64, 256, 256), + Fire(512, 64, 256, 256), + ) + + self.global_avgpool = nn.AdaptiveAvgPool2d(1) + self.fc = self._construct_fc_layer(fc_dims, 512, dropout_p) + self.classifier = nn.Linear(self.feature_dim, num_classes) + + self._init_params() + + def _construct_fc_layer(self, fc_dims, input_dim, dropout_p=None): + """Constructs fully connected layer + + Args: + fc_dims (list or tuple): dimensions of fc layers, if None, no fc layers are constructed + input_dim (int): input dimension + dropout_p (float): dropout probability, if None, dropout is unused + """ + if fc_dims is None: + self.feature_dim = input_dim + return None + + assert isinstance( + fc_dims, (list, tuple) + ), 'fc_dims must be either list or tuple, but got {}'.format( + type(fc_dims) + ) + + layers = [] + for dim in fc_dims: + layers.append(nn.Linear(input_dim, dim)) + layers.append(nn.BatchNorm1d(dim)) + layers.append(nn.ReLU(inplace=True)) + if dropout_p is not None: + layers.append(nn.Dropout(p=dropout_p)) + input_dim = dim + + self.feature_dim = fc_dims[-1] + + return nn.Sequential(*layers) + + def _init_params(self): + for m in self.modules(): + if isinstance(m, nn.Conv2d): + nn.init.kaiming_normal_( + m.weight, mode='fan_out', nonlinearity='relu' + ) + if m.bias is not None: + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.BatchNorm2d): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.BatchNorm1d): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.Linear): + nn.init.normal_(m.weight, 0, 0.01) + if m.bias is not None: + nn.init.constant_(m.bias, 0) + + def forward(self, x): + f = self.features(x) + v = self.global_avgpool(f) + v = v.view(v.size(0), -1) + + if self.fc is not None: + v = self.fc(v) + + if not self.training: + return v + + y = self.classifier(v) + + if self.loss == 'softmax': + return y + elif self.loss == 'triplet': + return y, v + else: + raise KeyError('Unsupported loss: {}'.format(self.loss)) + + +def init_pretrained_weights(model, model_url): + """Initializes model with pretrained weights. + + Layers that don't match with pretrained layers in name or size are kept unchanged. + """ + pretrain_dict = model_zoo.load_url(model_url, map_location=None) + model_dict = model.state_dict() + pretrain_dict = { + k: v + for k, v in pretrain_dict.items() + if k in model_dict and model_dict[k].size() == v.size() + } + model_dict.update(pretrain_dict) + model.load_state_dict(model_dict) + + +def squeezenet1_0(num_classes, loss='softmax', pretrained=True, **kwargs): + model = SqueezeNet( + num_classes, loss, version=1.0, fc_dims=None, dropout_p=None, **kwargs + ) + if pretrained: + init_pretrained_weights(model, model_urls['squeezenet1_0']) + return model + + +def squeezenet1_0_fc512( + num_classes, loss='softmax', pretrained=True, **kwargs +): + model = SqueezeNet( + num_classes, + loss, + version=1.0, + fc_dims=[512], + dropout_p=None, + **kwargs + ) + if pretrained: + init_pretrained_weights(model, model_urls['squeezenet1_0']) + return model + + +def squeezenet1_1(num_classes, loss='softmax', pretrained=True, **kwargs): + model = SqueezeNet( + num_classes, loss, version=1.1, fc_dims=None, dropout_p=None, **kwargs + ) + if pretrained: + init_pretrained_weights(model, model_urls['squeezenet1_1']) + return model diff --git a/strong_sort/deep/reid/torchreid/models/xception.py b/strong_sort/deep/reid/torchreid/models/xception.py new file mode 100644 index 0000000..43db4ab --- /dev/null +++ b/strong_sort/deep/reid/torchreid/models/xception.py @@ -0,0 +1,344 @@ +from __future__ import division, absolute_import +import torch.nn as nn +import torch.nn.functional as F +import torch.utils.model_zoo as model_zoo + +__all__ = ['xception'] + +pretrained_settings = { + 'xception': { + 'imagenet': { + 'url': + 'http://data.lip6.fr/cadene/pretrainedmodels/xception-43020ad28.pth', + 'input_space': 'RGB', + 'input_size': [3, 299, 299], + 'input_range': [0, 1], + 'mean': [0.5, 0.5, 0.5], + 'std': [0.5, 0.5, 0.5], + 'num_classes': 1000, + 'scale': + 0.8975 # The resize parameter of the validation transform should be 333, and make sure to center crop at 299x299 + } + } +} + + +class SeparableConv2d(nn.Module): + + def __init__( + self, + in_channels, + out_channels, + kernel_size=1, + stride=1, + padding=0, + dilation=1, + bias=False + ): + super(SeparableConv2d, self).__init__() + + self.conv1 = nn.Conv2d( + in_channels, + in_channels, + kernel_size, + stride, + padding, + dilation, + groups=in_channels, + bias=bias + ) + self.pointwise = nn.Conv2d( + in_channels, out_channels, 1, 1, 0, 1, 1, bias=bias + ) + + def forward(self, x): + x = self.conv1(x) + x = self.pointwise(x) + return x + + +class Block(nn.Module): + + def __init__( + self, + in_filters, + out_filters, + reps, + strides=1, + start_with_relu=True, + grow_first=True + ): + super(Block, self).__init__() + + if out_filters != in_filters or strides != 1: + self.skip = nn.Conv2d( + in_filters, out_filters, 1, stride=strides, bias=False + ) + self.skipbn = nn.BatchNorm2d(out_filters) + else: + self.skip = None + + self.relu = nn.ReLU(inplace=True) + rep = [] + + filters = in_filters + if grow_first: + rep.append(self.relu) + rep.append( + SeparableConv2d( + in_filters, + out_filters, + 3, + stride=1, + padding=1, + bias=False + ) + ) + rep.append(nn.BatchNorm2d(out_filters)) + filters = out_filters + + for i in range(reps - 1): + rep.append(self.relu) + rep.append( + SeparableConv2d( + filters, filters, 3, stride=1, padding=1, bias=False + ) + ) + rep.append(nn.BatchNorm2d(filters)) + + if not grow_first: + rep.append(self.relu) + rep.append( + SeparableConv2d( + in_filters, + out_filters, + 3, + stride=1, + padding=1, + bias=False + ) + ) + rep.append(nn.BatchNorm2d(out_filters)) + + if not start_with_relu: + rep = rep[1:] + else: + rep[0] = nn.ReLU(inplace=False) + + if strides != 1: + rep.append(nn.MaxPool2d(3, strides, 1)) + self.rep = nn.Sequential(*rep) + + def forward(self, inp): + x = self.rep(inp) + + if self.skip is not None: + skip = self.skip(inp) + skip = self.skipbn(skip) + else: + skip = inp + + x += skip + return x + + +class Xception(nn.Module): + """Xception. + + Reference: + Chollet. Xception: Deep Learning with Depthwise + Separable Convolutions. CVPR 2017. + + Public keys: + - ``xception``: Xception. + """ + + def __init__( + self, num_classes, loss, fc_dims=None, dropout_p=None, **kwargs + ): + super(Xception, self).__init__() + self.loss = loss + + self.conv1 = nn.Conv2d(3, 32, 3, 2, 0, bias=False) + self.bn1 = nn.BatchNorm2d(32) + + self.conv2 = nn.Conv2d(32, 64, 3, bias=False) + self.bn2 = nn.BatchNorm2d(64) + + self.block1 = Block( + 64, 128, 2, 2, start_with_relu=False, grow_first=True + ) + self.block2 = Block( + 128, 256, 2, 2, start_with_relu=True, grow_first=True + ) + self.block3 = Block( + 256, 728, 2, 2, start_with_relu=True, grow_first=True + ) + + self.block4 = Block( + 728, 728, 3, 1, start_with_relu=True, grow_first=True + ) + self.block5 = Block( + 728, 728, 3, 1, start_with_relu=True, grow_first=True + ) + self.block6 = Block( + 728, 728, 3, 1, start_with_relu=True, grow_first=True + ) + self.block7 = Block( + 728, 728, 3, 1, start_with_relu=True, grow_first=True + ) + + self.block8 = Block( + 728, 728, 3, 1, start_with_relu=True, grow_first=True + ) + self.block9 = Block( + 728, 728, 3, 1, start_with_relu=True, grow_first=True + ) + self.block10 = Block( + 728, 728, 3, 1, start_with_relu=True, grow_first=True + ) + self.block11 = Block( + 728, 728, 3, 1, start_with_relu=True, grow_first=True + ) + + self.block12 = Block( + 728, 1024, 2, 2, start_with_relu=True, grow_first=False + ) + + self.conv3 = SeparableConv2d(1024, 1536, 3, 1, 1) + self.bn3 = nn.BatchNorm2d(1536) + + self.conv4 = SeparableConv2d(1536, 2048, 3, 1, 1) + self.bn4 = nn.BatchNorm2d(2048) + + self.global_avgpool = nn.AdaptiveAvgPool2d(1) + self.feature_dim = 2048 + self.fc = self._construct_fc_layer(fc_dims, 2048, dropout_p) + self.classifier = nn.Linear(self.feature_dim, num_classes) + + self._init_params() + + def _construct_fc_layer(self, fc_dims, input_dim, dropout_p=None): + """Constructs fully connected layer. + + Args: + fc_dims (list or tuple): dimensions of fc layers, if None, no fc layers are constructed + input_dim (int): input dimension + dropout_p (float): dropout probability, if None, dropout is unused + """ + if fc_dims is None: + self.feature_dim = input_dim + return None + + assert isinstance( + fc_dims, (list, tuple) + ), 'fc_dims must be either list or tuple, but got {}'.format( + type(fc_dims) + ) + + layers = [] + for dim in fc_dims: + layers.append(nn.Linear(input_dim, dim)) + layers.append(nn.BatchNorm1d(dim)) + layers.append(nn.ReLU(inplace=True)) + if dropout_p is not None: + layers.append(nn.Dropout(p=dropout_p)) + input_dim = dim + + self.feature_dim = fc_dims[-1] + + return nn.Sequential(*layers) + + def _init_params(self): + for m in self.modules(): + if isinstance(m, nn.Conv2d): + nn.init.kaiming_normal_( + m.weight, mode='fan_out', nonlinearity='relu' + ) + if m.bias is not None: + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.BatchNorm2d): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.BatchNorm1d): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.Linear): + nn.init.normal_(m.weight, 0, 0.01) + if m.bias is not None: + nn.init.constant_(m.bias, 0) + + def featuremaps(self, input): + x = self.conv1(input) + x = self.bn1(x) + x = F.relu(x, inplace=True) + + x = self.conv2(x) + x = self.bn2(x) + x = F.relu(x, inplace=True) + + x = self.block1(x) + x = self.block2(x) + x = self.block3(x) + x = self.block4(x) + x = self.block5(x) + x = self.block6(x) + x = self.block7(x) + x = self.block8(x) + x = self.block9(x) + x = self.block10(x) + x = self.block11(x) + x = self.block12(x) + + x = self.conv3(x) + x = self.bn3(x) + x = F.relu(x, inplace=True) + + x = self.conv4(x) + x = self.bn4(x) + x = F.relu(x, inplace=True) + return x + + def forward(self, x): + f = self.featuremaps(x) + v = self.global_avgpool(f) + v = v.view(v.size(0), -1) + + if self.fc is not None: + v = self.fc(v) + + if not self.training: + return v + + y = self.classifier(v) + + if self.loss == 'softmax': + return y + elif self.loss == 'triplet': + return y, v + else: + raise KeyError('Unsupported loss: {}'.format(self.loss)) + + +def init_pretrained_weights(model, model_url): + """Initialize models with pretrained weights. + + Layers that don't match with pretrained layers in name or size are kept unchanged. + """ + pretrain_dict = model_zoo.load_url(model_url) + model_dict = model.state_dict() + pretrain_dict = { + k: v + for k, v in pretrain_dict.items() + if k in model_dict and model_dict[k].size() == v.size() + } + model_dict.update(pretrain_dict) + model.load_state_dict(model_dict) + + +def xception(num_classes, loss='softmax', pretrained=True, **kwargs): + model = Xception(num_classes, loss, fc_dims=None, dropout_p=None, **kwargs) + if pretrained: + model_url = pretrained_settings['xception']['imagenet']['url'] + init_pretrained_weights(model, model_url) + return model diff --git a/strong_sort/deep/reid/torchreid/optim/__init__.py b/strong_sort/deep/reid/torchreid/optim/__init__.py new file mode 100644 index 0000000..1813e46 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/optim/__init__.py @@ -0,0 +1,4 @@ +from __future__ import absolute_import + +from .optimizer import build_optimizer +from .lr_scheduler import build_lr_scheduler diff --git a/strong_sort/deep/reid/torchreid/optim/lr_scheduler.py b/strong_sort/deep/reid/torchreid/optim/lr_scheduler.py new file mode 100644 index 0000000..d60bd1d --- /dev/null +++ b/strong_sort/deep/reid/torchreid/optim/lr_scheduler.py @@ -0,0 +1,68 @@ +from __future__ import print_function, absolute_import +import torch + +AVAI_SCH = ['single_step', 'multi_step', 'cosine'] + + +def build_lr_scheduler( + optimizer, lr_scheduler='single_step', stepsize=1, gamma=0.1, max_epoch=1 +): + """A function wrapper for building a learning rate scheduler. + + Args: + optimizer (Optimizer): an Optimizer. + lr_scheduler (str, optional): learning rate scheduler method. Default is single_step. + stepsize (int or list, optional): step size to decay learning rate. When ``lr_scheduler`` + is "single_step", ``stepsize`` should be an integer. When ``lr_scheduler`` is + "multi_step", ``stepsize`` is a list. Default is 1. + gamma (float, optional): decay rate. Default is 0.1. + max_epoch (int, optional): maximum epoch (for cosine annealing). Default is 1. + + Examples:: + >>> # Decay learning rate by every 20 epochs. + >>> scheduler = torchreid.optim.build_lr_scheduler( + >>> optimizer, lr_scheduler='single_step', stepsize=20 + >>> ) + >>> # Decay learning rate at 30, 50 and 55 epochs. + >>> scheduler = torchreid.optim.build_lr_scheduler( + >>> optimizer, lr_scheduler='multi_step', stepsize=[30, 50, 55] + >>> ) + """ + if lr_scheduler not in AVAI_SCH: + raise ValueError( + 'Unsupported scheduler: {}. Must be one of {}'.format( + lr_scheduler, AVAI_SCH + ) + ) + + if lr_scheduler == 'single_step': + if isinstance(stepsize, list): + stepsize = stepsize[-1] + + if not isinstance(stepsize, int): + raise TypeError( + 'For single_step lr_scheduler, stepsize must ' + 'be an integer, but got {}'.format(type(stepsize)) + ) + + scheduler = torch.optim.lr_scheduler.StepLR( + optimizer, step_size=stepsize, gamma=gamma + ) + + elif lr_scheduler == 'multi_step': + if not isinstance(stepsize, list): + raise TypeError( + 'For multi_step lr_scheduler, stepsize must ' + 'be a list, but got {}'.format(type(stepsize)) + ) + + scheduler = torch.optim.lr_scheduler.MultiStepLR( + optimizer, milestones=stepsize, gamma=gamma + ) + + elif lr_scheduler == 'cosine': + scheduler = torch.optim.lr_scheduler.CosineAnnealingLR( + optimizer, float(max_epoch) + ) + + return scheduler diff --git a/strong_sort/deep/reid/torchreid/optim/optimizer.py b/strong_sort/deep/reid/torchreid/optim/optimizer.py new file mode 100644 index 0000000..f57b03a --- /dev/null +++ b/strong_sort/deep/reid/torchreid/optim/optimizer.py @@ -0,0 +1,157 @@ +from __future__ import print_function, absolute_import +import warnings +import torch +import torch.nn as nn + +from .radam import RAdam + +AVAI_OPTIMS = ['adam', 'amsgrad', 'sgd', 'rmsprop', 'radam'] + + +def build_optimizer( + model, + optim='adam', + lr=0.0003, + weight_decay=5e-04, + momentum=0.9, + sgd_dampening=0, + sgd_nesterov=False, + rmsprop_alpha=0.99, + adam_beta1=0.9, + adam_beta2=0.99, + staged_lr=False, + new_layers='', + base_lr_mult=0.1 +): + """A function wrapper for building an optimizer. + + Args: + model (nn.Module): model. + optim (str, optional): optimizer. Default is "adam". + lr (float, optional): learning rate. Default is 0.0003. + weight_decay (float, optional): weight decay (L2 penalty). Default is 5e-04. + momentum (float, optional): momentum factor in sgd. Default is 0.9, + sgd_dampening (float, optional): dampening for momentum. Default is 0. + sgd_nesterov (bool, optional): enables Nesterov momentum. Default is False. + rmsprop_alpha (float, optional): smoothing constant for rmsprop. Default is 0.99. + adam_beta1 (float, optional): beta-1 value in adam. Default is 0.9. + adam_beta2 (float, optional): beta-2 value in adam. Default is 0.99, + staged_lr (bool, optional): uses different learning rates for base and new layers. Base + layers are pretrained layers while new layers are randomly initialized, e.g. the + identity classification layer. Enabling ``staged_lr`` can allow the base layers to + be trained with a smaller learning rate determined by ``base_lr_mult``, while the new + layers will take the ``lr``. Default is False. + new_layers (str or list): attribute names in ``model``. Default is empty. + base_lr_mult (float, optional): learning rate multiplier for base layers. Default is 0.1. + + Examples:: + >>> # A normal optimizer can be built by + >>> optimizer = torchreid.optim.build_optimizer(model, optim='sgd', lr=0.01) + >>> # If you want to use a smaller learning rate for pretrained layers + >>> # and the attribute name for the randomly initialized layer is 'classifier', + >>> # you can do + >>> optimizer = torchreid.optim.build_optimizer( + >>> model, optim='sgd', lr=0.01, staged_lr=True, + >>> new_layers='classifier', base_lr_mult=0.1 + >>> ) + >>> # Now the `classifier` has learning rate 0.01 but the base layers + >>> # have learning rate 0.01 * 0.1. + >>> # new_layers can also take multiple attribute names. Say the new layers + >>> # are 'fc' and 'classifier', you can do + >>> optimizer = torchreid.optim.build_optimizer( + >>> model, optim='sgd', lr=0.01, staged_lr=True, + >>> new_layers=['fc', 'classifier'], base_lr_mult=0.1 + >>> ) + """ + if optim not in AVAI_OPTIMS: + raise ValueError( + 'Unsupported optim: {}. Must be one of {}'.format( + optim, AVAI_OPTIMS + ) + ) + + if not isinstance(model, nn.Module): + raise TypeError( + 'model given to build_optimizer must be an instance of nn.Module' + ) + + if staged_lr: + if isinstance(new_layers, str): + if new_layers is None: + warnings.warn( + 'new_layers is empty, therefore, staged_lr is useless' + ) + new_layers = [new_layers] + + if isinstance(model, nn.DataParallel): + model = model.module + + base_params = [] + base_layers = [] + new_params = [] + + for name, module in model.named_children(): + if name in new_layers: + new_params += [p for p in module.parameters()] + else: + base_params += [p for p in module.parameters()] + base_layers.append(name) + + param_groups = [ + { + 'params': base_params, + 'lr': lr * base_lr_mult + }, + { + 'params': new_params + }, + ] + + else: + param_groups = model.parameters() + + if optim == 'adam': + optimizer = torch.optim.Adam( + param_groups, + lr=lr, + weight_decay=weight_decay, + betas=(adam_beta1, adam_beta2), + ) + + elif optim == 'amsgrad': + optimizer = torch.optim.Adam( + param_groups, + lr=lr, + weight_decay=weight_decay, + betas=(adam_beta1, adam_beta2), + amsgrad=True, + ) + + elif optim == 'sgd': + optimizer = torch.optim.SGD( + param_groups, + lr=lr, + momentum=momentum, + weight_decay=weight_decay, + dampening=sgd_dampening, + nesterov=sgd_nesterov, + ) + + elif optim == 'rmsprop': + optimizer = torch.optim.RMSprop( + param_groups, + lr=lr, + momentum=momentum, + weight_decay=weight_decay, + alpha=rmsprop_alpha, + ) + + elif optim == 'radam': + optimizer = RAdam( + param_groups, + lr=lr, + weight_decay=weight_decay, + betas=(adam_beta1, adam_beta2) + ) + + return optimizer diff --git a/strong_sort/deep/reid/torchreid/optim/radam.py b/strong_sort/deep/reid/torchreid/optim/radam.py new file mode 100644 index 0000000..f066c57 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/optim/radam.py @@ -0,0 +1,330 @@ +""" +Imported from: https://github.com/LiyuanLucasLiu/RAdam + +Paper: https://arxiv.org/abs/1908.03265 + +@article{liu2019radam, + title={On the Variance of the Adaptive Learning Rate and Beyond}, + author={Liu, Liyuan and Jiang, Haoming and He, Pengcheng and Chen, Weizhu and Liu, Xiaodong and Gao, Jianfeng and Han, Jiawei}, + journal={arXiv preprint arXiv:1908.03265}, + year={2019} +} +""" +from __future__ import print_function, absolute_import +import math +import torch +from torch.optim.optimizer import Optimizer + + +class RAdam(Optimizer): + + def __init__( + self, + params, + lr=1e-3, + betas=(0.9, 0.999), + eps=1e-8, + weight_decay=0, + degenerated_to_sgd=True + ): + if not 0.0 <= lr: + raise ValueError("Invalid learning rate: {}".format(lr)) + if not 0.0 <= eps: + raise ValueError("Invalid epsilon value: {}".format(eps)) + if not 0.0 <= betas[0] < 1.0: + raise ValueError( + "Invalid beta parameter at index 0: {}".format(betas[0]) + ) + if not 0.0 <= betas[1] < 1.0: + raise ValueError( + "Invalid beta parameter at index 1: {}".format(betas[1]) + ) + + self.degenerated_to_sgd = degenerated_to_sgd + defaults = dict(lr=lr, betas=betas, eps=eps, weight_decay=weight_decay) + self.buffer = [[None, None, None] for ind in range(10)] + super(RAdam, self).__init__(params, defaults) + + def __setstate__(self, state): + super(RAdam, self).__setstate__(state) + + def step(self, closure=None): + + loss = None + if closure is not None: + loss = closure() + + for group in self.param_groups: + + for p in group['params']: + if p.grad is None: + continue + grad = p.grad.data.float() + if grad.is_sparse: + raise RuntimeError( + 'RAdam does not support sparse gradients' + ) + + p_data_fp32 = p.data.float() + + state = self.state[p] + + if len(state) == 0: + state['step'] = 0 + state['exp_avg'] = torch.zeros_like(p_data_fp32) + state['exp_avg_sq'] = torch.zeros_like(p_data_fp32) + else: + state['exp_avg'] = state['exp_avg'].type_as(p_data_fp32) + state['exp_avg_sq'] = state['exp_avg_sq'].type_as( + p_data_fp32 + ) + + exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq'] + beta1, beta2 = group['betas'] + + exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad) + exp_avg.mul_(beta1).add_(1 - beta1, grad) + + state['step'] += 1 + buffered = self.buffer[int(state['step'] % 10)] + if state['step'] == buffered[0]: + N_sma, step_size = buffered[1], buffered[2] + else: + buffered[0] = state['step'] + beta2_t = beta2**state['step'] + N_sma_max = 2 / (1-beta2) - 1 + N_sma = N_sma_max - 2 * state['step' + ] * beta2_t / (1-beta2_t) + buffered[1] = N_sma + + # more conservative since it's an approximated value + if N_sma >= 5: + step_size = math.sqrt( + (1-beta2_t) * (N_sma-4) / (N_sma_max-4) * + (N_sma-2) / N_sma * N_sma_max / (N_sma_max-2) + ) / (1 - beta1**state['step']) + elif self.degenerated_to_sgd: + step_size = 1.0 / (1 - beta1**state['step']) + else: + step_size = -1 + buffered[2] = step_size + + # more conservative since it's an approximated value + if N_sma >= 5: + if group['weight_decay'] != 0: + p_data_fp32.add_( + -group['weight_decay'] * group['lr'], p_data_fp32 + ) + denom = exp_avg_sq.sqrt().add_(group['eps']) + p_data_fp32.addcdiv_( + -step_size * group['lr'], exp_avg, denom + ) + p.data.copy_(p_data_fp32) + elif step_size > 0: + if group['weight_decay'] != 0: + p_data_fp32.add_( + -group['weight_decay'] * group['lr'], p_data_fp32 + ) + p_data_fp32.add_(-step_size * group['lr'], exp_avg) + p.data.copy_(p_data_fp32) + + return loss + + +class PlainRAdam(Optimizer): + + def __init__( + self, + params, + lr=1e-3, + betas=(0.9, 0.999), + eps=1e-8, + weight_decay=0, + degenerated_to_sgd=True + ): + if not 0.0 <= lr: + raise ValueError("Invalid learning rate: {}".format(lr)) + if not 0.0 <= eps: + raise ValueError("Invalid epsilon value: {}".format(eps)) + if not 0.0 <= betas[0] < 1.0: + raise ValueError( + "Invalid beta parameter at index 0: {}".format(betas[0]) + ) + if not 0.0 <= betas[1] < 1.0: + raise ValueError( + "Invalid beta parameter at index 1: {}".format(betas[1]) + ) + + self.degenerated_to_sgd = degenerated_to_sgd + defaults = dict(lr=lr, betas=betas, eps=eps, weight_decay=weight_decay) + + super(PlainRAdam, self).__init__(params, defaults) + + def __setstate__(self, state): + super(PlainRAdam, self).__setstate__(state) + + def step(self, closure=None): + + loss = None + if closure is not None: + loss = closure() + + for group in self.param_groups: + + for p in group['params']: + if p.grad is None: + continue + grad = p.grad.data.float() + if grad.is_sparse: + raise RuntimeError( + 'RAdam does not support sparse gradients' + ) + + p_data_fp32 = p.data.float() + + state = self.state[p] + + if len(state) == 0: + state['step'] = 0 + state['exp_avg'] = torch.zeros_like(p_data_fp32) + state['exp_avg_sq'] = torch.zeros_like(p_data_fp32) + else: + state['exp_avg'] = state['exp_avg'].type_as(p_data_fp32) + state['exp_avg_sq'] = state['exp_avg_sq'].type_as( + p_data_fp32 + ) + + exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq'] + beta1, beta2 = group['betas'] + + exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad) + exp_avg.mul_(beta1).add_(1 - beta1, grad) + + state['step'] += 1 + beta2_t = beta2**state['step'] + N_sma_max = 2 / (1-beta2) - 1 + N_sma = N_sma_max - 2 * state['step'] * beta2_t / (1-beta2_t) + + # more conservative since it's an approximated value + if N_sma >= 5: + if group['weight_decay'] != 0: + p_data_fp32.add_( + -group['weight_decay'] * group['lr'], p_data_fp32 + ) + step_size = group['lr'] * math.sqrt( + (1-beta2_t) * (N_sma-4) / (N_sma_max-4) * + (N_sma-2) / N_sma * N_sma_max / (N_sma_max-2) + ) / (1 - beta1**state['step']) + denom = exp_avg_sq.sqrt().add_(group['eps']) + p_data_fp32.addcdiv_(-step_size, exp_avg, denom) + p.data.copy_(p_data_fp32) + elif self.degenerated_to_sgd: + if group['weight_decay'] != 0: + p_data_fp32.add_( + -group['weight_decay'] * group['lr'], p_data_fp32 + ) + step_size = group['lr'] / (1 - beta1**state['step']) + p_data_fp32.add_(-step_size, exp_avg) + p.data.copy_(p_data_fp32) + + return loss + + +class AdamW(Optimizer): + + def __init__( + self, + params, + lr=1e-3, + betas=(0.9, 0.999), + eps=1e-8, + weight_decay=0, + warmup=0 + ): + if not 0.0 <= lr: + raise ValueError("Invalid learning rate: {}".format(lr)) + if not 0.0 <= eps: + raise ValueError("Invalid epsilon value: {}".format(eps)) + if not 0.0 <= betas[0] < 1.0: + raise ValueError( + "Invalid beta parameter at index 0: {}".format(betas[0]) + ) + if not 0.0 <= betas[1] < 1.0: + raise ValueError( + "Invalid beta parameter at index 1: {}".format(betas[1]) + ) + + defaults = dict( + lr=lr, + betas=betas, + eps=eps, + weight_decay=weight_decay, + warmup=warmup + ) + super(AdamW, self).__init__(params, defaults) + + def __setstate__(self, state): + super(AdamW, self).__setstate__(state) + + def step(self, closure=None): + loss = None + if closure is not None: + loss = closure() + + for group in self.param_groups: + + for p in group['params']: + if p.grad is None: + continue + grad = p.grad.data.float() + if grad.is_sparse: + raise RuntimeError( + 'Adam does not support sparse gradients, please consider SparseAdam instead' + ) + + p_data_fp32 = p.data.float() + + state = self.state[p] + + if len(state) == 0: + state['step'] = 0 + state['exp_avg'] = torch.zeros_like(p_data_fp32) + state['exp_avg_sq'] = torch.zeros_like(p_data_fp32) + else: + state['exp_avg'] = state['exp_avg'].type_as(p_data_fp32) + state['exp_avg_sq'] = state['exp_avg_sq'].type_as( + p_data_fp32 + ) + + exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq'] + beta1, beta2 = group['betas'] + + state['step'] += 1 + + exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad) + exp_avg.mul_(beta1).add_(1 - beta1, grad) + + denom = exp_avg_sq.sqrt().add_(group['eps']) + bias_correction1 = 1 - beta1**state['step'] + bias_correction2 = 1 - beta2**state['step'] + + if group['warmup'] > state['step']: + scheduled_lr = 1e-8 + state['step'] * group['lr'] / group[ + 'warmup'] + else: + scheduled_lr = group['lr'] + + step_size = scheduled_lr * math.sqrt( + bias_correction2 + ) / bias_correction1 + + if group['weight_decay'] != 0: + p_data_fp32.add_( + -group['weight_decay'] * scheduled_lr, p_data_fp32 + ) + + p_data_fp32.addcdiv_(-step_size, exp_avg, denom) + + p.data.copy_(p_data_fp32) + + return loss diff --git a/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/README.md b/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/README.md new file mode 100644 index 0000000..349a9ef --- /dev/null +++ b/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/README.md @@ -0,0 +1,37 @@ +# Understanding Image Retrieval Re-Ranking: A Graph Neural Network Perspective + +[[Paper]](https://arxiv.org/abs/2012.07620v2) + +On the Market-1501 dataset, we accelerate the re-ranking processing from **89.2s** to **9.4ms** with one K40m GPU, facilitating the real-time post-processing. +Similarly, we observe that our method achieves comparable or even better retrieval results on the other four image retrieval benchmarks, +i.e., VeRi-776, Oxford-5k, Paris-6k and University-1652, with limited time cost. + +## Prerequisites + +The code was mainly developed and tested with python 3.7, PyTorch 1.4.1, CUDA 10.2, and CentOS release 6.10. + +The code has been included in `/extension`. To compile it: + +```shell +cd extension +sh make.sh +``` + +## Demo + +The demo script `main.py` provides the gnn re-ranking method using the prepared feature. + +```shell +python main.py --data_path PATH_TO_DATA --k1 26 --k2 7 +``` + +## Citation +```bibtex +@article{zhang2020understanding, + title={Understanding Image Retrieval Re-Ranking: A Graph Neural Network Perspective}, + author={Xuanmeng Zhang, Minyue Jiang, Zhedong Zheng, Xiao Tan, Errui Ding, Yi Yang}, + journal={arXiv preprint arXiv:2012.07620}, + year={2020} +} +``` + diff --git a/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/extension/adjacency_matrix/build_adjacency_matrix.cpp b/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/extension/adjacency_matrix/build_adjacency_matrix.cpp new file mode 100644 index 0000000..4c49604 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/extension/adjacency_matrix/build_adjacency_matrix.cpp @@ -0,0 +1,19 @@ +#include +#include +#include + +at::Tensor build_adjacency_matrix_forward(torch::Tensor initial_rank); + + +#define CHECK_CUDA(x) AT_ASSERTM(x.type().is_cuda(), #x " must be a CUDA tensor") +#define CHECK_CONTIGUOUS(x) AT_ASSERTM(x.is_contiguous(), #x " must be contiguous") +#define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x) + +at::Tensor build_adjacency_matrix(at::Tensor initial_rank) { + CHECK_INPUT(initial_rank); + return build_adjacency_matrix_forward(initial_rank); +} + +PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { + m.def("forward", &build_adjacency_matrix, "build_adjacency_matrix (CUDA)"); +} diff --git a/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/extension/adjacency_matrix/build_adjacency_matrix_kernel.cu b/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/extension/adjacency_matrix/build_adjacency_matrix_kernel.cu new file mode 100644 index 0000000..4973dde --- /dev/null +++ b/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/extension/adjacency_matrix/build_adjacency_matrix_kernel.cu @@ -0,0 +1,31 @@ +#include + +#include +#include +#include + +#define CUDA_1D_KERNEL_LOOP(i, n) for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < n; i += blockDim.x * gridDim.x) + + +__global__ void build_adjacency_matrix_kernel(float* initial_rank, float* A, const int total_num, const int topk, const int nthreads, const int all_num) { + int index = blockIdx.x * blockDim.x + threadIdx.x; + int stride = blockDim.x * gridDim.x; + for (int i = index; i < all_num; i += stride) { + int ii = i / topk; + A[ii * total_num + int(initial_rank[i])] = float(1.0); + } +} + +at::Tensor build_adjacency_matrix_forward(at::Tensor initial_rank) { + const auto total_num = initial_rank.size(0); + const auto topk = initial_rank.size(1); + const auto all_num = total_num * topk; + auto A = torch::zeros({total_num, total_num}, at::device(initial_rank.device()).dtype(at::ScalarType::Float)); + + const int threads = 1024; + const int blocks = (all_num + threads - 1) / threads; + + build_adjacency_matrix_kernel<<>>(initial_rank.data_ptr(), A.data_ptr(), total_num, topk, threads, all_num); + return A; + +} diff --git a/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/extension/adjacency_matrix/setup.py b/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/extension/adjacency_matrix/setup.py new file mode 100644 index 0000000..abd3e3a --- /dev/null +++ b/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/extension/adjacency_matrix/setup.py @@ -0,0 +1,36 @@ +""" + Understanding Image Retrieval Re-Ranking: A Graph Neural Network Perspective + + Xuanmeng Zhang, Minyue Jiang, Zhedong Zheng, Xiao Tan, Errui Ding, Yi Yang + + Project Page : https://github.com/Xuanmeng-Zhang/gnn-re-ranking + + Paper: https://arxiv.org/abs/2012.07620v2 + + ====================================================================== + + On the Market-1501 dataset, we accelerate the re-ranking processing from 89.2s to 9.4ms + with one K40m GPU, facilitating the real-time post-processing. Similarly, we observe + that our method achieves comparable or even better retrieval results on the other four + image retrieval benchmarks, i.e., VeRi-776, Oxford-5k, Paris-6k and University-1652, + with limited time cost. +""" + +from setuptools import Extension, setup +import torch +import torch.nn as nn +from torch.autograd import Function +from torch.utils.cpp_extension import CUDAExtension, BuildExtension + +setup( + name='build_adjacency_matrix', + ext_modules=[ + CUDAExtension( + 'build_adjacency_matrix', [ + 'build_adjacency_matrix.cpp', + 'build_adjacency_matrix_kernel.cu', + ] + ), + ], + cmdclass={'build_ext': BuildExtension} +) diff --git a/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/extension/make.sh b/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/extension/make.sh new file mode 100644 index 0000000..f0197ff --- /dev/null +++ b/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/extension/make.sh @@ -0,0 +1,4 @@ +cd adjacency_matrix +python setup.py install +cd ../propagation +python setup.py install \ No newline at end of file diff --git a/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/extension/propagation/gnn_propagate.cpp b/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/extension/propagation/gnn_propagate.cpp new file mode 100644 index 0000000..10a939f --- /dev/null +++ b/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/extension/propagation/gnn_propagate.cpp @@ -0,0 +1,21 @@ +#include +#include +#include + +at::Tensor gnn_propagate_forward(at::Tensor A, at::Tensor initial_rank, at::Tensor S); + + +#define CHECK_CUDA(x) AT_ASSERTM(x.type().is_cuda(), #x " must be a CUDA tensor") +#define CHECK_CONTIGUOUS(x) AT_ASSERTM(x.is_contiguous(), #x " must be contiguous") +#define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x) + +at::Tensor gnn_propagate(at::Tensor A ,at::Tensor initial_rank, at::Tensor S) { + CHECK_INPUT(A); + CHECK_INPUT(initial_rank); + CHECK_INPUT(S); + return gnn_propagate_forward(A, initial_rank, S); +} + +PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { + m.def("forward", &gnn_propagate, "gnn propagate (CUDA)"); +} \ No newline at end of file diff --git a/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/extension/propagation/gnn_propagate_kernel.cu b/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/extension/propagation/gnn_propagate_kernel.cu new file mode 100644 index 0000000..8bdebf1 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/extension/propagation/gnn_propagate_kernel.cu @@ -0,0 +1,36 @@ +#include + +#include +#include +#include +#include + +__global__ void gnn_propagate_forward_kernel(float* initial_rank, float* A, float* A_qe, float* S, const int sample_num, const int topk, const int total_num) { + int index = blockIdx.x * blockDim.x + threadIdx.x; + int stride = blockDim.x * gridDim.x; + for (int i = index; i < total_num; i += stride) { + int fea = i % sample_num; + int sample_index = i / sample_num; + float sum = 0.0; + for (int j = 0; j < topk ; j++) { + int topk_fea_index = int(initial_rank[sample_index*topk+j]) * sample_num + fea; + sum += A[ topk_fea_index] * S[sample_index*topk+j]; + } + A_qe[i] = sum; + } +} + +at::Tensor gnn_propagate_forward(at::Tensor A, at::Tensor initial_rank, at::Tensor S) { + const auto sample_num = A.size(0); + const auto topk = initial_rank.size(1); + + const auto total_num = sample_num * sample_num ; + auto A_qe = torch::zeros({sample_num, sample_num}, at::device(initial_rank.device()).dtype(at::ScalarType::Float)); + + const int threads = 1024; + const int blocks = (total_num + threads - 1) / threads; + + gnn_propagate_forward_kernel<<>>(initial_rank.data_ptr(), A.data_ptr(), A_qe.data_ptr(), S.data_ptr(), sample_num, topk, total_num); + return A_qe; + +} \ No newline at end of file diff --git a/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/extension/propagation/setup.py b/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/extension/propagation/setup.py new file mode 100644 index 0000000..1f7b43b --- /dev/null +++ b/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/extension/propagation/setup.py @@ -0,0 +1,36 @@ +""" + Understanding Image Retrieval Re-Ranking: A Graph Neural Network Perspective + + Xuanmeng Zhang, Minyue Jiang, Zhedong Zheng, Xiao Tan, Errui Ding, Yi Yang + + Project Page : https://github.com/Xuanmeng-Zhang/gnn-re-ranking + + Paper: https://arxiv.org/abs/2012.07620v2 + + ====================================================================== + + On the Market-1501 dataset, we accelerate the re-ranking processing from 89.2s to 9.4ms + with one K40m GPU, facilitating the real-time post-processing. Similarly, we observe + that our method achieves comparable or even better retrieval results on the other four + image retrieval benchmarks, i.e., VeRi-776, Oxford-5k, Paris-6k and University-1652, + with limited time cost. +""" + +from setuptools import Extension, setup +import torch +import torch.nn as nn +from torch.autograd import Function +from torch.utils.cpp_extension import CUDAExtension, BuildExtension + +setup( + name='gnn_propagate', + ext_modules=[ + CUDAExtension( + 'gnn_propagate', [ + 'gnn_propagate.cpp', + 'gnn_propagate_kernel.cu', + ] + ), + ], + cmdclass={'build_ext': BuildExtension} +) diff --git a/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/gnn_reranking.py b/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/gnn_reranking.py new file mode 100644 index 0000000..2c8cc53 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/gnn_reranking.py @@ -0,0 +1,59 @@ +""" + Understanding Image Retrieval Re-Ranking: A Graph Neural Network Perspective + + Xuanmeng Zhang, Minyue Jiang, Zhedong Zheng, Xiao Tan, Errui Ding, Yi Yang + + Project Page : https://github.com/Xuanmeng-Zhang/gnn-re-ranking + + Paper: https://arxiv.org/abs/2012.07620v2 + + ====================================================================== + + On the Market-1501 dataset, we accelerate the re-ranking processing from 89.2s to 9.4ms + with one K40m GPU, facilitating the real-time post-processing. Similarly, we observe + that our method achieves comparable or even better retrieval results on the other four + image retrieval benchmarks, i.e., VeRi-776, Oxford-5k, Paris-6k and University-1652, + with limited time cost. +""" + +import numpy as np +import torch + +import gnn_propagate +import build_adjacency_matrix +from utils import * + + +def gnn_reranking(X_q, X_g, k1, k2): + query_num, gallery_num = X_q.shape[0], X_g.shape[0] + + X_u = torch.cat((X_q, X_g), axis=0) + original_score = torch.mm(X_u, X_u.t()) + del X_u, X_q, X_g + + # initial ranking list + S, initial_rank = original_score.topk( + k=k1, dim=-1, largest=True, sorted=True + ) + + # stage 1 + A = build_adjacency_matrix.forward(initial_rank.float()) + S = S * S + + # stage 2 + if k2 != 1: + for i in range(2): + A = A + A.T + A = gnn_propagate.forward( + A, initial_rank[:, :k2].contiguous().float(), + S[:, :k2].contiguous().float() + ) + A_norm = torch.norm(A, p=2, dim=1, keepdim=True) + A = A.div(A_norm.expand_as(A)) + + cosine_similarity = torch.mm(A[:query_num, ], A[query_num:, ].t()) + del A, S + + L = torch.sort(-cosine_similarity, dim=1)[1] + L = L.data.cpu().numpy() + return L diff --git a/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/main.py b/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/main.py new file mode 100644 index 0000000..53ef6ac --- /dev/null +++ b/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/main.py @@ -0,0 +1,72 @@ +""" + Understanding Image Retrieval Re-Ranking: A Graph Neural Network Perspective + + Xuanmeng Zhang, Minyue Jiang, Zhedong Zheng, Xiao Tan, Errui Ding, Yi Yang + + Project Page : https://github.com/Xuanmeng-Zhang/gnn-re-ranking + + Paper: https://arxiv.org/abs/2012.07620v2 + + ====================================================================== + + On the Market-1501 dataset, we accelerate the re-ranking processing from 89.2s to 9.4ms + with one K40m GPU, facilitating the real-time post-processing. Similarly, we observe + that our method achieves comparable or even better retrieval results on the other four + image retrieval benchmarks, i.e., VeRi-776, Oxford-5k, Paris-6k and University-1652, + with limited time cost. +""" + +import os +import numpy as np +import argparse +import torch + +from utils import * +from gnn_reranking import * + +parser = argparse.ArgumentParser(description='Reranking_is_GNN') +parser.add_argument( + '--data_path', + type=str, + default='../xm_rerank_gpu_2/features/market_88_test.pkl', + help='path to dataset' +) +parser.add_argument( + '--k1', + type=int, + default=26, # Market-1501 + # default=60, # Veri-776 + help='parameter k1' +) +parser.add_argument( + '--k2', + type=int, + default=7, # Market-1501 + # default=10, # Veri-776 + help='parameter k2' +) + +args = parser.parse_args() + + +def main(): + data = load_pickle(args.data_path) + + query_cam = data['query_cam'] + query_label = data['query_label'] + gallery_cam = data['gallery_cam'] + gallery_label = data['gallery_label'] + + gallery_feature = torch.FloatTensor(data['gallery_f']) + query_feature = torch.FloatTensor(data['query_f']) + query_feature = query_feature.cuda() + gallery_feature = gallery_feature.cuda() + + indices = gnn_reranking(query_feature, gallery_feature, args.k1, args.k2) + evaluate_ranking_list( + indices, query_label, query_cam, gallery_label, gallery_cam + ) + + +if __name__ == '__main__': + main() diff --git a/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/utils.py b/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/utils.py new file mode 100644 index 0000000..5f1ed9f --- /dev/null +++ b/strong_sort/deep/reid/torchreid/utils/GPU-Re-Ranking/utils.py @@ -0,0 +1,137 @@ +""" + Understanding Image Retrieval Re-Ranking: A Graph Neural Network Perspective + + Xuanmeng Zhang, Minyue Jiang, Zhedong Zheng, Xiao Tan, Errui Ding, Yi Yang + + Project Page : https://github.com/Xuanmeng-Zhang/gnn-re-ranking + + Paper: https://arxiv.org/abs/2012.07620v2 + + ====================================================================== + + On the Market-1501 dataset, we accelerate the re-ranking processing from 89.2s to 9.4ms + with one K40m GPU, facilitating the real-time post-processing. Similarly, we observe + that our method achieves comparable or even better retrieval results on the other four + image retrieval benchmarks, i.e., VeRi-776, Oxford-5k, Paris-6k and University-1652, + with limited time cost. +""" + +import os +import numpy as np +import pickle +import torch + + +def load_pickle(pickle_path): + with open(pickle_path, 'rb') as f: + data = pickle.load(f) + return data + + +def save_pickle(pickle_path, data): + with open(pickle_path, 'wb') as f: + pickle.dump(data, f, protocol=pickle.HIGHEST_PROTOCOL) + + +def pairwise_squared_distance(x): + ''' + x : (n_samples, n_points, dims) + return : (n_samples, n_points, n_points) + ''' + x2s = (x * x).sum(-1, keepdim=True) + return x2s + x2s.transpose(-1, -2) - 2 * x @ x.transpose(-1, -2) + + +def pairwise_distance(x, y): + m, n = x.size(0), y.size(0) + + x = x.view(m, -1) + y = y.view(n, -1) + + dist = torch.pow(x, 2).sum( + dim=1, keepdim=True + ).expand(m, n) + torch.pow(y, 2).sum( + dim=1, keepdim=True + ).expand(n, m).t() + dist.addmm_(1, -2, x, y.t()) + + return dist + + +def cosine_similarity(x, y): + m, n = x.size(0), y.size(0) + + x = x.view(m, -1) + y = y.view(n, -1) + + y = y.t() + score = torch.mm(x, y) + + return score + + +def evaluate_ranking_list( + indices, query_label, query_cam, gallery_label, gallery_cam +): + CMC = np.zeros((len(gallery_label)), dtype=np.int) + ap = 0.0 + + for i in range(len(query_label)): + ap_tmp, CMC_tmp = evaluate( + indices[i], query_label[i], query_cam[i], gallery_label, + gallery_cam + ) + if CMC_tmp[0] == -1: + continue + CMC = CMC + CMC_tmp + ap += ap_tmp + + CMC = CMC.astype(np.float32) + CMC = CMC / len(query_label) #average CMC + print( + 'Rank@1:%f Rank@5:%f Rank@10:%f mAP:%f' % + (CMC[0], CMC[4], CMC[9], ap / len(query_label)) + ) + + +def evaluate(index, ql, qc, gl, gc): + query_index = np.argwhere(gl == ql) + camera_index = np.argwhere(gc == qc) + + good_index = np.setdiff1d(query_index, camera_index, assume_unique=True) + junk_index1 = np.argwhere(gl == -1) + junk_index2 = np.intersect1d(query_index, camera_index) + junk_index = np.append(junk_index2, junk_index1) #.flatten()) + + CMC_tmp = compute_mAP(index, good_index, junk_index) + return CMC_tmp + + +def compute_mAP(index, good_index, junk_index): + ap = 0 + cmc = np.zeros((len(index)), dtype=np.int) + if good_index.size == 0: # if empty + cmc[0] = -1 + return ap, cmc + + # remove junk_index + mask = np.in1d(index, junk_index, invert=True) + index = index[mask] + + # find good_index index + ngood = len(good_index) + mask = np.in1d(index, good_index) + rows_good = np.argwhere(mask == True) + rows_good = rows_good.flatten() + + cmc[rows_good[0]:] = 1 + for i in range(ngood): + d_recall = 1.0 / ngood + precision = (i+1) * 1.0 / (rows_good[i] + 1) + if rows_good[i] != 0: + old_precision = i * 1.0 / rows_good[i] + else: + old_precision = 1.0 + ap = ap + d_recall * (old_precision+precision) / 2 + + return ap, cmc diff --git a/strong_sort/deep/reid/torchreid/utils/__init__.py b/strong_sort/deep/reid/torchreid/utils/__init__.py new file mode 100644 index 0000000..50167c3 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/utils/__init__.py @@ -0,0 +1,10 @@ +from __future__ import absolute_import + +from .tools import * +from .rerank import re_ranking +from .loggers import * +from .avgmeter import * +from .reidtools import * +from .torchtools import * +from .model_complexity import compute_model_complexity +from .feature_extractor import FeatureExtractor diff --git a/strong_sort/deep/reid/torchreid/utils/avgmeter.py b/strong_sort/deep/reid/torchreid/utils/avgmeter.py new file mode 100644 index 0000000..b62d26d --- /dev/null +++ b/strong_sort/deep/reid/torchreid/utils/avgmeter.py @@ -0,0 +1,73 @@ +from __future__ import division, absolute_import +from collections import defaultdict +import torch + +__all__ = ['AverageMeter', 'MetricMeter'] + + +class AverageMeter(object): + """Computes and stores the average and current value. + + Examples:: + >>> # Initialize a meter to record loss + >>> losses = AverageMeter() + >>> # Update meter after every minibatch update + >>> losses.update(loss_value, batch_size) + """ + + def __init__(self): + self.reset() + + def reset(self): + self.val = 0 + self.avg = 0 + self.sum = 0 + self.count = 0 + + def update(self, val, n=1): + self.val = val + self.sum += val * n + self.count += n + self.avg = self.sum / self.count + + +class MetricMeter(object): + """A collection of metrics. + + Source: https://github.com/KaiyangZhou/Dassl.pytorch + + Examples:: + >>> # 1. Create an instance of MetricMeter + >>> metric = MetricMeter() + >>> # 2. Update using a dictionary as input + >>> input_dict = {'loss_1': value_1, 'loss_2': value_2} + >>> metric.update(input_dict) + >>> # 3. Convert to string and print + >>> print(str(metric)) + """ + + def __init__(self, delimiter='\t'): + self.meters = defaultdict(AverageMeter) + self.delimiter = delimiter + + def update(self, input_dict): + if input_dict is None: + return + + if not isinstance(input_dict, dict): + raise TypeError( + 'Input to MetricMeter.update() must be a dictionary' + ) + + for k, v in input_dict.items(): + if isinstance(v, torch.Tensor): + v = v.item() + self.meters[k].update(v) + + def __str__(self): + output_str = [] + for name, meter in self.meters.items(): + output_str.append( + '{} {:.4f} ({:.4f})'.format(name, meter.val, meter.avg) + ) + return self.delimiter.join(output_str) diff --git a/strong_sort/deep/reid/torchreid/utils/feature_extractor.py b/strong_sort/deep/reid/torchreid/utils/feature_extractor.py new file mode 100644 index 0000000..9c147e2 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/utils/feature_extractor.py @@ -0,0 +1,157 @@ +from __future__ import absolute_import +import numpy as np +import torch +import torchvision.transforms as T +from PIL import Image +from mb_models import mb_CA_auto_same_depth_build_model + +from torchreid.utils import ( + check_isfile, load_pretrained_weights, compute_model_complexity +) +from torchreid.models import build_model + + +class FeatureExtractor(object): + """A simple API for feature extraction. + + FeatureExtractor can be used like a python function, which + accepts input of the following types: + - a list of strings (image paths) + - a list of numpy.ndarray each with shape (H, W, C) + - a single string (image path) + - a single numpy.ndarray with shape (H, W, C) + - a torch.Tensor with shape (B, C, H, W) or (C, H, W) + + Returned is a torch tensor with shape (B, D) where D is the + feature dimension. + + Args: + model_name (str): model name. + model_path (str): path to model weights. + image_size (sequence or int): image height and width. + pixel_mean (list): pixel mean for normalization. + pixel_std (list): pixel std for normalization. + pixel_norm (bool): whether to normalize pixels. + device (str): 'cpu' or 'cuda' (could be specific gpu devices). + verbose (bool): show model details. + + Examples:: + + from torchreid.utils import FeatureExtractor + + extractor = FeatureExtractor( + model_name='osnet_x1_0', + model_path='a/b/c/model.pth.tar', + device='cuda' + ) + + image_list = [ + 'a/b/c/image001.jpg', + 'a/b/c/image002.jpg', + 'a/b/c/image003.jpg', + 'a/b/c/image004.jpg', + 'a/b/c/image005.jpg' + ] + + features = extractor(image_list) + print(features.shape) # output (5, 512) + """ + + def __init__( + self, + model_name='', + model_path='', + image_size=(256, 256), + pixel_mean=[0.4611, 0.4658, 0.4728], + pixel_std=[0.2552, 0.2502, 0.2520], + pixel_norm=True, + device='cuda', + verbose=True + ): + + model = build_model( + name='resnet50', + num_classes=29, + loss='softmax', + pretrained=False + ) + + trained_net = torch.load('./best_attr_net.pth') + model.load_state_dict(trained_net) + + model.eval() + + if verbose: + num_params, flops = compute_model_complexity( + model, (1, 3, image_size[0], image_size[1]) + ) + print('Model: {}'.format(model_name)) + print('- params: {:,}'.format(num_params)) + print('- flops: {:,}'.format(flops)) + + if model_path and check_isfile(model_path): + load_pretrained_weights(model, model_path) + + # Build transform functions + transforms = [] + transforms += [T.Resize(image_size)] + transforms += [T.ToTensor()] + if pixel_norm: + transforms += [T.Normalize(mean=pixel_mean, std=pixel_std)] + preprocess = T.Compose(transforms) + + to_pil = T.ToPILImage() + + device = torch.device(device) + model.to(device) + + # Class attributes + self.model = model + self.preprocess = preprocess + self.to_pil = to_pil + self.device = device + + def __call__(self, input): + if isinstance(input, list): + images = [] + + for element in input: + if isinstance(element, str): + image = Image.open(element).convert('RGB') + + elif isinstance(element, np.ndarray): + image = self.to_pil(element) + + else: + raise TypeError( + 'Type of each element must belong to [str | numpy.ndarray]' + ) + + image = self.preprocess(image) + images.append(image) + + images = torch.stack(images, dim=0) + images = images.to(self.device) + + elif isinstance(input, str): + image = Image.open(input).convert('RGB') + image = self.preprocess(image) + images = image.unsqueeze(0).to(self.device) + + elif isinstance(input, np.ndarray): + image = self.to_pil(input) + image = self.preprocess(image) + images = image.unsqueeze(0).to(self.device) + + elif isinstance(input, torch.Tensor): + if input.dim() == 3: + input = input.unsqueeze(0) + images = input.to(self.device) + + else: + raise NotImplementedError + + with torch.no_grad(): + features = self.model(images) + + return features diff --git a/strong_sort/deep/reid/torchreid/utils/loggers.py b/strong_sort/deep/reid/torchreid/utils/loggers.py new file mode 100644 index 0000000..f7fae3c --- /dev/null +++ b/strong_sort/deep/reid/torchreid/utils/loggers.py @@ -0,0 +1,146 @@ +from __future__ import absolute_import +import os +import sys +import os.path as osp + +from .tools import mkdir_if_missing + +__all__ = ['Logger', 'RankLogger'] + + +class Logger(object): + """Writes console output to external text file. + + Imported from ``_ + + Args: + fpath (str): directory to save logging file. + + Examples:: + >>> import sys + >>> import os + >>> import os.path as osp + >>> from torchreid.utils import Logger + >>> save_dir = 'log/resnet50-softmax-market1501' + >>> log_name = 'train.log' + >>> sys.stdout = Logger(osp.join(args.save_dir, log_name)) + """ + + def __init__(self, fpath=None): + self.console = sys.stdout + self.file = None + if fpath is not None: + mkdir_if_missing(osp.dirname(fpath)) + self.file = open(fpath, 'w') + + def __del__(self): + self.close() + + def __enter__(self): + pass + + def __exit__(self, *args): + self.close() + + def write(self, msg): + self.console.write(msg) + if self.file is not None: + self.file.write(msg) + + def flush(self): + self.console.flush() + if self.file is not None: + self.file.flush() + os.fsync(self.file.fileno()) + + def close(self): + self.console.close() + if self.file is not None: + self.file.close() + + +class RankLogger(object): + """Records the rank1 matching accuracy obtained for each + test dataset at specified evaluation steps and provides a function + to show the summarized results, which are convenient for analysis. + + Args: + sources (str or list): source dataset name(s). + targets (str or list): target dataset name(s). + + Examples:: + >>> from torchreid.utils import RankLogger + >>> s = 'market1501' + >>> t = 'market1501' + >>> ranklogger = RankLogger(s, t) + >>> ranklogger.write(t, 10, 0.5) + >>> ranklogger.write(t, 20, 0.7) + >>> ranklogger.write(t, 30, 0.9) + >>> ranklogger.show_summary() + >>> # You will see: + >>> # => Show performance summary + >>> # market1501 (source) + >>> # - epoch 10 rank1 50.0% + >>> # - epoch 20 rank1 70.0% + >>> # - epoch 30 rank1 90.0% + >>> # If there are multiple test datasets + >>> t = ['market1501', 'dukemtmcreid'] + >>> ranklogger = RankLogger(s, t) + >>> ranklogger.write(t[0], 10, 0.5) + >>> ranklogger.write(t[0], 20, 0.7) + >>> ranklogger.write(t[0], 30, 0.9) + >>> ranklogger.write(t[1], 10, 0.1) + >>> ranklogger.write(t[1], 20, 0.2) + >>> ranklogger.write(t[1], 30, 0.3) + >>> ranklogger.show_summary() + >>> # You can see: + >>> # => Show performance summary + >>> # market1501 (source) + >>> # - epoch 10 rank1 50.0% + >>> # - epoch 20 rank1 70.0% + >>> # - epoch 30 rank1 90.0% + >>> # dukemtmcreid (target) + >>> # - epoch 10 rank1 10.0% + >>> # - epoch 20 rank1 20.0% + >>> # - epoch 30 rank1 30.0% + """ + + def __init__(self, sources, targets): + self.sources = sources + self.targets = targets + + if isinstance(self.sources, str): + self.sources = [self.sources] + + if isinstance(self.targets, str): + self.targets = [self.targets] + + self.logger = { + name: { + 'epoch': [], + 'rank1': [] + } + for name in self.targets + } + + def write(self, name, epoch, rank1): + """Writes result. + + Args: + name (str): dataset name. + epoch (int): current epoch. + rank1 (float): rank1 result. + """ + self.logger[name]['epoch'].append(epoch) + self.logger[name]['rank1'].append(rank1) + + def show_summary(self): + """Shows saved results.""" + print('=> Show performance summary') + for name in self.targets: + from_where = 'source' if name in self.sources else 'target' + print('{} ({})'.format(name, from_where)) + for epoch, rank1 in zip( + self.logger[name]['epoch'], self.logger[name]['rank1'] + ): + print('- epoch {}\t rank1 {:.1%}'.format(epoch, rank1)) diff --git a/strong_sort/deep/reid/torchreid/utils/model_complexity.py b/strong_sort/deep/reid/torchreid/utils/model_complexity.py new file mode 100644 index 0000000..7d1dc1e --- /dev/null +++ b/strong_sort/deep/reid/torchreid/utils/model_complexity.py @@ -0,0 +1,363 @@ +from __future__ import division, print_function, absolute_import +import math +import numpy as np +from itertools import repeat +from collections import namedtuple, defaultdict +import torch + +__all__ = ['compute_model_complexity'] +""" +Utility +""" + + +def _ntuple(n): + + def parse(x): + if isinstance(x, int): + return tuple(repeat(x, n)) + return x + + return parse + + +_single = _ntuple(1) +_pair = _ntuple(2) +_triple = _ntuple(3) +""" +Convolution +""" + + +def hook_convNd(m, x, y): + k = torch.prod(torch.Tensor(m.kernel_size)).item() + cin = m.in_channels + flops_per_ele = k * cin # + (k*cin-1) + if m.bias is not None: + flops_per_ele += 1 + flops = flops_per_ele * y.numel() / m.groups + return int(flops) + + +""" +Pooling +""" + + +def hook_maxpool1d(m, x, y): + flops_per_ele = m.kernel_size - 1 + flops = flops_per_ele * y.numel() + return int(flops) + + +def hook_maxpool2d(m, x, y): + k = _pair(m.kernel_size) + k = torch.prod(torch.Tensor(k)).item() + # ops: compare + flops_per_ele = k - 1 + flops = flops_per_ele * y.numel() + return int(flops) + + +def hook_maxpool3d(m, x, y): + k = _triple(m.kernel_size) + k = torch.prod(torch.Tensor(k)).item() + flops_per_ele = k - 1 + flops = flops_per_ele * y.numel() + return int(flops) + + +def hook_avgpool1d(m, x, y): + flops_per_ele = m.kernel_size + flops = flops_per_ele * y.numel() + return int(flops) + + +def hook_avgpool2d(m, x, y): + k = _pair(m.kernel_size) + k = torch.prod(torch.Tensor(k)).item() + flops_per_ele = k + flops = flops_per_ele * y.numel() + return int(flops) + + +def hook_avgpool3d(m, x, y): + k = _triple(m.kernel_size) + k = torch.prod(torch.Tensor(k)).item() + flops_per_ele = k + flops = flops_per_ele * y.numel() + return int(flops) + + +def hook_adapmaxpool1d(m, x, y): + x = x[0] + out_size = m.output_size + k = math.ceil(x.size(2) / out_size) + flops_per_ele = k - 1 + flops = flops_per_ele * y.numel() + return int(flops) + + +def hook_adapmaxpool2d(m, x, y): + x = x[0] + out_size = _pair(m.output_size) + k = torch.Tensor(list(x.size()[2:])) / torch.Tensor(out_size) + k = torch.prod(torch.ceil(k)).item() + flops_per_ele = k - 1 + flops = flops_per_ele * y.numel() + return int(flops) + + +def hook_adapmaxpool3d(m, x, y): + x = x[0] + out_size = _triple(m.output_size) + k = torch.Tensor(list(x.size()[2:])) / torch.Tensor(out_size) + k = torch.prod(torch.ceil(k)).item() + flops_per_ele = k - 1 + flops = flops_per_ele * y.numel() + return int(flops) + + +def hook_adapavgpool1d(m, x, y): + x = x[0] + out_size = m.output_size + k = math.ceil(x.size(2) / out_size) + flops_per_ele = k + flops = flops_per_ele * y.numel() + return int(flops) + + +def hook_adapavgpool2d(m, x, y): + x = x[0] + out_size = _pair(m.output_size) + k = torch.Tensor(list(x.size()[2:])) / torch.Tensor(out_size) + k = torch.prod(torch.ceil(k)).item() + flops_per_ele = k + flops = flops_per_ele * y.numel() + return int(flops) + + +def hook_adapavgpool3d(m, x, y): + x = x[0] + out_size = _triple(m.output_size) + k = torch.Tensor(list(x.size()[2:])) / torch.Tensor(out_size) + k = torch.prod(torch.ceil(k)).item() + flops_per_ele = k + flops = flops_per_ele * y.numel() + return int(flops) + + +""" +Non-linear activations +""" + + +def hook_relu(m, x, y): + # eq: max(0, x) + num_ele = y.numel() + return int(num_ele) + + +def hook_leakyrelu(m, x, y): + # eq: max(0, x) + negative_slope*min(0, x) + num_ele = y.numel() + flops = 3 * num_ele + return int(flops) + + +""" +Normalization +""" + + +def hook_batchnormNd(m, x, y): + num_ele = y.numel() + flops = 2 * num_ele # mean and std + if m.affine: + flops += 2 * num_ele # gamma and beta + return int(flops) + + +def hook_instancenormNd(m, x, y): + return hook_batchnormNd(m, x, y) + + +def hook_groupnorm(m, x, y): + return hook_batchnormNd(m, x, y) + + +def hook_layernorm(m, x, y): + num_ele = y.numel() + flops = 2 * num_ele # mean and std + if m.elementwise_affine: + flops += 2 * num_ele # gamma and beta + return int(flops) + + +""" +Linear +""" + + +def hook_linear(m, x, y): + flops_per_ele = m.in_features # + (m.in_features-1) + if m.bias is not None: + flops_per_ele += 1 + flops = flops_per_ele * y.numel() + return int(flops) + + +__generic_flops_counter = { + # Convolution + 'Conv1d': hook_convNd, + 'Conv2d': hook_convNd, + 'Conv3d': hook_convNd, + # Pooling + 'MaxPool1d': hook_maxpool1d, + 'MaxPool2d': hook_maxpool2d, + 'MaxPool3d': hook_maxpool3d, + 'AvgPool1d': hook_avgpool1d, + 'AvgPool2d': hook_avgpool2d, + 'AvgPool3d': hook_avgpool3d, + 'AdaptiveMaxPool1d': hook_adapmaxpool1d, + 'AdaptiveMaxPool2d': hook_adapmaxpool2d, + 'AdaptiveMaxPool3d': hook_adapmaxpool3d, + 'AdaptiveAvgPool1d': hook_adapavgpool1d, + 'AdaptiveAvgPool2d': hook_adapavgpool2d, + 'AdaptiveAvgPool3d': hook_adapavgpool3d, + # Non-linear activations + 'ReLU': hook_relu, + 'ReLU6': hook_relu, + 'LeakyReLU': hook_leakyrelu, + # Normalization + 'BatchNorm1d': hook_batchnormNd, + 'BatchNorm2d': hook_batchnormNd, + 'BatchNorm3d': hook_batchnormNd, + 'InstanceNorm1d': hook_instancenormNd, + 'InstanceNorm2d': hook_instancenormNd, + 'InstanceNorm3d': hook_instancenormNd, + 'GroupNorm': hook_groupnorm, + 'LayerNorm': hook_layernorm, + # Linear + 'Linear': hook_linear, +} + +__conv_linear_flops_counter = { + # Convolution + 'Conv1d': hook_convNd, + 'Conv2d': hook_convNd, + 'Conv3d': hook_convNd, + # Linear + 'Linear': hook_linear, +} + + +def _get_flops_counter(only_conv_linear): + if only_conv_linear: + return __conv_linear_flops_counter + return __generic_flops_counter + + +def compute_model_complexity( + model, input_size, verbose=False, only_conv_linear=True +): + """Returns number of parameters and FLOPs. + + .. note:: + (1) this function only provides an estimate of the theoretical time complexity + rather than the actual running time which depends on implementations and hardware, + and (2) the FLOPs is only counted for layers that are used at test time. This means + that redundant layers such as person ID classification layer will be ignored as it + is discarded when doing feature extraction. Note that the inference graph depends on + how you construct the computations in ``forward()``. + + Args: + model (nn.Module): network model. + input_size (tuple): input size, e.g. (1, 3, 256, 128). + verbose (bool, optional): shows detailed complexity of + each module. Default is False. + only_conv_linear (bool, optional): only considers convolution + and linear layers when counting flops. Default is True. + If set to False, flops of all layers will be counted. + + Examples:: + >>> from torchreid import models, utils + >>> model = models.build_model(name='resnet50', num_classes=1000) + >>> num_params, flops = utils.compute_model_complexity(model, (1, 3, 256, 128), verbose=True) + """ + registered_handles = [] + layer_list = [] + layer = namedtuple('layer', ['class_name', 'params', 'flops']) + + def _add_hooks(m): + + def _has_submodule(m): + return len(list(m.children())) > 0 + + def _hook(m, x, y): + params = sum(p.numel() for p in m.parameters()) + class_name = str(m.__class__.__name__) + flops_counter = _get_flops_counter(only_conv_linear) + if class_name in flops_counter: + flops = flops_counter[class_name](m, x, y) + else: + flops = 0 + layer_list.append( + layer(class_name=class_name, params=params, flops=flops) + ) + + # only consider the very basic nn layer + if _has_submodule(m): + return + + handle = m.register_forward_hook(_hook) + registered_handles.append(handle) + + default_train_mode = model.training + + model.eval().apply(_add_hooks) + input = torch.rand(input_size) + if next(model.parameters()).is_cuda: + input = input.cuda() + model(input) # forward + + for handle in registered_handles: + handle.remove() + + model.train(default_train_mode) + + if verbose: + per_module_params = defaultdict(list) + per_module_flops = defaultdict(list) + + total_params, total_flops = 0, 0 + + for layer in layer_list: + total_params += layer.params + total_flops += layer.flops + if verbose: + per_module_params[layer.class_name].append(layer.params) + per_module_flops[layer.class_name].append(layer.flops) + + if verbose: + num_udscore = 55 + print(' {}'.format('-' * num_udscore)) + print(' Model complexity with input size {}'.format(input_size)) + print(' {}'.format('-' * num_udscore)) + for class_name in per_module_params: + params = int(np.sum(per_module_params[class_name])) + flops = int(np.sum(per_module_flops[class_name])) + print( + ' {} (params={:,}, flops={:,})'.format( + class_name, params, flops + ) + ) + print(' {}'.format('-' * num_udscore)) + print( + ' Total (params={:,}, flops={:,})'.format( + total_params, total_flops + ) + ) + print(' {}'.format('-' * num_udscore)) + + return total_params, total_flops diff --git a/strong_sort/deep/reid/torchreid/utils/reidtools.py b/strong_sort/deep/reid/torchreid/utils/reidtools.py new file mode 100644 index 0000000..acb8760 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/utils/reidtools.py @@ -0,0 +1,154 @@ +from __future__ import print_function, absolute_import +import numpy as np +import shutil +import os.path as osp +import cv2 + +from .tools import mkdir_if_missing + +__all__ = ['visualize_ranked_results'] + +GRID_SPACING = 10 +QUERY_EXTRA_SPACING = 90 +BW = 5 # border width +GREEN = (0, 255, 0) +RED = (0, 0, 255) + + +def visualize_ranked_results( + distmat, dataset, data_type, width=128, height=256, save_dir='', topk=10 +): + """Visualizes ranked results. + + Supports both image-reid and video-reid. + + For image-reid, ranks will be plotted in a single figure. For video-reid, ranks will be + saved in folders each containing a tracklet. + + Args: + distmat (numpy.ndarray): distance matrix of shape (num_query, num_gallery). + dataset (tuple): a 2-tuple containing (query, gallery), each of which contains + tuples of (img_path(s), pid, camid, dsetid). + data_type (str): "image" or "video". + width (int, optional): resized image width. Default is 128. + height (int, optional): resized image height. Default is 256. + save_dir (str): directory to save output images. + topk (int, optional): denoting top-k images in the rank list to be visualized. + Default is 10. + """ + num_q, num_g = distmat.shape + mkdir_if_missing(save_dir) + + print('# query: {}\n# gallery {}'.format(num_q, num_g)) + print('Visualizing top-{} ranks ...'.format(topk)) + + query, gallery = dataset + assert num_q == len(query) + assert num_g == len(gallery) + + indices = np.argsort(distmat, axis=1) + + def _cp_img_to(src, dst, rank, prefix, matched=False): + """ + Args: + src: image path or tuple (for vidreid) + dst: target directory + rank: int, denoting ranked position, starting from 1 + prefix: string + matched: bool + """ + if isinstance(src, (tuple, list)): + if prefix == 'gallery': + suffix = 'TRUE' if matched else 'FALSE' + dst = osp.join( + dst, prefix + '_top' + str(rank).zfill(3) + ) + '_' + suffix + else: + dst = osp.join(dst, prefix + '_top' + str(rank).zfill(3)) + mkdir_if_missing(dst) + for img_path in src: + shutil.copy(img_path, dst) + else: + dst = osp.join( + dst, prefix + '_top' + str(rank).zfill(3) + '_name_' + + osp.basename(src) + ) + shutil.copy(src, dst) + + for q_idx in range(num_q): + qimg_path, qpid, qcamid = query[q_idx][:3] + qimg_path_name = qimg_path[0] if isinstance( + qimg_path, (tuple, list) + ) else qimg_path + + if data_type == 'image': + qimg = cv2.imread(qimg_path) + qimg = cv2.resize(qimg, (width, height)) + qimg = cv2.copyMakeBorder( + qimg, BW, BW, BW, BW, cv2.BORDER_CONSTANT, value=(0, 0, 0) + ) + # resize twice to ensure that the border width is consistent across images + qimg = cv2.resize(qimg, (width, height)) + num_cols = topk + 1 + grid_img = 255 * np.ones( + ( + height, + num_cols*width + topk*GRID_SPACING + QUERY_EXTRA_SPACING, 3 + ), + dtype=np.uint8 + ) + grid_img[:, :width, :] = qimg + else: + qdir = osp.join( + save_dir, osp.basename(osp.splitext(qimg_path_name)[0]) + ) + mkdir_if_missing(qdir) + _cp_img_to(qimg_path, qdir, rank=0, prefix='query') + + rank_idx = 1 + for g_idx in indices[q_idx, :]: + gimg_path, gpid, gcamid = gallery[g_idx][:3] + invalid = (qpid == gpid) & (qcamid == gcamid) + + if not invalid: + matched = gpid == qpid + if data_type == 'image': + border_color = GREEN if matched else RED + gimg = cv2.imread(gimg_path) + gimg = cv2.resize(gimg, (width, height)) + gimg = cv2.copyMakeBorder( + gimg, + BW, + BW, + BW, + BW, + cv2.BORDER_CONSTANT, + value=border_color + ) + gimg = cv2.resize(gimg, (width, height)) + start = rank_idx*width + rank_idx*GRID_SPACING + QUERY_EXTRA_SPACING + end = ( + rank_idx+1 + ) * width + rank_idx*GRID_SPACING + QUERY_EXTRA_SPACING + grid_img[:, start:end, :] = gimg + else: + _cp_img_to( + gimg_path, + qdir, + rank=rank_idx, + prefix='gallery', + matched=matched + ) + + rank_idx += 1 + if rank_idx > topk: + break + + if data_type == 'image': + imname = osp.basename(osp.splitext(qimg_path_name)[0]) + cv2.imwrite(osp.join(save_dir, imname + '.jpg'), grid_img) + + if (q_idx+1) % 100 == 0: + print('- done {}/{}'.format(q_idx + 1, num_q)) + + print('Done. Images have been saved to "{}" ...'.format(save_dir)) diff --git a/strong_sort/deep/reid/torchreid/utils/rerank.py b/strong_sort/deep/reid/torchreid/utils/rerank.py new file mode 100644 index 0000000..efadf5a --- /dev/null +++ b/strong_sort/deep/reid/torchreid/utils/rerank.py @@ -0,0 +1,113 @@ +#!/usr/bin/env python2/python3 +# -*- coding: utf-8 -*- +""" +Source: https://github.com/zhunzhong07/person-re-ranking + +Created on Mon Jun 26 14:46:56 2017 +@author: luohao +Modified by Houjing Huang, 2017-12-22. +- This version accepts distance matrix instead of raw features. +- The difference of `/` division between python 2 and 3 is handled. +- numpy.float16 is replaced by numpy.float32 for numerical precision. + +CVPR2017 paper:Zhong Z, Zheng L, Cao D, et al. Re-ranking Person Re-identification with k-reciprocal Encoding[J]. 2017. +url:http://openaccess.thecvf.com/content_cvpr_2017/papers/Zhong_Re-Ranking_Person_Re-Identification_CVPR_2017_paper.pdf +Matlab version: https://github.com/zhunzhong07/person-re-ranking + +API +q_g_dist: query-gallery distance matrix, numpy array, shape [num_query, num_gallery] +q_q_dist: query-query distance matrix, numpy array, shape [num_query, num_query] +g_g_dist: gallery-gallery distance matrix, numpy array, shape [num_gallery, num_gallery] +k1, k2, lambda_value: parameters, the original paper is (k1=20, k2=6, lambda_value=0.3) +Returns: + final_dist: re-ranked distance, numpy array, shape [num_query, num_gallery] +""" +from __future__ import division, print_function, absolute_import +import numpy as np + +__all__ = ['re_ranking'] + + +def re_ranking(q_g_dist, q_q_dist, g_g_dist, k1=20, k2=6, lambda_value=0.3): + + # The following naming, e.g. gallery_num, is different from outer scope. + # Don't care about it. + + original_dist = np.concatenate( + [ + np.concatenate([q_q_dist, q_g_dist], axis=1), + np.concatenate([q_g_dist.T, g_g_dist], axis=1) + ], + axis=0 + ) + original_dist = np.power(original_dist, 2).astype(np.float32) + original_dist = np.transpose( + 1. * original_dist / np.max(original_dist, axis=0) + ) + V = np.zeros_like(original_dist).astype(np.float32) + initial_rank = np.argsort(original_dist).astype(np.int32) + + query_num = q_g_dist.shape[0] + gallery_num = q_g_dist.shape[0] + q_g_dist.shape[1] + all_num = gallery_num + + for i in range(all_num): + # k-reciprocal neighbors + forward_k_neigh_index = initial_rank[i, :k1 + 1] + backward_k_neigh_index = initial_rank[forward_k_neigh_index, :k1 + 1] + fi = np.where(backward_k_neigh_index == i)[0] + k_reciprocal_index = forward_k_neigh_index[fi] + k_reciprocal_expansion_index = k_reciprocal_index + for j in range(len(k_reciprocal_index)): + candidate = k_reciprocal_index[j] + candidate_forward_k_neigh_index = initial_rank[ + candidate, :int(np.around(k1 / 2.)) + 1] + candidate_backward_k_neigh_index = initial_rank[ + candidate_forward_k_neigh_index, :int(np.around(k1 / 2.)) + 1] + fi_candidate = np.where( + candidate_backward_k_neigh_index == candidate + )[0] + candidate_k_reciprocal_index = candidate_forward_k_neigh_index[ + fi_candidate] + if len( + np. + intersect1d(candidate_k_reciprocal_index, k_reciprocal_index) + ) > 2. / 3 * len(candidate_k_reciprocal_index): + k_reciprocal_expansion_index = np.append( + k_reciprocal_expansion_index, candidate_k_reciprocal_index + ) + + k_reciprocal_expansion_index = np.unique(k_reciprocal_expansion_index) + weight = np.exp(-original_dist[i, k_reciprocal_expansion_index]) + V[i, k_reciprocal_expansion_index] = 1. * weight / np.sum(weight) + original_dist = original_dist[:query_num, ] + if k2 != 1: + V_qe = np.zeros_like(V, dtype=np.float32) + for i in range(all_num): + V_qe[i, :] = np.mean(V[initial_rank[i, :k2], :], axis=0) + V = V_qe + del V_qe + del initial_rank + invIndex = [] + for i in range(gallery_num): + invIndex.append(np.where(V[:, i] != 0)[0]) + + jaccard_dist = np.zeros_like(original_dist, dtype=np.float32) + + for i in range(query_num): + temp_min = np.zeros(shape=[1, gallery_num], dtype=np.float32) + indNonZero = np.where(V[i, :] != 0)[0] + indImages = [] + indImages = [invIndex[ind] for ind in indNonZero] + for j in range(len(indNonZero)): + temp_min[0, indImages[j]] = temp_min[0, indImages[j]] + np.minimum( + V[i, indNonZero[j]], V[indImages[j], indNonZero[j]] + ) + jaccard_dist[i] = 1 - temp_min / (2.-temp_min) + + final_dist = jaccard_dist * (1-lambda_value) + original_dist*lambda_value + del original_dist + del V + del jaccard_dist + final_dist = final_dist[:query_num, query_num:] + return final_dist diff --git a/strong_sort/deep/reid/torchreid/utils/tools.py b/strong_sort/deep/reid/torchreid/utils/tools.py new file mode 100644 index 0000000..518fa18 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/utils/tools.py @@ -0,0 +1,143 @@ +from __future__ import division, print_function, absolute_import +import os +import sys +import json +import time +import errno +import numpy as np +import random +import os.path as osp +import warnings +import PIL +import torch +from PIL import Image + +__all__ = [ + 'mkdir_if_missing', 'check_isfile', 'read_json', 'write_json', + 'set_random_seed', 'download_url', 'read_image', 'collect_env_info', + 'listdir_nohidden' +] + + +def mkdir_if_missing(dirname): + """Creates dirname if it is missing.""" + if not osp.exists(dirname): + try: + os.makedirs(dirname) + except OSError as e: + if e.errno != errno.EEXIST: + raise + + +def check_isfile(fpath): + """Checks if the given path is a file. + + Args: + fpath (str): file path. + + Returns: + bool + """ + isfile = osp.isfile(fpath) + if not isfile: + warnings.warn('No file found at "{}"'.format(fpath)) + return isfile + + +def read_json(fpath): + """Reads json file from a path.""" + with open(fpath, 'r') as f: + obj = json.load(f) + return obj + + +def write_json(obj, fpath): + """Writes to a json file.""" + mkdir_if_missing(osp.dirname(fpath)) + with open(fpath, 'w') as f: + json.dump(obj, f, indent=4, separators=(',', ': ')) + + +def set_random_seed(seed): + random.seed(seed) + np.random.seed(seed) + torch.manual_seed(seed) + torch.cuda.manual_seed_all(seed) + + +def download_url(url, dst): + """Downloads file from a url to a destination. + + Args: + url (str): url to download file. + dst (str): destination path. + """ + from six.moves import urllib + print('* url="{}"'.format(url)) + print('* destination="{}"'.format(dst)) + + def _reporthook(count, block_size, total_size): + global start_time + if count == 0: + start_time = time.time() + return + duration = time.time() - start_time + progress_size = int(count * block_size) + speed = int(progress_size / (1024*duration)) + percent = int(count * block_size * 100 / total_size) + sys.stdout.write( + '\r...%d%%, %d MB, %d KB/s, %d seconds passed' % + (percent, progress_size / (1024*1024), speed, duration) + ) + sys.stdout.flush() + + urllib.request.urlretrieve(url, dst, _reporthook) + sys.stdout.write('\n') + + +def read_image(path): + """Reads image from path using ``PIL.Image``. + + Args: + path (str): path to an image. + + Returns: + PIL image + """ + got_img = False + if not osp.exists(path): + raise IOError('"{}" does not exist'.format(path)) + while not got_img: + try: + img = Image.open(path).convert('RGB') + got_img = True + except IOError: + print( + 'IOError incurred when reading "{}". Will redo. Don\'t worry. Just chill.' + .format(path) + ) + return img + + +def collect_env_info(): + """Returns env info as a string. + + Code source: github.com/facebookresearch/maskrcnn-benchmark + """ + from torch.utils.collect_env import get_pretty_env_info + env_str = get_pretty_env_info() + env_str += '\n Pillow ({})'.format(PIL.__version__) + return env_str + + +def listdir_nohidden(path, sort=False): + """List non-hidden items in a directory. + + Args: + path (str): directory path. + sort (bool): sort the items. + """ + items = [f for f in os.listdir(path) if not f.startswith('.')] + if sort: + items.sort() + return items diff --git a/strong_sort/deep/reid/torchreid/utils/torchtools.py b/strong_sort/deep/reid/torchreid/utils/torchtools.py new file mode 100644 index 0000000..e854278 --- /dev/null +++ b/strong_sort/deep/reid/torchreid/utils/torchtools.py @@ -0,0 +1,312 @@ +from __future__ import division, print_function, absolute_import +import pickle +import shutil +import os.path as osp +import warnings +from functools import partial +from collections import OrderedDict +import torch +import torch.nn as nn + +from .tools import mkdir_if_missing + +__all__ = [ + 'save_checkpoint', 'load_checkpoint', 'resume_from_checkpoint', + 'open_all_layers', 'open_specified_layers', 'count_num_param', + 'load_pretrained_weights' +] + + +def save_checkpoint( + state, save_dir, is_best=False, remove_module_from_keys=False +): + r"""Saves checkpoint. + + Args: + state (dict): dictionary. + save_dir (str): directory to save checkpoint. + is_best (bool, optional): if True, this checkpoint will be copied and named + ``model-best.pth.tar``. Default is False. + remove_module_from_keys (bool, optional): whether to remove "module." + from layer names. Default is False. + + Examples:: + >>> state = { + >>> 'state_dict': model.state_dict(), + >>> 'epoch': 10, + >>> 'rank1': 0.5, + >>> 'optimizer': optimizer.state_dict() + >>> } + >>> save_checkpoint(state, 'log/my_model') + """ + mkdir_if_missing(save_dir) + if remove_module_from_keys: + # remove 'module.' in state_dict's keys + state_dict = state['state_dict'] + new_state_dict = OrderedDict() + for k, v in state_dict.items(): + if k.startswith('module.'): + k = k[7:] + new_state_dict[k] = v + state['state_dict'] = new_state_dict + # save + epoch = state['epoch'] + fpath = osp.join(save_dir, 'model.pth.tar-' + str(epoch)) + torch.save(state, fpath) + print('Checkpoint saved to "{}"'.format(fpath)) + if is_best: + shutil.copy(fpath, osp.join(osp.dirname(fpath), 'model-best.pth.tar')) + + +def load_checkpoint(fpath): + r"""Loads checkpoint. + + ``UnicodeDecodeError`` can be well handled, which means + python2-saved files can be read from python3. + + Args: + fpath (str): path to checkpoint. + + Returns: + dict + + Examples:: + >>> from torchreid.utils import load_checkpoint + >>> fpath = 'log/my_model/model.pth.tar-10' + >>> checkpoint = load_checkpoint(fpath) + """ + if fpath is None: + raise ValueError('File path is None') + fpath = osp.abspath(osp.expanduser(fpath)) + if not osp.exists(fpath): + raise FileNotFoundError('File is not found at "{}"'.format(fpath)) + map_location = None if torch.cuda.is_available() else 'cpu' + try: + checkpoint = torch.load(fpath, map_location=map_location) + except UnicodeDecodeError: + pickle.load = partial(pickle.load, encoding="latin1") + pickle.Unpickler = partial(pickle.Unpickler, encoding="latin1") + checkpoint = torch.load( + fpath, pickle_module=pickle, map_location=map_location + ) + except Exception: + print('Unable to load checkpoint from "{}"'.format(fpath)) + raise + return checkpoint + + +def resume_from_checkpoint(fpath, model, optimizer=None, scheduler=None): + r"""Resumes training from a checkpoint. + + This will load (1) model weights and (2) ``state_dict`` + of optimizer if ``optimizer`` is not None. + + Args: + fpath (str): path to checkpoint. + model (nn.Module): model. + optimizer (Optimizer, optional): an Optimizer. + scheduler (LRScheduler, optional): an LRScheduler. + + Returns: + int: start_epoch. + + Examples:: + >>> from torchreid.utils import resume_from_checkpoint + >>> fpath = 'log/my_model/model.pth.tar-10' + >>> start_epoch = resume_from_checkpoint( + >>> fpath, model, optimizer, scheduler + >>> ) + """ + print('Loading checkpoint from "{}"'.format(fpath)) + checkpoint = load_checkpoint(fpath) + model.load_state_dict(checkpoint['state_dict']) + print('Loaded model weights') + if optimizer is not None and 'optimizer' in checkpoint.keys(): + optimizer.load_state_dict(checkpoint['optimizer']) + print('Loaded optimizer') + if scheduler is not None and 'scheduler' in checkpoint.keys(): + scheduler.load_state_dict(checkpoint['scheduler']) + print('Loaded scheduler') + start_epoch = checkpoint['epoch'] + print('Last epoch = {}'.format(start_epoch)) + if 'rank1' in checkpoint.keys(): + print('Last rank1 = {:.1%}'.format(checkpoint['rank1'])) + return start_epoch + + +def adjust_learning_rate( + optimizer, + base_lr, + epoch, + stepsize=20, + gamma=0.1, + linear_decay=False, + final_lr=0, + max_epoch=100 +): + r"""Adjusts learning rate. + + Deprecated. + """ + if linear_decay: + # linearly decay learning rate from base_lr to final_lr + frac_done = epoch / max_epoch + lr = frac_done*final_lr + (1.-frac_done) * base_lr + else: + # decay learning rate by gamma for every stepsize + lr = base_lr * (gamma**(epoch // stepsize)) + + for param_group in optimizer.param_groups: + param_group['lr'] = lr + + +def set_bn_to_eval(m): + r"""Sets BatchNorm layers to eval mode.""" + # 1. no update for running mean and var + # 2. scale and shift parameters are still trainable + classname = m.__class__.__name__ + if classname.find('BatchNorm') != -1: + m.eval() + + +def open_all_layers(model): + r"""Opens all layers in model for training. + + Examples:: + >>> from torchreid.utils import open_all_layers + >>> open_all_layers(model) + """ + model.train() + for p in model.parameters(): + p.requires_grad = True + + +def open_specified_layers(model, open_layers): + r"""Opens specified layers in model for training while keeping + other layers frozen. + + Args: + model (nn.Module): neural net model. + open_layers (str or list): layers open for training. + + Examples:: + >>> from torchreid.utils import open_specified_layers + >>> # Only model.classifier will be updated. + >>> open_layers = 'classifier' + >>> open_specified_layers(model, open_layers) + >>> # Only model.fc and model.classifier will be updated. + >>> open_layers = ['fc', 'classifier'] + >>> open_specified_layers(model, open_layers) + """ + if isinstance(model, nn.DataParallel): + model = model.module + + if isinstance(open_layers, str): + open_layers = [open_layers] + + for layer in open_layers: + assert hasattr( + model, layer + ), '"{}" is not an attribute of the model, please provide the correct name'.format( + layer + ) + + for name, module in model.named_children(): + if name in open_layers: + module.train() + for p in module.parameters(): + p.requires_grad = True + else: + module.eval() + for p in module.parameters(): + p.requires_grad = False + + +def count_num_param(model): + r"""Counts number of parameters in a model while ignoring ``self.classifier``. + + Args: + model (nn.Module): network model. + + Examples:: + >>> from torchreid.utils import count_num_param + >>> model_size = count_num_param(model) + + .. warning:: + + This method is deprecated in favor of + ``torchreid.utils.compute_model_complexity``. + """ + warnings.warn( + 'This method is deprecated and will be removed in the future.' + ) + + num_param = sum(p.numel() for p in model.parameters()) + + if isinstance(model, nn.DataParallel): + model = model.module + + if hasattr(model, + 'classifier') and isinstance(model.classifier, nn.Module): + # we ignore the classifier because it is unused at test time + num_param -= sum(p.numel() for p in model.classifier.parameters()) + + return num_param + + +def load_pretrained_weights(model, weight_path): + r"""Loads pretrianed weights to model. + + Features:: + - Incompatible layers (unmatched in name or size) will be ignored. + - Can automatically deal with keys containing "module.". + + Args: + model (nn.Module): network model. + weight_path (str): path to pretrained weights. + + Examples:: + >>> from torchreid.utils import load_pretrained_weights + >>> weight_path = 'log/my_model/model-best.pth.tar' + >>> load_pretrained_weights(model, weight_path) + """ + checkpoint = load_checkpoint(weight_path) + if 'state_dict' in checkpoint: + state_dict = checkpoint['state_dict'] + else: + state_dict = checkpoint + + model_dict = model.state_dict() + new_state_dict = OrderedDict() + matched_layers, discarded_layers = [], [] + + for k, v in state_dict.items(): + if k.startswith('module.'): + k = k[7:] # discard module. + + if k in model_dict and model_dict[k].size() == v.size(): + new_state_dict[k] = v + matched_layers.append(k) + else: + discarded_layers.append(k) + + model_dict.update(new_state_dict) + model.load_state_dict(model_dict) + + if len(matched_layers) == 0: + warnings.warn( + 'The pretrained weights "{}" cannot be loaded, ' + 'please check the key names manually ' + '(** ignored and continue **)'.format(weight_path) + ) + else: + print( + 'Successfully loaded pretrained weights from "{}"'. + format(weight_path) + ) + if len(discarded_layers) > 0: + print( + '** The following layers are discarded ' + 'due to unmatched keys or layer size: {}'. + format(discarded_layers) + ) diff --git a/strong_sort/deep/reid_model_factory.py b/strong_sort/deep/reid_model_factory.py new file mode 100644 index 0000000..335fd1d --- /dev/null +++ b/strong_sort/deep/reid_model_factory.py @@ -0,0 +1,125 @@ +__model_types = [ + 'resnet50', 'mlfn', 'hacnn', 'mobilenetv2_x1_0', 'mobilenetv2_x1_4', + 'osnet_x1_0', 'osnet_x0_75', 'osnet_x0_5', 'osnet_x0_25', + 'osnet_ibn_x1_0', 'osnet_ain_x1_0'] + +__trained_urls = { + + # market1501 models ######################################################## + 'resnet50_market1501.pt': + 'https://drive.google.com/uc?id=1dUUZ4rHDWohmsQXCRe2C_HbYkzz94iBV', + 'resnet50_dukemtmcreid.pt': + 'https://drive.google.com/uc?id=17ymnLglnc64NRvGOitY3BqMRS9UWd1wg', + 'resnet50_msmt17.pt': + 'https://drive.google.com/uc?id=1ep7RypVDOthCRIAqDnn4_N-UhkkFHJsj', + + 'resnet50_fc512_market1501.pt': + 'https://drive.google.com/uc?id=1kv8l5laX_YCdIGVCetjlNdzKIA3NvsSt', + 'resnet50_fc512_dukemtmcreid.pt': + 'https://drive.google.com/uc?id=13QN8Mp3XH81GK4BPGXobKHKyTGH50Rtx', + 'resnet50_fc512_msmt17.pt': + 'https://drive.google.com/uc?id=1fDJLcz4O5wxNSUvImIIjoaIF9u1Rwaud', + + 'mlfn_market1501.pt': + 'https://drive.google.com/uc?id=1wXcvhA_b1kpDfrt9s2Pma-MHxtj9pmvS', + 'mlfn_dukemtmcreid.pt': + 'https://drive.google.com/uc?id=1rExgrTNb0VCIcOnXfMsbwSUW1h2L1Bum', + 'mlfn_msmt17.pt': + 'https://drive.google.com/uc?id=18JzsZlJb3Wm7irCbZbZ07TN4IFKvR6p-', + + 'hacnn_market1501.pt': + 'https://drive.google.com/uc?id=1LRKIQduThwGxMDQMiVkTScBwR7WidmYF', + 'hacnn_dukemtmcreid.pt': + 'https://drive.google.com/uc?id=1zNm6tP4ozFUCUQ7Sv1Z98EAJWXJEhtYH', + 'hacnn_msmt17.pt': + 'https://drive.google.com/uc?id=1MsKRtPM5WJ3_Tk2xC0aGOO7pM3VaFDNZ', + + 'mobilenetv2_x1_0_market1501.pt': + 'https://drive.google.com/uc?id=18DgHC2ZJkjekVoqBWszD8_Xiikz-fewp', + 'mobilenetv2_x1_0_dukemtmcreid.pt': + 'https://drive.google.com/uc?id=1q1WU2FETRJ3BXcpVtfJUuqq4z3psetds', + 'mobilenetv2_x1_0_msmt17.pt': + 'https://drive.google.com/uc?id=1j50Hv14NOUAg7ZeB3frzfX-WYLi7SrhZ', + + 'mobilenetv2_x1_4_market1501.pt': + 'https://drive.google.com/uc?id=1t6JCqphJG-fwwPVkRLmGGyEBhGOf2GO5', + 'mobilenetv2_x1_4_dukemtmcreid.pt': + 'https://drive.google.com/uc?id=12uD5FeVqLg9-AFDju2L7SQxjmPb4zpBN', + 'mobilenetv2_x1_4_msmt17.pt': + 'https://drive.google.com/uc?id=1ZY5P2Zgm-3RbDpbXM0kIBMPvspeNIbXz', + + 'osnet_x1_0_market1501.pt': + 'https://drive.google.com/uc?id=1vduhq5DpN2q1g4fYEZfPI17MJeh9qyrA', + 'osnet_x1_0_dukemtmcreid.pt': + 'https://drive.google.com/uc?id=1QZO_4sNf4hdOKKKzKc-TZU9WW1v6zQbq', + 'osnet_x1_0_msmt17.pt': + 'https://drive.google.com/uc?id=112EMUfBPYeYg70w-syK6V6Mx8-Qb9Q1M', + + 'osnet_x0_75_market1501.pt': + 'https://drive.google.com/uc?id=1ozRaDSQw_EQ8_93OUmjDbvLXw9TnfPer', + 'osnet_x0_75_dukemtmcreid.pt': + 'https://drive.google.com/uc?id=1IE3KRaTPp4OUa6PGTFL_d5_KQSJbP0Or', + 'osnet_x0_75_msmt17.pt': + 'https://drive.google.com/uc?id=1QEGO6WnJ-BmUzVPd3q9NoaO_GsPNlmWc', + + 'osnet_x0_5_market1501.pt': + 'https://drive.google.com/uc?id=1PLB9rgqrUM7blWrg4QlprCuPT7ILYGKT', + 'osnet_x0_5_dukemtmcreid.pt': + 'https://drive.google.com/uc?id=1KoUVqmiST175hnkALg9XuTi1oYpqcyTu', + 'osnet_x0_5_msmt17.pt': + 'https://drive.google.com/uc?id=1UT3AxIaDvS2PdxzZmbkLmjtiqq7AIKCv', + + 'osnet_x0_25_market1501.pt': + 'https://drive.google.com/uc?id=1z1UghYvOTtjx7kEoRfmqSMu-z62J6MAj', + 'osnet_x0_25_dukemtmcreid.pt': + 'https://drive.google.com/uc?id=1eumrtiXT4NOspjyEV4j8cHmlOaaCGk5l', + 'osnet_x0_25_msmt17.pt': + 'https://drive.google.com/uc?id=1sSwXSUlj4_tHZequ_iZ8w_Jh0VaRQMqF', + + ####### market1501 models ################################################## + 'resnet50_msmt17.pt': + 'https://drive.google.com/uc?id=1yiBteqgIZoOeywE8AhGmEQl7FTVwrQmf', + 'osnet_x1_0_msmt17.pt': + 'https://drive.google.com/uc?id=1IosIFlLiulGIjwW3H8uMRmx3MzPwf86x', + 'osnet_x0_75_msmt17.pt': + 'https://drive.google.com/uc?id=1fhjSS_7SUGCioIf2SWXaRGPqIY9j7-uw', + + 'osnet_x0_5_msmt17.pt': + 'https://drive.google.com/uc?id=1DHgmb6XV4fwG3n-CnCM0zdL9nMsZ9_RF', + 'osnet_x0_25_msmt17.pt': + 'https://drive.google.com/uc?id=1Kkx2zW89jq_NETu4u42CFZTMVD5Hwm6e', + 'osnet_ibn_x1_0_msmt17.pt': + 'https://drive.google.com/uc?id=1q3Sj2ii34NlfxA4LvmHdWO_75NDRmECJ', + 'osnet_ain_x1_0_msmt17.pt': + 'https://drive.google.com/uc?id=1SigwBE6mPdqiJMqhuIY4aqC7--5CsMal', +} + + +def show_downloadeable_models(): + print('\nAvailable ReID models for automatic download') + print(list(__trained_urls.keys())) + + +def get_model_url(model): + model = str(model).rsplit('/', 1)[-1] + if model in __trained_urls: + return __trained_urls[model] + else: + None + + +def is_model_in_model_types(model): + model = str(model).rsplit('/', 1)[-1].split('.')[0] + if model in __model_types: + return True + else: + return False + + +def get_model_name(model): + model = str(model).rsplit('/', 1)[-1].split('.')[0] + for x in __model_types: + if x in model: + return x + return None + diff --git a/strong_sort/sort/__init__.py b/strong_sort/sort/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/strong_sort/sort/detection.py b/strong_sort/sort/detection.py new file mode 100644 index 0000000..22fbd60 --- /dev/null +++ b/strong_sort/sort/detection.py @@ -0,0 +1,49 @@ +# vim: expandtab:ts=4:sw=4 +import numpy as np + + +class Detection(object): + """ + This class represents a bounding box detection in a single image. + + Parameters + ---------- + tlwh : array_like + Bounding box in format `(x, y, w, h)`. + confidence : float + Detector confidence score. + feature : array_like + A feature vector that describes the object contained in this image. + + Attributes + ---------- + tlwh : ndarray + Bounding box in format `(top left x, top left y, width, height)`. + confidence : ndarray + Detector confidence score. + feature : ndarray | NoneType + A feature vector that describes the object contained in this image. + + """ + + def __init__(self, tlwh, confidence, feature): + self.tlwh = np.asarray(tlwh, dtype=np.float) + self.confidence = float(confidence) + self.feature = np.asarray(feature.cpu(), dtype=np.float32) + + def to_tlbr(self): + """Convert bounding box to format `(min x, min y, max x, max y)`, i.e., + `(top left, bottom right)`. + """ + ret = self.tlwh.copy() + ret[2:] += ret[:2] + return ret + + def to_xyah(self): + """Convert bounding box to format `(center x, center y, aspect ratio, + height)`, where the aspect ratio is `width / height`. + """ + ret = self.tlwh.copy() + ret[:2] += ret[2:] / 2 + ret[2] /= ret[3] + return ret diff --git a/strong_sort/sort/iou_matching.py b/strong_sort/sort/iou_matching.py new file mode 100644 index 0000000..62d5a3f --- /dev/null +++ b/strong_sort/sort/iou_matching.py @@ -0,0 +1,82 @@ +# vim: expandtab:ts=4:sw=4 +from __future__ import absolute_import +import numpy as np +from . import linear_assignment + + +def iou(bbox, candidates): + """Computer intersection over union. + + Parameters + ---------- + bbox : ndarray + A bounding box in format `(top left x, top left y, width, height)`. + candidates : ndarray + A matrix of candidate bounding boxes (one per row) in the same format + as `bbox`. + + Returns + ------- + ndarray + The intersection over union in [0, 1] between the `bbox` and each + candidate. A higher score means a larger fraction of the `bbox` is + occluded by the candidate. + + """ + bbox_tl, bbox_br = bbox[:2], bbox[:2] + bbox[2:] + candidates_tl = candidates[:, :2] + candidates_br = candidates[:, :2] + candidates[:, 2:] + + tl = np.c_[np.maximum(bbox_tl[0], candidates_tl[:, 0])[:, np.newaxis], + np.maximum(bbox_tl[1], candidates_tl[:, 1])[:, np.newaxis]] + br = np.c_[np.minimum(bbox_br[0], candidates_br[:, 0])[:, np.newaxis], + np.minimum(bbox_br[1], candidates_br[:, 1])[:, np.newaxis]] + wh = np.maximum(0., br - tl) + + area_intersection = wh.prod(axis=1) + area_bbox = bbox[2:].prod() + area_candidates = candidates[:, 2:].prod(axis=1) + return area_intersection / (area_bbox + area_candidates - area_intersection) + + +def iou_cost(tracks, detections, track_indices=None, + detection_indices=None): + """An intersection over union distance metric. + + Parameters + ---------- + tracks : List[deep_sort.track.Track] + A list of tracks. + detections : List[deep_sort.detection.Detection] + A list of detections. + track_indices : Optional[List[int]] + A list of indices to tracks that should be matched. Defaults to + all `tracks`. + detection_indices : Optional[List[int]] + A list of indices to detections that should be matched. Defaults + to all `detections`. + + Returns + ------- + ndarray + Returns a cost matrix of shape + len(track_indices), len(detection_indices) where entry (i, j) is + `1 - iou(tracks[track_indices[i]], detections[detection_indices[j]])`. + + """ + if track_indices is None: + track_indices = np.arange(len(tracks)) + if detection_indices is None: + detection_indices = np.arange(len(detections)) + + cost_matrix = np.zeros((len(track_indices), len(detection_indices))) + for row, track_idx in enumerate(track_indices): + if tracks[track_idx].time_since_update > 1: + cost_matrix[row, :] = linear_assignment.INFTY_COST + continue + + bbox = tracks[track_idx].to_tlwh() + candidates = np.asarray( + [detections[i].tlwh for i in detection_indices]) + cost_matrix[row, :] = 1. - iou(bbox, candidates) + return cost_matrix diff --git a/strong_sort/sort/kalman_filter.py b/strong_sort/sort/kalman_filter.py new file mode 100644 index 0000000..87c48d7 --- /dev/null +++ b/strong_sort/sort/kalman_filter.py @@ -0,0 +1,214 @@ +# vim: expandtab:ts=4:sw=4 +import numpy as np +import scipy.linalg +""" +Table for the 0.95 quantile of the chi-square distribution with N degrees of +freedom (contains values for N=1, ..., 9). Taken from MATLAB/Octave's chi2inv +function and used as Mahalanobis gating threshold. +""" +chi2inv95 = { + 1: 3.8415, + 2: 5.9915, + 3: 7.8147, + 4: 9.4877, + 5: 11.070, + 6: 12.592, + 7: 14.067, + 8: 15.507, + 9: 16.919} + + +class KalmanFilter(object): + """ + A simple Kalman filter for tracking bounding boxes in image space. + The 8-dimensional state space + x, y, a, h, vx, vy, va, vh + contains the bounding box center position (x, y), aspect ratio a, height h, + and their respective velocities. + Object motion follows a constant velocity model. The bounding box location + (x, y, a, h) is taken as direct observation of the state space (linear + observation model). + """ + + def __init__(self): + ndim, dt = 4, 1. + + # Create Kalman filter model matrices. + self._motion_mat = np.eye(2 * ndim, 2 * ndim) + for i in range(ndim): + self._motion_mat[i, ndim + i] = dt + + self._update_mat = np.eye(ndim, 2 * ndim) + + # Motion and observation uncertainty are chosen relative to the current + # state estimate. These weights control the amount of uncertainty in + # the model. This is a bit hacky. + self._std_weight_position = 1. / 20 + self._std_weight_velocity = 1. / 160 + + def initiate(self, measurement): + """Create track from unassociated measurement. + Parameters + ---------- + measurement : ndarray + Bounding box coordinates (x, y, a, h) with center position (x, y), + aspect ratio a, and height h. + Returns + ------- + (ndarray, ndarray) + Returns the mean vector (8 dimensional) and covariance matrix (8x8 + dimensional) of the new track. Unobserved velocities are initialized + to 0 mean. + """ + mean_pos = measurement + mean_vel = np.zeros_like(mean_pos) + mean = np.r_[mean_pos, mean_vel] + + std = [ + 2 * self._std_weight_position * measurement[0], # the center point x + 2 * self._std_weight_position * measurement[1], # the center point y + 1 * measurement[2], # the ratio of width/height + 2 * self._std_weight_position * measurement[3], # the height + 10 * self._std_weight_velocity * measurement[0], + 10 * self._std_weight_velocity * measurement[1], + 0.1 * measurement[2], + 10 * self._std_weight_velocity * measurement[3]] + covariance = np.diag(np.square(std)) + return mean, covariance + + def predict(self, mean, covariance): + """Run Kalman filter prediction step. + Parameters + ---------- + mean : ndarray + The 8 dimensional mean vector of the object state at the previous + time step. + covariance : ndarray + The 8x8 dimensional covariance matrix of the object state at the + previous time step. + Returns + ------- + (ndarray, ndarray) + Returns the mean vector and covariance matrix of the predicted + state. Unobserved velocities are initialized to 0 mean. + """ + std_pos = [ + self._std_weight_position * mean[0], + self._std_weight_position * mean[1], + 1 * mean[2], + self._std_weight_position * mean[3]] + std_vel = [ + self._std_weight_velocity * mean[0], + self._std_weight_velocity * mean[1], + 0.1 * mean[2], + self._std_weight_velocity * mean[3]] + motion_cov = np.diag(np.square(np.r_[std_pos, std_vel])) + + mean = np.dot(self._motion_mat, mean) + covariance = np.linalg.multi_dot(( + self._motion_mat, covariance, self._motion_mat.T)) + motion_cov + + return mean, covariance + + def project(self, mean, covariance, confidence=.0): + """Project state distribution to measurement space. + Parameters + ---------- + mean : ndarray + The state's mean vector (8 dimensional array). + covariance : ndarray + The state's covariance matrix (8x8 dimensional). + confidence: (dyh) 检测框置信度 + Returns + ------- + (ndarray, ndarray) + Returns the projected mean and covariance matrix of the given state + estimate. + """ + std = [ + self._std_weight_position * mean[3], + self._std_weight_position * mean[3], + 1e-1, + self._std_weight_position * mean[3]] + + + std = [(1 - confidence) * x for x in std] + + innovation_cov = np.diag(np.square(std)) + + mean = np.dot(self._update_mat, mean) + covariance = np.linalg.multi_dot(( + self._update_mat, covariance, self._update_mat.T)) + return mean, covariance + innovation_cov + + def update(self, mean, covariance, measurement, confidence=.0): + """Run Kalman filter correction step. + Parameters + ---------- + mean : ndarray + The predicted state's mean vector (8 dimensional). + covariance : ndarray + The state's covariance matrix (8x8 dimensional). + measurement : ndarray + The 4 dimensional measurement vector (x, y, a, h), where (x, y) + is the center position, a the aspect ratio, and h the height of the + bounding box. + confidence: (dyh)检测框置信度 + Returns + ------- + (ndarray, ndarray) + Returns the measurement-corrected state distribution. + """ + projected_mean, projected_cov = self.project(mean, covariance, confidence) + + chol_factor, lower = scipy.linalg.cho_factor( + projected_cov, lower=True, check_finite=False) + kalman_gain = scipy.linalg.cho_solve( + (chol_factor, lower), np.dot(covariance, self._update_mat.T).T, + check_finite=False).T + innovation = measurement - projected_mean + + new_mean = mean + np.dot(innovation, kalman_gain.T) + new_covariance = covariance - np.linalg.multi_dot(( + kalman_gain, projected_cov, kalman_gain.T)) + return new_mean, new_covariance + + def gating_distance(self, mean, covariance, measurements, + only_position=False): + """Compute gating distance between state distribution and measurements. + A suitable distance threshold can be obtained from `chi2inv95`. If + `only_position` is False, the chi-square distribution has 4 degrees of + freedom, otherwise 2. + Parameters + ---------- + mean : ndarray + Mean vector over the state distribution (8 dimensional). + covariance : ndarray + Covariance of the state distribution (8x8 dimensional). + measurements : ndarray + An Nx4 dimensional matrix of N measurements, each in + format (x, y, a, h) where (x, y) is the bounding box center + position, a the aspect ratio, and h the height. + only_position : Optional[bool] + If True, distance computation is done with respect to the bounding + box center position only. + Returns + ------- + ndarray + Returns an array of length N, where the i-th element contains the + squared Mahalanobis distance between (mean, covariance) and + `measurements[i]`. + """ + mean, covariance = self.project(mean, covariance) + + if only_position: + mean, covariance = mean[:2], covariance[:2, :2] + measurements = measurements[:, :2] + + cholesky_factor = np.linalg.cholesky(covariance) + d = measurements - mean + z = scipy.linalg.solve_triangular( + cholesky_factor, d.T, lower=True, check_finite=False, + overwrite_b=True) + squared_maha = np.sum(z * z, axis=0) + return squared_maha \ No newline at end of file diff --git a/strong_sort/sort/linear_assignment.py b/strong_sort/sort/linear_assignment.py new file mode 100644 index 0000000..924895f --- /dev/null +++ b/strong_sort/sort/linear_assignment.py @@ -0,0 +1,174 @@ +# vim: expandtab:ts=4:sw=4 +from __future__ import absolute_import +import numpy as np +from scipy.optimize import linear_sum_assignment +from . import kalman_filter + + +INFTY_COST = 1e+5 + + +def min_cost_matching( + distance_metric, max_distance, tracks, detections, track_indices=None, + detection_indices=None): + """Solve linear assignment problem. + Parameters + ---------- + distance_metric : Callable[List[Track], List[Detection], List[int], List[int]) -> ndarray + The distance metric is given a list of tracks and detections as well as + a list of N track indices and M detection indices. The metric should + return the NxM dimensional cost matrix, where element (i, j) is the + association cost between the i-th track in the given track indices and + the j-th detection in the given detection_indices. + max_distance : float + Gating threshold. Associations with cost larger than this value are + disregarded. + tracks : List[track.Track] + A list of predicted tracks at the current time step. + detections : List[detection.Detection] + A list of detections at the current time step. + track_indices : List[int] + List of track indices that maps rows in `cost_matrix` to tracks in + `tracks` (see description above). + detection_indices : List[int] + List of detection indices that maps columns in `cost_matrix` to + detections in `detections` (see description above). + Returns + ------- + (List[(int, int)], List[int], List[int]) + Returns a tuple with the following three entries: + * A list of matched track and detection indices. + * A list of unmatched track indices. + * A list of unmatched detection indices. + """ + if track_indices is None: + track_indices = np.arange(len(tracks)) + if detection_indices is None: + detection_indices = np.arange(len(detections)) + + if len(detection_indices) == 0 or len(track_indices) == 0: + return [], track_indices, detection_indices # Nothing to match. + + cost_matrix = distance_metric( + tracks, detections, track_indices, detection_indices) + cost_matrix[cost_matrix > max_distance] = max_distance + 1e-5 + row_indices, col_indices = linear_sum_assignment(cost_matrix) + + matches, unmatched_tracks, unmatched_detections = [], [], [] + for col, detection_idx in enumerate(detection_indices): + if col not in col_indices: + unmatched_detections.append(detection_idx) + for row, track_idx in enumerate(track_indices): + if row not in row_indices: + unmatched_tracks.append(track_idx) + for row, col in zip(row_indices, col_indices): + track_idx = track_indices[row] + detection_idx = detection_indices[col] + if cost_matrix[row, col] > max_distance: + unmatched_tracks.append(track_idx) + unmatched_detections.append(detection_idx) + else: + matches.append((track_idx, detection_idx)) + return matches, unmatched_tracks, unmatched_detections + + +def matching_cascade( + distance_metric, max_distance, cascade_depth, tracks, detections, + track_indices=None, detection_indices=None): + """Run matching cascade. + Parameters + ---------- + distance_metric : Callable[List[Track], List[Detection], List[int], List[int]) -> ndarray + The distance metric is given a list of tracks and detections as well as + a list of N track indices and M detection indices. The metric should + return the NxM dimensional cost matrix, where element (i, j) is the + association cost between the i-th track in the given track indices and + the j-th detection in the given detection indices. + max_distance : float + Gating threshold. Associations with cost larger than this value are + disregarded. + cascade_depth: int + The cascade depth, should be se to the maximum track age. + tracks : List[track.Track] + A list of predicted tracks at the current time step. + detections : List[detection.Detection] + A list of detections at the current time step. + track_indices : Optional[List[int]] + List of track indices that maps rows in `cost_matrix` to tracks in + `tracks` (see description above). Defaults to all tracks. + detection_indices : Optional[List[int]] + List of detection indices that maps columns in `cost_matrix` to + detections in `detections` (see description above). Defaults to all + detections. + Returns + ------- + (List[(int, int)], List[int], List[int]) + Returns a tuple with the following three entries: + * A list of matched track and detection indices. + * A list of unmatched track indices. + * A list of unmatched detection indices. + """ + if track_indices is None: + track_indices = list(range(len(tracks))) + if detection_indices is None: + detection_indices = list(range(len(detections))) + + unmatched_detections = detection_indices + matches = [] + track_indices_l = [ + k for k in track_indices + # if tracks[k].time_since_update == 1 + level + ] + matches_l, _, unmatched_detections = \ + min_cost_matching( + distance_metric, max_distance, tracks, detections, + track_indices_l, unmatched_detections) + matches += matches_l + unmatched_tracks = list(set(track_indices) - set(k for k, _ in matches)) + return matches, unmatched_tracks, unmatched_detections + + +def gate_cost_matrix( + cost_matrix, tracks, detections, track_indices, detection_indices, + gated_cost=INFTY_COST, only_position=False): + """Invalidate infeasible entries in cost matrix based on the state + distributions obtained by Kalman filtering. + Parameters + ---------- + kf : The Kalman filter. + cost_matrix : ndarray + The NxM dimensional cost matrix, where N is the number of track indices + and M is the number of detection indices, such that entry (i, j) is the + association cost between `tracks[track_indices[i]]` and + `detections[detection_indices[j]]`. + tracks : List[track.Track] + A list of predicted tracks at the current time step. + detections : List[detection.Detection] + A list of detections at the current time step. + track_indices : List[int] + List of track indices that maps rows in `cost_matrix` to tracks in + `tracks` (see description above). + detection_indices : List[int] + List of detection indices that maps columns in `cost_matrix` to + detections in `detections` (see description above). + gated_cost : Optional[float] + Entries in the cost matrix corresponding to infeasible associations are + set this value. Defaults to a very large value. + only_position : Optional[bool] + If True, only the x, y position of the state distribution is considered + during gating. Defaults to False. + Returns + ------- + ndarray + Returns the modified cost matrix. + """ + gating_dim = 2 if only_position else 4 + gating_threshold = kalman_filter.chi2inv95[gating_dim] + measurements = np.asarray( + [detections[i].to_xyah() for i in detection_indices]) + for row, track_idx in enumerate(track_indices): + track = tracks[track_idx] + gating_distance = track.kf.gating_distance(track.mean, track.covariance, measurements, only_position) + cost_matrix[row, gating_distance > gating_threshold] = gated_cost + cost_matrix[row] = 0.995 * cost_matrix[row] + (1 - 0.995) * gating_distance + return cost_matrix \ No newline at end of file diff --git a/strong_sort/sort/nn_matching.py b/strong_sort/sort/nn_matching.py new file mode 100644 index 0000000..fa5cd08 --- /dev/null +++ b/strong_sort/sort/nn_matching.py @@ -0,0 +1,164 @@ +# vim: expandtab:ts=4:sw=4 +import numpy as np +import sys +import torch +sys.path.append('strong_sort/deep/reid') +from torchreid.metrics.distance import compute_distance_matrix + + +def _pdist(a, b): + """Compute pair-wise squared distance between points in `a` and `b`. + Parameters + ---------- + a : array_like + An NxM matrix of N samples of dimensionality M. + b : array_like + An LxM matrix of L samples of dimensionality M. + Returns + ------- + ndarray + Returns a matrix of size len(a), len(b) such that eleement (i, j) + contains the squared distance between `a[i]` and `b[j]`. + """ + a, b = np.asarray(a), np.asarray(b) + if len(a) == 0 or len(b) == 0: + return np.zeros((len(a), len(b))) + a2, b2 = np.square(a).sum(axis=1), np.square(b).sum(axis=1) + r2 = -2. * np.dot(a, b.T) + a2[:, None] + b2[None, :] + r2 = np.clip(r2, 0., float(np.inf)) + return r2 + + +def _cosine_distance(a, b, data_is_normalized=False): + """Compute pair-wise cosine distance between points in `a` and `b`. + Parameters + ---------- + a : array_like + An NxM matrix of N samples of dimensionality M. + b : array_like + An LxM matrix of L samples of dimensionality M. + data_is_normalized : Optional[bool] + If True, assumes rows in a and b are unit length vectors. + Otherwise, a and b are explicitly normalized to lenght 1. + Returns + ------- + ndarray + Returns a matrix of size len(a), len(b) such that eleement (i, j) + contains the squared distance between `a[i]` and `b[j]`. + """ + if not data_is_normalized: + a = np.asarray(a) / np.linalg.norm(a, axis=1, keepdims=True) + b = np.asarray(b) / np.linalg.norm(b, axis=1, keepdims=True) + return 1. - np.dot(a, b.T) + + +def _nn_euclidean_distance(x, y): + """ Helper function for nearest neighbor distance metric (Euclidean). + Parameters + ---------- + x : ndarray + A matrix of N row-vectors (sample points). + y : ndarray + A matrix of M row-vectors (query points). + Returns + ------- + ndarray + A vector of length M that contains for each entry in `y` the + smallest Euclidean distance to a sample in `x`. + """ + x_ = torch.from_numpy(np.asarray(x) / np.linalg.norm(x, axis=1, keepdims=True)) + y_ = torch.from_numpy(np.asarray(y) / np.linalg.norm(y, axis=1, keepdims=True)) + distances = compute_distance_matrix(x_, y_, metric='euclidean') + return np.maximum(0.0, torch.min(distances, axis=0)[0].numpy()) + + +def _nn_cosine_distance(x, y): + """ Helper function for nearest neighbor distance metric (cosine). + Parameters + ---------- + x : ndarray + A matrix of N row-vectors (sample points). + y : ndarray + A matrix of M row-vectors (query points). + Returns + ------- + ndarray + A vector of length M that contains for each entry in `y` the + smallest cosine distance to a sample in `x`. + """ + x_ = torch.from_numpy(np.asarray(x)) + y_ = torch.from_numpy(np.asarray(y)) + distances = compute_distance_matrix(x_, y_, metric='cosine') + distances = distances.cpu().detach().numpy() + return distances.min(axis=0) + + +class NearestNeighborDistanceMetric(object): + """ + A nearest neighbor distance metric that, for each target, returns + the closest distance to any sample that has been observed so far. + Parameters + ---------- + metric : str + Either "euclidean" or "cosine". + matching_threshold: float + The matching threshold. Samples with larger distance are considered an + invalid match. + budget : Optional[int] + If not None, fix samples per class to at most this number. Removes + the oldest samples when the budget is reached. + Attributes + ---------- + samples : Dict[int -> List[ndarray]] + A dictionary that maps from target identities to the list of samples + that have been observed so far. + """ + + def __init__(self, metric, matching_threshold, budget=None): + if metric == "euclidean": + self._metric = _nn_euclidean_distance + elif metric == "cosine": + self._metric = _nn_cosine_distance + else: + raise ValueError( + "Invalid metric; must be either 'euclidean' or 'cosine'") + self.matching_threshold = matching_threshold + self.budget = budget + self.samples = {} + + def partial_fit(self, features, targets, active_targets): + """Update the distance metric with new data. + Parameters + ---------- + features : ndarray + An NxM matrix of N features of dimensionality M. + targets : ndarray + An integer array of associated target identities. + active_targets : List[int] + A list of targets that are currently present in the scene. + """ + for feature, target in zip(features, targets): + self.samples.setdefault(target, []).append(feature) + if self.budget is not None: + self.samples[target] = self.samples[target][-self.budget:] + self.samples = {k: self.samples[k] for k in active_targets} + + def distance(self, features, targets): + """Compute distance between features and targets. + Parameters + ---------- + features : ndarray + An NxM matrix of N features of dimensionality M. + targets : List[int] + A list of targets to match the given `features` against. + Returns + ------- + ndarray + Returns a cost matrix of shape len(targets), len(features), where + element (i, j) contains the closest squared distance between + `targets[i]` and `features[j]`. + """ + cost_matrix = np.zeros((len(targets), len(features))) + for i, target in enumerate(targets): + cost_matrix[i, :] = self._metric(self.samples[target], features) + return cost_matrix \ No newline at end of file diff --git a/strong_sort/sort/preprocessing.py b/strong_sort/sort/preprocessing.py new file mode 100644 index 0000000..5493b12 --- /dev/null +++ b/strong_sort/sort/preprocessing.py @@ -0,0 +1,73 @@ +# vim: expandtab:ts=4:sw=4 +import numpy as np +import cv2 + + +def non_max_suppression(boxes, max_bbox_overlap, scores=None): + """Suppress overlapping detections. + + Original code from [1]_ has been adapted to include confidence score. + + .. [1] http://www.pyimagesearch.com/2015/02/16/ + faster-non-maximum-suppression-python/ + + Examples + -------- + + >>> boxes = [d.roi for d in detections] + >>> scores = [d.confidence for d in detections] + >>> indices = non_max_suppression(boxes, max_bbox_overlap, scores) + >>> detections = [detections[i] for i in indices] + + Parameters + ---------- + boxes : ndarray + Array of ROIs (x, y, width, height). + max_bbox_overlap : float + ROIs that overlap more than this values are suppressed. + scores : Optional[array_like] + Detector confidence score. + + Returns + ------- + List[int] + Returns indices of detections that have survived non-maxima suppression. + + """ + if len(boxes) == 0: + return [] + + boxes = boxes.astype(np.float) + pick = [] + + x1 = boxes[:, 0] + y1 = boxes[:, 1] + x2 = boxes[:, 2] + boxes[:, 0] + y2 = boxes[:, 3] + boxes[:, 1] + + area = (x2 - x1 + 1) * (y2 - y1 + 1) + if scores is not None: + idxs = np.argsort(scores) + else: + idxs = np.argsort(y2) + + while len(idxs) > 0: + last = len(idxs) - 1 + i = idxs[last] + pick.append(i) + + xx1 = np.maximum(x1[i], x1[idxs[:last]]) + yy1 = np.maximum(y1[i], y1[idxs[:last]]) + xx2 = np.minimum(x2[i], x2[idxs[:last]]) + yy2 = np.minimum(y2[i], y2[idxs[:last]]) + + w = np.maximum(0, xx2 - xx1 + 1) + h = np.maximum(0, yy2 - yy1 + 1) + + overlap = (w * h) / area[idxs[:last]] + + idxs = np.delete( + idxs, np.concatenate( + ([last], np.where(overlap > max_bbox_overlap)[0]))) + + return pick diff --git a/strong_sort/sort/track.py b/strong_sort/sort/track.py new file mode 100644 index 0000000..9f3d0e1 --- /dev/null +++ b/strong_sort/sort/track.py @@ -0,0 +1,305 @@ +# vim: expandtab:ts=4:sw=4 +import cv2 +import numpy as np +from strong_sort.sort.kalman_filter import KalmanFilter + + +class TrackState: + """ + Enumeration type for the single target track state. Newly created tracks are + classified as `tentative` until enough evidence has been collected. Then, + the track state is changed to `confirmed`. Tracks that are no longer alive + are classified as `deleted` to mark them for removal from the set of active + tracks. + + """ + + Tentative = 1 + Confirmed = 2 + Deleted = 3 + + +class Track: + """ + A single target track with state space `(x, y, a, h)` and associated + velocities, where `(x, y)` is the center of the bounding box, `a` is the + aspect ratio and `h` is the height. + + Parameters + ---------- + mean : ndarray + Mean vector of the initial state distribution. + covariance : ndarray + Covariance matrix of the initial state distribution. + track_id : int + A unique track identifier. + n_init : int + Number of consecutive detections before the track is confirmed. The + track state is set to `Deleted` if a miss occurs within the first + `n_init` frames. + max_age : int + The maximum number of consecutive misses before the track state is + set to `Deleted`. + feature : Optional[ndarray] + Feature vector of the detection this track originates from. If not None, + this feature is added to the `features` cache. + + Attributes + ---------- + mean : ndarray + Mean vector of the initial state distribution. + covariance : ndarray + Covariance matrix of the initial state distribution. + track_id : int + A unique track identifier. + hits : int + Total number of measurement updates. + age : int + Total number of frames since first occurance. + time_since_update : int + Total number of frames since last measurement update. + state : TrackState + The current track state. + features : List[ndarray] + A cache of features. On each measurement update, the associated feature + vector is added to this list. + + """ + + def __init__(self, detection, track_id, class_id, conf, n_init, max_age, ema_alpha, + feature=None): + self.track_id = track_id + self.class_id = int(class_id) + self.hits = 1 + self.age = 1 + self.time_since_update = 0 + self.ema_alpha = ema_alpha + + self.state = TrackState.Tentative + self.features = [] + if feature is not None: + feature /= np.linalg.norm(feature) + self.features.append(feature) + + self.conf = conf + self._n_init = n_init + self._max_age = max_age + + self.kf = KalmanFilter() + self.mean, self.covariance = self.kf.initiate(detection) + + def to_tlwh(self): + """Get current position in bounding box format `(top left x, top left y, + width, height)`. + + Returns + ------- + ndarray + The bounding box. + + """ + ret = self.mean[:4].copy() + ret[2] *= ret[3] + ret[:2] -= ret[2:] / 2 + return ret + + def to_tlbr(self): + """Get kf estimated current position in bounding box format `(min x, miny, max x, + max y)`. + + Returns + ------- + ndarray + The predicted kf bounding box. + + """ + ret = self.to_tlwh() + ret[2:] = ret[:2] + ret[2:] + return ret + + + def ECC(self, src, dst, warp_mode = cv2.MOTION_EUCLIDEAN, eps = 1e-5, + max_iter = 100, scale = 0.1, align = False): + """Compute the warp matrix from src to dst. + Parameters + ---------- + src : ndarray + An NxM matrix of source img(BGR or Gray), it must be the same format as dst. + dst : ndarray + An NxM matrix of target img(BGR or Gray). + warp_mode: flags of opencv + translation: cv2.MOTION_TRANSLATION + rotated and shifted: cv2.MOTION_EUCLIDEAN + affine(shift,rotated,shear): cv2.MOTION_AFFINE + homography(3d): cv2.MOTION_HOMOGRAPHY + eps: float + the threshold of the increment in the correlation coefficient between two iterations + max_iter: int + the number of iterations. + scale: float or [int, int] + scale_ratio: float + scale_size: [W, H] + align: bool + whether to warp affine or perspective transforms to the source image + Returns + ------- + warp matrix : ndarray + Returns the warp matrix from src to dst. + if motion models is homography, the warp matrix will be 3x3, otherwise 2x3 + src_aligned: ndarray + aligned source image of gray + """ + + # skip if current and previous frame are not initialized (1st inference) + if (src.any() or dst.any() is None): + return None, None + # skip if current and previous fames are not the same size + elif (src.shape != dst.shape): + return None, None + + # BGR2GRAY + if src.ndim == 3: + # Convert images to grayscale + src = cv2.cvtColor(src, cv2.COLOR_BGR2GRAY) + dst = cv2.cvtColor(dst, cv2.COLOR_BGR2GRAY) + + # make the imgs smaller to speed up + if scale is not None: + if isinstance(scale, float) or isinstance(scale, int): + if scale != 1: + src_r = cv2.resize(src, (0, 0), fx = scale, fy = scale,interpolation = cv2.INTER_LINEAR) + dst_r = cv2.resize(dst, (0, 0), fx = scale, fy = scale,interpolation = cv2.INTER_LINEAR) + scale = [scale, scale] + else: + src_r, dst_r = src, dst + scale = None + else: + if scale[0] != src.shape[1] and scale[1] != src.shape[0]: + src_r = cv2.resize(src, (scale[0], scale[1]), interpolation = cv2.INTER_LINEAR) + dst_r = cv2.resize(dst, (scale[0], scale[1]), interpolation=cv2.INTER_LINEAR) + scale = [scale[0] / src.shape[1], scale[1] / src.shape[0]] + else: + src_r, dst_r = src, dst + scale = None + else: + src_r, dst_r = src, dst + + # Define 2x3 or 3x3 matrices and initialize the matrix to identity + if warp_mode == cv2.MOTION_HOMOGRAPHY : + warp_matrix = np.eye(3, 3, dtype=np.float32) + else : + warp_matrix = np.eye(2, 3, dtype=np.float32) + + # Define termination criteria + criteria = (cv2.TERM_CRITERIA_EPS | cv2.TERM_CRITERIA_COUNT, max_iter, eps) + + # Run the ECC algorithm. The results are stored in warp_matrix. + try: + (cc, warp_matrix) = cv2.findTransformECC (src_r, dst_r, warp_matrix, warp_mode, criteria, None, 1) + except cv2.error as e: + return None, None + + + if scale is not None: + warp_matrix[0, 2] = warp_matrix[0, 2] / scale[0] + warp_matrix[1, 2] = warp_matrix[1, 2] / scale[1] + + if align: + sz = src.shape + if warp_mode == cv2.MOTION_HOMOGRAPHY: + # Use warpPerspective for Homography + src_aligned = cv2.warpPerspective(src, warp_matrix, (sz[1],sz[0]), flags=cv2.INTER_LINEAR) + else : + # Use warpAffine for Translation, Euclidean and Affine + src_aligned = cv2.warpAffine(src, warp_matrix, (sz[1],sz[0]), flags=cv2.INTER_LINEAR) + return warp_matrix, src_aligned + else: + return warp_matrix, None + + + def get_matrix(self, matrix): + eye = np.eye(3) + dist = np.linalg.norm(eye - matrix) + if dist < 100: + return matrix + else: + return eye + + def camera_update(self, previous_frame, next_frame): + warp_matrix, src_aligned = self.ECC(previous_frame, next_frame) + if warp_matrix is None and src_aligned is None: + return + [a,b] = warp_matrix + warp_matrix=np.array([a,b,[0,0,1]]) + warp_matrix = warp_matrix.tolist() + matrix = self.get_matrix(warp_matrix) + + x1, y1, x2, y2 = self.to_tlbr() + x1_, y1_, _ = matrix @ np.array([x1, y1, 1]).T + x2_, y2_, _ = matrix @ np.array([x2, y2, 1]).T + w, h = x2_ - x1_, y2_ - y1_ + cx, cy = x1_ + w / 2, y1_ + h / 2 + self.mean[:4] = [cx, cy, w / h, h] + + + def increment_age(self): + self.age += 1 + self.time_since_update += 1 + + def predict(self, kf): + """Propagate the state distribution to the current time step using a + Kalman filter prediction step. + + Parameters + ---------- + kf : kalman_filter.KalmanFilter + The Kalman filter. + + """ + self.mean, self.covariance = self.kf.predict(self.mean, self.covariance) + self.age += 1 + self.time_since_update += 1 + + def update(self, detection, class_id, conf): + """Perform Kalman filter measurement update step and update the feature + cache. + Parameters + ---------- + detection : Detection + The associated detection. + """ + self.conf = conf + self.class_id = class_id.int() + self.mean, self.covariance = self.kf.update(self.mean, self.covariance, detection.to_xyah(), detection.confidence) + + feature = detection.feature / np.linalg.norm(detection.feature) + + smooth_feat = self.ema_alpha * self.features[-1] + (1 - self.ema_alpha) * feature + smooth_feat /= np.linalg.norm(smooth_feat) + self.features = [smooth_feat] + + self.hits += 1 + self.time_since_update = 0 + if self.state == TrackState.Tentative and self.hits >= self._n_init: + self.state = TrackState.Confirmed + + def mark_missed(self): + """Mark this track as missed (no association at the current time step). + """ + if self.state == TrackState.Tentative: + self.state = TrackState.Deleted + elif self.time_since_update > self._max_age: + self.state = TrackState.Deleted + + def is_tentative(self): + """Returns True if this track is tentative (unconfirmed). + """ + return self.state == TrackState.Tentative + + def is_confirmed(self): + """Returns True if this track is confirmed.""" + return self.state == TrackState.Confirmed + + def is_deleted(self): + """Returns True if this track is dead and should be deleted.""" + return self.state == TrackState.Deleted diff --git a/strong_sort/sort/tracker.py b/strong_sort/sort/tracker.py new file mode 100644 index 0000000..e0bcce6 --- /dev/null +++ b/strong_sort/sort/tracker.py @@ -0,0 +1,177 @@ +# vim: expandtab:ts=4:sw=4 +from __future__ import absolute_import +import numpy as np +from . import kalman_filter +from . import linear_assignment +from . import iou_matching +from .track import Track + + +class Tracker: + """ + This is the multi-target tracker. + Parameters + ---------- + metric : nn_matching.NearestNeighborDistanceMetric + A distance metric for measurement-to-track association. + max_age : int + Maximum number of missed misses before a track is deleted. + n_init : int + Number of consecutive detections before the track is confirmed. The + track state is set to `Deleted` if a miss occurs within the first + `n_init` frames. + Attributes + ---------- + metric : nn_matching.NearestNeighborDistanceMetric + The distance metric used for measurement to track association. + max_age : int + Maximum number of missed misses before a track is deleted. + n_init : int + Number of frames that a track remains in initialization phase. + kf : kalman_filter.KalmanFilter + A Kalman filter to filter target trajectories in image space. + tracks : List[Track] + The list of active tracks at the current time step. + """ + GATING_THRESHOLD = np.sqrt(kalman_filter.chi2inv95[4]) + + def __init__(self, metric, max_iou_distance=0.9, max_age=30, n_init=3, _lambda=0, ema_alpha=0.9, mc_lambda=0.995): + self.metric = metric + self.max_iou_distance = max_iou_distance + self.max_age = max_age + self.n_init = n_init + self._lambda = _lambda + self.ema_alpha = ema_alpha + self.mc_lambda = mc_lambda + + self.kf = kalman_filter.KalmanFilter() + self.tracks = [] + self._next_id = 1 + + def predict(self): + """Propagate track state distributions one time step forward. + + This function should be called once every time step, before `update`. + """ + for track in self.tracks: + track.predict(self.kf) + + def increment_ages(self): + for track in self.tracks: + track.increment_age() + track.mark_missed() + + def camera_update(self, previous_img, current_img): + for track in self.tracks: + track.camera_update(previous_img, current_img) + + def update(self, detections, classes, confidences): + """Perform measurement update and track management. + + Parameters + ---------- + detections : List[deep_sort.detection.Detection] + A list of detections at the current time step. + + """ + # Run matching cascade. + matches, unmatched_tracks, unmatched_detections = \ + self._match(detections) + + # Update track set. + for track_idx, detection_idx in matches: + self.tracks[track_idx].update( + detections[detection_idx], classes[detection_idx], confidences[detection_idx]) + for track_idx in unmatched_tracks: + self.tracks[track_idx].mark_missed() + for detection_idx in unmatched_detections: + self._initiate_track(detections[detection_idx], classes[detection_idx].item(), confidences[detection_idx].item()) + self.tracks = [t for t in self.tracks if not t.is_deleted()] + + # Update distance metric. + active_targets = [t.track_id for t in self.tracks if t.is_confirmed()] + features, targets = [], [] + for track in self.tracks: + if not track.is_confirmed(): + continue + features += track.features + targets += [track.track_id for _ in track.features] + self.metric.partial_fit(np.asarray(features), np.asarray(targets), active_targets) + + def _full_cost_metric(self, tracks, dets, track_indices, detection_indices): + """ + This implements the full lambda-based cost-metric. However, in doing so, it disregards + the possibility to gate the position only which is provided by + linear_assignment.gate_cost_matrix(). Instead, I gate by everything. + Note that the Mahalanobis distance is itself an unnormalised metric. Given the cosine + distance being normalised, we employ a quick and dirty normalisation based on the + threshold: that is, we divide the positional-cost by the gating threshold, thus ensuring + that the valid values range 0-1. + Note also that the authors work with the squared distance. I also sqrt this, so that it + is more intuitive in terms of values. + """ + # Compute First the Position-based Cost Matrix + pos_cost = np.empty([len(track_indices), len(detection_indices)]) + msrs = np.asarray([dets[i].to_xyah() for i in detection_indices]) + for row, track_idx in enumerate(track_indices): + pos_cost[row, :] = np.sqrt( + self.kf.gating_distance( + tracks[track_idx].mean, tracks[track_idx].covariance, msrs, False + ) + ) / self.GATING_THRESHOLD + pos_gate = pos_cost > 1.0 + # Now Compute the Appearance-based Cost Matrix + app_cost = self.metric.distance( + np.array([dets[i].feature for i in detection_indices]), + np.array([tracks[i].track_id for i in track_indices]), + ) + app_gate = app_cost > self.metric.matching_threshold + # Now combine and threshold + cost_matrix = self._lambda * pos_cost + (1 - self._lambda) * app_cost + cost_matrix[np.logical_or(pos_gate, app_gate)] = linear_assignment.INFTY_COST + # Return Matrix + return cost_matrix + + def _match(self, detections): + + def gated_metric(tracks, dets, track_indices, detection_indices): + features = np.array([dets[i].feature for i in detection_indices]) + targets = np.array([tracks[i].track_id for i in track_indices]) + cost_matrix = self.metric.distance(features, targets) + cost_matrix = linear_assignment.gate_cost_matrix(cost_matrix, tracks, dets, track_indices, detection_indices) + + return cost_matrix + + # Split track set into confirmed and unconfirmed tracks. + confirmed_tracks = [ + i for i, t in enumerate(self.tracks) if t.is_confirmed()] + unconfirmed_tracks = [ + i for i, t in enumerate(self.tracks) if not t.is_confirmed()] + + # Associate confirmed tracks using appearance features. + matches_a, unmatched_tracks_a, unmatched_detections = \ + linear_assignment.matching_cascade( + gated_metric, self.metric.matching_threshold, self.max_age, + self.tracks, detections, confirmed_tracks) + + # Associate remaining tracks together with unconfirmed tracks using IOU. + iou_track_candidates = unconfirmed_tracks + [ + k for k in unmatched_tracks_a if + self.tracks[k].time_since_update == 1] + unmatched_tracks_a = [ + k for k in unmatched_tracks_a if + self.tracks[k].time_since_update != 1] + matches_b, unmatched_tracks_b, unmatched_detections = \ + linear_assignment.min_cost_matching( + iou_matching.iou_cost, self.max_iou_distance, self.tracks, + detections, iou_track_candidates, unmatched_detections) + + matches = matches_a + matches_b + unmatched_tracks = list(set(unmatched_tracks_a + unmatched_tracks_b)) + return matches, unmatched_tracks, unmatched_detections + + def _initiate_track(self, detection, class_id, conf): + self.tracks.append(Track( + detection.to_xyah(), self._next_id, class_id, conf, self.n_init, self.max_age, self.ema_alpha, + detection.feature)) + self._next_id += 1 diff --git a/strong_sort/strong_sort.py b/strong_sort/strong_sort.py new file mode 100644 index 0000000..0e879be --- /dev/null +++ b/strong_sort/strong_sort.py @@ -0,0 +1,148 @@ +import numpy as np +import torch +import sys +import gdown +from os.path import exists as file_exists, join + +from .sort.nn_matching import NearestNeighborDistanceMetric +from .sort.detection import Detection +from .sort.tracker import Tracker +from .deep.reid_model_factory import show_downloadeable_models, get_model_url, get_model_name + +from torchreid.utils import FeatureExtractor +from torchreid.utils.tools import download_url + +__all__ = ['StrongSORT'] + + +class StrongSORT(object): + def __init__(self, + model_weights, + device, max_dist=0.2, + max_iou_distance=0.7, + max_age=70, n_init=3, + nn_budget=100, + mc_lambda=0.995, + ema_alpha=0.9 + ): + model_name = get_model_name(model_weights) + model_url = get_model_url(model_weights) + + if not file_exists(model_weights) and model_url is not None: + gdown.download(model_url, str(model_weights), quiet=False) + elif file_exists(model_weights): + pass + elif model_url is None: + print('No URL associated to the chosen DeepSort weights. Choose between:') + show_downloadeable_models() + exit() + + self.extractor = FeatureExtractor( + # get rid of dataset information DeepSort model name + model_name=model_name, + model_path=model_weights, + image_size=(256, 256), + device=str(device) + ) + + self.max_dist = max_dist + metric = NearestNeighborDistanceMetric( + "cosine", self.max_dist, nn_budget) + self.tracker = Tracker( + metric, max_iou_distance=max_iou_distance, max_age=max_age, n_init=n_init) + + def update(self, bbox_xywh, confidences, classes, ori_img): + self.height, self.width = ori_img.shape[:2] + # generate detections + features = self._get_features(bbox_xywh, ori_img) + attrs = features[0] + features = features[1] + bbox_tlwh = self._xywh_to_tlwh(bbox_xywh) + detections = [Detection(bbox_tlwh[i], conf, features[i]) for i, conf in enumerate( + confidences)] + + # run on non-maximum supression + boxes = np.array([d.tlwh for d in detections]) + scores = np.array([d.confidence for d in detections]) + + # update tracker + self.tracker.predict() + self.tracker.update(detections, classes, confidences) + + # output bbox identities + outputs = [] + for track in self.tracker.tracks: + if not track.is_confirmed() or track.time_since_update > 1: + continue + + box = track.to_tlwh() + x1, y1, x2, y2 = self._tlwh_to_xyxy(box) + + track_id = track.track_id + class_id = track.class_id + conf = track.conf + outputs.append(np.array([x1, y1, x2, y2, track_id, class_id, conf])) + if len(outputs) > 0: + outputs = np.stack(outputs, axis=0) + return outputs, attrs + + """ + TODO: + Convert bbox from xc_yc_w_h to xtl_ytl_w_h + Thanks JieChen91@github.com for reporting this bug! + """ + @staticmethod + def _xywh_to_tlwh(bbox_xywh): + if isinstance(bbox_xywh, np.ndarray): + bbox_tlwh = bbox_xywh.copy() + elif isinstance(bbox_xywh, torch.Tensor): + bbox_tlwh = bbox_xywh.clone() + bbox_tlwh[:, 0] = bbox_xywh[:, 0] - bbox_xywh[:, 2] / 2. + bbox_tlwh[:, 1] = bbox_xywh[:, 1] - bbox_xywh[:, 3] / 2. + return bbox_tlwh + + def _xywh_to_xyxy(self, bbox_xywh): + x, y, w, h = bbox_xywh + x1 = max(int(x - w / 2), 0) + x2 = min(int(x + w / 2), self.width - 1) + y1 = max(int(y - h / 2), 0) + y2 = min(int(y + h / 2), self.height - 1) + return x1, y1, x2, y2 + + def _tlwh_to_xyxy(self, bbox_tlwh): + """ + TODO: + Convert bbox from xtl_ytl_w_h to xc_yc_w_h + Thanks JieChen91@github.com for reporting this bug! + """ + x, y, w, h = bbox_tlwh + x1 = max(int(x), 0) + x2 = min(int(x+w), self.width - 1) + y1 = max(int(y), 0) + y2 = min(int(y+h), self.height - 1) + return x1, y1, x2, y2 + + def increment_ages(self): + self.tracker.increment_ages() + + def _xyxy_to_tlwh(self, bbox_xyxy): + x1, y1, x2, y2 = bbox_xyxy + + t = x1 + l = y1 + w = int(x2 - x1) + h = int(y2 - y1) + return t, l, w, h + + def _get_features(self, bbox_xywh, ori_img): + im_crops = [] + for box in bbox_xywh: + x1, y1, x2, y2 = self._xywh_to_xyxy(box) + im = ori_img[y1:y2, x1:x2] + im_crops.append(im) + if im_crops: + features = self.extractor(im_crops) + else: + features = np.array([]) + + return features diff --git a/strong_sort/utils/__init__.py b/strong_sort/utils/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/strong_sort/utils/asserts.py b/strong_sort/utils/asserts.py new file mode 100644 index 0000000..59a73cc --- /dev/null +++ b/strong_sort/utils/asserts.py @@ -0,0 +1,13 @@ +from os import environ + + +def assert_in(file, files_to_check): + if file not in files_to_check: + raise AssertionError("{} does not exist in the list".format(str(file))) + return True + + +def assert_in_env(check_list: list): + for item in check_list: + assert_in(item, environ.keys()) + return True diff --git a/strong_sort/utils/draw.py b/strong_sort/utils/draw.py new file mode 100644 index 0000000..bc7cb53 --- /dev/null +++ b/strong_sort/utils/draw.py @@ -0,0 +1,36 @@ +import numpy as np +import cv2 + +palette = (2 ** 11 - 1, 2 ** 15 - 1, 2 ** 20 - 1) + + +def compute_color_for_labels(label): + """ + Simple function that adds fixed color depending on the class + """ + color = [int((p * (label ** 2 - label + 1)) % 255) for p in palette] + return tuple(color) + + +def draw_boxes(img, bbox, identities=None, offset=(0,0)): + for i,box in enumerate(bbox): + x1,y1,x2,y2 = [int(i) for i in box] + x1 += offset[0] + x2 += offset[0] + y1 += offset[1] + y2 += offset[1] + # box text and bar + id = int(identities[i]) if identities is not None else 0 + color = compute_color_for_labels(id) + label = '{}{:d}'.format("", id) + t_size = cv2.getTextSize(label, cv2.FONT_HERSHEY_PLAIN, 2 , 2)[0] + cv2.rectangle(img,(x1, y1),(x2,y2),color,3) + cv2.rectangle(img,(x1, y1),(x1+t_size[0]+3,y1+t_size[1]+4), color,-1) + cv2.putText(img,label,(x1,y1+t_size[1]+4), cv2.FONT_HERSHEY_PLAIN, 2, [255,255,255], 2) + return img + + + +if __name__ == '__main__': + for i in range(82): + print(compute_color_for_labels(i)) diff --git a/strong_sort/utils/evaluation.py b/strong_sort/utils/evaluation.py new file mode 100644 index 0000000..1001794 --- /dev/null +++ b/strong_sort/utils/evaluation.py @@ -0,0 +1,103 @@ +import os +import numpy as np +import copy +import motmetrics as mm +mm.lap.default_solver = 'lap' +from utils.io import read_results, unzip_objs + + +class Evaluator(object): + + def __init__(self, data_root, seq_name, data_type): + self.data_root = data_root + self.seq_name = seq_name + self.data_type = data_type + + self.load_annotations() + self.reset_accumulator() + + def load_annotations(self): + assert self.data_type == 'mot' + + gt_filename = os.path.join(self.data_root, self.seq_name, 'gt', 'gt.txt') + self.gt_frame_dict = read_results(gt_filename, self.data_type, is_gt=True) + self.gt_ignore_frame_dict = read_results(gt_filename, self.data_type, is_ignore=True) + + def reset_accumulator(self): + self.acc = mm.MOTAccumulator(auto_id=True) + + def eval_frame(self, frame_id, trk_tlwhs, trk_ids, rtn_events=False): + # results + trk_tlwhs = np.copy(trk_tlwhs) + trk_ids = np.copy(trk_ids) + + # gts + gt_objs = self.gt_frame_dict.get(frame_id, []) + gt_tlwhs, gt_ids = unzip_objs(gt_objs)[:2] + + # ignore boxes + ignore_objs = self.gt_ignore_frame_dict.get(frame_id, []) + ignore_tlwhs = unzip_objs(ignore_objs)[0] + + + # remove ignored results + keep = np.ones(len(trk_tlwhs), dtype=bool) + iou_distance = mm.distances.iou_matrix(ignore_tlwhs, trk_tlwhs, max_iou=0.5) + if len(iou_distance) > 0: + match_is, match_js = mm.lap.linear_sum_assignment(iou_distance) + match_is, match_js = map(lambda a: np.asarray(a, dtype=int), [match_is, match_js]) + match_ious = iou_distance[match_is, match_js] + + match_js = np.asarray(match_js, dtype=int) + match_js = match_js[np.logical_not(np.isnan(match_ious))] + keep[match_js] = False + trk_tlwhs = trk_tlwhs[keep] + trk_ids = trk_ids[keep] + + # get distance matrix + iou_distance = mm.distances.iou_matrix(gt_tlwhs, trk_tlwhs, max_iou=0.5) + + # acc + self.acc.update(gt_ids, trk_ids, iou_distance) + + if rtn_events and iou_distance.size > 0 and hasattr(self.acc, 'last_mot_events'): + events = self.acc.last_mot_events # only supported by https://github.com/longcw/py-motmetrics + else: + events = None + return events + + def eval_file(self, filename): + self.reset_accumulator() + + result_frame_dict = read_results(filename, self.data_type, is_gt=False) + frames = sorted(list(set(self.gt_frame_dict.keys()) | set(result_frame_dict.keys()))) + for frame_id in frames: + trk_objs = result_frame_dict.get(frame_id, []) + trk_tlwhs, trk_ids = unzip_objs(trk_objs)[:2] + self.eval_frame(frame_id, trk_tlwhs, trk_ids, rtn_events=False) + + return self.acc + + @staticmethod + def get_summary(accs, names, metrics=('mota', 'num_switches', 'idp', 'idr', 'idf1', 'precision', 'recall')): + names = copy.deepcopy(names) + if metrics is None: + metrics = mm.metrics.motchallenge_metrics + metrics = copy.deepcopy(metrics) + + mh = mm.metrics.create() + summary = mh.compute_many( + accs, + metrics=metrics, + names=names, + generate_overall=True + ) + + return summary + + @staticmethod + def save_summary(summary, filename): + import pandas as pd + writer = pd.ExcelWriter(filename) + summary.to_excel(writer) + writer.save() diff --git a/strong_sort/utils/io.py b/strong_sort/utils/io.py new file mode 100644 index 0000000..2dc9afd --- /dev/null +++ b/strong_sort/utils/io.py @@ -0,0 +1,133 @@ +import os +from typing import Dict +import numpy as np + +# from utils.log import get_logger + + +def write_results(filename, results, data_type): + if data_type == 'mot': + save_format = '{frame},{id},{x1},{y1},{w},{h},-1,-1,-1,-1\n' + elif data_type == 'kitti': + save_format = '{frame} {id} pedestrian 0 0 -10 {x1} {y1} {x2} {y2} -10 -10 -10 -1000 -1000 -1000 -10\n' + else: + raise ValueError(data_type) + + with open(filename, 'w') as f: + for frame_id, tlwhs, track_ids in results: + if data_type == 'kitti': + frame_id -= 1 + for tlwh, track_id in zip(tlwhs, track_ids): + if track_id < 0: + continue + x1, y1, w, h = tlwh + x2, y2 = x1 + w, y1 + h + line = save_format.format(frame=frame_id, id=track_id, x1=x1, y1=y1, x2=x2, y2=y2, w=w, h=h) + f.write(line) + + +# def write_results(filename, results_dict: Dict, data_type: str): +# if not filename: +# return +# path = os.path.dirname(filename) +# if not os.path.exists(path): +# os.makedirs(path) + +# if data_type in ('mot', 'mcmot', 'lab'): +# save_format = '{frame},{id},{x1},{y1},{w},{h},1,-1,-1,-1\n' +# elif data_type == 'kitti': +# save_format = '{frame} {id} pedestrian -1 -1 -10 {x1} {y1} {x2} {y2} -1 -1 -1 -1000 -1000 -1000 -10 {score}\n' +# else: +# raise ValueError(data_type) + +# with open(filename, 'w') as f: +# for frame_id, frame_data in results_dict.items(): +# if data_type == 'kitti': +# frame_id -= 1 +# for tlwh, track_id in frame_data: +# if track_id < 0: +# continue +# x1, y1, w, h = tlwh +# x2, y2 = x1 + w, y1 + h +# line = save_format.format(frame=frame_id, id=track_id, x1=x1, y1=y1, x2=x2, y2=y2, w=w, h=h, score=1.0) +# f.write(line) +# logger.info('Save results to {}'.format(filename)) + + +def read_results(filename, data_type: str, is_gt=False, is_ignore=False): + if data_type in ('mot', 'lab'): + read_fun = read_mot_results + else: + raise ValueError('Unknown data type: {}'.format(data_type)) + + return read_fun(filename, is_gt, is_ignore) + + +""" +labels={'ped', ... % 1 +'person_on_vhcl', ... % 2 +'car', ... % 3 +'bicycle', ... % 4 +'mbike', ... % 5 +'non_mot_vhcl', ... % 6 +'static_person', ... % 7 +'distractor', ... % 8 +'occluder', ... % 9 +'occluder_on_grnd', ... %10 +'occluder_full', ... % 11 +'reflection', ... % 12 +'crowd' ... % 13 +}; +""" + + +def read_mot_results(filename, is_gt, is_ignore): + valid_labels = {1} + ignore_labels = {2, 7, 8, 12} + results_dict = dict() + if os.path.isfile(filename): + with open(filename, 'r') as f: + for line in f.readlines(): + linelist = line.split(',') + if len(linelist) < 7: + continue + fid = int(linelist[0]) + if fid < 1: + continue + results_dict.setdefault(fid, list()) + + if is_gt: + if 'MOT16-' in filename or 'MOT17-' in filename: + label = int(float(linelist[7])) + mark = int(float(linelist[6])) + if mark == 0 or label not in valid_labels: + continue + score = 1 + elif is_ignore: + if 'MOT16-' in filename or 'MOT17-' in filename: + label = int(float(linelist[7])) + vis_ratio = float(linelist[8]) + if label not in ignore_labels and vis_ratio >= 0: + continue + else: + continue + score = 1 + else: + score = float(linelist[6]) + + tlwh = tuple(map(float, linelist[2:6])) + target_id = int(linelist[1]) + + results_dict[fid].append((tlwh, target_id, score)) + + return results_dict + + +def unzip_objs(objs): + if len(objs) > 0: + tlwhs, ids, scores = zip(*objs) + else: + tlwhs, ids, scores = [], [], [] + tlwhs = np.asarray(tlwhs, dtype=float).reshape(-1, 4) + + return tlwhs, ids, scores \ No newline at end of file diff --git a/strong_sort/utils/json_logger.py b/strong_sort/utils/json_logger.py new file mode 100644 index 0000000..0afd0b4 --- /dev/null +++ b/strong_sort/utils/json_logger.py @@ -0,0 +1,383 @@ +""" +References: + https://medium.com/analytics-vidhya/creating-a-custom-logging-mechanism-for-real-time-object-detection-using-tdd-4ca2cfcd0a2f +""" +import json +from os import makedirs +from os.path import exists, join +from datetime import datetime + + +class JsonMeta(object): + HOURS = 3 + MINUTES = 59 + SECONDS = 59 + PATH_TO_SAVE = 'LOGS' + DEFAULT_FILE_NAME = 'remaining' + + +class BaseJsonLogger(object): + """ + This is the base class that returns __dict__ of its own + it also returns the dicts of objects in the attributes that are list instances + + """ + + def dic(self): + # returns dicts of objects + out = {} + for k, v in self.__dict__.items(): + if hasattr(v, 'dic'): + out[k] = v.dic() + elif isinstance(v, list): + out[k] = self.list(v) + else: + out[k] = v + return out + + @staticmethod + def list(values): + # applies the dic method on items in the list + return [v.dic() if hasattr(v, 'dic') else v for v in values] + + +class Label(BaseJsonLogger): + """ + For each bounding box there are various categories with confidences. Label class keeps track of that information. + """ + + def __init__(self, category: str, confidence: float): + self.category = category + self.confidence = confidence + + +class Bbox(BaseJsonLogger): + """ + This module stores the information for each frame and use them in JsonParser + Attributes: + labels (list): List of label module. + top (int): + left (int): + width (int): + height (int): + + Args: + bbox_id (float): + top (int): + left (int): + width (int): + height (int): + + References: + Check Label module for better understanding. + + + """ + + def __init__(self, bbox_id, top, left, width, height): + self.labels = [] + self.bbox_id = bbox_id + self.top = top + self.left = left + self.width = width + self.height = height + + def add_label(self, category, confidence): + # adds category and confidence only if top_k is not exceeded. + self.labels.append(Label(category, confidence)) + + def labels_full(self, value): + return len(self.labels) == value + + +class Frame(BaseJsonLogger): + """ + This module stores the information for each frame and use them in JsonParser + Attributes: + timestamp (float): The elapsed time of captured frame + frame_id (int): The frame number of the captured video + bboxes (list of Bbox objects): Stores the list of bbox objects. + + References: + Check Bbox class for better information + + Args: + timestamp (float): + frame_id (int): + + """ + + def __init__(self, frame_id: int, timestamp: float = None): + self.frame_id = frame_id + self.timestamp = timestamp + self.bboxes = [] + + def add_bbox(self, bbox_id: int, top: int, left: int, width: int, height: int): + bboxes_ids = [bbox.bbox_id for bbox in self.bboxes] + if bbox_id not in bboxes_ids: + self.bboxes.append(Bbox(bbox_id, top, left, width, height)) + else: + raise ValueError("Frame with id: {} already has a Bbox with id: {}".format(self.frame_id, bbox_id)) + + def add_label_to_bbox(self, bbox_id: int, category: str, confidence: float): + bboxes = {bbox.id: bbox for bbox in self.bboxes} + if bbox_id in bboxes.keys(): + res = bboxes.get(bbox_id) + res.add_label(category, confidence) + else: + raise ValueError('the bbox with id: {} does not exists!'.format(bbox_id)) + + +class BboxToJsonLogger(BaseJsonLogger): + """ + ُ This module is designed to automate the task of logging jsons. An example json is used + to show the contents of json file shortly + Example: + { + "video_details": { + "frame_width": 1920, + "frame_height": 1080, + "frame_rate": 20, + "video_name": "/home/gpu/codes/MSD/pedestrian_2/project/public/camera1.avi" + }, + "frames": [ + { + "frame_id": 329, + "timestamp": 3365.1254 + "bboxes": [ + { + "labels": [ + { + "category": "pedestrian", + "confidence": 0.9 + } + ], + "bbox_id": 0, + "top": 1257, + "left": 138, + "width": 68, + "height": 109 + } + ] + }], + + Attributes: + frames (dict): It's a dictionary that maps each frame_id to json attributes. + video_details (dict): information about video file. + top_k_labels (int): shows the allowed number of labels + start_time (datetime object): we use it to automate the json output by time. + + Args: + top_k_labels (int): shows the allowed number of labels + + """ + + def __init__(self, top_k_labels: int = 1): + self.frames = {} + self.video_details = self.video_details = dict(frame_width=None, frame_height=None, frame_rate=None, + video_name=None) + self.top_k_labels = top_k_labels + self.start_time = datetime.now() + + def set_top_k(self, value): + self.top_k_labels = value + + def frame_exists(self, frame_id: int) -> bool: + """ + Args: + frame_id (int): + + Returns: + bool: true if frame_id is recognized + """ + return frame_id in self.frames.keys() + + def add_frame(self, frame_id: int, timestamp: float = None) -> None: + """ + Args: + frame_id (int): + timestamp (float): opencv captured frame time property + + Raises: + ValueError: if frame_id would not exist in class frames attribute + + Returns: + None + + """ + if not self.frame_exists(frame_id): + self.frames[frame_id] = Frame(frame_id, timestamp) + else: + raise ValueError("Frame id: {} already exists".format(frame_id)) + + def bbox_exists(self, frame_id: int, bbox_id: int) -> bool: + """ + Args: + frame_id: + bbox_id: + + Returns: + bool: if bbox exists in frame bboxes list + """ + bboxes = [] + if self.frame_exists(frame_id=frame_id): + bboxes = [bbox.bbox_id for bbox in self.frames[frame_id].bboxes] + return bbox_id in bboxes + + def find_bbox(self, frame_id: int, bbox_id: int): + """ + + Args: + frame_id: + bbox_id: + + Returns: + bbox_id (int): + + Raises: + ValueError: if bbox_id does not exist in the bbox list of specific frame. + """ + if not self.bbox_exists(frame_id, bbox_id): + raise ValueError("frame with id: {} does not contain bbox with id: {}".format(frame_id, bbox_id)) + bboxes = {bbox.bbox_id: bbox for bbox in self.frames[frame_id].bboxes} + return bboxes.get(bbox_id) + + def add_bbox_to_frame(self, frame_id: int, bbox_id: int, top: int, left: int, width: int, height: int) -> None: + """ + + Args: + frame_id (int): + bbox_id (int): + top (int): + left (int): + width (int): + height (int): + + Returns: + None + + Raises: + ValueError: if bbox_id already exist in frame information with frame_id + ValueError: if frame_id does not exist in frames attribute + """ + if self.frame_exists(frame_id): + frame = self.frames[frame_id] + if not self.bbox_exists(frame_id, bbox_id): + frame.add_bbox(bbox_id, top, left, width, height) + else: + raise ValueError( + "frame with frame_id: {} already contains the bbox with id: {} ".format(frame_id, bbox_id)) + else: + raise ValueError("frame with frame_id: {} does not exist".format(frame_id)) + + def add_label_to_bbox(self, frame_id: int, bbox_id: int, category: str, confidence: float): + """ + Args: + frame_id: + bbox_id: + category: + confidence: the confidence value returned from yolo detection + + Returns: + None + + Raises: + ValueError: if labels quota (top_k_labels) exceeds. + """ + bbox = self.find_bbox(frame_id, bbox_id) + if not bbox.labels_full(self.top_k_labels): + bbox.add_label(category, confidence) + else: + raise ValueError("labels in frame_id: {}, bbox_id: {} is fulled".format(frame_id, bbox_id)) + + def add_video_details(self, frame_width: int = None, frame_height: int = None, frame_rate: int = None, + video_name: str = None): + self.video_details['frame_width'] = frame_width + self.video_details['frame_height'] = frame_height + self.video_details['frame_rate'] = frame_rate + self.video_details['video_name'] = video_name + + def output(self): + output = {'video_details': self.video_details} + result = list(self.frames.values()) + output['frames'] = [item.dic() for item in result] + return output + + def json_output(self, output_name): + """ + Args: + output_name: + + Returns: + None + + Notes: + It creates the json output with `output_name` name. + """ + if not output_name.endswith('.json'): + output_name += '.json' + with open(output_name, 'w') as file: + json.dump(self.output(), file) + file.close() + + def set_start(self): + self.start_time = datetime.now() + + def schedule_output_by_time(self, output_dir=JsonMeta.PATH_TO_SAVE, hours: int = 0, minutes: int = 0, + seconds: int = 60) -> None: + """ + Notes: + Creates folder and then periodically stores the jsons on that address. + + Args: + output_dir (str): the directory where output files will be stored + hours (int): + minutes (int): + seconds (int): + + Returns: + None + + """ + end = datetime.now() + interval = 0 + interval += abs(min([hours, JsonMeta.HOURS]) * 3600) + interval += abs(min([minutes, JsonMeta.MINUTES]) * 60) + interval += abs(min([seconds, JsonMeta.SECONDS])) + diff = (end - self.start_time).seconds + + if diff > interval: + output_name = self.start_time.strftime('%Y-%m-%d %H-%M-%S') + '.json' + if not exists(output_dir): + makedirs(output_dir) + output = join(output_dir, output_name) + self.json_output(output_name=output) + self.frames = {} + self.start_time = datetime.now() + + def schedule_output_by_frames(self, frames_quota, frame_counter, output_dir=JsonMeta.PATH_TO_SAVE): + """ + saves as the number of frames quota increases higher. + :param frames_quota: + :param frame_counter: + :param output_dir: + :return: + """ + pass + + def flush(self, output_dir): + """ + Notes: + We use this function to output jsons whenever possible. + like the time that we exit the while loop of opencv. + + Args: + output_dir: + + Returns: + None + + """ + filename = self.start_time.strftime('%Y-%m-%d %H-%M-%S') + '-remaining.json' + output = join(output_dir, filename) + self.json_output(output_name=output) diff --git a/strong_sort/utils/log.py b/strong_sort/utils/log.py new file mode 100644 index 0000000..0d48757 --- /dev/null +++ b/strong_sort/utils/log.py @@ -0,0 +1,17 @@ +import logging + + +def get_logger(name='root'): + formatter = logging.Formatter( + # fmt='%(asctime)s [%(levelname)s]: %(filename)s(%(funcName)s:%(lineno)s) >> %(message)s') + fmt='%(asctime)s [%(levelname)s]: %(message)s', datefmt='%Y-%m-%d %H:%M:%S') + + handler = logging.StreamHandler() + handler.setFormatter(formatter) + + logger = logging.getLogger(name) + logger.setLevel(logging.INFO) + logger.addHandler(handler) + return logger + + diff --git a/strong_sort/utils/parser.py b/strong_sort/utils/parser.py new file mode 100644 index 0000000..c29ed84 --- /dev/null +++ b/strong_sort/utils/parser.py @@ -0,0 +1,41 @@ +import os +import yaml +from easydict import EasyDict as edict + + +class YamlParser(edict): + """ + This is yaml parser based on EasyDict. + """ + + def __init__(self, cfg_dict=None, config_file=None): + if cfg_dict is None: + cfg_dict = {} + + if config_file is not None: + assert(os.path.isfile(config_file)) + with open(config_file, 'r') as fo: + yaml_ = yaml.load(fo.read(), Loader=yaml.FullLoader) + cfg_dict.update(yaml_) + + super(YamlParser, self).__init__(cfg_dict) + + def merge_from_file(self, config_file): + with open(config_file, 'r') as fo: + yaml_ = yaml.load(fo.read(), Loader=yaml.FullLoader) + self.update(yaml_) + + def merge_from_dict(self, config_dict): + self.update(config_dict) + + +def get_config(config_file=None): + return YamlParser(config_file=config_file) + + +if __name__ == "__main__": + cfg = YamlParser(config_file="../configs/yolov3.yaml") + cfg.merge_from_file("../configs/strong_sort.yaml") + + import ipdb + ipdb.set_trace() diff --git a/strong_sort/utils/tools.py b/strong_sort/utils/tools.py new file mode 100644 index 0000000..965fb69 --- /dev/null +++ b/strong_sort/utils/tools.py @@ -0,0 +1,39 @@ +from functools import wraps +from time import time + + +def is_video(ext: str): + """ + Returns true if ext exists in + allowed_exts for video files. + + Args: + ext: + + Returns: + + """ + + allowed_exts = ('.mp4', '.webm', '.ogg', '.avi', '.wmv', '.mkv', '.3gp') + return any((ext.endswith(x) for x in allowed_exts)) + + +def tik_tok(func): + """ + keep track of time for each process. + Args: + func: + + Returns: + + """ + @wraps(func) + def _time_it(*args, **kwargs): + start = time() + try: + return func(*args, **kwargs) + finally: + end_ = time() + print("time: {:.03f}s, fps: {:.03f}".format(end_ - start, 1 / (end_ - start))) + + return _time_it diff --git a/test.mp4 b/test.mp4 new file mode 100644 index 0000000..f5a2e6b Binary files /dev/null and b/test.mp4 differ diff --git a/test2.mp4 b/test2.mp4 new file mode 100644 index 0000000..b26429a Binary files /dev/null and b/test2.mp4 differ diff --git a/testir.mp4 b/testir.mp4 new file mode 100644 index 0000000..eb8e94b Binary files /dev/null and b/testir.mp4 differ diff --git a/track.py b/track.py new file mode 100644 index 0000000..ba76875 --- /dev/null +++ b/track.py @@ -0,0 +1,355 @@ +import argparse + +import os +# limit the number of cpus used by high performance libraries +os.environ["OMP_NUM_THREADS"] = "1" +os.environ["OPENBLAS_NUM_THREADS"] = "1" +os.environ["MKL_NUM_THREADS"] = "1" +os.environ["VECLIB_MAXIMUM_THREADS"] = "1" +os.environ["NUMEXPR_NUM_THREADS"] = "1" + +import sys +import numpy as np +from pathlib import Path +import torch +import torch.backends.cudnn as cudnn + +FILE = Path(__file__).resolve() +ROOT = FILE.parents[0] # yolov5 strongsort root directory +WEIGHTS = ROOT / 'weights' + +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) # add ROOT to PATH +if str(ROOT / 'yolov5') not in sys.path: + sys.path.append(str(ROOT / 'yolov5')) # add yolov5 ROOT to PATH +if str(ROOT / 'strong_sort') not in sys.path: + sys.path.append(str(ROOT / 'strong_sort')) # add strong_sort ROOT to PATH +ROOT = Path(os.path.relpath(ROOT, Path.cwd())) # relative + +import logging +from yolov5.models.common import DetectMultiBackend +from yolov5.utils.dataloaders import VID_FORMATS, LoadImages, LoadStreams +from yolov5.utils.general import (LOGGER, check_img_size, non_max_suppression, scale_coords, check_requirements, cv2, + check_imshow, xyxy2xywh, increment_path, strip_optimizer, colorstr, print_args, check_file) +from yolov5.utils.torch_utils import select_device, time_sync +from yolov5.utils.plots import Annotator, colors, save_one_box +from strong_sort.utils.parser import get_config +from strong_sort.strong_sort import StrongSORT + +# remove duplicated stream handler to avoid duplicated logging +logging.getLogger().removeHandler(logging.getLogger().handlers[0]) + +softmax = torch.nn.Softmax(dim=1) + +def tensor_max(tensor): + + idx = torch.argmax(tensor, dim=1, keepdim=True) + y = torch.zeros(tensor.size(), device='cuda').scatter_(1, idx, 1.) + return y + +def tensor_thresh(tensor, thr=0.5): + out = (tensor>thr).float() + return out + +@torch.no_grad() +def run( + source='0', + yolo_weights=WEIGHTS / 'yolov5m.pt', # model.pt path(s), + strong_sort_weights=WEIGHTS / 'osnet_x0_25_msmt17.pt', # model.pt path, + config_strongsort=ROOT / 'strong_sort/configs/strong_sort.yaml', + imgsz=(640, 640), # inference size (height, width) + conf_thres=0.25, # confidence threshold + iou_thres=0.45, # NMS IOU threshold + max_det=1000, # maximum detections per image + device='', # cuda device, i.e. 0 or 0,1,2,3 or cpu + show_vid=False, # show results + save_txt=False, # save results to *.txt + save_conf=False, # save confidences in --save-txt labels + save_crop=False, # save cropped prediction boxes + save_vid=False, # save confidences in --save-txt labels + nosave=False, # do not save images/videos + classes=None, # filter by class: --class 0, or --class 0 2 3 + agnostic_nms=False, # class-agnostic NMS + augment=False, # augmented inference + visualize=False, # visualize features + update=False, # update all models + project=ROOT / 'runs/track', # save results to project/name + name='exp', # save results to project/name + exist_ok=False, # existing project/name ok, do not increment + line_thickness=3, # bounding box thickness (pixels) + hide_labels=False, # hide labels + hide_conf=False, # hide confidences + hide_class=False, # hide IDs + half=False, # use FP16 half-precision inference + dnn=False, # use OpenCV DNN for ONNX inference +): + + source = str(source) + save_img = not nosave and not source.endswith('.txt') # save inference images + is_file = Path(source).suffix[1:] in (VID_FORMATS) + is_url = source.lower().startswith(('rtsp://', 'rtmp://', 'http://', 'https://')) + webcam = source.isnumeric() or source.endswith('.txt') or (is_url and not is_file) + if is_url and is_file: + source = check_file(source) # download + + # Directories + if not isinstance(yolo_weights, list): # single yolo model + exp_name = yolo_weights.stem + elif type(yolo_weights) is list and len(yolo_weights) == 1: # single models after --yolo_weights + exp_name = Path(yolo_weights[0]).stem + else: # multiple models after --yolo_weights + exp_name = 'ensemble' + exp_name = name if name else exp_name + "_" + strong_sort_weights.stem + save_dir = increment_path(Path(project) / exp_name, exist_ok=exist_ok) # increment run + (save_dir / 'tracks' if save_txt else save_dir).mkdir(parents=True, exist_ok=True) # make dir + + # Load model + device = select_device(device) + model = DetectMultiBackend(yolo_weights, device=device, dnn=dnn, data=None, fp16=half) + stride, names, pt = model.stride, model.names, model.pt + imgsz = check_img_size(imgsz, s=stride) # check image size + + # Dataloader + if webcam: + show_vid = check_imshow() + cudnn.benchmark = True # set True to speed up constant image size inference + dataset = LoadStreams(source, img_size=imgsz, stride=stride, auto=pt) + nr_sources = len(dataset) + else: + dataset = LoadImages(source, img_size=imgsz, stride=stride, auto=pt) + nr_sources = 1 + vid_path, vid_writer, txt_path = [None] * nr_sources, [None] * nr_sources, [None] * nr_sources + + # initialize StrongSORT + cfg = get_config() + cfg.merge_from_file(opt.config_strongsort) + dict_results = dict() + # Create as many strong sort instances as there are video sources + strongsort_list = [] + for i in range(nr_sources): + strongsort_list.append( + StrongSORT( + strong_sort_weights, + device, + max_dist=cfg.STRONGSORT.MAX_DIST, + max_iou_distance=cfg.STRONGSORT.MAX_IOU_DISTANCE, + max_age=cfg.STRONGSORT.MAX_AGE, + n_init=cfg.STRONGSORT.N_INIT, + nn_budget=cfg.STRONGSORT.NN_BUDGET, + mc_lambda=cfg.STRONGSORT.MC_LAMBDA, + ema_alpha=cfg.STRONGSORT.EMA_ALPHA, + + ) + ) + outputs = [None] * nr_sources + + # Run tracking + model.warmup(imgsz=(1 if pt else nr_sources, 3, *imgsz)) # warmup + dt, seen = [0.0, 0.0, 0.0, 0.0], 0 + curr_frames, prev_frames = [None] * nr_sources, [None] * nr_sources + for frame_idx, (path, im, im0s, vid_cap, s) in enumerate(dataset): + t1 = time_sync() + im = torch.from_numpy(im).to(device) + im = im.half() if half else im.float() # uint8 to fp16/32 + im /= 255.0 # 0 - 255 to 0.0 - 1.0 + if len(im.shape) == 3: + im = im[None] # expand for batch dim + t2 = time_sync() + dt[0] += t2 - t1 + + # Inference + visualize = increment_path(save_dir / Path(path[0]).stem, mkdir=True) if visualize else False + pred = model(im, augment=augment, visualize=visualize) + t3 = time_sync() + dt[1] += t3 - t2 + + # Apply NMS + pred = non_max_suppression(pred, conf_thres, iou_thres, classes, agnostic_nms, max_det=max_det) + dt[2] += time_sync() - t3 + + # Process detections + for i, det in enumerate(pred): # detections per image + seen += 1 + if webcam: # nr_sources >= 1 + p, im0, _ = path[i], im0s[i].copy(), dataset.count + p = Path(p) # to Path + s += f'{i}: ' + txt_file_name = p.name + save_path = str(save_dir / p.name) # im.jpg, vid.mp4, .. + else: + p, im0, _ = path, im0s.copy(), getattr(dataset, 'frame', 0) + p = Path(p) # to Path + # video file + if source.endswith(VID_FORMATS): + txt_file_name = p.stem + save_path = str(save_dir / p.name) # im.jpg, vid.mp4, ... + # folder with imgs + else: + txt_file_name = p.parent.name # get folder name containing current img + save_path = str(save_dir / p.parent.name) # im.jpg, vid.mp4, ... + curr_frames[i] = im0 + + txt_path = str(save_dir / 'tracks' / txt_file_name) # im.txt + s += '%gx%g ' % im.shape[2:] # print string + imc = im0.copy() if save_crop else im0 # for save_crop + + annotator = Annotator(im0, line_width=2, pil=not ascii) + if cfg.STRONGSORT.ECC: # camera motion compensation + strongsort_list[i].tracker.camera_update(prev_frames[i], curr_frames[i]) + + if det is not None and len(det): + # Rescale boxes from img_size to im0 size + det[:, :4] = scale_coords(im.shape[2:], det[:, :4], im0.shape).round() + + # Print results + for c in det[:, -1].unique(): + n = (det[:, -1] == c).sum() # detections per class + s += f"{n} {names[int(c)]}{'s' * (n > 1)}, " # add to string + + + xywhs = xyxy2xywh(det[:, 0:4]) + confs = det[:, 4] + clss = det[:, 5] + + # pass detections to strongsort + t4 = time_sync() + outputs[i], attrs = strongsort_list[i].update(xywhs.cpu(), confs.cpu(), clss.cpu(), im0) + attrs = softmax(attrs) + prob, preds = torch.max(attrs, 1) + names = ['206','207i','405','Arisun','Dena','HcCross','JackS5','KaraMazdaPickup','L90','MVM315H', + 'MVMX22','NeissanVanet','Pars','PeykanSavari','PeykanVanet','Pride131nasimsaba', + 'Pride132and111','Pride141','PrideVanet151','Quik','RenaultPK','RioSD','Runna','Saina', + 'Samand','SamandSoren','Shahin','Tiba','Xantia'] + + t5 = time_sync() + dt[3] += t5 - t4 + + # draw boxes for visualization + if len(outputs[i]) > 0: + for j, (output, conf) in enumerate(zip(outputs[i], confs)): + + bboxes = output[0:4] + id = output[4] + if id not in dict_results.keys(): + dict_results[int(id)] = [] + if len(dict_results[int(id)]) > 50: + dict_results[int(id)].pop(0) + + if prob[j].item() > 0.9: + dict_results[int(id)].append(preds[j].item()) + + cls = output[5] + + if save_txt: + # to MOT format + bbox_left = output[0] + bbox_top = output[1] + bbox_w = output[2] - output[0] + bbox_h = output[3] - output[1] + # Write MOT compliant results to file + with open(txt_path + '.txt', 'a') as f: + f.write(('%g ' * 10 + '\n') % (frame_idx + 1, id, bbox_left, # MOT format + bbox_top, bbox_w, bbox_h, -1, -1, -1, i)) + + if save_vid or save_crop or show_vid: # Add bbox to image + c = int(cls) # integer class + id = int(id) # integer id + if len(dict_results[int(id)]) > 3: + clsss = max(set(dict_results[int(id)]), key=dict_results[int(id)].count) + + label = (str(id)+'-'+f'{names[clsss]}') + + else: + label = str(id)+'-'+'N' + + annotator.box_label(bboxes, label , color=colors(c, True)) + if save_crop: + txt_file_name = txt_file_name if (isinstance(path, list) and len(path) > 1) else '' + save_one_box(bboxes, imc, file=save_dir / 'crops' / txt_file_name / names[c] / f'{id}' / f'{p.stem}.jpg', BGR=True) + + LOGGER.info(f'{s}Done. YOLO:({t3 - t2:.3f}s), StrongSORT:({t5 - t4:.3f}s)') + + else: + strongsort_list[i].increment_ages() + LOGGER.info('No detections') + + # Stream results + im0 = annotator.result() + if show_vid: + cv2.imshow(str(p), im0) + cv2.waitKey(1) # 1 millisecond + + # Save results (image with detections) + if save_vid: + if vid_path[i] != save_path: # new video + vid_path[i] = save_path + if isinstance(vid_writer[i], cv2.VideoWriter): + vid_writer[i].release() # release previous video writer + if vid_cap: # video + fps = vid_cap.get(cv2.CAP_PROP_FPS) + w = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH)) + h = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT)) + else: # stream + fps, w, h = 30, im0.shape[1], im0.shape[0] + save_path = str(Path(save_path).with_suffix('.mp4')) # force *.mp4 suffix on results videos + vid_writer[i] = cv2.VideoWriter(save_path, cv2.VideoWriter_fourcc(*'mp4v'), fps, (w, h)) + vid_writer[i].write(im0) + + prev_frames[i] = curr_frames[i] + + # Print results + t = tuple(x / seen * 1E3 for x in dt) # speeds per image + LOGGER.info(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS, %.1fms strong sort update per image at shape {(1, 3, *imgsz)}' % t) + if save_txt or save_vid: + s = f"\n{len(list(save_dir.glob('tracks/*.txt')))} tracks saved to {save_dir / 'tracks'}" if save_txt else '' + LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}{s}") + if update: + strip_optimizer(yolo_weights) # update model (to fix SourceChangeWarning) + + +def parse_opt(): + parser = argparse.ArgumentParser() + parser.add_argument('--yolo-weights', nargs='+', type=str, default=WEIGHTS / './yolov5/yolov5s.pt', help='model.pt path(s)') + parser.add_argument('--strong-sort-weights', type=str, default='osnet_x1_0_msmt17.pt') + parser.add_argument('--config-strongsort', type=str, default='strong_sort/configs/strong_sort.yaml') + parser.add_argument('--source', type=str, default="eee.mp4", help='file/dir/URL/glob, 0 for webcam') + parser.add_argument('--imgsz', '--img', '--img-size', nargs='+', type=int, default=[640], help='inference size h,w') + parser.add_argument('--conf-thres', type=float, default=0.5, help='confidence threshold') + parser.add_argument('--iou-thres', type=float, default=0.5, help='NMS IoU threshold') + parser.add_argument('--max-det', type=int, default=1000, help='maximum detections per image') + parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') + parser.add_argument('--show-vid', default=True, action='store_true', help='display tracking video results') + parser.add_argument('--save-txt', action='store_true', help='save results to *.txt') + parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels') + parser.add_argument('--save-crop', action='store_true', help='save cropped prediction boxes') + parser.add_argument('--save-vid', action='store_true', help='save video tracking results') + parser.add_argument('--nosave', action='store_true', help='do not save images/videos') + # class 0 is person, 1 is bycicle, 2 is car... 79 is oven + parser.add_argument('--classes', nargs='+', type=int, default=2, help='filter by class: --classes 0, or --classes 0 2 3') + parser.add_argument('--agnostic-nms', action='store_true', help='class-agnostic NMS') + parser.add_argument('--augment', action='store_true', help='augmented inference') + parser.add_argument('--visualize', action='store_true', help='visualize features') + parser.add_argument('--update', action='store_true', help='update all models') + parser.add_argument('--project', default=ROOT / 'runs/track', help='save results to project/name') + parser.add_argument('--name', default='exp', help='save results to project/name') + parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment') + parser.add_argument('--line-thickness', default=3, type=int, help='bounding box thickness (pixels)') + parser.add_argument('--hide-labels', default=False, action='store_true', help='hide labels') + parser.add_argument('--hide-conf', default=False, action='store_true', help='hide confidences') + parser.add_argument('--hide-class', default=False, action='store_true', help='hide IDs') + parser.add_argument('--half', action='store_true', help='use FP16 half-precision inference') + parser.add_argument('--dnn', action='store_true', help='use OpenCV DNN for ONNX inference') + opt = parser.parse_args() + opt.imgsz *= 2 if len(opt.imgsz) == 1 else 1 # expand + print_args(vars(opt)) + return opt + + +def main(opt): + check_requirements(requirements=ROOT / 'requirements.txt', exclude=('tensorboard', 'thop')) + run(**vars(opt)) + + +if __name__ == "__main__": + opt = parse_opt() + main(opt) \ No newline at end of file diff --git a/track_V1.py b/track_V1.py new file mode 100644 index 0000000..fddbcce --- /dev/null +++ b/track_V1.py @@ -0,0 +1,349 @@ +import argparse + +import os +# limit the number of cpus used by high performance libraries +os.environ["OMP_NUM_THREADS"] = "1" +os.environ["OPENBLAS_NUM_THREADS"] = "1" +os.environ["MKL_NUM_THREADS"] = "1" +os.environ["VECLIB_MAXIMUM_THREADS"] = "1" +os.environ["NUMEXPR_NUM_THREADS"] = "1" + +import sys +import numpy as np +from pathlib import Path +import torch +import torch.backends.cudnn as cudnn + +FILE = Path(__file__).resolve() +ROOT = FILE.parents[0] # yolov5 strongsort root directory +WEIGHTS = ROOT / 'weights' + +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) # add ROOT to PATH +if str(ROOT / 'yolov5') not in sys.path: + sys.path.append(str(ROOT / 'yolov5')) # add yolov5 ROOT to PATH +if str(ROOT / 'strong_sort') not in sys.path: + sys.path.append(str(ROOT / 'strong_sort')) # add strong_sort ROOT to PATH +ROOT = Path(os.path.relpath(ROOT, Path.cwd())) # relative + +import logging +from yolov5.models.common import DetectMultiBackend +from yolov5.utils.dataloaders import VID_FORMATS, LoadImages, LoadStreams +from yolov5.utils.general import (LOGGER, check_img_size, non_max_suppression, scale_coords, check_requirements, cv2, + check_imshow, xyxy2xywh, increment_path, strip_optimizer, colorstr, print_args, check_file) +from yolov5.utils.torch_utils import select_device, time_sync +from yolov5.utils.plots import Annotator, colors, save_one_box +from strong_sort.utils.parser import get_config +from strong_sort.strong_sort import StrongSORT + +# remove duplicated stream handler to avoid duplicated logging +logging.getLogger().removeHandler(logging.getLogger().handlers[0]) + +softmax = torch.nn.Softmax(dim=1) + +def tensor_max(tensor): + + idx = torch.argmax(tensor, dim=1, keepdim=True) + y = torch.zeros(tensor.size(), device='cuda').scatter_(1, idx, 1.) + return y + +def tensor_thresh(tensor, thr=0.5): + out = (tensor>thr).float() + return out + +@torch.no_grad() +def run( + source='0', + yolo_weights=WEIGHTS / 'yolov5m.pt', # model.pt path(s), + strong_sort_weights=WEIGHTS / 'osnet_x0_25_msmt17.pt', # model.pt path, + config_strongsort=ROOT / 'strong_sort/configs/strong_sort.yaml', + imgsz=(640, 640), # inference size (height, width) + conf_thres=0.25, # confidence threshold + iou_thres=0.45, # NMS IOU threshold + max_det=1000, # maximum detections per image + device='', # cuda device, i.e. 0 or 0,1,2,3 or cpu + show_vid=False, # show results + save_txt=False, # save results to *.txt + save_conf=False, # save confidences in --save-txt labels + save_crop=False, # save cropped prediction boxes + save_vid=False, # save confidences in --save-txt labels + nosave=False, # do not save images/videos + classes=None, # filter by class: --class 0, or --class 0 2 3 + agnostic_nms=False, # class-agnostic NMS + augment=False, # augmented inference + visualize=False, # visualize features + update=False, # update all models + project=ROOT / 'runs/track', # save results to project/name + name='exp', # save results to project/name + exist_ok=False, # existing project/name ok, do not increment + line_thickness=3, # bounding box thickness (pixels) + hide_labels=False, # hide labels + hide_conf=False, # hide confidences + hide_class=False, # hide IDs + half=False, # use FP16 half-precision inference + dnn=False, # use OpenCV DNN for ONNX inference +): + + source = str(source) + save_img = not nosave and not source.endswith('.txt') # save inference images + is_file = Path(source).suffix[1:] in (VID_FORMATS) + is_url = source.lower().startswith(('rtsp://', 'rtmp://', 'http://', 'https://')) + webcam = source.isnumeric() or source.endswith('.txt') or (is_url and not is_file) + if is_url and is_file: + source = check_file(source) # download + + # Directories + if not isinstance(yolo_weights, list): # single yolo model + exp_name = yolo_weights.stem + elif type(yolo_weights) is list and len(yolo_weights) == 1: # single models after --yolo_weights + exp_name = Path(yolo_weights[0]).stem + else: # multiple models after --yolo_weights + exp_name = 'ensemble' + exp_name = name if name else exp_name + "_" + strong_sort_weights.stem + save_dir = increment_path(Path(project) / exp_name, exist_ok=exist_ok) # increment run + (save_dir / 'tracks' if save_txt else save_dir).mkdir(parents=True, exist_ok=True) # make dir + + # Load model + device = select_device(device) + model = DetectMultiBackend(yolo_weights, device=device, dnn=dnn, data=None, fp16=half) + stride, names, pt = model.stride, model.names, model.pt + imgsz = check_img_size(imgsz, s=stride) # check image size + + # Dataloader + if webcam: + show_vid = check_imshow() + cudnn.benchmark = True # set True to speed up constant image size inference + dataset = LoadStreams(source, img_size=imgsz, stride=stride, auto=pt) + nr_sources = len(dataset) + else: + dataset = LoadImages(source, img_size=imgsz, stride=stride, auto=pt) + nr_sources = 1 + vid_path, vid_writer, txt_path = [None] * nr_sources, [None] * nr_sources, [None] * nr_sources + + # initialize StrongSORT + cfg = get_config() + cfg.merge_from_file(opt.config_strongsort) + + # Create as many strong sort instances as there are video sources + strongsort_list = [] + for i in range(nr_sources): + strongsort_list.append( + StrongSORT( + strong_sort_weights, + device, + max_dist=cfg.STRONGSORT.MAX_DIST, + max_iou_distance=cfg.STRONGSORT.MAX_IOU_DISTANCE, + max_age=cfg.STRONGSORT.MAX_AGE, + n_init=cfg.STRONGSORT.N_INIT, + nn_budget=cfg.STRONGSORT.NN_BUDGET, + mc_lambda=cfg.STRONGSORT.MC_LAMBDA, + ema_alpha=cfg.STRONGSORT.EMA_ALPHA, + + ) + ) + outputs = [None] * nr_sources + + # Run tracking + model.warmup(imgsz=(1 if pt else nr_sources, 3, *imgsz)) # warmup + dt, seen = [0.0, 0.0, 0.0, 0.0], 0 + curr_frames, prev_frames = [None] * nr_sources, [None] * nr_sources + for frame_idx, (path, im, im0s, vid_cap, s) in enumerate(dataset): + t1 = time_sync() + im = torch.from_numpy(im).to(device) + im = im.half() if half else im.float() # uint8 to fp16/32 + im /= 255.0 # 0 - 255 to 0.0 - 1.0 + if len(im.shape) == 3: + im = im[None] # expand for batch dim + t2 = time_sync() + dt[0] += t2 - t1 + + # Inference + visualize = increment_path(save_dir / Path(path[0]).stem, mkdir=True) if visualize else False + pred = model(im, augment=augment, visualize=visualize) + t3 = time_sync() + dt[1] += t3 - t2 + + # Apply NMS + pred = non_max_suppression(pred, conf_thres, iou_thres, classes, agnostic_nms, max_det=max_det) + dt[2] += time_sync() - t3 + + # Process detections + for i, det in enumerate(pred): # detections per image + seen += 1 + if webcam: # nr_sources >= 1 + p, im0, _ = path[i], im0s[i].copy(), dataset.count + p = Path(p) # to Path + s += f'{i}: ' + txt_file_name = p.name + save_path = str(save_dir / p.name) # im.jpg, vid.mp4, .. + else: + p, im0, _ = path, im0s.copy(), getattr(dataset, 'frame', 0) + p = Path(p) # to Path + # video file + if source.endswith(VID_FORMATS): + txt_file_name = p.stem + save_path = str(save_dir / p.name) # im.jpg, vid.mp4, ... + # folder with imgs + else: + txt_file_name = p.parent.name # get folder name containing current img + save_path = str(save_dir / p.parent.name) # im.jpg, vid.mp4, ... + curr_frames[i] = im0 + + txt_path = str(save_dir / 'tracks' / txt_file_name) # im.txt + s += '%gx%g ' % im.shape[2:] # print string + imc = im0.copy() if save_crop else im0 # for save_crop + + annotator = Annotator(im0, line_width=2, pil=not ascii) + if cfg.STRONGSORT.ECC: # camera motion compensation + strongsort_list[i].tracker.camera_update(prev_frames[i], curr_frames[i]) + + if det is not None and len(det): + # Rescale boxes from img_size to im0 size + det[:, :4] = scale_coords(im.shape[2:], det[:, :4], im0.shape).round() + + # Print results + for c in det[:, -1].unique(): + n = (det[:, -1] == c).sum() # detections per class + s += f"{n} {names[int(c)]}{'s' * (n > 1)}, " # add to string + + + xywhs = xyxy2xywh(det[:, 0:4]) + confs = det[:, 4] + clss = det[:, 5] + + # pass detections to strongsort + t4 = time_sync() + outputs[i], attrs = strongsort_list[i].update(xywhs.cpu(), confs.cpu(), clss.cpu(), im0) + gender = tensor_thresh(torch.sigmoid(attrs['gender']), 0.5) + attrs['gender'] = torch.sigmoid(attrs['gender']) + #body_colors = ["b_white","b_blue","b_green","b_red","b_brown","b_yellow","b_gray","b_black"] + #attrs['body_color'] = [torch.argmax(atts) for atts in tensor_max(softmax(attrs['body_color']))] + + t5 = time_sync() + dt[3] += t5 - t4 + + # draw boxes for visualization + if len(outputs[i]) > 0: + for j, (output, conf) in enumerate(zip(outputs[i], confs)): + + bboxes = output[0:4] + id = output[4] + cls = output[5] + + if save_txt: + # to MOT format + bbox_left = output[0] + bbox_top = output[1] + bbox_w = output[2] - output[0] + bbox_h = output[3] - output[1] + # Write MOT compliant results to file + with open(txt_path + '.txt', 'a') as f: + f.write(('%g ' * 10 + '\n') % (frame_idx + 1, id, bbox_left, # MOT format + bbox_top, bbox_w, bbox_h, -1, -1, -1, i)) + + if save_vid or save_crop or show_vid: # Add bbox to image + c = int(cls) # integer class + id = int(id) # integer id + + if gender[j] == 0: + label = None if hide_labels else (f'{id} M' if hide_conf else \ + (f'{id} M {attrs["gender"][j].item():.2f}' if hide_class else f'{id} M {attrs["gender"][j].item():.2f}')) + + elif gender[j] == 1: + label = None if hide_labels else (f'{id} F' if hide_conf else \ + (f'{id} F {attrs["gender"][j].item():.2f}' if hide_class else f'{id} F {attrs["gender"][j].item():.2f}')) + + #label = None if hide_labels else (f'{id} {attrs["body_color"][j]}' if hide_conf else \ + # (f'{id} {conf:.2f}' if hide_class else f'{id} {attrs["body_color"][j]} {conf:.2f}')) + + annotator.box_label(bboxes, label , color=colors(c, True)) + if save_crop: + txt_file_name = txt_file_name if (isinstance(path, list) and len(path) > 1) else '' + save_one_box(bboxes, imc, file=save_dir / 'crops' / txt_file_name / names[c] / f'{id}' / f'{p.stem}.jpg', BGR=True) + + LOGGER.info(f'{s}Done. YOLO:({t3 - t2:.3f}s), StrongSORT:({t5 - t4:.3f}s)') + + else: + strongsort_list[i].increment_ages() + LOGGER.info('No detections') + + # Stream results + im0 = annotator.result() + if show_vid: + cv2.imshow(str(p), im0) + cv2.waitKey(1) # 1 millisecond + + # Save results (image with detections) + if save_vid: + if vid_path[i] != save_path: # new video + vid_path[i] = save_path + if isinstance(vid_writer[i], cv2.VideoWriter): + vid_writer[i].release() # release previous video writer + if vid_cap: # video + fps = vid_cap.get(cv2.CAP_PROP_FPS) + w = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH)) + h = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT)) + else: # stream + fps, w, h = 30, im0.shape[1], im0.shape[0] + save_path = str(Path(save_path).with_suffix('.mp4')) # force *.mp4 suffix on results videos + vid_writer[i] = cv2.VideoWriter(save_path, cv2.VideoWriter_fourcc(*'mp4v'), fps, (w, h)) + vid_writer[i].write(im0) + + prev_frames[i] = curr_frames[i] + + # Print results + t = tuple(x / seen * 1E3 for x in dt) # speeds per image + LOGGER.info(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS, %.1fms strong sort update per image at shape {(1, 3, *imgsz)}' % t) + if save_txt or save_vid: + s = f"\n{len(list(save_dir.glob('tracks/*.txt')))} tracks saved to {save_dir / 'tracks'}" if save_txt else '' + LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}{s}") + if update: + strip_optimizer(yolo_weights) # update model (to fix SourceChangeWarning) + + +def parse_opt(): + parser = argparse.ArgumentParser() + parser.add_argument('--yolo-weights', nargs='+', type=str, default=WEIGHTS / '../yolov5/crowdhuman_yolov5m.pt', help='model.pt path(s)') + parser.add_argument('--strong-sort-weights', type=str, default='osnet_x1_0_msmt17.pt') + parser.add_argument('--config-strongsort', type=str, default='strong_sort/configs/strong_sort.yaml') + parser.add_argument('--source', type=str, default='test2.mp4', help='file/dir/URL/glob, 0 for webcam') + parser.add_argument('--imgsz', '--img', '--img-size', nargs='+', type=int, default=[640], help='inference size h,w') + parser.add_argument('--conf-thres', type=float, default=0.5, help='confidence threshold') + parser.add_argument('--iou-thres', type=float, default=0.5, help='NMS IoU threshold') + parser.add_argument('--max-det', type=int, default=1000, help='maximum detections per image') + parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') + parser.add_argument('--show-vid', default=True, action='store_true', help='display tracking video results') + parser.add_argument('--save-txt', action='store_true', help='save results to *.txt') + parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels') + parser.add_argument('--save-crop', action='store_true', help='save cropped prediction boxes') + parser.add_argument('--save-vid', action='store_true', help='save video tracking results') + parser.add_argument('--nosave', action='store_true', help='do not save images/videos') + # class 0 is person, 1 is bycicle, 2 is car... 79 is oven + parser.add_argument('--classes', nargs='+', type=int, default=0, help='filter by class: --classes 0, or --classes 0 2 3') + parser.add_argument('--agnostic-nms', action='store_true', help='class-agnostic NMS') + parser.add_argument('--augment', action='store_true', help='augmented inference') + parser.add_argument('--visualize', action='store_true', help='visualize features') + parser.add_argument('--update', action='store_true', help='update all models') + parser.add_argument('--project', default=ROOT / 'runs/track', help='save results to project/name') + parser.add_argument('--name', default='exp', help='save results to project/name') + parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment') + parser.add_argument('--line-thickness', default=3, type=int, help='bounding box thickness (pixels)') + parser.add_argument('--hide-labels', default=False, action='store_true', help='hide labels') + parser.add_argument('--hide-conf', default=False, action='store_true', help='hide confidences') + parser.add_argument('--hide-class', default=False, action='store_true', help='hide IDs') + parser.add_argument('--half', action='store_true', help='use FP16 half-precision inference') + parser.add_argument('--dnn', action='store_true', help='use OpenCV DNN for ONNX inference') + opt = parser.parse_args() + opt.imgsz *= 2 if len(opt.imgsz) == 1 else 1 # expand + print_args(vars(opt)) + return opt + + +def main(opt): + check_requirements(requirements=ROOT / 'requirements.txt', exclude=('tensorboard', 'thop')) + run(**vars(opt)) + + +if __name__ == "__main__": + opt = parse_opt() + main(opt) \ No newline at end of file diff --git a/www.mp4 b/www.mp4 new file mode 100644 index 0000000..ab86fc9 Binary files /dev/null and b/www.mp4 differ diff --git a/yolov5/.gitattributes b/yolov5/.gitattributes new file mode 100644 index 0000000..dad4239 --- /dev/null +++ b/yolov5/.gitattributes @@ -0,0 +1,2 @@ +# this drop notebooks from GitHub language stats +*.ipynb linguist-vendored diff --git a/yolov5/.github/CODE_OF_CONDUCT.md b/yolov5/.github/CODE_OF_CONDUCT.md new file mode 100644 index 0000000..27e59e9 --- /dev/null +++ b/yolov5/.github/CODE_OF_CONDUCT.md @@ -0,0 +1,128 @@ +# YOLOv5 🚀 Contributor Covenant Code of Conduct + +## Our Pledge + +We as members, contributors, and leaders pledge to make participation in our +community a harassment-free experience for everyone, regardless of age, body +size, visible or invisible disability, ethnicity, sex characteristics, gender +identity and expression, level of experience, education, socio-economic status, +nationality, personal appearance, race, religion, or sexual identity +and orientation. + +We pledge to act and interact in ways that contribute to an open, welcoming, +diverse, inclusive, and healthy community. + +## Our Standards + +Examples of behavior that contributes to a positive environment for our +community include: + +- Demonstrating empathy and kindness toward other people +- Being respectful of differing opinions, viewpoints, and experiences +- Giving and gracefully accepting constructive feedback +- Accepting responsibility and apologizing to those affected by our mistakes, + and learning from the experience +- Focusing on what is best not just for us as individuals, but for the + overall community + +Examples of unacceptable behavior include: + +- The use of sexualized language or imagery, and sexual attention or + advances of any kind +- Trolling, insulting or derogatory comments, and personal or political attacks +- Public or private harassment +- Publishing others' private information, such as a physical or email + address, without their explicit permission +- Other conduct which could reasonably be considered inappropriate in a + professional setting + +## Enforcement Responsibilities + +Community leaders are responsible for clarifying and enforcing our standards of +acceptable behavior and will take appropriate and fair corrective action in +response to any behavior that they deem inappropriate, threatening, offensive, +or harmful. + +Community leaders have the right and responsibility to remove, edit, or reject +comments, commits, code, wiki edits, issues, and other contributions that are +not aligned to this Code of Conduct, and will communicate reasons for moderation +decisions when appropriate. + +## Scope + +This Code of Conduct applies within all community spaces, and also applies when +an individual is officially representing the community in public spaces. +Examples of representing our community include using an official e-mail address, +posting via an official social media account, or acting as an appointed +representative at an online or offline event. + +## Enforcement + +Instances of abusive, harassing, or otherwise unacceptable behavior may be +reported to the community leaders responsible for enforcement at +hello@ultralytics.com. +All complaints will be reviewed and investigated promptly and fairly. + +All community leaders are obligated to respect the privacy and security of the +reporter of any incident. + +## Enforcement Guidelines + +Community leaders will follow these Community Impact Guidelines in determining +the consequences for any action they deem in violation of this Code of Conduct: + +### 1. Correction + +**Community Impact**: Use of inappropriate language or other behavior deemed +unprofessional or unwelcome in the community. + +**Consequence**: A private, written warning from community leaders, providing +clarity around the nature of the violation and an explanation of why the +behavior was inappropriate. A public apology may be requested. + +### 2. Warning + +**Community Impact**: A violation through a single incident or series +of actions. + +**Consequence**: A warning with consequences for continued behavior. No +interaction with the people involved, including unsolicited interaction with +those enforcing the Code of Conduct, for a specified period of time. This +includes avoiding interactions in community spaces as well as external channels +like social media. Violating these terms may lead to a temporary or +permanent ban. + +### 3. Temporary Ban + +**Community Impact**: A serious violation of community standards, including +sustained inappropriate behavior. + +**Consequence**: A temporary ban from any sort of interaction or public +communication with the community for a specified period of time. No public or +private interaction with the people involved, including unsolicited interaction +with those enforcing the Code of Conduct, is allowed during this period. +Violating these terms may lead to a permanent ban. + +### 4. Permanent Ban + +**Community Impact**: Demonstrating a pattern of violation of community +standards, including sustained inappropriate behavior, harassment of an +individual, or aggression toward or disparagement of classes of individuals. + +**Consequence**: A permanent ban from any sort of public interaction within +the community. + +## Attribution + +This Code of Conduct is adapted from the [Contributor Covenant][homepage], +version 2.0, available at +https://www.contributor-covenant.org/version/2/0/code_of_conduct.html. + +Community Impact Guidelines were inspired by [Mozilla's code of conduct +enforcement ladder](https://github.com/mozilla/diversity). + +For answers to common questions about this code of conduct, see the FAQ at +https://www.contributor-covenant.org/faq. Translations are available at +https://www.contributor-covenant.org/translations. + +[homepage]: https://www.contributor-covenant.org diff --git a/yolov5/.github/ISSUE_TEMPLATE/bug-report.yml b/yolov5/.github/ISSUE_TEMPLATE/bug-report.yml new file mode 100644 index 0000000..fcb6413 --- /dev/null +++ b/yolov5/.github/ISSUE_TEMPLATE/bug-report.yml @@ -0,0 +1,85 @@ +name: 🐛 Bug Report +# title: " " +description: Problems with YOLOv5 +labels: [bug, triage] +body: + - type: markdown + attributes: + value: | + Thank you for submitting a YOLOv5 🐛 Bug Report! + + - type: checkboxes + attributes: + label: Search before asking + description: > + Please search the [issues](https://github.com/ultralytics/yolov5/issues) to see if a similar bug report already exists. + options: + - label: > + I have searched the YOLOv5 [issues](https://github.com/ultralytics/yolov5/issues) and found no similar bug report. + required: true + + - type: dropdown + attributes: + label: YOLOv5 Component + description: | + Please select the part of YOLOv5 where you found the bug. + multiple: true + options: + - "Training" + - "Validation" + - "Detection" + - "Export" + - "PyTorch Hub" + - "Multi-GPU" + - "Evolution" + - "Integrations" + - "Other" + validations: + required: false + + - type: textarea + attributes: + label: Bug + description: Provide console output with error messages and/or screenshots of the bug. + placeholder: | + 💡 ProTip! Include as much information as possible (screenshots, logs, tracebacks etc.) to receive the most helpful response. + validations: + required: true + + - type: textarea + attributes: + label: Environment + description: Please specify the software and hardware you used to produce the bug. + placeholder: | + - YOLO: YOLOv5 🚀 v6.0-67-g60e42e1 torch 1.9.0+cu111 CUDA:0 (A100-SXM4-40GB, 40536MiB) + - OS: Ubuntu 20.04 + - Python: 3.9.0 + validations: + required: false + + - type: textarea + attributes: + label: Minimal Reproducible Example + description: > + When asking a question, people will be better able to provide help if you provide code that they can easily understand and use to **reproduce** the problem. + This is referred to by community members as creating a [minimal reproducible example](https://stackoverflow.com/help/minimal-reproducible-example). + placeholder: | + ``` + # Code to reproduce your issue here + ``` + validations: + required: false + + - type: textarea + attributes: + label: Additional + description: Anything else you would like to share? + + - type: checkboxes + attributes: + label: Are you willing to submit a PR? + description: > + (Optional) We encourage you to submit a [Pull Request](https://github.com/ultralytics/yolov5/pulls) (PR) to help improve YOLOv5 for everyone, especially if you have a good understanding of how to implement a fix or feature. + See the YOLOv5 [Contributing Guide](https://github.com/ultralytics/yolov5/blob/master/CONTRIBUTING.md) to get started. + options: + - label: Yes I'd like to help by submitting a PR! diff --git a/yolov5/.github/ISSUE_TEMPLATE/config.yml b/yolov5/.github/ISSUE_TEMPLATE/config.yml new file mode 100644 index 0000000..4db7cef --- /dev/null +++ b/yolov5/.github/ISSUE_TEMPLATE/config.yml @@ -0,0 +1,8 @@ +blank_issues_enabled: true +contact_links: + - name: 💬 Forum + url: https://community.ultralytics.com/ + about: Ask on Ultralytics Community Forum + - name: Stack Overflow + url: https://stackoverflow.com/search?q=YOLOv5 + about: Ask on Stack Overflow with 'YOLOv5' tag diff --git a/yolov5/.github/ISSUE_TEMPLATE/feature-request.yml b/yolov5/.github/ISSUE_TEMPLATE/feature-request.yml new file mode 100644 index 0000000..68ef985 --- /dev/null +++ b/yolov5/.github/ISSUE_TEMPLATE/feature-request.yml @@ -0,0 +1,50 @@ +name: 🚀 Feature Request +description: Suggest a YOLOv5 idea +# title: " " +labels: [enhancement] +body: + - type: markdown + attributes: + value: | + Thank you for submitting a YOLOv5 🚀 Feature Request! + + - type: checkboxes + attributes: + label: Search before asking + description: > + Please search the [issues](https://github.com/ultralytics/yolov5/issues) to see if a similar feature request already exists. + options: + - label: > + I have searched the YOLOv5 [issues](https://github.com/ultralytics/yolov5/issues) and found no similar feature requests. + required: true + + - type: textarea + attributes: + label: Description + description: A short description of your feature. + placeholder: | + What new feature would you like to see in YOLOv5? + validations: + required: true + + - type: textarea + attributes: + label: Use case + description: | + Describe the use case of your feature request. It will help us understand and prioritize the feature request. + placeholder: | + How would this feature be used, and who would use it? + + - type: textarea + attributes: + label: Additional + description: Anything else you would like to share? + + - type: checkboxes + attributes: + label: Are you willing to submit a PR? + description: > + (Optional) We encourage you to submit a [Pull Request](https://github.com/ultralytics/yolov5/pulls) (PR) to help improve YOLOv5 for everyone, especially if you have a good understanding of how to implement a fix or feature. + See the YOLOv5 [Contributing Guide](https://github.com/ultralytics/yolov5/blob/master/CONTRIBUTING.md) to get started. + options: + - label: Yes I'd like to help by submitting a PR! diff --git a/yolov5/.github/ISSUE_TEMPLATE/question.yml b/yolov5/.github/ISSUE_TEMPLATE/question.yml new file mode 100644 index 0000000..8e0993c --- /dev/null +++ b/yolov5/.github/ISSUE_TEMPLATE/question.yml @@ -0,0 +1,33 @@ +name: ❓ Question +description: Ask a YOLOv5 question +# title: " " +labels: [question] +body: + - type: markdown + attributes: + value: | + Thank you for asking a YOLOv5 ❓ Question! + + - type: checkboxes + attributes: + label: Search before asking + description: > + Please search the [issues](https://github.com/ultralytics/yolov5/issues) and [discussions](https://github.com/ultralytics/yolov5/discussions) to see if a similar question already exists. + options: + - label: > + I have searched the YOLOv5 [issues](https://github.com/ultralytics/yolov5/issues) and [discussions](https://github.com/ultralytics/yolov5/discussions) and found no similar questions. + required: true + + - type: textarea + attributes: + label: Question + description: What is your question? + placeholder: | + 💡 ProTip! Include as much information as possible (screenshots, logs, tracebacks etc.) to receive the most helpful response. + validations: + required: true + + - type: textarea + attributes: + label: Additional + description: Anything else you would like to share? diff --git a/yolov5/.github/PULL_REQUEST_TEMPLATE.md b/yolov5/.github/PULL_REQUEST_TEMPLATE.md new file mode 100644 index 0000000..f25b017 --- /dev/null +++ b/yolov5/.github/PULL_REQUEST_TEMPLATE.md @@ -0,0 +1,9 @@ + diff --git a/yolov5/.github/SECURITY.md b/yolov5/.github/SECURITY.md new file mode 100644 index 0000000..aa3e840 --- /dev/null +++ b/yolov5/.github/SECURITY.md @@ -0,0 +1,7 @@ +# Security Policy + +We aim to make YOLOv5 🚀 as secure as possible! If you find potential vulnerabilities or have any concerns please let us know so we can investigate and take corrective action if needed. + +### Reporting a Vulnerability + +To report vulnerabilities please email us at hello@ultralytics.com or visit https://ultralytics.com/contact. Thank you! diff --git a/yolov5/.github/dependabot.yml b/yolov5/.github/dependabot.yml new file mode 100644 index 0000000..c1b3d5d --- /dev/null +++ b/yolov5/.github/dependabot.yml @@ -0,0 +1,23 @@ +version: 2 +updates: + - package-ecosystem: pip + directory: "/" + schedule: + interval: weekly + time: "04:00" + open-pull-requests-limit: 10 + reviewers: + - glenn-jocher + labels: + - dependencies + + - package-ecosystem: github-actions + directory: "/" + schedule: + interval: weekly + time: "04:00" + open-pull-requests-limit: 5 + reviewers: + - glenn-jocher + labels: + - dependencies diff --git a/yolov5/.github/workflows/ci-testing.yml b/yolov5/.github/workflows/ci-testing.yml new file mode 100644 index 0000000..6a73985 --- /dev/null +++ b/yolov5/.github/workflows/ci-testing.yml @@ -0,0 +1,121 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# YOLOv5 Continuous Integration (CI) GitHub Actions tests + +name: YOLOv5 CI + +on: + push: + branches: [master] + pull_request: + branches: [master] + schedule: + - cron: '0 0 * * *' # runs at 00:00 UTC every day + +jobs: + Benchmarks: + runs-on: ${{ matrix.os }} + strategy: + matrix: + os: [ubuntu-latest] + python-version: [3.9] + model: [yolov5n] + steps: + - uses: actions/checkout@v3 + - uses: actions/setup-python@v3 + with: + python-version: ${{ matrix.python-version }} + #- name: Cache pip + # uses: actions/cache@v3 + # with: + # path: ~/.cache/pip + # key: ${{ runner.os }}-Benchmarks-${{ hashFiles('requirements.txt') }} + # restore-keys: ${{ runner.os }}-Benchmarks- + - name: Install requirements + run: | + python -m pip install --upgrade pip + pip install -r requirements.txt coremltools openvino-dev tensorflow-cpu --extra-index-url https://download.pytorch.org/whl/cpu + python --version + pip --version + pip list + - name: Run benchmarks + run: | + python utils/benchmarks.py --weights ${{ matrix.model }}.pt --img 320 + + Tests: + timeout-minutes: 60 + runs-on: ${{ matrix.os }} + strategy: + fail-fast: false + matrix: + os: [ubuntu-latest, macos-latest, windows-latest] + python-version: [3.9] + model: [yolov5n] + include: + - os: ubuntu-latest + python-version: '3.7' # '3.6.8' min + model: yolov5n + - os: ubuntu-latest + python-version: '3.8' + model: yolov5n + - os: ubuntu-latest + python-version: '3.10' + model: yolov5n + steps: + - uses: actions/checkout@v3 + - uses: actions/setup-python@v3 + with: + python-version: ${{ matrix.python-version }} + - name: Get cache dir + # https://github.com/actions/cache/blob/master/examples.md#multiple-oss-in-a-workflow + id: pip-cache + run: echo "::set-output name=dir::$(pip cache dir)" + - name: Cache pip + uses: actions/cache@v3 + with: + path: ${{ steps.pip-cache.outputs.dir }} + key: ${{ runner.os }}-${{ matrix.python-version }}-pip-${{ hashFiles('requirements.txt') }} + restore-keys: ${{ runner.os }}-${{ matrix.python-version }}-pip- + - name: Install requirements + run: | + python -m pip install --upgrade pip + pip install -r requirements.txt --extra-index-url https://download.pytorch.org/whl/cpu + python --version + pip --version + pip list + - name: Check environment + run: | + python -c "import utils; utils.notebook_init()" + echo "RUNNER_OS is $RUNNER_OS" + echo "GITHUB_EVENT_NAME is $GITHUB_EVENT_NAME" + echo "GITHUB_WORKFLOW is $GITHUB_WORKFLOW" + echo "GITHUB_ACTOR is $GITHUB_ACTOR" + echo "GITHUB_REPOSITORY is $GITHUB_REPOSITORY" + echo "GITHUB_REPOSITORY_OWNER is $GITHUB_REPOSITORY_OWNER" + - name: Run tests + shell: bash + run: | + # export PYTHONPATH="$PWD" # to run '$ python *.py' files in subdirectories + d=cpu # device + model=${{ matrix.model }} + best=runs/train/exp/weights/best.pt + # Train + python train.py --img 64 --batch 32 --weights $model.pt --cfg $model.yaml --epochs 1 --device $d + # Val + python val.py --img 64 --batch 32 --weights $model.pt --device $d + python val.py --img 64 --batch 32 --weights $best --device $d + # Detect + python detect.py --weights $model.pt --device $d + python detect.py --weights $best --device $d + python hubconf.py # hub + # Export + # python models/tf.py --weights $model.pt # build TF model + python models/yolo.py --cfg $model.yaml # build PyTorch model + python export.py --weights $model.pt --img 64 --include torchscript # export + # Python + python - <=3.7.0**](https://www.python.org/) with all [requirements.txt](https://github.com/ultralytics/yolov5/blob/master/requirements.txt) installed including [**PyTorch>=1.7**](https://pytorch.org/get-started/locally/). To get started: + ```bash + git clone https://github.com/ultralytics/yolov5 # clone + cd yolov5 + pip install -r requirements.txt # install + ``` + + ## Environments + + YOLOv5 may be run in any of the following up-to-date verified environments (with all dependencies including [CUDA](https://developer.nvidia.com/cuda)/[CUDNN](https://developer.nvidia.com/cudnn), [Python](https://www.python.org/) and [PyTorch](https://pytorch.org/) preinstalled): + + - **Google Colab and Kaggle** notebooks with free GPU: Open In Colab Open In Kaggle + - **Google Cloud** Deep Learning VM. See [GCP Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/GCP-Quickstart) + - **Amazon** Deep Learning AMI. See [AWS Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/AWS-Quickstart) + - **Docker Image**. See [Docker Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/Docker-Quickstart) Docker Pulls + + + ## Status + + CI CPU testing + + If this badge is green, all [YOLOv5 GitHub Actions](https://github.com/ultralytics/yolov5/actions) Continuous Integration (CI) tests are currently passing. CI tests verify correct operation of YOLOv5 training ([train.py](https://github.com/ultralytics/yolov5/blob/master/train.py)), validation ([val.py](https://github.com/ultralytics/yolov5/blob/master/val.py)), inference ([detect.py](https://github.com/ultralytics/yolov5/blob/master/detect.py)) and export ([export.py](https://github.com/ultralytics/yolov5/blob/master/export.py)) on macOS, Windows, and Ubuntu every 24 hours and on every commit. diff --git a/yolov5/.github/workflows/rebase.yml b/yolov5/.github/workflows/rebase.yml new file mode 100644 index 0000000..a4dc9e5 --- /dev/null +++ b/yolov5/.github/workflows/rebase.yml @@ -0,0 +1,21 @@ +# https://github.com/marketplace/actions/automatic-rebase + +name: Automatic Rebase +on: + issue_comment: + types: [created] +jobs: + rebase: + name: Rebase + if: github.event.issue.pull_request != '' && contains(github.event.comment.body, '/rebase') + runs-on: ubuntu-latest + steps: + - name: Checkout the latest code + uses: actions/checkout@v3 + with: + token: ${{ secrets.ACTIONS_TOKEN }} + fetch-depth: 0 # otherwise, you will fail to push refs to dest repo + - name: Automatic Rebase + uses: cirrus-actions/rebase@1.7 + env: + GITHUB_TOKEN: ${{ secrets.ACTIONS_TOKEN }} diff --git a/yolov5/.github/workflows/stale.yml b/yolov5/.github/workflows/stale.yml new file mode 100644 index 0000000..ee08510 --- /dev/null +++ b/yolov5/.github/workflows/stale.yml @@ -0,0 +1,38 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +name: Close stale issues +on: + schedule: + - cron: '0 0 * * *' # Runs at 00:00 UTC every day + +jobs: + stale: + runs-on: ubuntu-latest + steps: + - uses: actions/stale@v5 + with: + repo-token: ${{ secrets.GITHUB_TOKEN }} + stale-issue-message: | + 👋 Hello, this issue has been automatically marked as stale because it has not had recent activity. Please note it will be closed if no further activity occurs. + + Access additional [YOLOv5](https://ultralytics.com/yolov5) 🚀 resources: + - **Wiki** – https://github.com/ultralytics/yolov5/wiki + - **Tutorials** – https://github.com/ultralytics/yolov5#tutorials + - **Docs** – https://docs.ultralytics.com + + Access additional [Ultralytics](https://ultralytics.com) ⚡ resources: + - **Ultralytics HUB** – https://ultralytics.com/hub + - **Vision API** – https://ultralytics.com/yolov5 + - **About Us** – https://ultralytics.com/about + - **Join Our Team** – https://ultralytics.com/work + - **Contact Us** – https://ultralytics.com/contact + + Feel free to inform us of any other **issues** you discover or **feature requests** that come to mind in the future. Pull Requests (PRs) are also always welcomed! + + Thank you for your contributions to YOLOv5 🚀 and Vision AI ⭐! + + stale-pr-message: 'This pull request has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions YOLOv5 🚀 and Vision AI ⭐.' + days-before-stale: 30 + days-before-close: 5 + exempt-issue-labels: 'documentation,tutorial,TODO' + operations-per-run: 300 # The maximum number of operations per run, used to control rate limiting. diff --git a/yolov5/.gitignore b/yolov5/.gitignore new file mode 100644 index 0000000..69a0084 --- /dev/null +++ b/yolov5/.gitignore @@ -0,0 +1,256 @@ +# Repo-specific GitIgnore ---------------------------------------------------------------------------------------------- +*.jpg +*.jpeg +*.png +*.bmp +*.tif +*.tiff +*.heic +*.JPG +*.JPEG +*.PNG +*.BMP +*.TIF +*.TIFF +*.HEIC +*.mp4 +*.mov +*.MOV +*.avi +*.data +*.json +*.cfg +!setup.cfg +!cfg/yolov3*.cfg + +storage.googleapis.com +runs/* +data/* +data/images/* +!data/*.yaml +!data/hyps +!data/scripts +!data/images +!data/images/zidane.jpg +!data/images/bus.jpg +!data/*.sh + +results*.csv + +# Datasets ------------------------------------------------------------------------------------------------------------- +coco/ +coco128/ +VOC/ + +# MATLAB GitIgnore ----------------------------------------------------------------------------------------------------- +*.m~ +*.mat +!targets*.mat + +# Neural Network weights ----------------------------------------------------------------------------------------------- +*.weights +*.pt +*.pb +*.onnx +*.engine +*.mlmodel +*.torchscript +*.tflite +*.h5 +*_saved_model/ +*_web_model/ +*_openvino_model/ +darknet53.conv.74 +yolov3-tiny.conv.15 + +# GitHub Python GitIgnore ---------------------------------------------------------------------------------------------- +# Byte-compiled / optimized / DLL files +__pycache__/ +*.py[cod] +*$py.class + +# C extensions +*.so + +# Distribution / packaging +.Python +env/ +build/ +develop-eggs/ +dist/ +downloads/ +eggs/ +.eggs/ +lib/ +lib64/ +parts/ +sdist/ +var/ +wheels/ +*.egg-info/ +/wandb/ +.installed.cfg +*.egg + + +# PyInstaller +# Usually these files are written by a python script from a template +# before PyInstaller builds the exe, so as to inject date/other infos into it. +*.manifest +*.spec + +# Installer logs +pip-log.txt +pip-delete-this-directory.txt + +# Unit test / coverage reports +htmlcov/ +.tox/ +.coverage +.coverage.* +.cache +nosetests.xml +coverage.xml +*.cover +.hypothesis/ + +# Translations +*.mo +*.pot + +# Django stuff: +*.log +local_settings.py + +# Flask stuff: +instance/ +.webassets-cache + +# Scrapy stuff: +.scrapy + +# Sphinx documentation +docs/_build/ + +# PyBuilder +target/ + +# Jupyter Notebook +.ipynb_checkpoints + +# pyenv +.python-version + +# celery beat schedule file +celerybeat-schedule + +# SageMath parsed files +*.sage.py + +# dotenv +.env + +# virtualenv +.venv* +venv*/ +ENV*/ + +# Spyder project settings +.spyderproject +.spyproject + +# Rope project settings +.ropeproject + +# mkdocs documentation +/site + +# mypy +.mypy_cache/ + + +# https://github.com/github/gitignore/blob/master/Global/macOS.gitignore ----------------------------------------------- + +# General +.DS_Store +.AppleDouble +.LSOverride + +# Icon must end with two \r +Icon +Icon? + +# Thumbnails +._* + +# Files that might appear in the root of a volume +.DocumentRevisions-V100 +.fseventsd +.Spotlight-V100 +.TemporaryItems +.Trashes +.VolumeIcon.icns +.com.apple.timemachine.donotpresent + +# Directories potentially created on remote AFP share +.AppleDB +.AppleDesktop +Network Trash Folder +Temporary Items +.apdisk + + +# https://github.com/github/gitignore/blob/master/Global/JetBrains.gitignore +# Covers JetBrains IDEs: IntelliJ, RubyMine, PhpStorm, AppCode, PyCharm, CLion, Android Studio and WebStorm +# Reference: https://intellij-support.jetbrains.com/hc/en-us/articles/206544839 + +# User-specific stuff: +.idea/* +.idea/**/workspace.xml +.idea/**/tasks.xml +.idea/dictionaries +.html # Bokeh Plots +.pg # TensorFlow Frozen Graphs +.avi # videos + +# Sensitive or high-churn files: +.idea/**/dataSources/ +.idea/**/dataSources.ids +.idea/**/dataSources.local.xml +.idea/**/sqlDataSources.xml +.idea/**/dynamic.xml +.idea/**/uiDesigner.xml + +# Gradle: +.idea/**/gradle.xml +.idea/**/libraries + +# CMake +cmake-build-debug/ +cmake-build-release/ + +# Mongo Explorer plugin: +.idea/**/mongoSettings.xml + +## File-based project format: +*.iws + +## Plugin-specific files: + +# IntelliJ +out/ + +# mpeltonen/sbt-idea plugin +.idea_modules/ + +# JIRA plugin +atlassian-ide-plugin.xml + +# Cursive Clojure plugin +.idea/replstate.xml + +# Crashlytics plugin (for Android Studio and IntelliJ) +com_crashlytics_export_strings.xml +crashlytics.properties +crashlytics-build.properties +fabric.properties diff --git a/yolov5/.pre-commit-config.yaml b/yolov5/.pre-commit-config.yaml new file mode 100644 index 0000000..828dc91 --- /dev/null +++ b/yolov5/.pre-commit-config.yaml @@ -0,0 +1,67 @@ +# Define hooks for code formations +# Will be applied on any updated commit files if a user has installed and linked commit hook + +default_language_version: + python: python3.8 + +# Define bot property if installed via https://github.com/marketplace/pre-commit-ci +ci: + autofix_prs: true + autoupdate_commit_msg: '[pre-commit.ci] pre-commit suggestions' + autoupdate_schedule: monthly + # submodules: true + +repos: + - repo: https://github.com/pre-commit/pre-commit-hooks + rev: v4.2.0 + hooks: + - id: end-of-file-fixer + - id: trailing-whitespace + - id: check-case-conflict + - id: check-yaml + - id: check-toml + - id: pretty-format-json + - id: check-docstring-first + + - repo: https://github.com/asottile/pyupgrade + rev: v2.32.0 + hooks: + - id: pyupgrade + name: Upgrade code + args: [ --py37-plus ] + + - repo: https://github.com/PyCQA/isort + rev: 5.10.1 + hooks: + - id: isort + name: Sort imports + + - repo: https://github.com/pre-commit/mirrors-yapf + rev: v0.32.0 + hooks: + - id: yapf + name: YAPF formatting + + - repo: https://github.com/executablebooks/mdformat + rev: 0.7.14 + hooks: + - id: mdformat + name: MD formatting + additional_dependencies: + - mdformat-gfm + - mdformat-black + exclude: | + (?x)^( + README.md + )$ + + - repo: https://github.com/asottile/yesqa + rev: v1.3.0 + hooks: + - id: yesqa + + - repo: https://github.com/PyCQA/flake8 + rev: 4.0.1 + hooks: + - id: flake8 + name: PEP8 diff --git a/yolov5/CONTRIBUTING.md b/yolov5/CONTRIBUTING.md new file mode 100644 index 0000000..13b9b73 --- /dev/null +++ b/yolov5/CONTRIBUTING.md @@ -0,0 +1,98 @@ +## Contributing to YOLOv5 🚀 + +We love your input! We want to make contributing to YOLOv5 as easy and transparent as possible, whether it's: + +- Reporting a bug +- Discussing the current state of the code +- Submitting a fix +- Proposing a new feature +- Becoming a maintainer + +YOLOv5 works so well due to our combined community effort, and for every small improvement you contribute you will be +helping push the frontiers of what's possible in AI 😃! + +## Submitting a Pull Request (PR) 🛠️ + +Submitting a PR is easy! This example shows how to submit a PR for updating `requirements.txt` in 4 steps: + +### 1. Select File to Update + +Select `requirements.txt` to update by clicking on it in GitHub. + +

PR_step1

+ +### 2. Click 'Edit this file' + +Button is in top-right corner. + +

PR_step2

+ +### 3. Make Changes + +Change `matplotlib` version from `3.2.2` to `3.3`. + +

PR_step3

+ +### 4. Preview Changes and Submit PR + +Click on the **Preview changes** tab to verify your updates. At the bottom of the screen select 'Create a **new branch** +for this commit', assign your branch a descriptive name such as `fix/matplotlib_version` and click the green **Propose +changes** button. All done, your PR is now submitted to YOLOv5 for review and approval 😃! + +

PR_step4

+ +### PR recommendations + +To allow your work to be integrated as seamlessly as possible, we advise you to: + +- ✅ Verify your PR is **up-to-date with upstream/master.** If your PR is behind upstream/master an + automatic [GitHub Actions](https://github.com/ultralytics/yolov5/blob/master/.github/workflows/rebase.yml) merge may + be attempted by writing /rebase in a new comment, or by running the following code, replacing 'feature' with the name + of your local branch: + +```bash +git remote add upstream https://github.com/ultralytics/yolov5.git +git fetch upstream +# git checkout feature # <--- replace 'feature' with local branch name +git merge upstream/master +git push -u origin -f +``` + +- ✅ Verify all Continuous Integration (CI) **checks are passing**. +- ✅ Reduce changes to the absolute **minimum** required for your bug fix or feature addition. _"It is not daily increase + but daily decrease, hack away the unessential. The closer to the source, the less wastage there is."_ — Bruce Lee + +## Submitting a Bug Report 🐛 + +If you spot a problem with YOLOv5 please submit a Bug Report! + +For us to start investigating a possible problem we need to be able to reproduce it ourselves first. We've created a few +short guidelines below to help users provide what we need in order to get started. + +When asking a question, people will be better able to provide help if you provide **code** that they can easily +understand and use to **reproduce** the problem. This is referred to by community members as creating +a [minimum reproducible example](https://stackoverflow.com/help/minimal-reproducible-example). Your code that reproduces +the problem should be: + +- ✅ **Minimal** – Use as little code as possible that still produces the same problem +- ✅ **Complete** – Provide **all** parts someone else needs to reproduce your problem in the question itself +- ✅ **Reproducible** – Test the code you're about to provide to make sure it reproduces the problem + +In addition to the above requirements, for [Ultralytics](https://ultralytics.com/) to provide assistance your code +should be: + +- ✅ **Current** – Verify that your code is up-to-date with current + GitHub [master](https://github.com/ultralytics/yolov5/tree/master), and if necessary `git pull` or `git clone` a new + copy to ensure your problem has not already been resolved by previous commits. +- ✅ **Unmodified** – Your problem must be reproducible without any modifications to the codebase in this + repository. [Ultralytics](https://ultralytics.com/) does not provide support for custom code ⚠️. + +If you believe your problem meets all of the above criteria, please close this issue and raise a new one using the 🐛 +**Bug Report** [template](https://github.com/ultralytics/yolov5/issues/new/choose) and providing +a [minimum reproducible example](https://stackoverflow.com/help/minimal-reproducible-example) to help us better +understand and diagnose your problem. + +## License + +By contributing, you agree that your contributions will be licensed under +the [GPL-3.0 license](https://choosealicense.com/licenses/gpl-3.0/) diff --git a/yolov5/LICENSE b/yolov5/LICENSE new file mode 100644 index 0000000..92b370f --- /dev/null +++ b/yolov5/LICENSE @@ -0,0 +1,674 @@ +GNU GENERAL PUBLIC LICENSE + Version 3, 29 June 2007 + + Copyright (C) 2007 Free Software Foundation, Inc. + Everyone is permitted to copy and distribute verbatim copies + of this license document, but changing it is not allowed. + + Preamble + + The GNU General Public License is a free, copyleft license for +software and other kinds of works. + + The licenses for most software and other practical works are designed +to take away your freedom to share and change the works. By contrast, +the GNU General Public License is intended to guarantee your freedom to +share and change all versions of a program--to make sure it remains free +software for all its users. We, the Free Software Foundation, use the +GNU General Public License for most of our software; it applies also to +any other work released this way by its authors. You can apply it to +your programs, too. + + When we speak of free software, we are referring to freedom, not +price. Our General Public Licenses are designed to make sure that you +have the freedom to distribute copies of free software (and charge for +them if you wish), that you receive source code or can get it if you +want it, that you can change the software or use pieces of it in new +free programs, and that you know you can do these things. + + To protect your rights, we need to prevent others from denying you +these rights or asking you to surrender the rights. Therefore, you have +certain responsibilities if you distribute copies of the software, or if +you modify it: responsibilities to respect the freedom of others. + + For example, if you distribute copies of such a program, whether +gratis or for a fee, you must pass on to the recipients the same +freedoms that you received. You must make sure that they, too, receive +or can get the source code. And you must show them these terms so they +know their rights. + + Developers that use the GNU GPL protect your rights with two steps: +(1) assert copyright on the software, and (2) offer you this License +giving you legal permission to copy, distribute and/or modify it. + + For the developers' and authors' protection, the GPL clearly explains +that there is no warranty for this free software. For both users' and +authors' sake, the GPL requires that modified versions be marked as +changed, so that their problems will not be attributed erroneously to +authors of previous versions. + + Some devices are designed to deny users access to install or run +modified versions of the software inside them, although the manufacturer +can do so. This is fundamentally incompatible with the aim of +protecting users' freedom to change the software. The systematic +pattern of such abuse occurs in the area of products for individuals to +use, which is precisely where it is most unacceptable. Therefore, we +have designed this version of the GPL to prohibit the practice for those +products. If such problems arise substantially in other domains, we +stand ready to extend this provision to those domains in future versions +of the GPL, as needed to protect the freedom of users. + + Finally, every program is threatened constantly by software patents. +States should not allow patents to restrict development and use of +software on general-purpose computers, but in those that do, we wish to +avoid the special danger that patents applied to a free program could +make it effectively proprietary. To prevent this, the GPL assures that +patents cannot be used to render the program non-free. + + The precise terms and conditions for copying, distribution and +modification follow. + + TERMS AND CONDITIONS + + 0. Definitions. + + "This License" refers to version 3 of the GNU General Public License. + + "Copyright" also means copyright-like laws that apply to other kinds of +works, such as semiconductor masks. + + "The Program" refers to any copyrightable work licensed under this +License. Each licensee is addressed as "you". "Licensees" and +"recipients" may be individuals or organizations. + + To "modify" a work means to copy from or adapt all or part of the work +in a fashion requiring copyright permission, other than the making of an +exact copy. The resulting work is called a "modified version" of the +earlier work or a work "based on" the earlier work. + + A "covered work" means either the unmodified Program or a work based +on the Program. + + To "propagate" a work means to do anything with it that, without +permission, would make you directly or secondarily liable for +infringement under applicable copyright law, except executing it on a +computer or modifying a private copy. Propagation includes copying, +distribution (with or without modification), making available to the +public, and in some countries other activities as well. + + To "convey" a work means any kind of propagation that enables other +parties to make or receive copies. Mere interaction with a user through +a computer network, with no transfer of a copy, is not conveying. + + An interactive user interface displays "Appropriate Legal Notices" +to the extent that it includes a convenient and prominently visible +feature that (1) displays an appropriate copyright notice, and (2) +tells the user that there is no warranty for the work (except to the +extent that warranties are provided), that licensees may convey the +work under this License, and how to view a copy of this License. If +the interface presents a list of user commands or options, such as a +menu, a prominent item in the list meets this criterion. + + 1. Source Code. + + The "source code" for a work means the preferred form of the work +for making modifications to it. "Object code" means any non-source +form of a work. + + A "Standard Interface" means an interface that either is an official +standard defined by a recognized standards body, or, in the case of +interfaces specified for a particular programming language, one that +is widely used among developers working in that language. + + The "System Libraries" of an executable work include anything, other +than the work as a whole, that (a) is included in the normal form of +packaging a Major Component, but which is not part of that Major +Component, and (b) serves only to enable use of the work with that +Major Component, or to implement a Standard Interface for which an +implementation is available to the public in source code form. A +"Major Component", in this context, means a major essential component +(kernel, window system, and so on) of the specific operating system +(if any) on which the executable work runs, or a compiler used to +produce the work, or an object code interpreter used to run it. + + The "Corresponding Source" for a work in object code form means all +the source code needed to generate, install, and (for an executable +work) run the object code and to modify the work, including scripts to +control those activities. However, it does not include the work's +System Libraries, or general-purpose tools or generally available free +programs which are used unmodified in performing those activities but +which are not part of the work. For example, Corresponding Source +includes interface definition files associated with source files for +the work, and the source code for shared libraries and dynamically +linked subprograms that the work is specifically designed to require, +such as by intimate data communication or control flow between those +subprograms and other parts of the work. + + The Corresponding Source need not include anything that users +can regenerate automatically from other parts of the Corresponding +Source. + + The Corresponding Source for a work in source code form is that +same work. + + 2. Basic Permissions. + + All rights granted under this License are granted for the term of +copyright on the Program, and are irrevocable provided the stated +conditions are met. This License explicitly affirms your unlimited +permission to run the unmodified Program. The output from running a +covered work is covered by this License only if the output, given its +content, constitutes a covered work. This License acknowledges your +rights of fair use or other equivalent, as provided by copyright law. + + You may make, run and propagate covered works that you do not +convey, without conditions so long as your license otherwise remains +in force. You may convey covered works to others for the sole purpose +of having them make modifications exclusively for you, or provide you +with facilities for running those works, provided that you comply with +the terms of this License in conveying all material for which you do +not control copyright. Those thus making or running the covered works +for you must do so exclusively on your behalf, under your direction +and control, on terms that prohibit them from making any copies of +your copyrighted material outside their relationship with you. + + Conveying under any other circumstances is permitted solely under +the conditions stated below. Sublicensing is not allowed; section 10 +makes it unnecessary. + + 3. Protecting Users' Legal Rights From Anti-Circumvention Law. + + No covered work shall be deemed part of an effective technological +measure under any applicable law fulfilling obligations under article +11 of the WIPO copyright treaty adopted on 20 December 1996, or +similar laws prohibiting or restricting circumvention of such +measures. + + When you convey a covered work, you waive any legal power to forbid +circumvention of technological measures to the extent such circumvention +is effected by exercising rights under this License with respect to +the covered work, and you disclaim any intention to limit operation or +modification of the work as a means of enforcing, against the work's +users, your or third parties' legal rights to forbid circumvention of +technological measures. + + 4. Conveying Verbatim Copies. + + You may convey verbatim copies of the Program's source code as you +receive it, in any medium, provided that you conspicuously and +appropriately publish on each copy an appropriate copyright notice; +keep intact all notices stating that this License and any +non-permissive terms added in accord with section 7 apply to the code; +keep intact all notices of the absence of any warranty; and give all +recipients a copy of this License along with the Program. + + You may charge any price or no price for each copy that you convey, +and you may offer support or warranty protection for a fee. + + 5. Conveying Modified Source Versions. + + You may convey a work based on the Program, or the modifications to +produce it from the Program, in the form of source code under the +terms of section 4, provided that you also meet all of these conditions: + + a) The work must carry prominent notices stating that you modified + it, and giving a relevant date. + + b) The work must carry prominent notices stating that it is + released under this License and any conditions added under section + 7. This requirement modifies the requirement in section 4 to + "keep intact all notices". + + c) You must license the entire work, as a whole, under this + License to anyone who comes into possession of a copy. This + License will therefore apply, along with any applicable section 7 + additional terms, to the whole of the work, and all its parts, + regardless of how they are packaged. This License gives no + permission to license the work in any other way, but it does not + invalidate such permission if you have separately received it. + + d) If the work has interactive user interfaces, each must display + Appropriate Legal Notices; however, if the Program has interactive + interfaces that do not display Appropriate Legal Notices, your + work need not make them do so. + + A compilation of a covered work with other separate and independent +works, which are not by their nature extensions of the covered work, +and which are not combined with it such as to form a larger program, +in or on a volume of a storage or distribution medium, is called an +"aggregate" if the compilation and its resulting copyright are not +used to limit the access or legal rights of the compilation's users +beyond what the individual works permit. Inclusion of a covered work +in an aggregate does not cause this License to apply to the other +parts of the aggregate. + + 6. Conveying Non-Source Forms. + + You may convey a covered work in object code form under the terms +of sections 4 and 5, provided that you also convey the +machine-readable Corresponding Source under the terms of this License, +in one of these ways: + + a) Convey the object code in, or embodied in, a physical product + (including a physical distribution medium), accompanied by the + Corresponding Source fixed on a durable physical medium + customarily used for software interchange. + + b) Convey the object code in, or embodied in, a physical product + (including a physical distribution medium), accompanied by a + written offer, valid for at least three years and valid for as + long as you offer spare parts or customer support for that product + model, to give anyone who possesses the object code either (1) a + copy of the Corresponding Source for all the software in the + product that is covered by this License, on a durable physical + medium customarily used for software interchange, for a price no + more than your reasonable cost of physically performing this + conveying of source, or (2) access to copy the + Corresponding Source from a network server at no charge. + + c) Convey individual copies of the object code with a copy of the + written offer to provide the Corresponding Source. This + alternative is allowed only occasionally and noncommercially, and + only if you received the object code with such an offer, in accord + with subsection 6b. + + d) Convey the object code by offering access from a designated + place (gratis or for a charge), and offer equivalent access to the + Corresponding Source in the same way through the same place at no + further charge. You need not require recipients to copy the + Corresponding Source along with the object code. If the place to + copy the object code is a network server, the Corresponding Source + may be on a different server (operated by you or a third party) + that supports equivalent copying facilities, provided you maintain + clear directions next to the object code saying where to find the + Corresponding Source. Regardless of what server hosts the + Corresponding Source, you remain obligated to ensure that it is + available for as long as needed to satisfy these requirements. + + e) Convey the object code using peer-to-peer transmission, provided + you inform other peers where the object code and Corresponding + Source of the work are being offered to the general public at no + charge under subsection 6d. + + A separable portion of the object code, whose source code is excluded +from the Corresponding Source as a System Library, need not be +included in conveying the object code work. + + A "User Product" is either (1) a "consumer product", which means any +tangible personal property which is normally used for personal, family, +or household purposes, or (2) anything designed or sold for incorporation +into a dwelling. In determining whether a product is a consumer product, +doubtful cases shall be resolved in favor of coverage. For a particular +product received by a particular user, "normally used" refers to a +typical or common use of that class of product, regardless of the status +of the particular user or of the way in which the particular user +actually uses, or expects or is expected to use, the product. A product +is a consumer product regardless of whether the product has substantial +commercial, industrial or non-consumer uses, unless such uses represent +the only significant mode of use of the product. + + "Installation Information" for a User Product means any methods, +procedures, authorization keys, or other information required to install +and execute modified versions of a covered work in that User Product from +a modified version of its Corresponding Source. The information must +suffice to ensure that the continued functioning of the modified object +code is in no case prevented or interfered with solely because +modification has been made. + + If you convey an object code work under this section in, or with, or +specifically for use in, a User Product, and the conveying occurs as +part of a transaction in which the right of possession and use of the +User Product is transferred to the recipient in perpetuity or for a +fixed term (regardless of how the transaction is characterized), the +Corresponding Source conveyed under this section must be accompanied +by the Installation Information. But this requirement does not apply +if neither you nor any third party retains the ability to install +modified object code on the User Product (for example, the work has +been installed in ROM). + + The requirement to provide Installation Information does not include a +requirement to continue to provide support service, warranty, or updates +for a work that has been modified or installed by the recipient, or for +the User Product in which it has been modified or installed. Access to a +network may be denied when the modification itself materially and +adversely affects the operation of the network or violates the rules and +protocols for communication across the network. + + Corresponding Source conveyed, and Installation Information provided, +in accord with this section must be in a format that is publicly +documented (and with an implementation available to the public in +source code form), and must require no special password or key for +unpacking, reading or copying. + + 7. Additional Terms. + + "Additional permissions" are terms that supplement the terms of this +License by making exceptions from one or more of its conditions. +Additional permissions that are applicable to the entire Program shall +be treated as though they were included in this License, to the extent +that they are valid under applicable law. If additional permissions +apply only to part of the Program, that part may be used separately +under those permissions, but the entire Program remains governed by +this License without regard to the additional permissions. + + When you convey a copy of a covered work, you may at your option +remove any additional permissions from that copy, or from any part of +it. (Additional permissions may be written to require their own +removal in certain cases when you modify the work.) You may place +additional permissions on material, added by you to a covered work, +for which you have or can give appropriate copyright permission. + + Notwithstanding any other provision of this License, for material you +add to a covered work, you may (if authorized by the copyright holders of +that material) supplement the terms of this License with terms: + + a) Disclaiming warranty or limiting liability differently from the + terms of sections 15 and 16 of this License; or + + b) Requiring preservation of specified reasonable legal notices or + author attributions in that material or in the Appropriate Legal + Notices displayed by works containing it; or + + c) Prohibiting misrepresentation of the origin of that material, or + requiring that modified versions of such material be marked in + reasonable ways as different from the original version; or + + d) Limiting the use for publicity purposes of names of licensors or + authors of the material; or + + e) Declining to grant rights under trademark law for use of some + trade names, trademarks, or service marks; or + + f) Requiring indemnification of licensors and authors of that + material by anyone who conveys the material (or modified versions of + it) with contractual assumptions of liability to the recipient, for + any liability that these contractual assumptions directly impose on + those licensors and authors. + + All other non-permissive additional terms are considered "further +restrictions" within the meaning of section 10. If the Program as you +received it, or any part of it, contains a notice stating that it is +governed by this License along with a term that is a further +restriction, you may remove that term. If a license document contains +a further restriction but permits relicensing or conveying under this +License, you may add to a covered work material governed by the terms +of that license document, provided that the further restriction does +not survive such relicensing or conveying. + + If you add terms to a covered work in accord with this section, you +must place, in the relevant source files, a statement of the +additional terms that apply to those files, or a notice indicating +where to find the applicable terms. + + Additional terms, permissive or non-permissive, may be stated in the +form of a separately written license, or stated as exceptions; +the above requirements apply either way. + + 8. Termination. + + You may not propagate or modify a covered work except as expressly +provided under this License. Any attempt otherwise to propagate or +modify it is void, and will automatically terminate your rights under +this License (including any patent licenses granted under the third +paragraph of section 11). + + However, if you cease all violation of this License, then your +license from a particular copyright holder is reinstated (a) +provisionally, unless and until the copyright holder explicitly and +finally terminates your license, and (b) permanently, if the copyright +holder fails to notify you of the violation by some reasonable means +prior to 60 days after the cessation. + + Moreover, your license from a particular copyright holder is +reinstated permanently if the copyright holder notifies you of the +violation by some reasonable means, this is the first time you have +received notice of violation of this License (for any work) from that +copyright holder, and you cure the violation prior to 30 days after +your receipt of the notice. + + Termination of your rights under this section does not terminate the +licenses of parties who have received copies or rights from you under +this License. If your rights have been terminated and not permanently +reinstated, you do not qualify to receive new licenses for the same +material under section 10. + + 9. Acceptance Not Required for Having Copies. + + You are not required to accept this License in order to receive or +run a copy of the Program. Ancillary propagation of a covered work +occurring solely as a consequence of using peer-to-peer transmission +to receive a copy likewise does not require acceptance. However, +nothing other than this License grants you permission to propagate or +modify any covered work. These actions infringe copyright if you do +not accept this License. Therefore, by modifying or propagating a +covered work, you indicate your acceptance of this License to do so. + + 10. Automatic Licensing of Downstream Recipients. + + Each time you convey a covered work, the recipient automatically +receives a license from the original licensors, to run, modify and +propagate that work, subject to this License. You are not responsible +for enforcing compliance by third parties with this License. + + An "entity transaction" is a transaction transferring control of an +organization, or substantially all assets of one, or subdividing an +organization, or merging organizations. If propagation of a covered +work results from an entity transaction, each party to that +transaction who receives a copy of the work also receives whatever +licenses to the work the party's predecessor in interest had or could +give under the previous paragraph, plus a right to possession of the +Corresponding Source of the work from the predecessor in interest, if +the predecessor has it or can get it with reasonable efforts. + + You may not impose any further restrictions on the exercise of the +rights granted or affirmed under this License. For example, you may +not impose a license fee, royalty, or other charge for exercise of +rights granted under this License, and you may not initiate litigation +(including a cross-claim or counterclaim in a lawsuit) alleging that +any patent claim is infringed by making, using, selling, offering for +sale, or importing the Program or any portion of it. + + 11. Patents. + + A "contributor" is a copyright holder who authorizes use under this +License of the Program or a work on which the Program is based. The +work thus licensed is called the contributor's "contributor version". + + A contributor's "essential patent claims" are all patent claims +owned or controlled by the contributor, whether already acquired or +hereafter acquired, that would be infringed by some manner, permitted +by this License, of making, using, or selling its contributor version, +but do not include claims that would be infringed only as a +consequence of further modification of the contributor version. For +purposes of this definition, "control" includes the right to grant +patent sublicenses in a manner consistent with the requirements of +this License. + + Each contributor grants you a non-exclusive, worldwide, royalty-free +patent license under the contributor's essential patent claims, to +make, use, sell, offer for sale, import and otherwise run, modify and +propagate the contents of its contributor version. + + In the following three paragraphs, a "patent license" is any express +agreement or commitment, however denominated, not to enforce a patent +(such as an express permission to practice a patent or covenant not to +sue for patent infringement). To "grant" such a patent license to a +party means to make such an agreement or commitment not to enforce a +patent against the party. + + If you convey a covered work, knowingly relying on a patent license, +and the Corresponding Source of the work is not available for anyone +to copy, free of charge and under the terms of this License, through a +publicly available network server or other readily accessible means, +then you must either (1) cause the Corresponding Source to be so +available, or (2) arrange to deprive yourself of the benefit of the +patent license for this particular work, or (3) arrange, in a manner +consistent with the requirements of this License, to extend the patent +license to downstream recipients. "Knowingly relying" means you have +actual knowledge that, but for the patent license, your conveying the +covered work in a country, or your recipient's use of the covered work +in a country, would infringe one or more identifiable patents in that +country that you have reason to believe are valid. + + If, pursuant to or in connection with a single transaction or +arrangement, you convey, or propagate by procuring conveyance of, a +covered work, and grant a patent license to some of the parties +receiving the covered work authorizing them to use, propagate, modify +or convey a specific copy of the covered work, then the patent license +you grant is automatically extended to all recipients of the covered +work and works based on it. + + A patent license is "discriminatory" if it does not include within +the scope of its coverage, prohibits the exercise of, or is +conditioned on the non-exercise of one or more of the rights that are +specifically granted under this License. You may not convey a covered +work if you are a party to an arrangement with a third party that is +in the business of distributing software, under which you make payment +to the third party based on the extent of your activity of conveying +the work, and under which the third party grants, to any of the +parties who would receive the covered work from you, a discriminatory +patent license (a) in connection with copies of the covered work +conveyed by you (or copies made from those copies), or (b) primarily +for and in connection with specific products or compilations that +contain the covered work, unless you entered into that arrangement, +or that patent license was granted, prior to 28 March 2007. + + Nothing in this License shall be construed as excluding or limiting +any implied license or other defenses to infringement that may +otherwise be available to you under applicable patent law. + + 12. No Surrender of Others' Freedom. + + If conditions are imposed on you (whether by court order, agreement or +otherwise) that contradict the conditions of this License, they do not +excuse you from the conditions of this License. If you cannot convey a +covered work so as to satisfy simultaneously your obligations under this +License and any other pertinent obligations, then as a consequence you may +not convey it at all. For example, if you agree to terms that obligate you +to collect a royalty for further conveying from those to whom you convey +the Program, the only way you could satisfy both those terms and this +License would be to refrain entirely from conveying the Program. + + 13. Use with the GNU Affero General Public License. + + Notwithstanding any other provision of this License, you have +permission to link or combine any covered work with a work licensed +under version 3 of the GNU Affero General Public License into a single +combined work, and to convey the resulting work. The terms of this +License will continue to apply to the part which is the covered work, +but the special requirements of the GNU Affero General Public License, +section 13, concerning interaction through a network will apply to the +combination as such. + + 14. Revised Versions of this License. + + The Free Software Foundation may publish revised and/or new versions of +the GNU General Public License from time to time. Such new versions will +be similar in spirit to the present version, but may differ in detail to +address new problems or concerns. + + Each version is given a distinguishing version number. If the +Program specifies that a certain numbered version of the GNU General +Public License "or any later version" applies to it, you have the +option of following the terms and conditions either of that numbered +version or of any later version published by the Free Software +Foundation. If the Program does not specify a version number of the +GNU General Public License, you may choose any version ever published +by the Free Software Foundation. + + If the Program specifies that a proxy can decide which future +versions of the GNU General Public License can be used, that proxy's +public statement of acceptance of a version permanently authorizes you +to choose that version for the Program. + + Later license versions may give you additional or different +permissions. However, no additional obligations are imposed on any +author or copyright holder as a result of your choosing to follow a +later version. + + 15. Disclaimer of Warranty. + + THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY +APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT +HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY +OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, +THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR +PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM +IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF +ALL NECESSARY SERVICING, REPAIR OR CORRECTION. + + 16. Limitation of Liability. + + IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING +WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS +THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY +GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE +USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF +DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD +PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), +EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF +SUCH DAMAGES. + + 17. Interpretation of Sections 15 and 16. + + If the disclaimer of warranty and limitation of liability provided +above cannot be given local legal effect according to their terms, +reviewing courts shall apply local law that most closely approximates +an absolute waiver of all civil liability in connection with the +Program, unless a warranty or assumption of liability accompanies a +copy of the Program in return for a fee. + + END OF TERMS AND CONDITIONS + + How to Apply These Terms to Your New Programs + + If you develop a new program, and you want it to be of the greatest +possible use to the public, the best way to achieve this is to make it +free software which everyone can redistribute and change under these terms. + + To do so, attach the following notices to the program. It is safest +to attach them to the start of each source file to most effectively +state the exclusion of warranty; and each file should have at least +the "copyright" line and a pointer to where the full notice is found. + + + Copyright (C) + + This program is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + This program is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . + +Also add information on how to contact you by electronic and paper mail. + + If the program does terminal interaction, make it output a short +notice like this when it starts in an interactive mode: + + Copyright (C) + This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. + This is free software, and you are welcome to redistribute it + under certain conditions; type `show c' for details. + +The hypothetical commands `show w' and `show c' should show the appropriate +parts of the General Public License. Of course, your program's commands +might be different; for a GUI interface, you would use an "about box". + + You should also get your employer (if you work as a programmer) or school, +if any, to sign a "copyright disclaimer" for the program, if necessary. +For more information on this, and how to apply and follow the GNU GPL, see +. + + The GNU General Public License does not permit incorporating your program +into proprietary programs. If your program is a subroutine library, you +may consider it more useful to permit linking proprietary applications with +the library. If this is what you want to do, use the GNU Lesser General +Public License instead of this License. But first, please read +. diff --git a/yolov5/README.md b/yolov5/README.md new file mode 100644 index 0000000..9537612 --- /dev/null +++ b/yolov5/README.md @@ -0,0 +1,300 @@ +
+

+ + +

+
+
+ CI CPU testing + YOLOv5 Citation + Docker Pulls +
+ Open In Colab + Open In Kaggle + Join Forum +
+ +
+

+YOLOv5 🚀 is a family of object detection architectures and models pretrained on the COCO dataset, and represents Ultralytics + open-source research into future vision AI methods, incorporating lessons learned and best practices evolved over thousands of hours of research and development. +

+ + + + + +
+ +##
Documentation
+ +See the [YOLOv5 Docs](https://docs.ultralytics.com) for full documentation on training, testing and deployment. + +##
Quick Start Examples
+ +
+Install + +Clone repo and install [requirements.txt](https://github.com/ultralytics/yolov5/blob/master/requirements.txt) in a +[**Python>=3.7.0**](https://www.python.org/) environment, including +[**PyTorch>=1.7**](https://pytorch.org/get-started/locally/). + +```bash +git clone https://github.com/ultralytics/yolov5 # clone +cd yolov5 +pip install -r requirements.txt # install +``` + +
+ +
+Inference + +YOLOv5 [PyTorch Hub](https://github.com/ultralytics/yolov5/issues/36) inference. [Models](https://github.com/ultralytics/yolov5/tree/master/models) download automatically from the latest +YOLOv5 [release](https://github.com/ultralytics/yolov5/releases). + +```python +import torch + +# Model +model = torch.hub.load('ultralytics/yolov5', 'yolov5s') # or yolov5n - yolov5x6, custom + +# Images +img = 'https://ultralytics.com/images/zidane.jpg' # or file, Path, PIL, OpenCV, numpy, list + +# Inference +results = model(img) + +# Results +results.print() # or .show(), .save(), .crop(), .pandas(), etc. +``` + +
+ +
+Inference with detect.py + +`detect.py` runs inference on a variety of sources, downloading [models](https://github.com/ultralytics/yolov5/tree/master/models) automatically from +the latest YOLOv5 [release](https://github.com/ultralytics/yolov5/releases) and saving results to `runs/detect`. + +```bash +python detect.py --source 0 # webcam + img.jpg # image + vid.mp4 # video + path/ # directory + path/*.jpg # glob + 'https://youtu.be/Zgi9g1ksQHc' # YouTube + 'rtsp://example.com/media.mp4' # RTSP, RTMP, HTTP stream +``` + +
+ +
+Training + +The commands below reproduce YOLOv5 [COCO](https://github.com/ultralytics/yolov5/blob/master/data/scripts/get_coco.sh) +results. [Models](https://github.com/ultralytics/yolov5/tree/master/models) +and [datasets](https://github.com/ultralytics/yolov5/tree/master/data) download automatically from the latest +YOLOv5 [release](https://github.com/ultralytics/yolov5/releases). Training times for YOLOv5n/s/m/l/x are +1/2/4/6/8 days on a V100 GPU ([Multi-GPU](https://github.com/ultralytics/yolov5/issues/475) times faster). Use the +largest `--batch-size` possible, or pass `--batch-size -1` for +YOLOv5 [AutoBatch](https://github.com/ultralytics/yolov5/pull/5092). Batch sizes shown for V100-16GB. + +```bash +python train.py --data coco.yaml --cfg yolov5n.yaml --weights '' --batch-size 128 + yolov5s 64 + yolov5m 40 + yolov5l 24 + yolov5x 16 +``` + + + +
+ +
+Tutorials + +- [Train Custom Data](https://github.com/ultralytics/yolov5/wiki/Train-Custom-Data)  🚀 RECOMMENDED +- [Tips for Best Training Results](https://github.com/ultralytics/yolov5/wiki/Tips-for-Best-Training-Results)  ☘️ + RECOMMENDED +- [Weights & Biases Logging](https://github.com/ultralytics/yolov5/issues/1289)  🌟 NEW +- [Roboflow for Datasets, Labeling, and Active Learning](https://github.com/ultralytics/yolov5/issues/4975)  🌟 NEW +- [Multi-GPU Training](https://github.com/ultralytics/yolov5/issues/475) +- [PyTorch Hub](https://github.com/ultralytics/yolov5/issues/36)  ⭐ NEW +- [TFLite, ONNX, CoreML, TensorRT Export](https://github.com/ultralytics/yolov5/issues/251) 🚀 +- [Test-Time Augmentation (TTA)](https://github.com/ultralytics/yolov5/issues/303) +- [Model Ensembling](https://github.com/ultralytics/yolov5/issues/318) +- [Model Pruning/Sparsity](https://github.com/ultralytics/yolov5/issues/304) +- [Hyperparameter Evolution](https://github.com/ultralytics/yolov5/issues/607) +- [Transfer Learning with Frozen Layers](https://github.com/ultralytics/yolov5/issues/1314)  ⭐ NEW +- [Architecture Summary](https://github.com/ultralytics/yolov5/issues/6998)  ⭐ NEW + +
+ +##
Environments
+ +Get started in seconds with our verified environments. Click each icon below for details. + + + +##
Integrations
+ + + +|Weights and Biases|Roboflow ⭐ NEW| +|:-:|:-:| +|Automatically track and visualize all your YOLOv5 training runs in the cloud with [Weights & Biases](https://wandb.ai/site?utm_campaign=repo_yolo_readme)|Label and export your custom datasets directly to YOLOv5 for training with [Roboflow](https://roboflow.com/?ref=ultralytics) | + + + +##
Why YOLOv5
+ +

+
+ YOLOv5-P5 640 Figure (click to expand) + +

+
+
+ Figure Notes (click to expand) + +- **COCO AP val** denotes mAP@0.5:0.95 metric measured on the 5000-image [COCO val2017](http://cocodataset.org) dataset over various inference sizes from 256 to 1536. +- **GPU Speed** measures average inference time per image on [COCO val2017](http://cocodataset.org) dataset using a [AWS p3.2xlarge](https://aws.amazon.com/ec2/instance-types/p3/) V100 instance at batch-size 32. +- **EfficientDet** data from [google/automl](https://github.com/google/automl) at batch size 8. +- **Reproduce** by `python val.py --task study --data coco.yaml --iou 0.7 --weights yolov5n6.pt yolov5s6.pt yolov5m6.pt yolov5l6.pt yolov5x6.pt` + +
+ +### Pretrained Checkpoints + +|Model |size
(pixels) |mAPval
0.5:0.95 |mAPval
0.5 |Speed
CPU b1
(ms) |Speed
V100 b1
(ms) |Speed
V100 b32
(ms) |params
(M) |FLOPs
@640 (B) +|--- |--- |--- |--- |--- |--- |--- |--- |--- +|[YOLOv5n][assets] |640 |28.0 |45.7 |**45** |**6.3**|**0.6**|**1.9**|**4.5** +|[YOLOv5s][assets] |640 |37.4 |56.8 |98 |6.4 |0.9 |7.2 |16.5 +|[YOLOv5m][assets] |640 |45.4 |64.1 |224 |8.2 |1.7 |21.2 |49.0 +|[YOLOv5l][assets] |640 |49.0 |67.3 |430 |10.1 |2.7 |46.5 |109.1 +|[YOLOv5x][assets] |640 |50.7 |68.9 |766 |12.1 |4.8 |86.7 |205.7 +| | | | | | | | | +|[YOLOv5n6][assets] |1280 |36.0 |54.4 |153 |8.1 |2.1 |3.2 |4.6 +|[YOLOv5s6][assets] |1280 |44.8 |63.7 |385 |8.2 |3.6 |12.6 |16.8 +|[YOLOv5m6][assets] |1280 |51.3 |69.3 |887 |11.1 |6.8 |35.7 |50.0 +|[YOLOv5l6][assets] |1280 |53.7 |71.3 |1784 |15.8 |10.5 |76.8 |111.4 +|[YOLOv5x6][assets]
+ [TTA][TTA]|1280
1536 |55.0
**55.8** |72.7
**72.7** |3136
- |26.2
- |19.4
- |140.7
- |209.8
- + +
+ Table Notes (click to expand) + +- All checkpoints are trained to 300 epochs with default settings. Nano and Small models use [hyp.scratch-low.yaml](https://github.com/ultralytics/yolov5/blob/master/data/hyps/hyp.scratch-low.yaml) hyps, all others use [hyp.scratch-high.yaml](https://github.com/ultralytics/yolov5/blob/master/data/hyps/hyp.scratch-high.yaml). +- **mAPval** values are for single-model single-scale on [COCO val2017](http://cocodataset.org) dataset.
Reproduce by `python val.py --data coco.yaml --img 640 --conf 0.001 --iou 0.65` +- **Speed** averaged over COCO val images using a [AWS p3.2xlarge](https://aws.amazon.com/ec2/instance-types/p3/) instance. NMS times (~1 ms/img) not included.
Reproduce by `python val.py --data coco.yaml --img 640 --task speed --batch 1` +- **TTA** [Test Time Augmentation](https://github.com/ultralytics/yolov5/issues/303) includes reflection and scale augmentations.
Reproduce by `python val.py --data coco.yaml --img 1536 --iou 0.7 --augment` + +
+ +##
Contribute
+ +We love your input! We want to make contributing to YOLOv5 as easy and transparent as possible. Please see our [Contributing Guide](CONTRIBUTING.md) to get started, and fill out the [YOLOv5 Survey](https://ultralytics.com/survey?utm_source=github&utm_medium=social&utm_campaign=Survey) to send us feedback on your experiences. Thank you to all our contributors! + + + +##
Contact
+ +For YOLOv5 bugs and feature requests please visit [GitHub Issues](https://github.com/ultralytics/yolov5/issues). For business inquiries or +professional support requests please visit [https://ultralytics.com/contact](https://ultralytics.com/contact). + +
+ + + +[assets]: https://github.com/ultralytics/yolov5/releases +[tta]: https://github.com/ultralytics/yolov5/issues/303 diff --git a/yolov5/data/Argoverse.yaml b/yolov5/data/Argoverse.yaml new file mode 100644 index 0000000..9d21296 --- /dev/null +++ b/yolov5/data/Argoverse.yaml @@ -0,0 +1,67 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# Argoverse-HD dataset (ring-front-center camera) http://www.cs.cmu.edu/~mengtial/proj/streaming/ by Argo AI +# Example usage: python train.py --data Argoverse.yaml +# parent +# ├── yolov5 +# └── datasets +# └── Argoverse ← downloads here (31.3 GB) + + +# Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..] +path: ../datasets/Argoverse # dataset root dir +train: Argoverse-1.1/images/train/ # train images (relative to 'path') 39384 images +val: Argoverse-1.1/images/val/ # val images (relative to 'path') 15062 images +test: Argoverse-1.1/images/test/ # test images (optional) https://eval.ai/web/challenges/challenge-page/800/overview + +# Classes +nc: 8 # number of classes +names: ['person', 'bicycle', 'car', 'motorcycle', 'bus', 'truck', 'traffic_light', 'stop_sign'] # class names + + +# Download script/URL (optional) --------------------------------------------------------------------------------------- +download: | + import json + + from tqdm import tqdm + from utils.general import download, Path + + + def argoverse2yolo(set): + labels = {} + a = json.load(open(set, "rb")) + for annot in tqdm(a['annotations'], desc=f"Converting {set} to YOLOv5 format..."): + img_id = annot['image_id'] + img_name = a['images'][img_id]['name'] + img_label_name = f'{img_name[:-3]}txt' + + cls = annot['category_id'] # instance class id + x_center, y_center, width, height = annot['bbox'] + x_center = (x_center + width / 2) / 1920.0 # offset and scale + y_center = (y_center + height / 2) / 1200.0 # offset and scale + width /= 1920.0 # scale + height /= 1200.0 # scale + + img_dir = set.parents[2] / 'Argoverse-1.1' / 'labels' / a['seq_dirs'][a['images'][annot['image_id']]['sid']] + if not img_dir.exists(): + img_dir.mkdir(parents=True, exist_ok=True) + + k = str(img_dir / img_label_name) + if k not in labels: + labels[k] = [] + labels[k].append(f"{cls} {x_center} {y_center} {width} {height}\n") + + for k in labels: + with open(k, "w") as f: + f.writelines(labels[k]) + + + # Download + dir = Path('../datasets/Argoverse') # dataset root dir + urls = ['https://argoverse-hd.s3.us-east-2.amazonaws.com/Argoverse-HD-Full.zip'] + download(urls, dir=dir, delete=False) + + # Convert + annotations_dir = 'Argoverse-HD/annotations/' + (dir / 'Argoverse-1.1' / 'tracking').rename(dir / 'Argoverse-1.1' / 'images') # rename 'tracking' to 'images' + for d in "train.json", "val.json": + argoverse2yolo(dir / annotations_dir / d) # convert VisDrone annotations to YOLO labels diff --git a/yolov5/data/GlobalWheat2020.yaml b/yolov5/data/GlobalWheat2020.yaml new file mode 100644 index 0000000..4c43693 --- /dev/null +++ b/yolov5/data/GlobalWheat2020.yaml @@ -0,0 +1,54 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# Global Wheat 2020 dataset http://www.global-wheat.com/ by University of Saskatchewan +# Example usage: python train.py --data GlobalWheat2020.yaml +# parent +# ├── yolov5 +# └── datasets +# └── GlobalWheat2020 ← downloads here (7.0 GB) + + +# Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..] +path: ../datasets/GlobalWheat2020 # dataset root dir +train: # train images (relative to 'path') 3422 images + - images/arvalis_1 + - images/arvalis_2 + - images/arvalis_3 + - images/ethz_1 + - images/rres_1 + - images/inrae_1 + - images/usask_1 +val: # val images (relative to 'path') 748 images (WARNING: train set contains ethz_1) + - images/ethz_1 +test: # test images (optional) 1276 images + - images/utokyo_1 + - images/utokyo_2 + - images/nau_1 + - images/uq_1 + +# Classes +nc: 1 # number of classes +names: ['wheat_head'] # class names + + +# Download script/URL (optional) --------------------------------------------------------------------------------------- +download: | + from utils.general import download, Path + + + # Download + dir = Path(yaml['path']) # dataset root dir + urls = ['https://zenodo.org/record/4298502/files/global-wheat-codalab-official.zip', + 'https://github.com/ultralytics/yolov5/releases/download/v1.0/GlobalWheat2020_labels.zip'] + download(urls, dir=dir) + + # Make Directories + for p in 'annotations', 'images', 'labels': + (dir / p).mkdir(parents=True, exist_ok=True) + + # Move + for p in 'arvalis_1', 'arvalis_2', 'arvalis_3', 'ethz_1', 'rres_1', 'inrae_1', 'usask_1', \ + 'utokyo_1', 'utokyo_2', 'nau_1', 'uq_1': + (dir / p).rename(dir / 'images' / p) # move to /images + f = (dir / p).with_suffix('.json') # json file + if f.exists(): + f.rename((dir / 'annotations' / p).with_suffix('.json')) # move to /annotations diff --git a/yolov5/data/Objects365.yaml b/yolov5/data/Objects365.yaml new file mode 100644 index 0000000..4cc9475 --- /dev/null +++ b/yolov5/data/Objects365.yaml @@ -0,0 +1,114 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# Objects365 dataset https://www.objects365.org/ by Megvii +# Example usage: python train.py --data Objects365.yaml +# parent +# ├── yolov5 +# └── datasets +# └── Objects365 ← downloads here (712 GB = 367G data + 345G zips) + + +# Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..] +path: ../datasets/Objects365 # dataset root dir +train: images/train # train images (relative to 'path') 1742289 images +val: images/val # val images (relative to 'path') 80000 images +test: # test images (optional) + +# Classes +nc: 365 # number of classes +names: ['Person', 'Sneakers', 'Chair', 'Other Shoes', 'Hat', 'Car', 'Lamp', 'Glasses', 'Bottle', 'Desk', 'Cup', + 'Street Lights', 'Cabinet/shelf', 'Handbag/Satchel', 'Bracelet', 'Plate', 'Picture/Frame', 'Helmet', 'Book', + 'Gloves', 'Storage box', 'Boat', 'Leather Shoes', 'Flower', 'Bench', 'Potted Plant', 'Bowl/Basin', 'Flag', + 'Pillow', 'Boots', 'Vase', 'Microphone', 'Necklace', 'Ring', 'SUV', 'Wine Glass', 'Belt', 'Monitor/TV', + 'Backpack', 'Umbrella', 'Traffic Light', 'Speaker', 'Watch', 'Tie', 'Trash bin Can', 'Slippers', 'Bicycle', + 'Stool', 'Barrel/bucket', 'Van', 'Couch', 'Sandals', 'Basket', 'Drum', 'Pen/Pencil', 'Bus', 'Wild Bird', + 'High Heels', 'Motorcycle', 'Guitar', 'Carpet', 'Cell Phone', 'Bread', 'Camera', 'Canned', 'Truck', + 'Traffic cone', 'Cymbal', 'Lifesaver', 'Towel', 'Stuffed Toy', 'Candle', 'Sailboat', 'Laptop', 'Awning', + 'Bed', 'Faucet', 'Tent', 'Horse', 'Mirror', 'Power outlet', 'Sink', 'Apple', 'Air Conditioner', 'Knife', + 'Hockey Stick', 'Paddle', 'Pickup Truck', 'Fork', 'Traffic Sign', 'Balloon', 'Tripod', 'Dog', 'Spoon', 'Clock', + 'Pot', 'Cow', 'Cake', 'Dinning Table', 'Sheep', 'Hanger', 'Blackboard/Whiteboard', 'Napkin', 'Other Fish', + 'Orange/Tangerine', 'Toiletry', 'Keyboard', 'Tomato', 'Lantern', 'Machinery Vehicle', 'Fan', + 'Green Vegetables', 'Banana', 'Baseball Glove', 'Airplane', 'Mouse', 'Train', 'Pumpkin', 'Soccer', 'Skiboard', + 'Luggage', 'Nightstand', 'Tea pot', 'Telephone', 'Trolley', 'Head Phone', 'Sports Car', 'Stop Sign', + 'Dessert', 'Scooter', 'Stroller', 'Crane', 'Remote', 'Refrigerator', 'Oven', 'Lemon', 'Duck', 'Baseball Bat', + 'Surveillance Camera', 'Cat', 'Jug', 'Broccoli', 'Piano', 'Pizza', 'Elephant', 'Skateboard', 'Surfboard', + 'Gun', 'Skating and Skiing shoes', 'Gas stove', 'Donut', 'Bow Tie', 'Carrot', 'Toilet', 'Kite', 'Strawberry', + 'Other Balls', 'Shovel', 'Pepper', 'Computer Box', 'Toilet Paper', 'Cleaning Products', 'Chopsticks', + 'Microwave', 'Pigeon', 'Baseball', 'Cutting/chopping Board', 'Coffee Table', 'Side Table', 'Scissors', + 'Marker', 'Pie', 'Ladder', 'Snowboard', 'Cookies', 'Radiator', 'Fire Hydrant', 'Basketball', 'Zebra', 'Grape', + 'Giraffe', 'Potato', 'Sausage', 'Tricycle', 'Violin', 'Egg', 'Fire Extinguisher', 'Candy', 'Fire Truck', + 'Billiards', 'Converter', 'Bathtub', 'Wheelchair', 'Golf Club', 'Briefcase', 'Cucumber', 'Cigar/Cigarette', + 'Paint Brush', 'Pear', 'Heavy Truck', 'Hamburger', 'Extractor', 'Extension Cord', 'Tong', 'Tennis Racket', + 'Folder', 'American Football', 'earphone', 'Mask', 'Kettle', 'Tennis', 'Ship', 'Swing', 'Coffee Machine', + 'Slide', 'Carriage', 'Onion', 'Green beans', 'Projector', 'Frisbee', 'Washing Machine/Drying Machine', + 'Chicken', 'Printer', 'Watermelon', 'Saxophone', 'Tissue', 'Toothbrush', 'Ice cream', 'Hot-air balloon', + 'Cello', 'French Fries', 'Scale', 'Trophy', 'Cabbage', 'Hot dog', 'Blender', 'Peach', 'Rice', 'Wallet/Purse', + 'Volleyball', 'Deer', 'Goose', 'Tape', 'Tablet', 'Cosmetics', 'Trumpet', 'Pineapple', 'Golf Ball', + 'Ambulance', 'Parking meter', 'Mango', 'Key', 'Hurdle', 'Fishing Rod', 'Medal', 'Flute', 'Brush', 'Penguin', + 'Megaphone', 'Corn', 'Lettuce', 'Garlic', 'Swan', 'Helicopter', 'Green Onion', 'Sandwich', 'Nuts', + 'Speed Limit Sign', 'Induction Cooker', 'Broom', 'Trombone', 'Plum', 'Rickshaw', 'Goldfish', 'Kiwi fruit', + 'Router/modem', 'Poker Card', 'Toaster', 'Shrimp', 'Sushi', 'Cheese', 'Notepaper', 'Cherry', 'Pliers', 'CD', + 'Pasta', 'Hammer', 'Cue', 'Avocado', 'Hamimelon', 'Flask', 'Mushroom', 'Screwdriver', 'Soap', 'Recorder', + 'Bear', 'Eggplant', 'Board Eraser', 'Coconut', 'Tape Measure/Ruler', 'Pig', 'Showerhead', 'Globe', 'Chips', + 'Steak', 'Crosswalk Sign', 'Stapler', 'Camel', 'Formula 1', 'Pomegranate', 'Dishwasher', 'Crab', + 'Hoverboard', 'Meat ball', 'Rice Cooker', 'Tuba', 'Calculator', 'Papaya', 'Antelope', 'Parrot', 'Seal', + 'Butterfly', 'Dumbbell', 'Donkey', 'Lion', 'Urinal', 'Dolphin', 'Electric Drill', 'Hair Dryer', 'Egg tart', + 'Jellyfish', 'Treadmill', 'Lighter', 'Grapefruit', 'Game board', 'Mop', 'Radish', 'Baozi', 'Target', 'French', + 'Spring Rolls', 'Monkey', 'Rabbit', 'Pencil Case', 'Yak', 'Red Cabbage', 'Binoculars', 'Asparagus', 'Barbell', + 'Scallop', 'Noddles', 'Comb', 'Dumpling', 'Oyster', 'Table Tennis paddle', 'Cosmetics Brush/Eyeliner Pencil', + 'Chainsaw', 'Eraser', 'Lobster', 'Durian', 'Okra', 'Lipstick', 'Cosmetics Mirror', 'Curling', 'Table Tennis'] + + +# Download script/URL (optional) --------------------------------------------------------------------------------------- +download: | + from tqdm import tqdm + + from utils.general import Path, check_requirements, download, np, xyxy2xywhn + + check_requirements(('pycocotools>=2.0',)) + from pycocotools.coco import COCO + + # Make Directories + dir = Path(yaml['path']) # dataset root dir + for p in 'images', 'labels': + (dir / p).mkdir(parents=True, exist_ok=True) + for q in 'train', 'val': + (dir / p / q).mkdir(parents=True, exist_ok=True) + + # Train, Val Splits + for split, patches in [('train', 50 + 1), ('val', 43 + 1)]: + print(f"Processing {split} in {patches} patches ...") + images, labels = dir / 'images' / split, dir / 'labels' / split + + # Download + url = f"https://dorc.ks3-cn-beijing.ksyun.com/data-set/2020Objects365%E6%95%B0%E6%8D%AE%E9%9B%86/{split}/" + if split == 'train': + download([f'{url}zhiyuan_objv2_{split}.tar.gz'], dir=dir, delete=False) # annotations json + download([f'{url}patch{i}.tar.gz' for i in range(patches)], dir=images, curl=True, delete=False, threads=8) + elif split == 'val': + download([f'{url}zhiyuan_objv2_{split}.json'], dir=dir, delete=False) # annotations json + download([f'{url}images/v1/patch{i}.tar.gz' for i in range(15 + 1)], dir=images, curl=True, delete=False, threads=8) + download([f'{url}images/v2/patch{i}.tar.gz' for i in range(16, patches)], dir=images, curl=True, delete=False, threads=8) + + # Move + for f in tqdm(images.rglob('*.jpg'), desc=f'Moving {split} images'): + f.rename(images / f.name) # move to /images/{split} + + # Labels + coco = COCO(dir / f'zhiyuan_objv2_{split}.json') + names = [x["name"] for x in coco.loadCats(coco.getCatIds())] + for cid, cat in enumerate(names): + catIds = coco.getCatIds(catNms=[cat]) + imgIds = coco.getImgIds(catIds=catIds) + for im in tqdm(coco.loadImgs(imgIds), desc=f'Class {cid + 1}/{len(names)} {cat}'): + width, height = im["width"], im["height"] + path = Path(im["file_name"]) # image filename + try: + with open(labels / path.with_suffix('.txt').name, 'a') as file: + annIds = coco.getAnnIds(imgIds=im["id"], catIds=catIds, iscrowd=None) + for a in coco.loadAnns(annIds): + x, y, w, h = a['bbox'] # bounding box in xywh (xy top-left corner) + xyxy = np.array([x, y, x + w, y + h])[None] # pixels(1,4) + x, y, w, h = xyxy2xywhn(xyxy, w=width, h=height, clip=True)[0] # normalized and clipped + file.write(f"{cid} {x:.5f} {y:.5f} {w:.5f} {h:.5f}\n") + except Exception as e: + print(e) diff --git a/yolov5/data/SKU-110K.yaml b/yolov5/data/SKU-110K.yaml new file mode 100644 index 0000000..2acf34d --- /dev/null +++ b/yolov5/data/SKU-110K.yaml @@ -0,0 +1,53 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# SKU-110K retail items dataset https://github.com/eg4000/SKU110K_CVPR19 by Trax Retail +# Example usage: python train.py --data SKU-110K.yaml +# parent +# ├── yolov5 +# └── datasets +# └── SKU-110K ← downloads here (13.6 GB) + + +# Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..] +path: ../datasets/SKU-110K # dataset root dir +train: train.txt # train images (relative to 'path') 8219 images +val: val.txt # val images (relative to 'path') 588 images +test: test.txt # test images (optional) 2936 images + +# Classes +nc: 1 # number of classes +names: ['object'] # class names + + +# Download script/URL (optional) --------------------------------------------------------------------------------------- +download: | + import shutil + from tqdm import tqdm + from utils.general import np, pd, Path, download, xyxy2xywh + + + # Download + dir = Path(yaml['path']) # dataset root dir + parent = Path(dir.parent) # download dir + urls = ['http://trax-geometry.s3.amazonaws.com/cvpr_challenge/SKU110K_fixed.tar.gz'] + download(urls, dir=parent, delete=False) + + # Rename directories + if dir.exists(): + shutil.rmtree(dir) + (parent / 'SKU110K_fixed').rename(dir) # rename dir + (dir / 'labels').mkdir(parents=True, exist_ok=True) # create labels dir + + # Convert labels + names = 'image', 'x1', 'y1', 'x2', 'y2', 'class', 'image_width', 'image_height' # column names + for d in 'annotations_train.csv', 'annotations_val.csv', 'annotations_test.csv': + x = pd.read_csv(dir / 'annotations' / d, names=names).values # annotations + images, unique_images = x[:, 0], np.unique(x[:, 0]) + with open((dir / d).with_suffix('.txt').__str__().replace('annotations_', ''), 'w') as f: + f.writelines(f'./images/{s}\n' for s in unique_images) + for im in tqdm(unique_images, desc=f'Converting {dir / d}'): + cls = 0 # single-class dataset + with open((dir / 'labels' / im).with_suffix('.txt'), 'a') as f: + for r in x[images == im]: + w, h = r[6], r[7] # image width, height + xywh = xyxy2xywh(np.array([[r[1] / w, r[2] / h, r[3] / w, r[4] / h]]))[0] # instance + f.write(f"{cls} {xywh[0]:.5f} {xywh[1]:.5f} {xywh[2]:.5f} {xywh[3]:.5f}\n") # write label diff --git a/yolov5/data/VOC.yaml b/yolov5/data/VOC.yaml new file mode 100644 index 0000000..636ddc4 --- /dev/null +++ b/yolov5/data/VOC.yaml @@ -0,0 +1,81 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# PASCAL VOC dataset http://host.robots.ox.ac.uk/pascal/VOC by University of Oxford +# Example usage: python train.py --data VOC.yaml +# parent +# ├── yolov5 +# └── datasets +# └── VOC ← downloads here (2.8 GB) + + +# Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..] +path: ../datasets/VOC +train: # train images (relative to 'path') 16551 images + - images/train2012 + - images/train2007 + - images/val2012 + - images/val2007 +val: # val images (relative to 'path') 4952 images + - images/test2007 +test: # test images (optional) + - images/test2007 + +# Classes +nc: 20 # number of classes +names: ['aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car', 'cat', 'chair', 'cow', 'diningtable', 'dog', + 'horse', 'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 'train', 'tvmonitor'] # class names + + +# Download script/URL (optional) --------------------------------------------------------------------------------------- +download: | + import xml.etree.ElementTree as ET + + from tqdm import tqdm + from utils.general import download, Path + + + def convert_label(path, lb_path, year, image_id): + def convert_box(size, box): + dw, dh = 1. / size[0], 1. / size[1] + x, y, w, h = (box[0] + box[1]) / 2.0 - 1, (box[2] + box[3]) / 2.0 - 1, box[1] - box[0], box[3] - box[2] + return x * dw, y * dh, w * dw, h * dh + + in_file = open(path / f'VOC{year}/Annotations/{image_id}.xml') + out_file = open(lb_path, 'w') + tree = ET.parse(in_file) + root = tree.getroot() + size = root.find('size') + w = int(size.find('width').text) + h = int(size.find('height').text) + + for obj in root.iter('object'): + cls = obj.find('name').text + if cls in yaml['names'] and not int(obj.find('difficult').text) == 1: + xmlbox = obj.find('bndbox') + bb = convert_box((w, h), [float(xmlbox.find(x).text) for x in ('xmin', 'xmax', 'ymin', 'ymax')]) + cls_id = yaml['names'].index(cls) # class id + out_file.write(" ".join([str(a) for a in (cls_id, *bb)]) + '\n') + + + # Download + dir = Path(yaml['path']) # dataset root dir + url = 'https://github.com/ultralytics/yolov5/releases/download/v1.0/' + urls = [f'{url}VOCtrainval_06-Nov-2007.zip', # 446MB, 5012 images + f'{url}VOCtest_06-Nov-2007.zip', # 438MB, 4953 images + f'{url}VOCtrainval_11-May-2012.zip'] # 1.95GB, 17126 images + download(urls, dir=dir / 'images', delete=False, curl=True, threads=3) + + # Convert + path = dir / 'images/VOCdevkit' + for year, image_set in ('2012', 'train'), ('2012', 'val'), ('2007', 'train'), ('2007', 'val'), ('2007', 'test'): + imgs_path = dir / 'images' / f'{image_set}{year}' + lbs_path = dir / 'labels' / f'{image_set}{year}' + imgs_path.mkdir(exist_ok=True, parents=True) + lbs_path.mkdir(exist_ok=True, parents=True) + + with open(path / f'VOC{year}/ImageSets/Main/{image_set}.txt') as f: + image_ids = f.read().strip().split() + for id in tqdm(image_ids, desc=f'{image_set}{year}'): + f = path / f'VOC{year}/JPEGImages/{id}.jpg' # old img path + lb_path = (lbs_path / f.name).with_suffix('.txt') # new label path + f.rename(imgs_path / f.name) # move image + convert_label(path, lb_path, year, id) # convert labels to YOLO format diff --git a/yolov5/data/VisDrone.yaml b/yolov5/data/VisDrone.yaml new file mode 100644 index 0000000..10337b4 --- /dev/null +++ b/yolov5/data/VisDrone.yaml @@ -0,0 +1,61 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# VisDrone2019-DET dataset https://github.com/VisDrone/VisDrone-Dataset by Tianjin University +# Example usage: python train.py --data VisDrone.yaml +# parent +# ├── yolov5 +# └── datasets +# └── VisDrone ← downloads here (2.3 GB) + + +# Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..] +path: ../datasets/VisDrone # dataset root dir +train: VisDrone2019-DET-train/images # train images (relative to 'path') 6471 images +val: VisDrone2019-DET-val/images # val images (relative to 'path') 548 images +test: VisDrone2019-DET-test-dev/images # test images (optional) 1610 images + +# Classes +nc: 10 # number of classes +names: ['pedestrian', 'people', 'bicycle', 'car', 'van', 'truck', 'tricycle', 'awning-tricycle', 'bus', 'motor'] + + +# Download script/URL (optional) --------------------------------------------------------------------------------------- +download: | + from utils.general import download, os, Path + + def visdrone2yolo(dir): + from PIL import Image + from tqdm import tqdm + + def convert_box(size, box): + # Convert VisDrone box to YOLO xywh box + dw = 1. / size[0] + dh = 1. / size[1] + return (box[0] + box[2] / 2) * dw, (box[1] + box[3] / 2) * dh, box[2] * dw, box[3] * dh + + (dir / 'labels').mkdir(parents=True, exist_ok=True) # make labels directory + pbar = tqdm((dir / 'annotations').glob('*.txt'), desc=f'Converting {dir}') + for f in pbar: + img_size = Image.open((dir / 'images' / f.name).with_suffix('.jpg')).size + lines = [] + with open(f, 'r') as file: # read annotation.txt + for row in [x.split(',') for x in file.read().strip().splitlines()]: + if row[4] == '0': # VisDrone 'ignored regions' class 0 + continue + cls = int(row[5]) - 1 + box = convert_box(img_size, tuple(map(int, row[:4]))) + lines.append(f"{cls} {' '.join(f'{x:.6f}' for x in box)}\n") + with open(str(f).replace(os.sep + 'annotations' + os.sep, os.sep + 'labels' + os.sep), 'w') as fl: + fl.writelines(lines) # write label.txt + + + # Download + dir = Path(yaml['path']) # dataset root dir + urls = ['https://github.com/ultralytics/yolov5/releases/download/v1.0/VisDrone2019-DET-train.zip', + 'https://github.com/ultralytics/yolov5/releases/download/v1.0/VisDrone2019-DET-val.zip', + 'https://github.com/ultralytics/yolov5/releases/download/v1.0/VisDrone2019-DET-test-dev.zip', + 'https://github.com/ultralytics/yolov5/releases/download/v1.0/VisDrone2019-DET-test-challenge.zip'] + download(urls, dir=dir, curl=True, threads=4) + + # Convert + for d in 'VisDrone2019-DET-train', 'VisDrone2019-DET-val', 'VisDrone2019-DET-test-dev': + visdrone2yolo(dir / d) # convert VisDrone annotations to YOLO labels diff --git a/yolov5/data/coco.yaml b/yolov5/data/coco.yaml new file mode 100644 index 0000000..0c0c4ad --- /dev/null +++ b/yolov5/data/coco.yaml @@ -0,0 +1,45 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# COCO 2017 dataset http://cocodataset.org by Microsoft +# Example usage: python train.py --data coco.yaml +# parent +# ├── yolov5 +# └── datasets +# └── coco ← downloads here (20.1 GB) + + +# Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..] +path: ../datasets/coco # dataset root dir +train: train2017.txt # train images (relative to 'path') 118287 images +val: val2017.txt # val images (relative to 'path') 5000 images +test: test-dev2017.txt # 20288 of 40670 images, submit to https://competitions.codalab.org/competitions/20794 + +# Classes +nc: 80 # number of classes +names: ['person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light', + 'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow', + 'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee', + 'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard', + 'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', + 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch', + 'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone', + 'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear', + 'hair drier', 'toothbrush'] # class names + + +# Download script/URL (optional) +download: | + from utils.general import download, Path + + + # Download labels + segments = False # segment or box labels + dir = Path(yaml['path']) # dataset root dir + url = 'https://github.com/ultralytics/yolov5/releases/download/v1.0/' + urls = [url + ('coco2017labels-segments.zip' if segments else 'coco2017labels.zip')] # labels + download(urls, dir=dir.parent) + + # Download data + urls = ['http://images.cocodataset.org/zips/train2017.zip', # 19G, 118k images + 'http://images.cocodataset.org/zips/val2017.zip', # 1G, 5k images + 'http://images.cocodataset.org/zips/test2017.zip'] # 7G, 41k images (optional) + download(urls, dir=dir / 'images', threads=3) diff --git a/yolov5/data/coco128.yaml b/yolov5/data/coco128.yaml new file mode 100644 index 0000000..2517d20 --- /dev/null +++ b/yolov5/data/coco128.yaml @@ -0,0 +1,30 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# COCO128 dataset https://www.kaggle.com/ultralytics/coco128 (first 128 images from COCO train2017) by Ultralytics +# Example usage: python train.py --data coco128.yaml +# parent +# ├── yolov5 +# └── datasets +# └── coco128 ← downloads here (7 MB) + + +# Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..] +path: ../datasets/coco128 # dataset root dir +train: images/train2017 # train images (relative to 'path') 128 images +val: images/train2017 # val images (relative to 'path') 128 images +test: # test images (optional) + +# Classes +nc: 80 # number of classes +names: ['person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light', + 'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow', + 'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee', + 'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard', + 'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', + 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch', + 'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone', + 'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear', + 'hair drier', 'toothbrush'] # class names + + +# Download script/URL (optional) +download: https://ultralytics.com/assets/coco128.zip diff --git a/yolov5/data/hyps/hyp.Objects365.yaml b/yolov5/data/hyps/hyp.Objects365.yaml new file mode 100644 index 0000000..7497174 --- /dev/null +++ b/yolov5/data/hyps/hyp.Objects365.yaml @@ -0,0 +1,34 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# Hyperparameters for Objects365 training +# python train.py --weights yolov5m.pt --data Objects365.yaml --evolve +# See Hyperparameter Evolution tutorial for details https://github.com/ultralytics/yolov5#tutorials + +lr0: 0.00258 +lrf: 0.17 +momentum: 0.779 +weight_decay: 0.00058 +warmup_epochs: 1.33 +warmup_momentum: 0.86 +warmup_bias_lr: 0.0711 +box: 0.0539 +cls: 0.299 +cls_pw: 0.825 +obj: 0.632 +obj_pw: 1.0 +iou_t: 0.2 +anchor_t: 3.44 +anchors: 3.2 +fl_gamma: 0.0 +hsv_h: 0.0188 +hsv_s: 0.704 +hsv_v: 0.36 +degrees: 0.0 +translate: 0.0902 +scale: 0.491 +shear: 0.0 +perspective: 0.0 +flipud: 0.0 +fliplr: 0.5 +mosaic: 1.0 +mixup: 0.0 +copy_paste: 0.0 diff --git a/yolov5/data/hyps/hyp.VOC.yaml b/yolov5/data/hyps/hyp.VOC.yaml new file mode 100644 index 0000000..0aa4e7d --- /dev/null +++ b/yolov5/data/hyps/hyp.VOC.yaml @@ -0,0 +1,40 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# Hyperparameters for VOC training +# python train.py --batch 128 --weights yolov5m6.pt --data VOC.yaml --epochs 50 --img 512 --hyp hyp.scratch-med.yaml --evolve +# See Hyperparameter Evolution tutorial for details https://github.com/ultralytics/yolov5#tutorials + +# YOLOv5 Hyperparameter Evolution Results +# Best generation: 467 +# Last generation: 996 +# metrics/precision, metrics/recall, metrics/mAP_0.5, metrics/mAP_0.5:0.95, val/box_loss, val/obj_loss, val/cls_loss +# 0.87729, 0.85125, 0.91286, 0.72664, 0.0076739, 0.0042529, 0.0013865 + +lr0: 0.00334 +lrf: 0.15135 +momentum: 0.74832 +weight_decay: 0.00025 +warmup_epochs: 3.3835 +warmup_momentum: 0.59462 +warmup_bias_lr: 0.18657 +box: 0.02 +cls: 0.21638 +cls_pw: 0.5 +obj: 0.51728 +obj_pw: 0.67198 +iou_t: 0.2 +anchor_t: 3.3744 +fl_gamma: 0.0 +hsv_h: 0.01041 +hsv_s: 0.54703 +hsv_v: 0.27739 +degrees: 0.0 +translate: 0.04591 +scale: 0.75544 +shear: 0.0 +perspective: 0.0 +flipud: 0.0 +fliplr: 0.5 +mosaic: 0.85834 +mixup: 0.04266 +copy_paste: 0.0 +anchors: 3.412 diff --git a/yolov5/data/hyps/hyp.scratch-high.yaml b/yolov5/data/hyps/hyp.scratch-high.yaml new file mode 100644 index 0000000..123cc84 --- /dev/null +++ b/yolov5/data/hyps/hyp.scratch-high.yaml @@ -0,0 +1,34 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# Hyperparameters for high-augmentation COCO training from scratch +# python train.py --batch 32 --cfg yolov5m6.yaml --weights '' --data coco.yaml --img 1280 --epochs 300 +# See tutorials for hyperparameter evolution https://github.com/ultralytics/yolov5#tutorials + +lr0: 0.01 # initial learning rate (SGD=1E-2, Adam=1E-3) +lrf: 0.1 # final OneCycleLR learning rate (lr0 * lrf) +momentum: 0.937 # SGD momentum/Adam beta1 +weight_decay: 0.0005 # optimizer weight decay 5e-4 +warmup_epochs: 3.0 # warmup epochs (fractions ok) +warmup_momentum: 0.8 # warmup initial momentum +warmup_bias_lr: 0.1 # warmup initial bias lr +box: 0.05 # box loss gain +cls: 0.3 # cls loss gain +cls_pw: 1.0 # cls BCELoss positive_weight +obj: 0.7 # obj loss gain (scale with pixels) +obj_pw: 1.0 # obj BCELoss positive_weight +iou_t: 0.20 # IoU training threshold +anchor_t: 4.0 # anchor-multiple threshold +# anchors: 3 # anchors per output layer (0 to ignore) +fl_gamma: 0.0 # focal loss gamma (efficientDet default gamma=1.5) +hsv_h: 0.015 # image HSV-Hue augmentation (fraction) +hsv_s: 0.7 # image HSV-Saturation augmentation (fraction) +hsv_v: 0.4 # image HSV-Value augmentation (fraction) +degrees: 0.0 # image rotation (+/- deg) +translate: 0.1 # image translation (+/- fraction) +scale: 0.9 # image scale (+/- gain) +shear: 0.0 # image shear (+/- deg) +perspective: 0.0 # image perspective (+/- fraction), range 0-0.001 +flipud: 0.0 # image flip up-down (probability) +fliplr: 0.5 # image flip left-right (probability) +mosaic: 1.0 # image mosaic (probability) +mixup: 0.1 # image mixup (probability) +copy_paste: 0.1 # segment copy-paste (probability) diff --git a/yolov5/data/hyps/hyp.scratch-low.yaml b/yolov5/data/hyps/hyp.scratch-low.yaml new file mode 100644 index 0000000..b9ef1d5 --- /dev/null +++ b/yolov5/data/hyps/hyp.scratch-low.yaml @@ -0,0 +1,34 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# Hyperparameters for low-augmentation COCO training from scratch +# python train.py --batch 64 --cfg yolov5n6.yaml --weights '' --data coco.yaml --img 640 --epochs 300 --linear +# See tutorials for hyperparameter evolution https://github.com/ultralytics/yolov5#tutorials + +lr0: 0.01 # initial learning rate (SGD=1E-2, Adam=1E-3) +lrf: 0.01 # final OneCycleLR learning rate (lr0 * lrf) +momentum: 0.937 # SGD momentum/Adam beta1 +weight_decay: 0.0005 # optimizer weight decay 5e-4 +warmup_epochs: 3.0 # warmup epochs (fractions ok) +warmup_momentum: 0.8 # warmup initial momentum +warmup_bias_lr: 0.1 # warmup initial bias lr +box: 0.05 # box loss gain +cls: 0.5 # cls loss gain +cls_pw: 1.0 # cls BCELoss positive_weight +obj: 1.0 # obj loss gain (scale with pixels) +obj_pw: 1.0 # obj BCELoss positive_weight +iou_t: 0.20 # IoU training threshold +anchor_t: 4.0 # anchor-multiple threshold +# anchors: 3 # anchors per output layer (0 to ignore) +fl_gamma: 0.0 # focal loss gamma (efficientDet default gamma=1.5) +hsv_h: 0.015 # image HSV-Hue augmentation (fraction) +hsv_s: 0.7 # image HSV-Saturation augmentation (fraction) +hsv_v: 0.4 # image HSV-Value augmentation (fraction) +degrees: 0.0 # image rotation (+/- deg) +translate: 0.1 # image translation (+/- fraction) +scale: 0.5 # image scale (+/- gain) +shear: 0.0 # image shear (+/- deg) +perspective: 0.0 # image perspective (+/- fraction), range 0-0.001 +flipud: 0.0 # image flip up-down (probability) +fliplr: 0.5 # image flip left-right (probability) +mosaic: 1.0 # image mosaic (probability) +mixup: 0.0 # image mixup (probability) +copy_paste: 0.0 # segment copy-paste (probability) diff --git a/yolov5/data/hyps/hyp.scratch-med.yaml b/yolov5/data/hyps/hyp.scratch-med.yaml new file mode 100644 index 0000000..d6867d7 --- /dev/null +++ b/yolov5/data/hyps/hyp.scratch-med.yaml @@ -0,0 +1,34 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# Hyperparameters for medium-augmentation COCO training from scratch +# python train.py --batch 32 --cfg yolov5m6.yaml --weights '' --data coco.yaml --img 1280 --epochs 300 +# See tutorials for hyperparameter evolution https://github.com/ultralytics/yolov5#tutorials + +lr0: 0.01 # initial learning rate (SGD=1E-2, Adam=1E-3) +lrf: 0.1 # final OneCycleLR learning rate (lr0 * lrf) +momentum: 0.937 # SGD momentum/Adam beta1 +weight_decay: 0.0005 # optimizer weight decay 5e-4 +warmup_epochs: 3.0 # warmup epochs (fractions ok) +warmup_momentum: 0.8 # warmup initial momentum +warmup_bias_lr: 0.1 # warmup initial bias lr +box: 0.05 # box loss gain +cls: 0.3 # cls loss gain +cls_pw: 1.0 # cls BCELoss positive_weight +obj: 0.7 # obj loss gain (scale with pixels) +obj_pw: 1.0 # obj BCELoss positive_weight +iou_t: 0.20 # IoU training threshold +anchor_t: 4.0 # anchor-multiple threshold +# anchors: 3 # anchors per output layer (0 to ignore) +fl_gamma: 0.0 # focal loss gamma (efficientDet default gamma=1.5) +hsv_h: 0.015 # image HSV-Hue augmentation (fraction) +hsv_s: 0.7 # image HSV-Saturation augmentation (fraction) +hsv_v: 0.4 # image HSV-Value augmentation (fraction) +degrees: 0.0 # image rotation (+/- deg) +translate: 0.1 # image translation (+/- fraction) +scale: 0.9 # image scale (+/- gain) +shear: 0.0 # image shear (+/- deg) +perspective: 0.0 # image perspective (+/- fraction), range 0-0.001 +flipud: 0.0 # image flip up-down (probability) +fliplr: 0.5 # image flip left-right (probability) +mosaic: 1.0 # image mosaic (probability) +mixup: 0.1 # image mixup (probability) +copy_paste: 0.0 # segment copy-paste (probability) diff --git a/yolov5/data/images/bus.jpg b/yolov5/data/images/bus.jpg new file mode 100644 index 0000000..b43e311 Binary files /dev/null and b/yolov5/data/images/bus.jpg differ diff --git a/yolov5/data/images/zidane.jpg b/yolov5/data/images/zidane.jpg new file mode 100644 index 0000000..92d72ea Binary files /dev/null and b/yolov5/data/images/zidane.jpg differ diff --git a/yolov5/data/scripts/download_weights.sh b/yolov5/data/scripts/download_weights.sh new file mode 100644 index 0000000..e9fa653 --- /dev/null +++ b/yolov5/data/scripts/download_weights.sh @@ -0,0 +1,20 @@ +#!/bin/bash +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# Download latest models from https://github.com/ultralytics/yolov5/releases +# Example usage: bash path/to/download_weights.sh +# parent +# └── yolov5 +# ├── yolov5s.pt ← downloads here +# ├── yolov5m.pt +# └── ... + +python - <= cls >= 0, f'incorrect class index {cls}' + + # Write YOLO label + if id not in shapes: + shapes[id] = Image.open(file).size + box = xyxy2xywhn(box[None].astype(np.float), w=shapes[id][0], h=shapes[id][1], clip=True) + with open((labels / id).with_suffix('.txt'), 'a') as f: + f.write(f"{cls} {' '.join(f'{x:.6f}' for x in box[0])}\n") # write label.txt + except Exception as e: + print(f'WARNING: skipping one label for {file}: {e}') + + + # Download manually from https://challenge.xviewdataset.org + dir = Path(yaml['path']) # dataset root dir + # urls = ['https://d307kc0mrhucc3.cloudfront.net/train_labels.zip', # train labels + # 'https://d307kc0mrhucc3.cloudfront.net/train_images.zip', # 15G, 847 train images + # 'https://d307kc0mrhucc3.cloudfront.net/val_images.zip'] # 5G, 282 val images (no labels) + # download(urls, dir=dir, delete=False) + + # Convert labels + convert_labels(dir / 'xView_train.geojson') + + # Move images + images = Path(dir / 'images') + images.mkdir(parents=True, exist_ok=True) + Path(dir / 'train_images').rename(dir / 'images' / 'train') + Path(dir / 'val_images').rename(dir / 'images' / 'val') + + # Split + autosplit(dir / 'images' / 'train') diff --git a/yolov5/detect.py b/yolov5/detect.py new file mode 100644 index 0000000..8feb07d --- /dev/null +++ b/yolov5/detect.py @@ -0,0 +1,252 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Run inference on images, videos, directories, streams, etc. + +Usage - sources: + $ python path/to/detect.py --weights yolov5s.pt --source 0 # webcam + img.jpg # image + vid.mp4 # video + path/ # directory + path/*.jpg # glob + 'https://youtu.be/Zgi9g1ksQHc' # YouTube + 'rtsp://example.com/media.mp4' # RTSP, RTMP, HTTP stream + +Usage - formats: + $ python path/to/detect.py --weights yolov5s.pt # PyTorch + yolov5s.torchscript # TorchScript + yolov5s.onnx # ONNX Runtime or OpenCV DNN with --dnn + yolov5s.xml # OpenVINO + yolov5s.engine # TensorRT + yolov5s.mlmodel # CoreML (macOS-only) + yolov5s_saved_model # TensorFlow SavedModel + yolov5s.pb # TensorFlow GraphDef + yolov5s.tflite # TensorFlow Lite + yolov5s_edgetpu.tflite # TensorFlow Edge TPU +""" + +import argparse +import os +import sys +from pathlib import Path + +import torch +import torch.backends.cudnn as cudnn + +FILE = Path(__file__).resolve() +ROOT = FILE.parents[0] # YOLOv5 root directory +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) # add ROOT to PATH +ROOT = Path(os.path.relpath(ROOT, Path.cwd())) # relative + +from models.common import DetectMultiBackend +from utils.dataloaders import IMG_FORMATS, VID_FORMATS, LoadImages, LoadStreams +from utils.general import (LOGGER, check_file, check_img_size, check_imshow, check_requirements, colorstr, cv2, + increment_path, non_max_suppression, print_args, scale_coords, strip_optimizer, xyxy2xywh) +from utils.plots import Annotator, colors, save_one_box +from utils.torch_utils import select_device, time_sync + + +@torch.no_grad() +def run( + weights=ROOT / 'yolov5s.pt', # model.pt path(s) + source=ROOT / 'data/images', # file/dir/URL/glob, 0 for webcam + data=ROOT / 'data/coco128.yaml', # dataset.yaml path + imgsz=(640, 640), # inference size (height, width) + conf_thres=0.25, # confidence threshold + iou_thres=0.45, # NMS IOU threshold + max_det=1000, # maximum detections per image + device='', # cuda device, i.e. 0 or 0,1,2,3 or cpu + view_img=False, # show results + save_txt=False, # save results to *.txt + save_conf=False, # save confidences in --save-txt labels + save_crop=False, # save cropped prediction boxes + nosave=False, # do not save images/videos + classes=None, # filter by class: --class 0, or --class 0 2 3 + agnostic_nms=False, # class-agnostic NMS + augment=False, # augmented inference + visualize=False, # visualize features + update=False, # update all models + project=ROOT / 'runs/detect', # save results to project/name + name='exp', # save results to project/name + exist_ok=False, # existing project/name ok, do not increment + line_thickness=3, # bounding box thickness (pixels) + hide_labels=False, # hide labels + hide_conf=False, # hide confidences + half=False, # use FP16 half-precision inference + dnn=False, # use OpenCV DNN for ONNX inference +): + source = str(source) + save_img = not nosave and not source.endswith('.txt') # save inference images + is_file = Path(source).suffix[1:] in (IMG_FORMATS + VID_FORMATS) + is_url = source.lower().startswith(('rtsp://', 'rtmp://', 'http://', 'https://')) + webcam = source.isnumeric() or source.endswith('.txt') or (is_url and not is_file) + if is_url and is_file: + source = check_file(source) # download + + # Directories + save_dir = increment_path(Path(project) / name, exist_ok=exist_ok) # increment run + (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True) # make dir + + # Load model + device = select_device(device) + model = DetectMultiBackend(weights, device=device, dnn=dnn, data=data, fp16=half) + stride, names, pt = model.stride, model.names, model.pt + imgsz = check_img_size(imgsz, s=stride) # check image size + + # Dataloader + if webcam: + view_img = check_imshow() + cudnn.benchmark = True # set True to speed up constant image size inference + dataset = LoadStreams(source, img_size=imgsz, stride=stride, auto=pt) + bs = len(dataset) # batch_size + else: + dataset = LoadImages(source, img_size=imgsz, stride=stride, auto=pt) + bs = 1 # batch_size + vid_path, vid_writer = [None] * bs, [None] * bs + + # Run inference + model.warmup(imgsz=(1 if pt else bs, 3, *imgsz)) # warmup + dt, seen = [0.0, 0.0, 0.0], 0 + for path, im, im0s, vid_cap, s in dataset: + t1 = time_sync() + im = torch.from_numpy(im).to(device) + im = im.half() if model.fp16 else im.float() # uint8 to fp16/32 + im /= 255 # 0 - 255 to 0.0 - 1.0 + if len(im.shape) == 3: + im = im[None] # expand for batch dim + t2 = time_sync() + dt[0] += t2 - t1 + + # Inference + visualize = increment_path(save_dir / Path(path).stem, mkdir=True) if visualize else False + pred = model(im, augment=augment, visualize=visualize) + t3 = time_sync() + dt[1] += t3 - t2 + + # NMS + pred = non_max_suppression(pred, conf_thres, iou_thres, classes, agnostic_nms, max_det=max_det) + dt[2] += time_sync() - t3 + + # Second-stage classifier (optional) + # pred = utils.general.apply_classifier(pred, classifier_model, im, im0s) + + # Process predictions + for i, det in enumerate(pred): # per image + seen += 1 + if webcam: # batch_size >= 1 + p, im0, frame = path[i], im0s[i].copy(), dataset.count + s += f'{i}: ' + else: + p, im0, frame = path, im0s.copy(), getattr(dataset, 'frame', 0) + + p = Path(p) # to Path + save_path = str(save_dir / p.name) # im.jpg + txt_path = str(save_dir / 'labels' / p.stem) + ('' if dataset.mode == 'image' else f'_{frame}') # im.txt + s += '%gx%g ' % im.shape[2:] # print string + gn = torch.tensor(im0.shape)[[1, 0, 1, 0]] # normalization gain whwh + imc = im0.copy() if save_crop else im0 # for save_crop + annotator = Annotator(im0, line_width=line_thickness, example=str(names)) + if len(det): + # Rescale boxes from img_size to im0 size + det[:, :4] = scale_coords(im.shape[2:], det[:, :4], im0.shape).round() + + # Print results + for c in det[:, -1].unique(): + n = (det[:, -1] == c).sum() # detections per class + s += f"{n} {names[int(c)]}{'s' * (n > 1)}, " # add to string + + # Write results + for *xyxy, conf, cls in reversed(det): + if save_txt: # Write to file + xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist() # normalized xywh + line = (cls, *xywh, conf) if save_conf else (cls, *xywh) # label format + with open(f'{txt_path}.txt', 'a') as f: + f.write(('%g ' * len(line)).rstrip() % line + '\n') + + if save_img or save_crop or view_img: # Add bbox to image + c = int(cls) # integer class + label = None if hide_labels else (names[c] if hide_conf else f'{names[c]} {conf:.2f}') + annotator.box_label(xyxy, label, color=colors(c, True)) + if save_crop: + save_one_box(xyxy, imc, file=save_dir / 'crops' / names[c] / f'{p.stem}.jpg', BGR=True) + + # Stream results + im0 = annotator.result() + if view_img: + cv2.imshow(str(p), im0) + cv2.waitKey(1) # 1 millisecond + + # Save results (image with detections) + if save_img: + if dataset.mode == 'image': + cv2.imwrite(save_path, im0) + else: # 'video' or 'stream' + if vid_path[i] != save_path: # new video + vid_path[i] = save_path + if isinstance(vid_writer[i], cv2.VideoWriter): + vid_writer[i].release() # release previous video writer + if vid_cap: # video + fps = vid_cap.get(cv2.CAP_PROP_FPS) + w = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH)) + h = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT)) + else: # stream + fps, w, h = 30, im0.shape[1], im0.shape[0] + save_path = str(Path(save_path).with_suffix('.mp4')) # force *.mp4 suffix on results videos + vid_writer[i] = cv2.VideoWriter(save_path, cv2.VideoWriter_fourcc(*'mp4v'), fps, (w, h)) + vid_writer[i].write(im0) + + # Print time (inference-only) + LOGGER.info(f'{s}Done. ({t3 - t2:.3f}s)') + + # Print results + t = tuple(x / seen * 1E3 for x in dt) # speeds per image + LOGGER.info(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {(1, 3, *imgsz)}' % t) + if save_txt or save_img: + s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else '' + LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}{s}") + if update: + strip_optimizer(weights) # update model (to fix SourceChangeWarning) + + +def parse_opt(): + parser = argparse.ArgumentParser() + parser.add_argument('--weights', nargs='+', type=str, default=ROOT / 'yolov5s.pt', help='model path(s)') + parser.add_argument('--source', type=str, default=ROOT / 'data/images', help='file/dir/URL/glob, 0 for webcam') + parser.add_argument('--data', type=str, default=ROOT / 'data/coco128.yaml', help='(optional) dataset.yaml path') + parser.add_argument('--imgsz', '--img', '--img-size', nargs='+', type=int, default=[640], help='inference size h,w') + parser.add_argument('--conf-thres', type=float, default=0.25, help='confidence threshold') + parser.add_argument('--iou-thres', type=float, default=0.45, help='NMS IoU threshold') + parser.add_argument('--max-det', type=int, default=1000, help='maximum detections per image') + parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') + parser.add_argument('--view-img', action='store_true', help='show results') + parser.add_argument('--save-txt', action='store_true', help='save results to *.txt') + parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels') + parser.add_argument('--save-crop', action='store_true', help='save cropped prediction boxes') + parser.add_argument('--nosave', action='store_true', help='do not save images/videos') + parser.add_argument('--classes', nargs='+', type=int, help='filter by class: --classes 0, or --classes 0 2 3') + parser.add_argument('--agnostic-nms', action='store_true', help='class-agnostic NMS') + parser.add_argument('--augment', action='store_true', help='augmented inference') + parser.add_argument('--visualize', action='store_true', help='visualize features') + parser.add_argument('--update', action='store_true', help='update all models') + parser.add_argument('--project', default=ROOT / 'runs/detect', help='save results to project/name') + parser.add_argument('--name', default='exp', help='save results to project/name') + parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment') + parser.add_argument('--line-thickness', default=3, type=int, help='bounding box thickness (pixels)') + parser.add_argument('--hide-labels', default=False, action='store_true', help='hide labels') + parser.add_argument('--hide-conf', default=False, action='store_true', help='hide confidences') + parser.add_argument('--half', action='store_true', help='use FP16 half-precision inference') + parser.add_argument('--dnn', action='store_true', help='use OpenCV DNN for ONNX inference') + opt = parser.parse_args() + opt.imgsz *= 2 if len(opt.imgsz) == 1 else 1 # expand + print_args(vars(opt)) + return opt + + +def main(opt): + check_requirements(exclude=('tensorboard', 'thop')) + run(**vars(opt)) + + +if __name__ == "__main__": + opt = parse_opt() + main(opt) diff --git a/yolov5/export.py b/yolov5/export.py new file mode 100644 index 0000000..72e170a --- /dev/null +++ b/yolov5/export.py @@ -0,0 +1,607 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Export a YOLOv5 PyTorch model to other formats. TensorFlow exports authored by https://github.com/zldrobit + +Format | `export.py --include` | Model +--- | --- | --- +PyTorch | - | yolov5s.pt +TorchScript | `torchscript` | yolov5s.torchscript +ONNX | `onnx` | yolov5s.onnx +OpenVINO | `openvino` | yolov5s_openvino_model/ +TensorRT | `engine` | yolov5s.engine +CoreML | `coreml` | yolov5s.mlmodel +TensorFlow SavedModel | `saved_model` | yolov5s_saved_model/ +TensorFlow GraphDef | `pb` | yolov5s.pb +TensorFlow Lite | `tflite` | yolov5s.tflite +TensorFlow Edge TPU | `edgetpu` | yolov5s_edgetpu.tflite +TensorFlow.js | `tfjs` | yolov5s_web_model/ + +Requirements: + $ pip install -r requirements.txt coremltools onnx onnx-simplifier onnxruntime openvino-dev tensorflow-cpu # CPU + $ pip install -r requirements.txt coremltools onnx onnx-simplifier onnxruntime-gpu openvino-dev tensorflow # GPU + +Usage: + $ python path/to/export.py --weights yolov5s.pt --include torchscript onnx openvino engine coreml tflite ... + +Inference: + $ python path/to/detect.py --weights yolov5s.pt # PyTorch + yolov5s.torchscript # TorchScript + yolov5s.onnx # ONNX Runtime or OpenCV DNN with --dnn + yolov5s.xml # OpenVINO + yolov5s.engine # TensorRT + yolov5s.mlmodel # CoreML (macOS-only) + yolov5s_saved_model # TensorFlow SavedModel + yolov5s.pb # TensorFlow GraphDef + yolov5s.tflite # TensorFlow Lite + yolov5s_edgetpu.tflite # TensorFlow Edge TPU + +TensorFlow.js: + $ cd .. && git clone https://github.com/zldrobit/tfjs-yolov5-example.git && cd tfjs-yolov5-example + $ npm install + $ ln -s ../../yolov5/yolov5s_web_model public/yolov5s_web_model + $ npm start +""" + +import argparse +import json +import os +import platform +import subprocess +import sys +import time +import warnings +from pathlib import Path + +import pandas as pd +import torch +import yaml +from torch.utils.mobile_optimizer import optimize_for_mobile + +FILE = Path(__file__).resolve() +ROOT = FILE.parents[0] # YOLOv5 root directory +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) # add ROOT to PATH +if platform.system() != 'Windows': + ROOT = Path(os.path.relpath(ROOT, Path.cwd())) # relative + +from models.experimental import attempt_load +from models.yolo import Detect +from utils.dataloaders import LoadImages +from utils.general import (LOGGER, check_dataset, check_img_size, check_requirements, check_version, colorstr, + file_size, print_args, url2file) +from utils.torch_utils import select_device + + +def export_formats(): + # YOLOv5 export formats + x = [ + ['PyTorch', '-', '.pt', True], + ['TorchScript', 'torchscript', '.torchscript', True], + ['ONNX', 'onnx', '.onnx', True], + ['OpenVINO', 'openvino', '_openvino_model', False], + ['TensorRT', 'engine', '.engine', True], + ['CoreML', 'coreml', '.mlmodel', False], + ['TensorFlow SavedModel', 'saved_model', '_saved_model', True], + ['TensorFlow GraphDef', 'pb', '.pb', True], + ['TensorFlow Lite', 'tflite', '.tflite', False], + ['TensorFlow Edge TPU', 'edgetpu', '_edgetpu.tflite', False], + ['TensorFlow.js', 'tfjs', '_web_model', False],] + return pd.DataFrame(x, columns=['Format', 'Argument', 'Suffix', 'GPU']) + + +def export_torchscript(model, im, file, optimize, prefix=colorstr('TorchScript:')): + # YOLOv5 TorchScript model export + try: + LOGGER.info(f'\n{prefix} starting export with torch {torch.__version__}...') + f = file.with_suffix('.torchscript') + + ts = torch.jit.trace(model, im, strict=False) + d = {"shape": im.shape, "stride": int(max(model.stride)), "names": model.names} + extra_files = {'config.txt': json.dumps(d)} # torch._C.ExtraFilesMap() + if optimize: # https://pytorch.org/tutorials/recipes/mobile_interpreter.html + optimize_for_mobile(ts)._save_for_lite_interpreter(str(f), _extra_files=extra_files) + else: + ts.save(str(f), _extra_files=extra_files) + + LOGGER.info(f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)') + return f + except Exception as e: + LOGGER.info(f'{prefix} export failure: {e}') + + +def export_onnx(model, im, file, opset, train, dynamic, simplify, prefix=colorstr('ONNX:')): + # YOLOv5 ONNX export + try: + check_requirements(('onnx',)) + import onnx + + LOGGER.info(f'\n{prefix} starting export with onnx {onnx.__version__}...') + f = file.with_suffix('.onnx') + + torch.onnx.export( + model, + im, + f, + verbose=False, + opset_version=opset, + training=torch.onnx.TrainingMode.TRAINING if train else torch.onnx.TrainingMode.EVAL, + do_constant_folding=not train, + input_names=['images'], + output_names=['output'], + dynamic_axes={ + 'images': { + 0: 'batch', + 2: 'height', + 3: 'width'}, # shape(1,3,640,640) + 'output': { + 0: 'batch', + 1: 'anchors'} # shape(1,25200,85) + } if dynamic else None) + + # Checks + model_onnx = onnx.load(f) # load onnx model + onnx.checker.check_model(model_onnx) # check onnx model + + # Metadata + d = {'stride': int(max(model.stride)), 'names': model.names} + for k, v in d.items(): + meta = model_onnx.metadata_props.add() + meta.key, meta.value = k, str(v) + onnx.save(model_onnx, f) + + # Simplify + if simplify: + try: + check_requirements(('onnx-simplifier',)) + import onnxsim + + LOGGER.info(f'{prefix} simplifying with onnx-simplifier {onnxsim.__version__}...') + model_onnx, check = onnxsim.simplify(model_onnx, + dynamic_input_shape=dynamic, + input_shapes={'images': list(im.shape)} if dynamic else None) + assert check, 'assert check failed' + onnx.save(model_onnx, f) + except Exception as e: + LOGGER.info(f'{prefix} simplifier failure: {e}') + LOGGER.info(f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)') + return f + except Exception as e: + LOGGER.info(f'{prefix} export failure: {e}') + + +def export_openvino(model, file, half, prefix=colorstr('OpenVINO:')): + # YOLOv5 OpenVINO export + try: + check_requirements(('openvino-dev',)) # requires openvino-dev: https://pypi.org/project/openvino-dev/ + import openvino.inference_engine as ie + + LOGGER.info(f'\n{prefix} starting export with openvino {ie.__version__}...') + f = str(file).replace('.pt', f'_openvino_model{os.sep}') + + cmd = f"mo --input_model {file.with_suffix('.onnx')} --output_dir {f} --data_type {'FP16' if half else 'FP32'}" + subprocess.check_output(cmd.split()) # export + with open(Path(f) / file.with_suffix('.yaml').name, 'w') as g: + yaml.dump({'stride': int(max(model.stride)), 'names': model.names}, g) # add metadata.yaml + + LOGGER.info(f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)') + return f + except Exception as e: + LOGGER.info(f'\n{prefix} export failure: {e}') + + +def export_coreml(model, im, file, int8, half, prefix=colorstr('CoreML:')): + # YOLOv5 CoreML export + try: + check_requirements(('coremltools',)) + import coremltools as ct + + LOGGER.info(f'\n{prefix} starting export with coremltools {ct.__version__}...') + f = file.with_suffix('.mlmodel') + + ts = torch.jit.trace(model, im, strict=False) # TorchScript model + ct_model = ct.convert(ts, inputs=[ct.ImageType('image', shape=im.shape, scale=1 / 255, bias=[0, 0, 0])]) + bits, mode = (8, 'kmeans_lut') if int8 else (16, 'linear') if half else (32, None) + if bits < 32: + if platform.system() == 'Darwin': # quantization only supported on macOS + with warnings.catch_warnings(): + warnings.filterwarnings("ignore", category=DeprecationWarning) # suppress numpy==1.20 float warning + ct_model = ct.models.neural_network.quantization_utils.quantize_weights(ct_model, bits, mode) + else: + print(f'{prefix} quantization only supported on macOS, skipping...') + ct_model.save(f) + + LOGGER.info(f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)') + return ct_model, f + except Exception as e: + LOGGER.info(f'\n{prefix} export failure: {e}') + return None, None + + +def export_engine(model, im, file, train, half, simplify, workspace=4, verbose=False, prefix=colorstr('TensorRT:')): + # YOLOv5 TensorRT export https://developer.nvidia.com/tensorrt + try: + assert im.device.type != 'cpu', 'export running on CPU but must be on GPU, i.e. `python export.py --device 0`' + try: + import tensorrt as trt + except Exception: + if platform.system() == 'Linux': + check_requirements(('nvidia-tensorrt',), cmds=('-U --index-url https://pypi.ngc.nvidia.com',)) + import tensorrt as trt + + if trt.__version__[0] == '7': # TensorRT 7 handling https://github.com/ultralytics/yolov5/issues/6012 + grid = model.model[-1].anchor_grid + model.model[-1].anchor_grid = [a[..., :1, :1, :] for a in grid] + export_onnx(model, im, file, 12, train, False, simplify) # opset 12 + model.model[-1].anchor_grid = grid + else: # TensorRT >= 8 + check_version(trt.__version__, '8.0.0', hard=True) # require tensorrt>=8.0.0 + export_onnx(model, im, file, 13, train, False, simplify) # opset 13 + onnx = file.with_suffix('.onnx') + + LOGGER.info(f'\n{prefix} starting export with TensorRT {trt.__version__}...') + assert onnx.exists(), f'failed to export ONNX file: {onnx}' + f = file.with_suffix('.engine') # TensorRT engine file + logger = trt.Logger(trt.Logger.INFO) + if verbose: + logger.min_severity = trt.Logger.Severity.VERBOSE + + builder = trt.Builder(logger) + config = builder.create_builder_config() + config.max_workspace_size = workspace * 1 << 30 + # config.set_memory_pool_limit(trt.MemoryPoolType.WORKSPACE, workspace << 30) # fix TRT 8.4 deprecation notice + + flag = (1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)) + network = builder.create_network(flag) + parser = trt.OnnxParser(network, logger) + if not parser.parse_from_file(str(onnx)): + raise RuntimeError(f'failed to load ONNX file: {onnx}') + + inputs = [network.get_input(i) for i in range(network.num_inputs)] + outputs = [network.get_output(i) for i in range(network.num_outputs)] + LOGGER.info(f'{prefix} Network Description:') + for inp in inputs: + LOGGER.info(f'{prefix}\tinput "{inp.name}" with shape {inp.shape} and dtype {inp.dtype}') + for out in outputs: + LOGGER.info(f'{prefix}\toutput "{out.name}" with shape {out.shape} and dtype {out.dtype}') + + LOGGER.info(f'{prefix} building FP{16 if builder.platform_has_fast_fp16 and half else 32} engine in {f}') + if builder.platform_has_fast_fp16 and half: + config.set_flag(trt.BuilderFlag.FP16) + with builder.build_engine(network, config) as engine, open(f, 'wb') as t: + t.write(engine.serialize()) + LOGGER.info(f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)') + return f + except Exception as e: + LOGGER.info(f'\n{prefix} export failure: {e}') + + +def export_saved_model(model, + im, + file, + dynamic, + tf_nms=False, + agnostic_nms=False, + topk_per_class=100, + topk_all=100, + iou_thres=0.45, + conf_thres=0.25, + keras=False, + prefix=colorstr('TensorFlow SavedModel:')): + # YOLOv5 TensorFlow SavedModel export + try: + import tensorflow as tf + from tensorflow.python.framework.convert_to_constants import convert_variables_to_constants_v2 + + from models.tf import TFDetect, TFModel + + LOGGER.info(f'\n{prefix} starting export with tensorflow {tf.__version__}...') + f = str(file).replace('.pt', '_saved_model') + batch_size, ch, *imgsz = list(im.shape) # BCHW + + tf_model = TFModel(cfg=model.yaml, model=model, nc=model.nc, imgsz=imgsz) + im = tf.zeros((batch_size, *imgsz, ch)) # BHWC order for TensorFlow + _ = tf_model.predict(im, tf_nms, agnostic_nms, topk_per_class, topk_all, iou_thres, conf_thres) + inputs = tf.keras.Input(shape=(*imgsz, ch), batch_size=None if dynamic else batch_size) + outputs = tf_model.predict(inputs, tf_nms, agnostic_nms, topk_per_class, topk_all, iou_thres, conf_thres) + keras_model = tf.keras.Model(inputs=inputs, outputs=outputs) + keras_model.trainable = False + keras_model.summary() + if keras: + keras_model.save(f, save_format='tf') + else: + spec = tf.TensorSpec(keras_model.inputs[0].shape, keras_model.inputs[0].dtype) + m = tf.function(lambda x: keras_model(x)) # full model + m = m.get_concrete_function(spec) + frozen_func = convert_variables_to_constants_v2(m) + tfm = tf.Module() + tfm.__call__ = tf.function(lambda x: frozen_func(x)[:4] if tf_nms else frozen_func(x)[0], [spec]) + tfm.__call__(im) + tf.saved_model.save(tfm, + f, + options=tf.saved_model.SaveOptions(experimental_custom_gradients=False) + if check_version(tf.__version__, '2.6') else tf.saved_model.SaveOptions()) + LOGGER.info(f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)') + return keras_model, f + except Exception as e: + LOGGER.info(f'\n{prefix} export failure: {e}') + return None, None + + +def export_pb(keras_model, file, prefix=colorstr('TensorFlow GraphDef:')): + # YOLOv5 TensorFlow GraphDef *.pb export https://github.com/leimao/Frozen_Graph_TensorFlow + try: + import tensorflow as tf + from tensorflow.python.framework.convert_to_constants import convert_variables_to_constants_v2 + + LOGGER.info(f'\n{prefix} starting export with tensorflow {tf.__version__}...') + f = file.with_suffix('.pb') + + m = tf.function(lambda x: keras_model(x)) # full model + m = m.get_concrete_function(tf.TensorSpec(keras_model.inputs[0].shape, keras_model.inputs[0].dtype)) + frozen_func = convert_variables_to_constants_v2(m) + frozen_func.graph.as_graph_def() + tf.io.write_graph(graph_or_graph_def=frozen_func.graph, logdir=str(f.parent), name=f.name, as_text=False) + + LOGGER.info(f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)') + return f + except Exception as e: + LOGGER.info(f'\n{prefix} export failure: {e}') + + +def export_tflite(keras_model, im, file, int8, data, nms, agnostic_nms, prefix=colorstr('TensorFlow Lite:')): + # YOLOv5 TensorFlow Lite export + try: + import tensorflow as tf + + LOGGER.info(f'\n{prefix} starting export with tensorflow {tf.__version__}...') + batch_size, ch, *imgsz = list(im.shape) # BCHW + f = str(file).replace('.pt', '-fp16.tflite') + + converter = tf.lite.TFLiteConverter.from_keras_model(keras_model) + converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS] + converter.target_spec.supported_types = [tf.float16] + converter.optimizations = [tf.lite.Optimize.DEFAULT] + if int8: + from models.tf import representative_dataset_gen + dataset = LoadImages(check_dataset(data)['train'], img_size=imgsz, auto=False) # representative data + converter.representative_dataset = lambda: representative_dataset_gen(dataset, ncalib=100) + converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8] + converter.target_spec.supported_types = [] + converter.inference_input_type = tf.uint8 # or tf.int8 + converter.inference_output_type = tf.uint8 # or tf.int8 + converter.experimental_new_quantizer = True + f = str(file).replace('.pt', '-int8.tflite') + if nms or agnostic_nms: + converter.target_spec.supported_ops.append(tf.lite.OpsSet.SELECT_TF_OPS) + + tflite_model = converter.convert() + open(f, "wb").write(tflite_model) + LOGGER.info(f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)') + return f + except Exception as e: + LOGGER.info(f'\n{prefix} export failure: {e}') + + +def export_edgetpu(file, prefix=colorstr('Edge TPU:')): + # YOLOv5 Edge TPU export https://coral.ai/docs/edgetpu/models-intro/ + try: + cmd = 'edgetpu_compiler --version' + help_url = 'https://coral.ai/docs/edgetpu/compiler/' + assert platform.system() == 'Linux', f'export only supported on Linux. See {help_url}' + if subprocess.run(f'{cmd} >/dev/null', shell=True).returncode != 0: + LOGGER.info(f'\n{prefix} export requires Edge TPU compiler. Attempting install from {help_url}') + sudo = subprocess.run('sudo --version >/dev/null', shell=True).returncode == 0 # sudo installed on system + for c in ( + 'curl https://packages.cloud.google.com/apt/doc/apt-key.gpg | sudo apt-key add -', + 'echo "deb https://packages.cloud.google.com/apt coral-edgetpu-stable main" | sudo tee /etc/apt/sources.list.d/coral-edgetpu.list', + 'sudo apt-get update', 'sudo apt-get install edgetpu-compiler'): + subprocess.run(c if sudo else c.replace('sudo ', ''), shell=True, check=True) + ver = subprocess.run(cmd, shell=True, capture_output=True, check=True).stdout.decode().split()[-1] + + LOGGER.info(f'\n{prefix} starting export with Edge TPU compiler {ver}...') + f = str(file).replace('.pt', '-int8_edgetpu.tflite') # Edge TPU model + f_tfl = str(file).replace('.pt', '-int8.tflite') # TFLite model + + cmd = f"edgetpu_compiler -s -o {file.parent} {f_tfl}" + subprocess.run(cmd.split(), check=True) + + LOGGER.info(f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)') + return f + except Exception as e: + LOGGER.info(f'\n{prefix} export failure: {e}') + + +def export_tfjs(file, prefix=colorstr('TensorFlow.js:')): + # YOLOv5 TensorFlow.js export + try: + check_requirements(('tensorflowjs',)) + import re + + import tensorflowjs as tfjs + + LOGGER.info(f'\n{prefix} starting export with tensorflowjs {tfjs.__version__}...') + f = str(file).replace('.pt', '_web_model') # js dir + f_pb = file.with_suffix('.pb') # *.pb path + f_json = f'{f}/model.json' # *.json path + + cmd = f'tensorflowjs_converter --input_format=tf_frozen_model ' \ + f'--output_node_names=Identity,Identity_1,Identity_2,Identity_3 {f_pb} {f}' + subprocess.run(cmd.split()) + + with open(f_json) as j: + json = j.read() + with open(f_json, 'w') as j: # sort JSON Identity_* in ascending order + subst = re.sub( + r'{"outputs": {"Identity.?.?": {"name": "Identity.?.?"}, ' + r'"Identity.?.?": {"name": "Identity.?.?"}, ' + r'"Identity.?.?": {"name": "Identity.?.?"}, ' + r'"Identity.?.?": {"name": "Identity.?.?"}}}', r'{"outputs": {"Identity": {"name": "Identity"}, ' + r'"Identity_1": {"name": "Identity_1"}, ' + r'"Identity_2": {"name": "Identity_2"}, ' + r'"Identity_3": {"name": "Identity_3"}}}', json) + j.write(subst) + + LOGGER.info(f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)') + return f + except Exception as e: + LOGGER.info(f'\n{prefix} export failure: {e}') + + +@torch.no_grad() +def run( + data=ROOT / 'data/coco128.yaml', # 'dataset.yaml path' + weights=ROOT / 'yolov5s.pt', # weights path + imgsz=(640, 640), # image (height, width) + batch_size=1, # batch size + device='cpu', # cuda device, i.e. 0 or 0,1,2,3 or cpu + include=('torchscript', 'onnx'), # include formats + half=False, # FP16 half-precision export + inplace=False, # set YOLOv5 Detect() inplace=True + train=False, # model.train() mode + keras=False, # use Keras + optimize=False, # TorchScript: optimize for mobile + int8=False, # CoreML/TF INT8 quantization + dynamic=False, # ONNX/TF: dynamic axes + simplify=False, # ONNX: simplify model + opset=12, # ONNX: opset version + verbose=False, # TensorRT: verbose log + workspace=4, # TensorRT: workspace size (GB) + nms=False, # TF: add NMS to model + agnostic_nms=False, # TF: add agnostic NMS to model + topk_per_class=100, # TF.js NMS: topk per class to keep + topk_all=100, # TF.js NMS: topk for all classes to keep + iou_thres=0.45, # TF.js NMS: IoU threshold + conf_thres=0.25, # TF.js NMS: confidence threshold +): + t = time.time() + include = [x.lower() for x in include] # to lowercase + fmts = tuple(export_formats()['Argument'][1:]) # --include arguments + flags = [x in include for x in fmts] + assert sum(flags) == len(include), f'ERROR: Invalid --include {include}, valid --include arguments are {fmts}' + jit, onnx, xml, engine, coreml, saved_model, pb, tflite, edgetpu, tfjs = flags # export booleans + file = Path(url2file(weights) if str(weights).startswith(('http:/', 'https:/')) else weights) # PyTorch weights + + # Load PyTorch model + device = select_device(device) + if half: + assert device.type != 'cpu' or coreml or xml, '--half only compatible with GPU export, i.e. use --device 0' + assert not dynamic, '--half not compatible with --dynamic, i.e. use either --half or --dynamic but not both' + model = attempt_load(weights, device=device, inplace=True, fuse=True) # load FP32 model + nc, names = model.nc, model.names # number of classes, class names + + # Checks + imgsz *= 2 if len(imgsz) == 1 else 1 # expand + assert nc == len(names), f'Model class count {nc} != len(names) {len(names)}' + + # Input + gs = int(max(model.stride)) # grid size (max stride) + imgsz = [check_img_size(x, gs) for x in imgsz] # verify img_size are gs-multiples + im = torch.zeros(batch_size, 3, *imgsz).to(device) # image size(1,3,320,192) BCHW iDetection + + # Update model + if half and not coreml and not xml: + im, model = im.half(), model.half() # to FP16 + model.train() if train else model.eval() # training mode = no Detect() layer grid construction + for k, m in model.named_modules(): + if isinstance(m, Detect): + m.inplace = inplace + m.onnx_dynamic = dynamic + m.export = True + + for _ in range(2): + y = model(im) # dry runs + shape = tuple(y[0].shape) # model output shape + LOGGER.info(f"\n{colorstr('PyTorch:')} starting from {file} with output shape {shape} ({file_size(file):.1f} MB)") + + # Exports + f = [''] * 10 # exported filenames + warnings.filterwarnings(action='ignore', category=torch.jit.TracerWarning) # suppress TracerWarning + if jit: + f[0] = export_torchscript(model, im, file, optimize) + if engine: # TensorRT required before ONNX + f[1] = export_engine(model, im, file, train, half, simplify, workspace, verbose) + if onnx or xml: # OpenVINO requires ONNX + f[2] = export_onnx(model, im, file, opset, train, dynamic, simplify) + if xml: # OpenVINO + f[3] = export_openvino(model, file, half) + if coreml: + _, f[4] = export_coreml(model, im, file, int8, half) + + # TensorFlow Exports + if any((saved_model, pb, tflite, edgetpu, tfjs)): + if int8 or edgetpu: # TFLite --int8 bug https://github.com/ultralytics/yolov5/issues/5707 + check_requirements(('flatbuffers==1.12',)) # required before `import tensorflow` + assert not tflite or not tfjs, 'TFLite and TF.js models must be exported separately, please pass only one type.' + model, f[5] = export_saved_model(model.cpu(), + im, + file, + dynamic, + tf_nms=nms or agnostic_nms or tfjs, + agnostic_nms=agnostic_nms or tfjs, + topk_per_class=topk_per_class, + topk_all=topk_all, + iou_thres=iou_thres, + conf_thres=conf_thres, + keras=keras) + if pb or tfjs: # pb prerequisite to tfjs + f[6] = export_pb(model, file) + if tflite or edgetpu: + f[7] = export_tflite(model, im, file, int8=int8 or edgetpu, data=data, nms=nms, agnostic_nms=agnostic_nms) + if edgetpu: + f[8] = export_edgetpu(file) + if tfjs: + f[9] = export_tfjs(file) + + # Finish + f = [str(x) for x in f if x] # filter out '' and None + if any(f): + LOGGER.info(f'\nExport complete ({time.time() - t:.2f}s)' + f"\nResults saved to {colorstr('bold', file.parent.resolve())}" + f"\nDetect: python detect.py --weights {f[-1]}" + f"\nPyTorch Hub: model = torch.hub.load('ultralytics/yolov5', 'custom', '{f[-1]}')" + f"\nValidate: python val.py --weights {f[-1]}" + f"\nVisualize: https://netron.app") + return f # return list of exported files/dirs + + +def parse_opt(): + parser = argparse.ArgumentParser() + parser.add_argument('--data', type=str, default=ROOT / 'data/coco128.yaml', help='dataset.yaml path') + parser.add_argument('--weights', nargs='+', type=str, default=ROOT / 'yolov5s.pt', help='model.pt path(s)') + parser.add_argument('--imgsz', '--img', '--img-size', nargs='+', type=int, default=[640, 640], help='image (h, w)') + parser.add_argument('--batch-size', type=int, default=1, help='batch size') + parser.add_argument('--device', default='cpu', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') + parser.add_argument('--half', action='store_true', help='FP16 half-precision export') + parser.add_argument('--inplace', action='store_true', help='set YOLOv5 Detect() inplace=True') + parser.add_argument('--train', action='store_true', help='model.train() mode') + parser.add_argument('--keras', action='store_true', help='TF: use Keras') + parser.add_argument('--optimize', action='store_true', help='TorchScript: optimize for mobile') + parser.add_argument('--int8', action='store_true', help='CoreML/TF INT8 quantization') + parser.add_argument('--dynamic', action='store_true', help='ONNX/TF: dynamic axes') + parser.add_argument('--simplify', action='store_true', help='ONNX: simplify model') + parser.add_argument('--opset', type=int, default=12, help='ONNX: opset version') + parser.add_argument('--verbose', action='store_true', help='TensorRT: verbose log') + parser.add_argument('--workspace', type=int, default=4, help='TensorRT: workspace size (GB)') + parser.add_argument('--nms', action='store_true', help='TF: add NMS to model') + parser.add_argument('--agnostic-nms', action='store_true', help='TF: add agnostic NMS to model') + parser.add_argument('--topk-per-class', type=int, default=100, help='TF.js NMS: topk per class to keep') + parser.add_argument('--topk-all', type=int, default=100, help='TF.js NMS: topk for all classes to keep') + parser.add_argument('--iou-thres', type=float, default=0.45, help='TF.js NMS: IoU threshold') + parser.add_argument('--conf-thres', type=float, default=0.25, help='TF.js NMS: confidence threshold') + parser.add_argument('--include', + nargs='+', + default=['torchscript', 'onnx'], + help='torchscript, onnx, openvino, engine, coreml, saved_model, pb, tflite, edgetpu, tfjs') + opt = parser.parse_args() + print_args(vars(opt)) + return opt + + +def main(opt): + for opt.weights in (opt.weights if isinstance(opt.weights, list) else [opt.weights]): + run(**vars(opt)) + + +if __name__ == "__main__": + opt = parse_opt() + main(opt) diff --git a/yolov5/hubconf.py b/yolov5/hubconf.py new file mode 100644 index 0000000..01f4eba --- /dev/null +++ b/yolov5/hubconf.py @@ -0,0 +1,146 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +PyTorch Hub models https://pytorch.org/hub/ultralytics_yolov5/ + +Usage: + import torch + model = torch.hub.load('ultralytics/yolov5', 'yolov5s') + model = torch.hub.load('ultralytics/yolov5:master', 'custom', 'path/to/yolov5s.onnx') # file from branch +""" + +import torch + + +def _create(name, pretrained=True, channels=3, classes=80, autoshape=True, verbose=True, device=None): + """Creates or loads a YOLOv5 model + + Arguments: + name (str): model name 'yolov5s' or path 'path/to/best.pt' + pretrained (bool): load pretrained weights into the model + channels (int): number of input channels + classes (int): number of model classes + autoshape (bool): apply YOLOv5 .autoshape() wrapper to model + verbose (bool): print all information to screen + device (str, torch.device, None): device to use for model parameters + + Returns: + YOLOv5 model + """ + from pathlib import Path + + from models.common import AutoShape, DetectMultiBackend + from models.yolo import Model + from utils.downloads import attempt_download + from utils.general import LOGGER, check_requirements, intersect_dicts, logging + from utils.torch_utils import select_device + + if not verbose: + LOGGER.setLevel(logging.WARNING) + + check_requirements(exclude=('tensorboard', 'thop', 'opencv-python')) + name = Path(name) + path = name.with_suffix('.pt') if name.suffix == '' and not name.is_dir() else name # checkpoint path + try: + device = select_device(('0' if torch.cuda.is_available() else 'cpu') if device is None else device) + + if pretrained and channels == 3 and classes == 80: + model = DetectMultiBackend(path, device=device) # download/load FP32 model + # model = models.experimental.attempt_load(path, map_location=device) # download/load FP32 model + else: + cfg = list((Path(__file__).parent / 'models').rglob(f'{path.stem}.yaml'))[0] # model.yaml path + model = Model(cfg, channels, classes) # create model + if pretrained: + ckpt = torch.load(attempt_download(path), map_location=device) # load + csd = ckpt['model'].float().state_dict() # checkpoint state_dict as FP32 + csd = intersect_dicts(csd, model.state_dict(), exclude=['anchors']) # intersect + model.load_state_dict(csd, strict=False) # load + if len(ckpt['model'].names) == classes: + model.names = ckpt['model'].names # set class names attribute + if autoshape: + model = AutoShape(model) # for file/URI/PIL/cv2/np inputs and NMS + return model.to(device) + + except Exception as e: + help_url = 'https://github.com/ultralytics/yolov5/issues/36' + s = f'{e}. Cache may be out of date, try `force_reload=True` or see {help_url} for help.' + raise Exception(s) from e + + +def custom(path='path/to/model.pt', autoshape=True, _verbose=True, device=None): + # YOLOv5 custom or local model + return _create(path, autoshape=autoshape, verbose=_verbose, device=device) + + +def yolov5n(pretrained=True, channels=3, classes=80, autoshape=True, _verbose=True, device=None): + # YOLOv5-nano model https://github.com/ultralytics/yolov5 + return _create('yolov5n', pretrained, channels, classes, autoshape, _verbose, device) + + +def yolov5s(pretrained=True, channels=3, classes=80, autoshape=True, _verbose=True, device=None): + # YOLOv5-small model https://github.com/ultralytics/yolov5 + return _create('yolov5s', pretrained, channels, classes, autoshape, _verbose, device) + + +def yolov5m(pretrained=True, channels=3, classes=80, autoshape=True, _verbose=True, device=None): + # YOLOv5-medium model https://github.com/ultralytics/yolov5 + return _create('yolov5m', pretrained, channels, classes, autoshape, _verbose, device) + + +def yolov5l(pretrained=True, channels=3, classes=80, autoshape=True, _verbose=True, device=None): + # YOLOv5-large model https://github.com/ultralytics/yolov5 + return _create('yolov5l', pretrained, channels, classes, autoshape, _verbose, device) + + +def yolov5x(pretrained=True, channels=3, classes=80, autoshape=True, _verbose=True, device=None): + # YOLOv5-xlarge model https://github.com/ultralytics/yolov5 + return _create('yolov5x', pretrained, channels, classes, autoshape, _verbose, device) + + +def yolov5n6(pretrained=True, channels=3, classes=80, autoshape=True, _verbose=True, device=None): + # YOLOv5-nano-P6 model https://github.com/ultralytics/yolov5 + return _create('yolov5n6', pretrained, channels, classes, autoshape, _verbose, device) + + +def yolov5s6(pretrained=True, channels=3, classes=80, autoshape=True, _verbose=True, device=None): + # YOLOv5-small-P6 model https://github.com/ultralytics/yolov5 + return _create('yolov5s6', pretrained, channels, classes, autoshape, _verbose, device) + + +def yolov5m6(pretrained=True, channels=3, classes=80, autoshape=True, _verbose=True, device=None): + # YOLOv5-medium-P6 model https://github.com/ultralytics/yolov5 + return _create('yolov5m6', pretrained, channels, classes, autoshape, _verbose, device) + + +def yolov5l6(pretrained=True, channels=3, classes=80, autoshape=True, _verbose=True, device=None): + # YOLOv5-large-P6 model https://github.com/ultralytics/yolov5 + return _create('yolov5l6', pretrained, channels, classes, autoshape, _verbose, device) + + +def yolov5x6(pretrained=True, channels=3, classes=80, autoshape=True, _verbose=True, device=None): + # YOLOv5-xlarge-P6 model https://github.com/ultralytics/yolov5 + return _create('yolov5x6', pretrained, channels, classes, autoshape, _verbose, device) + + +if __name__ == '__main__': + model = _create(name='yolov5s', pretrained=True, channels=3, classes=80, autoshape=True, verbose=True) + # model = custom(path='path/to/model.pt') # custom + + # Verify inference + from pathlib import Path + + import numpy as np + from PIL import Image + + from utils.general import cv2 + + imgs = [ + 'data/images/zidane.jpg', # filename + Path('data/images/zidane.jpg'), # Path + 'https://ultralytics.com/images/zidane.jpg', # URI + cv2.imread('data/images/bus.jpg')[:, :, ::-1], # OpenCV + Image.open('data/images/bus.jpg'), # PIL + np.zeros((320, 640, 3))] # numpy + + results = model(imgs, size=320) # batched inference + results.print() + results.save() diff --git a/yolov5/models/__init__.py b/yolov5/models/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/yolov5/models/common.py b/yolov5/models/common.py new file mode 100644 index 0000000..66467a0 --- /dev/null +++ b/yolov5/models/common.py @@ -0,0 +1,738 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Common modules +""" + +import json +import math +import platform +import warnings +from collections import OrderedDict, namedtuple +from copy import copy +from pathlib import Path + +import cv2 +import numpy as np +import pandas as pd +import requests +import torch +import torch.nn as nn +import yaml +from PIL import Image +from torch.cuda import amp + +from utils.dataloaders import exif_transpose, letterbox +from utils.general import (LOGGER, check_requirements, check_suffix, check_version, colorstr, increment_path, + make_divisible, non_max_suppression, scale_coords, xywh2xyxy, xyxy2xywh) +from utils.plots import Annotator, colors, save_one_box +from utils.torch_utils import copy_attr, time_sync + + +def autopad(k, p=None): # kernel, padding + # Pad to 'same' + if p is None: + p = k // 2 if isinstance(k, int) else [x // 2 for x in k] # auto-pad + return p + + +class Conv(nn.Module): + # Standard convolution + def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True): # ch_in, ch_out, kernel, stride, padding, groups + super().__init__() + self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g, bias=False) + self.bn = nn.BatchNorm2d(c2) + self.act = nn.SiLU() if act is True else (act if isinstance(act, nn.Module) else nn.Identity()) + + def forward(self, x): + return self.act(self.bn(self.conv(x))) + + def forward_fuse(self, x): + return self.act(self.conv(x)) + + +class DWConv(Conv): + # Depth-wise convolution class + def __init__(self, c1, c2, k=1, s=1, act=True): # ch_in, ch_out, kernel, stride, padding, groups + super().__init__(c1, c2, k, s, g=math.gcd(c1, c2), act=act) + + +class DWConvTranspose2d(nn.ConvTranspose2d): + # Depth-wise transpose convolution class + def __init__(self, c1, c2, k=1, s=1, p1=0, p2=0): # ch_in, ch_out, kernel, stride, padding, padding_out + super().__init__(c1, c2, k, s, p1, p2, groups=math.gcd(c1, c2)) + + +class TransformerLayer(nn.Module): + # Transformer layer https://arxiv.org/abs/2010.11929 (LayerNorm layers removed for better performance) + def __init__(self, c, num_heads): + super().__init__() + self.q = nn.Linear(c, c, bias=False) + self.k = nn.Linear(c, c, bias=False) + self.v = nn.Linear(c, c, bias=False) + self.ma = nn.MultiheadAttention(embed_dim=c, num_heads=num_heads) + self.fc1 = nn.Linear(c, c, bias=False) + self.fc2 = nn.Linear(c, c, bias=False) + + def forward(self, x): + x = self.ma(self.q(x), self.k(x), self.v(x))[0] + x + x = self.fc2(self.fc1(x)) + x + return x + + +class TransformerBlock(nn.Module): + # Vision Transformer https://arxiv.org/abs/2010.11929 + def __init__(self, c1, c2, num_heads, num_layers): + super().__init__() + self.conv = None + if c1 != c2: + self.conv = Conv(c1, c2) + self.linear = nn.Linear(c2, c2) # learnable position embedding + self.tr = nn.Sequential(*(TransformerLayer(c2, num_heads) for _ in range(num_layers))) + self.c2 = c2 + + def forward(self, x): + if self.conv is not None: + x = self.conv(x) + b, _, w, h = x.shape + p = x.flatten(2).permute(2, 0, 1) + return self.tr(p + self.linear(p)).permute(1, 2, 0).reshape(b, self.c2, w, h) + + +class Bottleneck(nn.Module): + # Standard bottleneck + def __init__(self, c1, c2, shortcut=True, g=1, e=0.5): # ch_in, ch_out, shortcut, groups, expansion + super().__init__() + c_ = int(c2 * e) # hidden channels + self.cv1 = Conv(c1, c_, 1, 1) + self.cv2 = Conv(c_, c2, 3, 1, g=g) + self.add = shortcut and c1 == c2 + + def forward(self, x): + return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x)) + + +class BottleneckCSP(nn.Module): + # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks + def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion + super().__init__() + c_ = int(c2 * e) # hidden channels + self.cv1 = Conv(c1, c_, 1, 1) + self.cv2 = nn.Conv2d(c1, c_, 1, 1, bias=False) + self.cv3 = nn.Conv2d(c_, c_, 1, 1, bias=False) + self.cv4 = Conv(2 * c_, c2, 1, 1) + self.bn = nn.BatchNorm2d(2 * c_) # applied to cat(cv2, cv3) + self.act = nn.SiLU() + self.m = nn.Sequential(*(Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n))) + + def forward(self, x): + y1 = self.cv3(self.m(self.cv1(x))) + y2 = self.cv2(x) + return self.cv4(self.act(self.bn(torch.cat((y1, y2), 1)))) + + +class CrossConv(nn.Module): + # Cross Convolution Downsample + def __init__(self, c1, c2, k=3, s=1, g=1, e=1.0, shortcut=False): + # ch_in, ch_out, kernel, stride, groups, expansion, shortcut + super().__init__() + c_ = int(c2 * e) # hidden channels + self.cv1 = Conv(c1, c_, (1, k), (1, s)) + self.cv2 = Conv(c_, c2, (k, 1), (s, 1), g=g) + self.add = shortcut and c1 == c2 + + def forward(self, x): + return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x)) + + +class C3(nn.Module): + # CSP Bottleneck with 3 convolutions + def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion + super().__init__() + c_ = int(c2 * e) # hidden channels + self.cv1 = Conv(c1, c_, 1, 1) + self.cv2 = Conv(c1, c_, 1, 1) + self.cv3 = Conv(2 * c_, c2, 1) # optional act=FReLU(c2) + self.m = nn.Sequential(*(Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n))) + + def forward(self, x): + return self.cv3(torch.cat((self.m(self.cv1(x)), self.cv2(x)), 1)) + + +class C3x(C3): + # C3 module with cross-convolutions + def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): + super().__init__(c1, c2, n, shortcut, g, e) + c_ = int(c2 * e) + self.m = nn.Sequential(*(CrossConv(c_, c_, 3, 1, g, 1.0, shortcut) for _ in range(n))) + + +class C3TR(C3): + # C3 module with TransformerBlock() + def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): + super().__init__(c1, c2, n, shortcut, g, e) + c_ = int(c2 * e) + self.m = TransformerBlock(c_, c_, 4, n) + + +class C3SPP(C3): + # C3 module with SPP() + def __init__(self, c1, c2, k=(5, 9, 13), n=1, shortcut=True, g=1, e=0.5): + super().__init__(c1, c2, n, shortcut, g, e) + c_ = int(c2 * e) + self.m = SPP(c_, c_, k) + + +class C3Ghost(C3): + # C3 module with GhostBottleneck() + def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): + super().__init__(c1, c2, n, shortcut, g, e) + c_ = int(c2 * e) # hidden channels + self.m = nn.Sequential(*(GhostBottleneck(c_, c_) for _ in range(n))) + + +class SPP(nn.Module): + # Spatial Pyramid Pooling (SPP) layer https://arxiv.org/abs/1406.4729 + def __init__(self, c1, c2, k=(5, 9, 13)): + super().__init__() + c_ = c1 // 2 # hidden channels + self.cv1 = Conv(c1, c_, 1, 1) + self.cv2 = Conv(c_ * (len(k) + 1), c2, 1, 1) + self.m = nn.ModuleList([nn.MaxPool2d(kernel_size=x, stride=1, padding=x // 2) for x in k]) + + def forward(self, x): + x = self.cv1(x) + with warnings.catch_warnings(): + warnings.simplefilter('ignore') # suppress torch 1.9.0 max_pool2d() warning + return self.cv2(torch.cat([x] + [m(x) for m in self.m], 1)) + + +class SPPF(nn.Module): + # Spatial Pyramid Pooling - Fast (SPPF) layer for YOLOv5 by Glenn Jocher + def __init__(self, c1, c2, k=5): # equivalent to SPP(k=(5, 9, 13)) + super().__init__() + c_ = c1 // 2 # hidden channels + self.cv1 = Conv(c1, c_, 1, 1) + self.cv2 = Conv(c_ * 4, c2, 1, 1) + self.m = nn.MaxPool2d(kernel_size=k, stride=1, padding=k // 2) + + def forward(self, x): + x = self.cv1(x) + with warnings.catch_warnings(): + warnings.simplefilter('ignore') # suppress torch 1.9.0 max_pool2d() warning + y1 = self.m(x) + y2 = self.m(y1) + return self.cv2(torch.cat((x, y1, y2, self.m(y2)), 1)) + + +class Focus(nn.Module): + # Focus wh information into c-space + def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True): # ch_in, ch_out, kernel, stride, padding, groups + super().__init__() + self.conv = Conv(c1 * 4, c2, k, s, p, g, act) + # self.contract = Contract(gain=2) + + def forward(self, x): # x(b,c,w,h) -> y(b,4c,w/2,h/2) + return self.conv(torch.cat((x[..., ::2, ::2], x[..., 1::2, ::2], x[..., ::2, 1::2], x[..., 1::2, 1::2]), 1)) + # return self.conv(self.contract(x)) + + +class GhostConv(nn.Module): + # Ghost Convolution https://github.com/huawei-noah/ghostnet + def __init__(self, c1, c2, k=1, s=1, g=1, act=True): # ch_in, ch_out, kernel, stride, groups + super().__init__() + c_ = c2 // 2 # hidden channels + self.cv1 = Conv(c1, c_, k, s, None, g, act) + self.cv2 = Conv(c_, c_, 5, 1, None, c_, act) + + def forward(self, x): + y = self.cv1(x) + return torch.cat((y, self.cv2(y)), 1) + + +class GhostBottleneck(nn.Module): + # Ghost Bottleneck https://github.com/huawei-noah/ghostnet + def __init__(self, c1, c2, k=3, s=1): # ch_in, ch_out, kernel, stride + super().__init__() + c_ = c2 // 2 + self.conv = nn.Sequential( + GhostConv(c1, c_, 1, 1), # pw + DWConv(c_, c_, k, s, act=False) if s == 2 else nn.Identity(), # dw + GhostConv(c_, c2, 1, 1, act=False)) # pw-linear + self.shortcut = nn.Sequential(DWConv(c1, c1, k, s, act=False), Conv(c1, c2, 1, 1, + act=False)) if s == 2 else nn.Identity() + + def forward(self, x): + return self.conv(x) + self.shortcut(x) + + +class Contract(nn.Module): + # Contract width-height into channels, i.e. x(1,64,80,80) to x(1,256,40,40) + def __init__(self, gain=2): + super().__init__() + self.gain = gain + + def forward(self, x): + b, c, h, w = x.size() # assert (h / s == 0) and (W / s == 0), 'Indivisible gain' + s = self.gain + x = x.view(b, c, h // s, s, w // s, s) # x(1,64,40,2,40,2) + x = x.permute(0, 3, 5, 1, 2, 4).contiguous() # x(1,2,2,64,40,40) + return x.view(b, c * s * s, h // s, w // s) # x(1,256,40,40) + + +class Expand(nn.Module): + # Expand channels into width-height, i.e. x(1,64,80,80) to x(1,16,160,160) + def __init__(self, gain=2): + super().__init__() + self.gain = gain + + def forward(self, x): + b, c, h, w = x.size() # assert C / s ** 2 == 0, 'Indivisible gain' + s = self.gain + x = x.view(b, s, s, c // s ** 2, h, w) # x(1,2,2,16,80,80) + x = x.permute(0, 3, 4, 1, 5, 2).contiguous() # x(1,16,80,2,80,2) + return x.view(b, c // s ** 2, h * s, w * s) # x(1,16,160,160) + + +class Concat(nn.Module): + # Concatenate a list of tensors along dimension + def __init__(self, dimension=1): + super().__init__() + self.d = dimension + + def forward(self, x): + return torch.cat(x, self.d) + + +class DetectMultiBackend(nn.Module): + # YOLOv5 MultiBackend class for python inference on various backends + def __init__(self, weights='yolov5s.pt', device=torch.device('cpu'), dnn=False, data=None, fp16=False): + # Usage: + # PyTorch: weights = *.pt + # TorchScript: *.torchscript + # ONNX Runtime: *.onnx + # ONNX OpenCV DNN: *.onnx with --dnn + # OpenVINO: *.xml + # CoreML: *.mlmodel + # TensorRT: *.engine + # TensorFlow SavedModel: *_saved_model + # TensorFlow GraphDef: *.pb + # TensorFlow Lite: *.tflite + # TensorFlow Edge TPU: *_edgetpu.tflite + from models.experimental import attempt_download, attempt_load # scoped to avoid circular import + + super().__init__() + w = str(weights[0] if isinstance(weights, list) else weights) + pt, jit, onnx, xml, engine, coreml, saved_model, pb, tflite, edgetpu, tfjs = self.model_type(w) # get backend + w = attempt_download(w) # download if not local + fp16 &= (pt or jit or onnx or engine) and device.type != 'cpu' # FP16 + stride, names = 32, [f'class{i}' for i in range(1000)] # assign defaults + if data: # assign class names (optional) + with open(data, errors='ignore') as f: + names = yaml.safe_load(f)['names'] + + if pt: # PyTorch + model = attempt_load(weights if isinstance(weights, list) else w, device=device) + stride = max(int(model.stride.max()), 32) # model stride + names = model.module.names if hasattr(model, 'module') else model.names # get class names + model.half() if fp16 else model.float() + self.model = model # explicitly assign for to(), cpu(), cuda(), half() + elif jit: # TorchScript + LOGGER.info(f'Loading {w} for TorchScript inference...') + extra_files = {'config.txt': ''} # model metadata + model = torch.jit.load(w, _extra_files=extra_files) + model.half() if fp16 else model.float() + if extra_files['config.txt']: + d = json.loads(extra_files['config.txt']) # extra_files dict + stride, names = int(d['stride']), d['names'] + elif dnn: # ONNX OpenCV DNN + LOGGER.info(f'Loading {w} for ONNX OpenCV DNN inference...') + check_requirements(('opencv-python>=4.5.4',)) + net = cv2.dnn.readNetFromONNX(w) + elif onnx: # ONNX Runtime + LOGGER.info(f'Loading {w} for ONNX Runtime inference...') + cuda = torch.cuda.is_available() + check_requirements(('onnx', 'onnxruntime-gpu' if cuda else 'onnxruntime')) + import onnxruntime + providers = ['CUDAExecutionProvider', 'CPUExecutionProvider'] if cuda else ['CPUExecutionProvider'] + session = onnxruntime.InferenceSession(w, providers=providers) + meta = session.get_modelmeta().custom_metadata_map # metadata + if 'stride' in meta: + stride, names = int(meta['stride']), eval(meta['names']) + elif xml: # OpenVINO + LOGGER.info(f'Loading {w} for OpenVINO inference...') + check_requirements(('openvino',)) # requires openvino-dev: https://pypi.org/project/openvino-dev/ + from openvino.runtime import Core + ie = Core() + if not Path(w).is_file(): # if not *.xml + w = next(Path(w).glob('*.xml')) # get *.xml file from *_openvino_model dir + network = ie.read_model(model=w, weights=Path(w).with_suffix('.bin')) + executable_network = ie.compile_model(model=network, device_name="CPU") + output_layer = next(iter(executable_network.outputs)) + meta = Path(w).with_suffix('.yaml') + if meta.exists(): + stride, names = self._load_metadata(meta) # load metadata + elif engine: # TensorRT + LOGGER.info(f'Loading {w} for TensorRT inference...') + import tensorrt as trt # https://developer.nvidia.com/nvidia-tensorrt-download + check_version(trt.__version__, '7.0.0', hard=True) # require tensorrt>=7.0.0 + Binding = namedtuple('Binding', ('name', 'dtype', 'shape', 'data', 'ptr')) + logger = trt.Logger(trt.Logger.INFO) + with open(w, 'rb') as f, trt.Runtime(logger) as runtime: + model = runtime.deserialize_cuda_engine(f.read()) + bindings = OrderedDict() + fp16 = False # default updated below + for index in range(model.num_bindings): + name = model.get_binding_name(index) + dtype = trt.nptype(model.get_binding_dtype(index)) + shape = tuple(model.get_binding_shape(index)) + data = torch.from_numpy(np.empty(shape, dtype=np.dtype(dtype))).to(device) + bindings[name] = Binding(name, dtype, shape, data, int(data.data_ptr())) + if model.binding_is_input(index) and dtype == np.float16: + fp16 = True + binding_addrs = OrderedDict((n, d.ptr) for n, d in bindings.items()) + context = model.create_execution_context() + batch_size = bindings['images'].shape[0] + elif coreml: # CoreML + LOGGER.info(f'Loading {w} for CoreML inference...') + import coremltools as ct + model = ct.models.MLModel(w) + else: # TensorFlow (SavedModel, GraphDef, Lite, Edge TPU) + if saved_model: # SavedModel + LOGGER.info(f'Loading {w} for TensorFlow SavedModel inference...') + import tensorflow as tf + keras = False # assume TF1 saved_model + model = tf.keras.models.load_model(w) if keras else tf.saved_model.load(w) + elif pb: # GraphDef https://www.tensorflow.org/guide/migrate#a_graphpb_or_graphpbtxt + LOGGER.info(f'Loading {w} for TensorFlow GraphDef inference...') + import tensorflow as tf + + def wrap_frozen_graph(gd, inputs, outputs): + x = tf.compat.v1.wrap_function(lambda: tf.compat.v1.import_graph_def(gd, name=""), []) # wrapped + ge = x.graph.as_graph_element + return x.prune(tf.nest.map_structure(ge, inputs), tf.nest.map_structure(ge, outputs)) + + gd = tf.Graph().as_graph_def() # graph_def + with open(w, 'rb') as f: + gd.ParseFromString(f.read()) + frozen_func = wrap_frozen_graph(gd, inputs="x:0", outputs="Identity:0") + elif tflite or edgetpu: # https://www.tensorflow.org/lite/guide/python#install_tensorflow_lite_for_python + try: # https://coral.ai/docs/edgetpu/tflite-python/#update-existing-tf-lite-code-for-the-edge-tpu + from tflite_runtime.interpreter import Interpreter, load_delegate + except ImportError: + import tensorflow as tf + Interpreter, load_delegate = tf.lite.Interpreter, tf.lite.experimental.load_delegate, + if edgetpu: # Edge TPU https://coral.ai/software/#edgetpu-runtime + LOGGER.info(f'Loading {w} for TensorFlow Lite Edge TPU inference...') + delegate = { + 'Linux': 'libedgetpu.so.1', + 'Darwin': 'libedgetpu.1.dylib', + 'Windows': 'edgetpu.dll'}[platform.system()] + interpreter = Interpreter(model_path=w, experimental_delegates=[load_delegate(delegate)]) + else: # Lite + LOGGER.info(f'Loading {w} for TensorFlow Lite inference...') + interpreter = Interpreter(model_path=w) # load TFLite model + interpreter.allocate_tensors() # allocate + input_details = interpreter.get_input_details() # inputs + output_details = interpreter.get_output_details() # outputs + elif tfjs: + raise Exception('ERROR: YOLOv5 TF.js inference is not supported') + self.__dict__.update(locals()) # assign all variables to self + + def forward(self, im, augment=False, visualize=False, val=False): + # YOLOv5 MultiBackend inference + b, ch, h, w = im.shape # batch, channel, height, width + if self.pt: # PyTorch + y = self.model(im, augment=augment, visualize=visualize)[0] + elif self.jit: # TorchScript + y = self.model(im)[0] + elif self.dnn: # ONNX OpenCV DNN + im = im.cpu().numpy() # torch to numpy + self.net.setInput(im) + y = self.net.forward() + elif self.onnx: # ONNX Runtime + im = im.cpu().numpy() # torch to numpy + y = self.session.run([self.session.get_outputs()[0].name], {self.session.get_inputs()[0].name: im})[0] + elif self.xml: # OpenVINO + im = im.cpu().numpy() # FP32 + y = self.executable_network([im])[self.output_layer] + elif self.engine: # TensorRT + assert im.shape == self.bindings['images'].shape, (im.shape, self.bindings['images'].shape) + self.binding_addrs['images'] = int(im.data_ptr()) + self.context.execute_v2(list(self.binding_addrs.values())) + y = self.bindings['output'].data + elif self.coreml: # CoreML + im = im.permute(0, 2, 3, 1).cpu().numpy() # torch BCHW to numpy BHWC shape(1,320,192,3) + im = Image.fromarray((im[0] * 255).astype('uint8')) + # im = im.resize((192, 320), Image.ANTIALIAS) + y = self.model.predict({'image': im}) # coordinates are xywh normalized + if 'confidence' in y: + box = xywh2xyxy(y['coordinates'] * [[w, h, w, h]]) # xyxy pixels + conf, cls = y['confidence'].max(1), y['confidence'].argmax(1).astype(np.float) + y = np.concatenate((box, conf.reshape(-1, 1), cls.reshape(-1, 1)), 1) + else: + k = 'var_' + str(sorted(int(k.replace('var_', '')) for k in y)[-1]) # output key + y = y[k] # output + else: # TensorFlow (SavedModel, GraphDef, Lite, Edge TPU) + im = im.permute(0, 2, 3, 1).cpu().numpy() # torch BCHW to numpy BHWC shape(1,320,192,3) + if self.saved_model: # SavedModel + y = (self.model(im, training=False) if self.keras else self.model(im)).numpy() + elif self.pb: # GraphDef + y = self.frozen_func(x=self.tf.constant(im)).numpy() + else: # Lite or Edge TPU + input, output = self.input_details[0], self.output_details[0] + int8 = input['dtype'] == np.uint8 # is TFLite quantized uint8 model + if int8: + scale, zero_point = input['quantization'] + im = (im / scale + zero_point).astype(np.uint8) # de-scale + self.interpreter.set_tensor(input['index'], im) + self.interpreter.invoke() + y = self.interpreter.get_tensor(output['index']) + if int8: + scale, zero_point = output['quantization'] + y = (y.astype(np.float32) - zero_point) * scale # re-scale + y[..., :4] *= [w, h, w, h] # xywh normalized to pixels + + if isinstance(y, np.ndarray): + y = torch.tensor(y, device=self.device) + return (y, []) if val else y + + def warmup(self, imgsz=(1, 3, 640, 640)): + # Warmup model by running inference once + warmup_types = self.pt, self.jit, self.onnx, self.engine, self.saved_model, self.pb + if any(warmup_types) and self.device.type != 'cpu': + im = torch.zeros(*imgsz, dtype=torch.half if self.fp16 else torch.float, device=self.device) # input + for _ in range(2 if self.jit else 1): # + self.forward(im) # warmup + + @staticmethod + def model_type(p='path/to/model.pt'): + # Return model type from model path, i.e. path='path/to/model.onnx' -> type=onnx + from export import export_formats + suffixes = list(export_formats().Suffix) + ['.xml'] # export suffixes + check_suffix(p, suffixes) # checks + p = Path(p).name # eliminate trailing separators + pt, jit, onnx, xml, engine, coreml, saved_model, pb, tflite, edgetpu, tfjs, xml2 = (s in p for s in suffixes) + xml |= xml2 # *_openvino_model or *.xml + tflite &= not edgetpu # *.tflite + return pt, jit, onnx, xml, engine, coreml, saved_model, pb, tflite, edgetpu, tfjs + + @staticmethod + def _load_metadata(f='path/to/meta.yaml'): + # Load metadata from meta.yaml if it exists + with open(f, errors='ignore') as f: + d = yaml.safe_load(f) + return d['stride'], d['names'] # assign stride, names + + +class AutoShape(nn.Module): + # YOLOv5 input-robust model wrapper for passing cv2/np/PIL/torch inputs. Includes preprocessing, inference and NMS + conf = 0.25 # NMS confidence threshold + iou = 0.45 # NMS IoU threshold + agnostic = False # NMS class-agnostic + multi_label = False # NMS multiple labels per box + classes = None # (optional list) filter by class, i.e. = [0, 15, 16] for COCO persons, cats and dogs + max_det = 1000 # maximum number of detections per image + amp = False # Automatic Mixed Precision (AMP) inference + + def __init__(self, model, verbose=True): + super().__init__() + if verbose: + LOGGER.info('Adding AutoShape... ') + copy_attr(self, model, include=('yaml', 'nc', 'hyp', 'names', 'stride', 'abc'), exclude=()) # copy attributes + self.dmb = isinstance(model, DetectMultiBackend) # DetectMultiBackend() instance + self.pt = not self.dmb or model.pt # PyTorch model + self.model = model.eval() + + def _apply(self, fn): + # Apply to(), cpu(), cuda(), half() to model tensors that are not parameters or registered buffers + self = super()._apply(fn) + if self.pt: + m = self.model.model.model[-1] if self.dmb else self.model.model[-1] # Detect() + m.stride = fn(m.stride) + m.grid = list(map(fn, m.grid)) + if isinstance(m.anchor_grid, list): + m.anchor_grid = list(map(fn, m.anchor_grid)) + return self + + @torch.no_grad() + def forward(self, imgs, size=640, augment=False, profile=False): + # Inference from various sources. For height=640, width=1280, RGB images example inputs are: + # file: imgs = 'data/images/zidane.jpg' # str or PosixPath + # URI: = 'https://ultralytics.com/images/zidane.jpg' + # OpenCV: = cv2.imread('image.jpg')[:,:,::-1] # HWC BGR to RGB x(640,1280,3) + # PIL: = Image.open('image.jpg') or ImageGrab.grab() # HWC x(640,1280,3) + # numpy: = np.zeros((640,1280,3)) # HWC + # torch: = torch.zeros(16,3,320,640) # BCHW (scaled to size=640, 0-1 values) + # multiple: = [Image.open('image1.jpg'), Image.open('image2.jpg'), ...] # list of images + + t = [time_sync()] + p = next(self.model.parameters()) if self.pt else torch.zeros(1, device=self.model.device) # for device, type + autocast = self.amp and (p.device.type != 'cpu') # Automatic Mixed Precision (AMP) inference + if isinstance(imgs, torch.Tensor): # torch + with amp.autocast(autocast): + return self.model(imgs.to(p.device).type_as(p), augment, profile) # inference + + # Pre-process + n, imgs = (len(imgs), list(imgs)) if isinstance(imgs, (list, tuple)) else (1, [imgs]) # number, list of images + shape0, shape1, files = [], [], [] # image and inference shapes, filenames + for i, im in enumerate(imgs): + f = f'image{i}' # filename + if isinstance(im, (str, Path)): # filename or uri + im, f = Image.open(requests.get(im, stream=True).raw if str(im).startswith('http') else im), im + im = np.asarray(exif_transpose(im)) + elif isinstance(im, Image.Image): # PIL Image + im, f = np.asarray(exif_transpose(im)), getattr(im, 'filename', f) or f + files.append(Path(f).with_suffix('.jpg').name) + if im.shape[0] < 5: # image in CHW + im = im.transpose((1, 2, 0)) # reverse dataloader .transpose(2, 0, 1) + im = im[..., :3] if im.ndim == 3 else np.tile(im[..., None], 3) # enforce 3ch input + s = im.shape[:2] # HWC + shape0.append(s) # image shape + g = (size / max(s)) # gain + shape1.append([y * g for y in s]) + imgs[i] = im if im.data.contiguous else np.ascontiguousarray(im) # update + shape1 = [make_divisible(x, self.stride) if self.pt else size for x in np.array(shape1).max(0)] # inf shape + x = [letterbox(im, shape1, auto=False)[0] for im in imgs] # pad + x = np.ascontiguousarray(np.array(x).transpose((0, 3, 1, 2))) # stack and BHWC to BCHW + x = torch.from_numpy(x).to(p.device).type_as(p) / 255 # uint8 to fp16/32 + t.append(time_sync()) + + with amp.autocast(autocast): + # Inference + y = self.model(x, augment, profile) # forward + t.append(time_sync()) + + # Post-process + y = non_max_suppression(y if self.dmb else y[0], + self.conf, + self.iou, + self.classes, + self.agnostic, + self.multi_label, + max_det=self.max_det) # NMS + for i in range(n): + scale_coords(shape1, y[i][:, :4], shape0[i]) + + t.append(time_sync()) + return Detections(imgs, y, files, t, self.names, x.shape) + + +class Detections: + # YOLOv5 detections class for inference results + def __init__(self, imgs, pred, files, times=(0, 0, 0, 0), names=None, shape=None): + super().__init__() + d = pred[0].device # device + gn = [torch.tensor([*(im.shape[i] for i in [1, 0, 1, 0]), 1, 1], device=d) for im in imgs] # normalizations + self.imgs = imgs # list of images as numpy arrays + self.pred = pred # list of tensors pred[0] = (xyxy, conf, cls) + self.names = names # class names + self.files = files # image filenames + self.times = times # profiling times + self.xyxy = pred # xyxy pixels + self.xywh = [xyxy2xywh(x) for x in pred] # xywh pixels + self.xyxyn = [x / g for x, g in zip(self.xyxy, gn)] # xyxy normalized + self.xywhn = [x / g for x, g in zip(self.xywh, gn)] # xywh normalized + self.n = len(self.pred) # number of images (batch size) + self.t = tuple((times[i + 1] - times[i]) * 1000 / self.n for i in range(3)) # timestamps (ms) + self.s = shape # inference BCHW shape + + def display(self, pprint=False, show=False, save=False, crop=False, render=False, labels=True, save_dir=Path('')): + crops = [] + for i, (im, pred) in enumerate(zip(self.imgs, self.pred)): + s = f'image {i + 1}/{len(self.pred)}: {im.shape[0]}x{im.shape[1]} ' # string + if pred.shape[0]: + for c in pred[:, -1].unique(): + n = (pred[:, -1] == c).sum() # detections per class + s += f"{n} {self.names[int(c)]}{'s' * (n > 1)}, " # add to string + if show or save or render or crop: + annotator = Annotator(im, example=str(self.names)) + for *box, conf, cls in reversed(pred): # xyxy, confidence, class + label = f'{self.names[int(cls)]} {conf:.2f}' + if crop: + file = save_dir / 'crops' / self.names[int(cls)] / self.files[i] if save else None + crops.append({ + 'box': box, + 'conf': conf, + 'cls': cls, + 'label': label, + 'im': save_one_box(box, im, file=file, save=save)}) + else: # all others + annotator.box_label(box, label if labels else '', color=colors(cls)) + im = annotator.im + else: + s += '(no detections)' + + im = Image.fromarray(im.astype(np.uint8)) if isinstance(im, np.ndarray) else im # from np + if pprint: + print(s.rstrip(', ')) + if show: + im.show(self.files[i]) # show + if save: + f = self.files[i] + im.save(save_dir / f) # save + if i == self.n - 1: + LOGGER.info(f"Saved {self.n} image{'s' * (self.n > 1)} to {colorstr('bold', save_dir)}") + if render: + self.imgs[i] = np.asarray(im) + if crop: + if save: + LOGGER.info(f'Saved results to {save_dir}\n') + return crops + + def print(self): + self.display(pprint=True) # print results + print(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {tuple(self.s)}' % self.t) + + def show(self, labels=True): + self.display(show=True, labels=labels) # show results + + def save(self, labels=True, save_dir='runs/detect/exp'): + save_dir = increment_path(save_dir, exist_ok=save_dir != 'runs/detect/exp', mkdir=True) # increment save_dir + self.display(save=True, labels=labels, save_dir=save_dir) # save results + + def crop(self, save=True, save_dir='runs/detect/exp'): + save_dir = increment_path(save_dir, exist_ok=save_dir != 'runs/detect/exp', mkdir=True) if save else None + return self.display(crop=True, save=save, save_dir=save_dir) # crop results + + def render(self, labels=True): + self.display(render=True, labels=labels) # render results + return self.imgs + + def pandas(self): + # return detections as pandas DataFrames, i.e. print(results.pandas().xyxy[0]) + new = copy(self) # return copy + ca = 'xmin', 'ymin', 'xmax', 'ymax', 'confidence', 'class', 'name' # xyxy columns + cb = 'xcenter', 'ycenter', 'width', 'height', 'confidence', 'class', 'name' # xywh columns + for k, c in zip(['xyxy', 'xyxyn', 'xywh', 'xywhn'], [ca, ca, cb, cb]): + a = [[x[:5] + [int(x[5]), self.names[int(x[5])]] for x in x.tolist()] for x in getattr(self, k)] # update + setattr(new, k, [pd.DataFrame(x, columns=c) for x in a]) + return new + + def tolist(self): + # return a list of Detections objects, i.e. 'for result in results.tolist():' + r = range(self.n) # iterable + x = [Detections([self.imgs[i]], [self.pred[i]], [self.files[i]], self.times, self.names, self.s) for i in r] + # for d in x: + # for k in ['imgs', 'pred', 'xyxy', 'xyxyn', 'xywh', 'xywhn']: + # setattr(d, k, getattr(d, k)[0]) # pop out of list + return x + + def __len__(self): + return self.n # override len(results) + + def __str__(self): + self.print() # override print(results) + return '' + + +class Classify(nn.Module): + # Classification head, i.e. x(b,c1,20,20) to x(b,c2) + def __init__(self, c1, c2, k=1, s=1, p=None, g=1): # ch_in, ch_out, kernel, stride, padding, groups + super().__init__() + self.aap = nn.AdaptiveAvgPool2d(1) # to x(b,c1,1,1) + self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g) # to x(b,c2,1,1) + self.flat = nn.Flatten() + + def forward(self, x): + z = torch.cat([self.aap(y) for y in (x if isinstance(x, list) else [x])], 1) # cat if list + return self.flat(self.conv(z)) # flatten to x(b,c2) diff --git a/yolov5/models/experimental.py b/yolov5/models/experimental.py new file mode 100644 index 0000000..c2d9345 --- /dev/null +++ b/yolov5/models/experimental.py @@ -0,0 +1,104 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Experimental modules +""" +import math + +import numpy as np +import torch +import torch.nn as nn + +from models.common import Conv +from utils.downloads import attempt_download + + +class Sum(nn.Module): + # Weighted sum of 2 or more layers https://arxiv.org/abs/1911.09070 + def __init__(self, n, weight=False): # n: number of inputs + super().__init__() + self.weight = weight # apply weights boolean + self.iter = range(n - 1) # iter object + if weight: + self.w = nn.Parameter(-torch.arange(1.0, n) / 2, requires_grad=True) # layer weights + + def forward(self, x): + y = x[0] # no weight + if self.weight: + w = torch.sigmoid(self.w) * 2 + for i in self.iter: + y = y + x[i + 1] * w[i] + else: + for i in self.iter: + y = y + x[i + 1] + return y + + +class MixConv2d(nn.Module): + # Mixed Depth-wise Conv https://arxiv.org/abs/1907.09595 + def __init__(self, c1, c2, k=(1, 3), s=1, equal_ch=True): # ch_in, ch_out, kernel, stride, ch_strategy + super().__init__() + n = len(k) # number of convolutions + if equal_ch: # equal c_ per group + i = torch.linspace(0, n - 1E-6, c2).floor() # c2 indices + c_ = [(i == g).sum() for g in range(n)] # intermediate channels + else: # equal weight.numel() per group + b = [c2] + [0] * n + a = np.eye(n + 1, n, k=-1) + a -= np.roll(a, 1, axis=1) + a *= np.array(k) ** 2 + a[0] = 1 + c_ = np.linalg.lstsq(a, b, rcond=None)[0].round() # solve for equal weight indices, ax = b + + self.m = nn.ModuleList([ + nn.Conv2d(c1, int(c_), k, s, k // 2, groups=math.gcd(c1, int(c_)), bias=False) for k, c_ in zip(k, c_)]) + self.bn = nn.BatchNorm2d(c2) + self.act = nn.SiLU() + + def forward(self, x): + return self.act(self.bn(torch.cat([m(x) for m in self.m], 1))) + + +class Ensemble(nn.ModuleList): + # Ensemble of models + def __init__(self): + super().__init__() + + def forward(self, x, augment=False, profile=False, visualize=False): + y = [module(x, augment, profile, visualize)[0] for module in self] + # y = torch.stack(y).max(0)[0] # max ensemble + # y = torch.stack(y).mean(0) # mean ensemble + y = torch.cat(y, 1) # nms ensemble + return y, None # inference, train output + + +def attempt_load(weights, device=None, inplace=True, fuse=True): + from models.yolo import Detect, Model + + # Loads an ensemble of models weights=[a,b,c] or a single model weights=[a] or weights=a + model = Ensemble() + for w in weights if isinstance(weights, list) else [weights]: + ckpt = torch.load(attempt_download(w), map_location=device) + ckpt = (ckpt.get('ema') or ckpt['model']).float() # FP32 model + model.append(ckpt.fuse().eval() if fuse else ckpt.eval()) # fused or un-fused model in eval mode + + # Compatibility updates + for m in model.modules(): + t = type(m) + if t in (nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6, nn.SiLU, Detect, Model): + m.inplace = inplace # torch 1.7.0 compatibility + if t is Detect and not isinstance(m.anchor_grid, list): + delattr(m, 'anchor_grid') + setattr(m, 'anchor_grid', [torch.zeros(1)] * m.nl) + elif t is Conv: + m._non_persistent_buffers_set = set() # torch 1.6.0 compatibility + elif t is nn.Upsample and not hasattr(m, 'recompute_scale_factor'): + m.recompute_scale_factor = None # torch 1.11.0 compatibility + + if len(model) == 1: + return model[-1] # return model + print(f'Ensemble created with {weights}\n') + for k in 'names', 'nc', 'yaml': + setattr(model, k, getattr(model[0], k)) + model.stride = model[torch.argmax(torch.tensor([m.stride.max() for m in model])).int()].stride # max stride + assert all(model[0].nc == m.nc for m in model), f'Models have different class counts: {[m.nc for m in model]}' + return model # return ensemble diff --git a/yolov5/models/hub/anchors.yaml b/yolov5/models/hub/anchors.yaml new file mode 100644 index 0000000..e4d7beb --- /dev/null +++ b/yolov5/models/hub/anchors.yaml @@ -0,0 +1,59 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# Default anchors for COCO data + + +# P5 ------------------------------------------------------------------------------------------------------------------- +# P5-640: +anchors_p5_640: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + + +# P6 ------------------------------------------------------------------------------------------------------------------- +# P6-640: thr=0.25: 0.9964 BPR, 5.54 anchors past thr, n=12, img_size=640, metric_all=0.281/0.716-mean/best, past_thr=0.469-mean: 9,11, 21,19, 17,41, 43,32, 39,70, 86,64, 65,131, 134,130, 120,265, 282,180, 247,354, 512,387 +anchors_p6_640: + - [9,11, 21,19, 17,41] # P3/8 + - [43,32, 39,70, 86,64] # P4/16 + - [65,131, 134,130, 120,265] # P5/32 + - [282,180, 247,354, 512,387] # P6/64 + +# P6-1280: thr=0.25: 0.9950 BPR, 5.55 anchors past thr, n=12, img_size=1280, metric_all=0.281/0.714-mean/best, past_thr=0.468-mean: 19,27, 44,40, 38,94, 96,68, 86,152, 180,137, 140,301, 303,264, 238,542, 436,615, 739,380, 925,792 +anchors_p6_1280: + - [19,27, 44,40, 38,94] # P3/8 + - [96,68, 86,152, 180,137] # P4/16 + - [140,301, 303,264, 238,542] # P5/32 + - [436,615, 739,380, 925,792] # P6/64 + +# P6-1920: thr=0.25: 0.9950 BPR, 5.55 anchors past thr, n=12, img_size=1920, metric_all=0.281/0.714-mean/best, past_thr=0.468-mean: 28,41, 67,59, 57,141, 144,103, 129,227, 270,205, 209,452, 455,396, 358,812, 653,922, 1109,570, 1387,1187 +anchors_p6_1920: + - [28,41, 67,59, 57,141] # P3/8 + - [144,103, 129,227, 270,205] # P4/16 + - [209,452, 455,396, 358,812] # P5/32 + - [653,922, 1109,570, 1387,1187] # P6/64 + + +# P7 ------------------------------------------------------------------------------------------------------------------- +# P7-640: thr=0.25: 0.9962 BPR, 6.76 anchors past thr, n=15, img_size=640, metric_all=0.275/0.733-mean/best, past_thr=0.466-mean: 11,11, 13,30, 29,20, 30,46, 61,38, 39,92, 78,80, 146,66, 79,163, 149,150, 321,143, 157,303, 257,402, 359,290, 524,372 +anchors_p7_640: + - [11,11, 13,30, 29,20] # P3/8 + - [30,46, 61,38, 39,92] # P4/16 + - [78,80, 146,66, 79,163] # P5/32 + - [149,150, 321,143, 157,303] # P6/64 + - [257,402, 359,290, 524,372] # P7/128 + +# P7-1280: thr=0.25: 0.9968 BPR, 6.71 anchors past thr, n=15, img_size=1280, metric_all=0.273/0.732-mean/best, past_thr=0.463-mean: 19,22, 54,36, 32,77, 70,83, 138,71, 75,173, 165,159, 148,334, 375,151, 334,317, 251,626, 499,474, 750,326, 534,814, 1079,818 +anchors_p7_1280: + - [19,22, 54,36, 32,77] # P3/8 + - [70,83, 138,71, 75,173] # P4/16 + - [165,159, 148,334, 375,151] # P5/32 + - [334,317, 251,626, 499,474] # P6/64 + - [750,326, 534,814, 1079,818] # P7/128 + +# P7-1920: thr=0.25: 0.9968 BPR, 6.71 anchors past thr, n=15, img_size=1920, metric_all=0.273/0.732-mean/best, past_thr=0.463-mean: 29,34, 81,55, 47,115, 105,124, 207,107, 113,259, 247,238, 222,500, 563,227, 501,476, 376,939, 749,711, 1126,489, 801,1222, 1618,1227 +anchors_p7_1920: + - [29,34, 81,55, 47,115] # P3/8 + - [105,124, 207,107, 113,259] # P4/16 + - [247,238, 222,500, 563,227] # P5/32 + - [501,476, 376,939, 749,711] # P6/64 + - [1126,489, 801,1222, 1618,1227] # P7/128 diff --git a/yolov5/models/hub/yolov3-spp.yaml b/yolov5/models/hub/yolov3-spp.yaml new file mode 100644 index 0000000..c669821 --- /dev/null +++ b/yolov5/models/hub/yolov3-spp.yaml @@ -0,0 +1,51 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# darknet53 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [32, 3, 1]], # 0 + [-1, 1, Conv, [64, 3, 2]], # 1-P1/2 + [-1, 1, Bottleneck, [64]], + [-1, 1, Conv, [128, 3, 2]], # 3-P2/4 + [-1, 2, Bottleneck, [128]], + [-1, 1, Conv, [256, 3, 2]], # 5-P3/8 + [-1, 8, Bottleneck, [256]], + [-1, 1, Conv, [512, 3, 2]], # 7-P4/16 + [-1, 8, Bottleneck, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 9-P5/32 + [-1, 4, Bottleneck, [1024]], # 10 + ] + +# YOLOv3-SPP head +head: + [[-1, 1, Bottleneck, [1024, False]], + [-1, 1, SPP, [512, [5, 9, 13]]], + [-1, 1, Conv, [1024, 3, 1]], + [-1, 1, Conv, [512, 1, 1]], + [-1, 1, Conv, [1024, 3, 1]], # 15 (P5/32-large) + + [-2, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 8], 1, Concat, [1]], # cat backbone P4 + [-1, 1, Bottleneck, [512, False]], + [-1, 1, Bottleneck, [512, False]], + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, Conv, [512, 3, 1]], # 22 (P4/16-medium) + + [-2, 1, Conv, [128, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P3 + [-1, 1, Bottleneck, [256, False]], + [-1, 2, Bottleneck, [256, False]], # 27 (P3/8-small) + + [[27, 22, 15], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/yolov5/models/hub/yolov3-tiny.yaml b/yolov5/models/hub/yolov3-tiny.yaml new file mode 100644 index 0000000..b28b443 --- /dev/null +++ b/yolov5/models/hub/yolov3-tiny.yaml @@ -0,0 +1,41 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple +anchors: + - [10,14, 23,27, 37,58] # P4/16 + - [81,82, 135,169, 344,319] # P5/32 + +# YOLOv3-tiny backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [16, 3, 1]], # 0 + [-1, 1, nn.MaxPool2d, [2, 2, 0]], # 1-P1/2 + [-1, 1, Conv, [32, 3, 1]], + [-1, 1, nn.MaxPool2d, [2, 2, 0]], # 3-P2/4 + [-1, 1, Conv, [64, 3, 1]], + [-1, 1, nn.MaxPool2d, [2, 2, 0]], # 5-P3/8 + [-1, 1, Conv, [128, 3, 1]], + [-1, 1, nn.MaxPool2d, [2, 2, 0]], # 7-P4/16 + [-1, 1, Conv, [256, 3, 1]], + [-1, 1, nn.MaxPool2d, [2, 2, 0]], # 9-P5/32 + [-1, 1, Conv, [512, 3, 1]], + [-1, 1, nn.ZeroPad2d, [[0, 1, 0, 1]]], # 11 + [-1, 1, nn.MaxPool2d, [2, 1, 0]], # 12 + ] + +# YOLOv3-tiny head +head: + [[-1, 1, Conv, [1024, 3, 1]], + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, Conv, [512, 3, 1]], # 15 (P5/32-large) + + [-2, 1, Conv, [128, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 8], 1, Concat, [1]], # cat backbone P4 + [-1, 1, Conv, [256, 3, 1]], # 19 (P4/16-medium) + + [[19, 15], 1, Detect, [nc, anchors]], # Detect(P4, P5) + ] diff --git a/yolov5/models/hub/yolov3.yaml b/yolov5/models/hub/yolov3.yaml new file mode 100644 index 0000000..d1ef912 --- /dev/null +++ b/yolov5/models/hub/yolov3.yaml @@ -0,0 +1,51 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# darknet53 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [32, 3, 1]], # 0 + [-1, 1, Conv, [64, 3, 2]], # 1-P1/2 + [-1, 1, Bottleneck, [64]], + [-1, 1, Conv, [128, 3, 2]], # 3-P2/4 + [-1, 2, Bottleneck, [128]], + [-1, 1, Conv, [256, 3, 2]], # 5-P3/8 + [-1, 8, Bottleneck, [256]], + [-1, 1, Conv, [512, 3, 2]], # 7-P4/16 + [-1, 8, Bottleneck, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 9-P5/32 + [-1, 4, Bottleneck, [1024]], # 10 + ] + +# YOLOv3 head +head: + [[-1, 1, Bottleneck, [1024, False]], + [-1, 1, Conv, [512, 1, 1]], + [-1, 1, Conv, [1024, 3, 1]], + [-1, 1, Conv, [512, 1, 1]], + [-1, 1, Conv, [1024, 3, 1]], # 15 (P5/32-large) + + [-2, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 8], 1, Concat, [1]], # cat backbone P4 + [-1, 1, Bottleneck, [512, False]], + [-1, 1, Bottleneck, [512, False]], + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, Conv, [512, 3, 1]], # 22 (P4/16-medium) + + [-2, 1, Conv, [128, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P3 + [-1, 1, Bottleneck, [256, False]], + [-1, 2, Bottleneck, [256, False]], # 27 (P3/8-small) + + [[27, 22, 15], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/yolov5/models/hub/yolov5-bifpn.yaml b/yolov5/models/hub/yolov5-bifpn.yaml new file mode 100644 index 0000000..504815f --- /dev/null +++ b/yolov5/models/hub/yolov5-bifpn.yaml @@ -0,0 +1,48 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 9 + ] + +# YOLOv5 v6.0 BiFPN head +head: + [[-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 13 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 17 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 14, 6], 1, Concat, [1]], # cat P4 <--- BiFPN change + [-1, 3, C3, [512, False]], # 20 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 10], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [1024, False]], # 23 (P5/32-large) + + [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/yolov5/models/hub/yolov5-fpn.yaml b/yolov5/models/hub/yolov5-fpn.yaml new file mode 100644 index 0000000..a23e9c6 --- /dev/null +++ b/yolov5/models/hub/yolov5-fpn.yaml @@ -0,0 +1,42 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 9 + ] + +# YOLOv5 v6.0 FPN head +head: + [[-1, 3, C3, [1024, False]], # 10 (P5/32-large) + + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 1, Conv, [512, 1, 1]], + [-1, 3, C3, [512, False]], # 14 (P4/16-medium) + + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 1, Conv, [256, 1, 1]], + [-1, 3, C3, [256, False]], # 18 (P3/8-small) + + [[18, 14, 10], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/yolov5/models/hub/yolov5-p2.yaml b/yolov5/models/hub/yolov5-p2.yaml new file mode 100644 index 0000000..554117d --- /dev/null +++ b/yolov5/models/hub/yolov5-p2.yaml @@ -0,0 +1,54 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple +anchors: 3 # AutoAnchor evolves 3 anchors per P output layer + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 9 + ] + +# YOLOv5 v6.0 head with (P2, P3, P4, P5) outputs +head: + [[-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 13 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 17 (P3/8-small) + + [-1, 1, Conv, [128, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 2], 1, Concat, [1]], # cat backbone P2 + [-1, 1, C3, [128, False]], # 21 (P2/4-xsmall) + + [-1, 1, Conv, [128, 3, 2]], + [[-1, 18], 1, Concat, [1]], # cat head P3 + [-1, 3, C3, [256, False]], # 24 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 14], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 27 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 10], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [1024, False]], # 30 (P5/32-large) + + [[21, 24, 27, 30], 1, Detect, [nc, anchors]], # Detect(P2, P3, P4, P5) + ] diff --git a/yolov5/models/hub/yolov5-p34.yaml b/yolov5/models/hub/yolov5-p34.yaml new file mode 100644 index 0000000..dbf0f85 --- /dev/null +++ b/yolov5/models/hub/yolov5-p34.yaml @@ -0,0 +1,41 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 0.33 # model depth multiple +width_multiple: 0.50 # layer channel multiple +anchors: 3 # AutoAnchor evolves 3 anchors per P output layer + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [ [ -1, 1, Conv, [ 64, 6, 2, 2 ] ], # 0-P1/2 + [ -1, 1, Conv, [ 128, 3, 2 ] ], # 1-P2/4 + [ -1, 3, C3, [ 128 ] ], + [ -1, 1, Conv, [ 256, 3, 2 ] ], # 3-P3/8 + [ -1, 6, C3, [ 256 ] ], + [ -1, 1, Conv, [ 512, 3, 2 ] ], # 5-P4/16 + [ -1, 9, C3, [ 512 ] ], + [ -1, 1, Conv, [ 1024, 3, 2 ] ], # 7-P5/32 + [ -1, 3, C3, [ 1024 ] ], + [ -1, 1, SPPF, [ 1024, 5 ] ], # 9 + ] + +# YOLOv5 v6.0 head with (P3, P4) outputs +head: + [ [ -1, 1, Conv, [ 512, 1, 1 ] ], + [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], + [ [ -1, 6 ], 1, Concat, [ 1 ] ], # cat backbone P4 + [ -1, 3, C3, [ 512, False ] ], # 13 + + [ -1, 1, Conv, [ 256, 1, 1 ] ], + [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], + [ [ -1, 4 ], 1, Concat, [ 1 ] ], # cat backbone P3 + [ -1, 3, C3, [ 256, False ] ], # 17 (P3/8-small) + + [ -1, 1, Conv, [ 256, 3, 2 ] ], + [ [ -1, 14 ], 1, Concat, [ 1 ] ], # cat head P4 + [ -1, 3, C3, [ 512, False ] ], # 20 (P4/16-medium) + + [ [ 17, 20 ], 1, Detect, [ nc, anchors ] ], # Detect(P3, P4) + ] diff --git a/yolov5/models/hub/yolov5-p6.yaml b/yolov5/models/hub/yolov5-p6.yaml new file mode 100644 index 0000000..a17202f --- /dev/null +++ b/yolov5/models/hub/yolov5-p6.yaml @@ -0,0 +1,56 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple +anchors: 3 # AutoAnchor evolves 3 anchors per P output layer + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [768, 3, 2]], # 7-P5/32 + [-1, 3, C3, [768]], + [-1, 1, Conv, [1024, 3, 2]], # 9-P6/64 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 11 + ] + +# YOLOv5 v6.0 head with (P3, P4, P5, P6) outputs +head: + [[-1, 1, Conv, [768, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 8], 1, Concat, [1]], # cat backbone P5 + [-1, 3, C3, [768, False]], # 15 + + [-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 19 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 23 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 20], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 26 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 16], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [768, False]], # 29 (P5/32-large) + + [-1, 1, Conv, [768, 3, 2]], + [[-1, 12], 1, Concat, [1]], # cat head P6 + [-1, 3, C3, [1024, False]], # 32 (P6/64-xlarge) + + [[23, 26, 29, 32], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5, P6) + ] diff --git a/yolov5/models/hub/yolov5-p7.yaml b/yolov5/models/hub/yolov5-p7.yaml new file mode 100644 index 0000000..edd7d13 --- /dev/null +++ b/yolov5/models/hub/yolov5-p7.yaml @@ -0,0 +1,67 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple +anchors: 3 # AutoAnchor evolves 3 anchors per P output layer + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [768, 3, 2]], # 7-P5/32 + [-1, 3, C3, [768]], + [-1, 1, Conv, [1024, 3, 2]], # 9-P6/64 + [-1, 3, C3, [1024]], + [-1, 1, Conv, [1280, 3, 2]], # 11-P7/128 + [-1, 3, C3, [1280]], + [-1, 1, SPPF, [1280, 5]], # 13 + ] + +# YOLOv5 v6.0 head with (P3, P4, P5, P6, P7) outputs +head: + [[-1, 1, Conv, [1024, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 10], 1, Concat, [1]], # cat backbone P6 + [-1, 3, C3, [1024, False]], # 17 + + [-1, 1, Conv, [768, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 8], 1, Concat, [1]], # cat backbone P5 + [-1, 3, C3, [768, False]], # 21 + + [-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 25 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 29 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 26], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 32 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 22], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [768, False]], # 35 (P5/32-large) + + [-1, 1, Conv, [768, 3, 2]], + [[-1, 18], 1, Concat, [1]], # cat head P6 + [-1, 3, C3, [1024, False]], # 38 (P6/64-xlarge) + + [-1, 1, Conv, [1024, 3, 2]], + [[-1, 14], 1, Concat, [1]], # cat head P7 + [-1, 3, C3, [1280, False]], # 41 (P7/128-xxlarge) + + [[29, 32, 35, 38, 41], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5, P6, P7) + ] diff --git a/yolov5/models/hub/yolov5-panet.yaml b/yolov5/models/hub/yolov5-panet.yaml new file mode 100644 index 0000000..ccfbf90 --- /dev/null +++ b/yolov5/models/hub/yolov5-panet.yaml @@ -0,0 +1,48 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 9 + ] + +# YOLOv5 v6.0 PANet head +head: + [[-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 13 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 17 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 14], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 20 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 10], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [1024, False]], # 23 (P5/32-large) + + [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/yolov5/models/hub/yolov5l6.yaml b/yolov5/models/hub/yolov5l6.yaml new file mode 100644 index 0000000..632c2cb --- /dev/null +++ b/yolov5/models/hub/yolov5l6.yaml @@ -0,0 +1,60 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple +anchors: + - [19,27, 44,40, 38,94] # P3/8 + - [96,68, 86,152, 180,137] # P4/16 + - [140,301, 303,264, 238,542] # P5/32 + - [436,615, 739,380, 925,792] # P6/64 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [768, 3, 2]], # 7-P5/32 + [-1, 3, C3, [768]], + [-1, 1, Conv, [1024, 3, 2]], # 9-P6/64 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 11 + ] + +# YOLOv5 v6.0 head +head: + [[-1, 1, Conv, [768, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 8], 1, Concat, [1]], # cat backbone P5 + [-1, 3, C3, [768, False]], # 15 + + [-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 19 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 23 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 20], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 26 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 16], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [768, False]], # 29 (P5/32-large) + + [-1, 1, Conv, [768, 3, 2]], + [[-1, 12], 1, Concat, [1]], # cat head P6 + [-1, 3, C3, [1024, False]], # 32 (P6/64-xlarge) + + [[23, 26, 29, 32], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5, P6) + ] diff --git a/yolov5/models/hub/yolov5m6.yaml b/yolov5/models/hub/yolov5m6.yaml new file mode 100644 index 0000000..ecc53fd --- /dev/null +++ b/yolov5/models/hub/yolov5m6.yaml @@ -0,0 +1,60 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 0.67 # model depth multiple +width_multiple: 0.75 # layer channel multiple +anchors: + - [19,27, 44,40, 38,94] # P3/8 + - [96,68, 86,152, 180,137] # P4/16 + - [140,301, 303,264, 238,542] # P5/32 + - [436,615, 739,380, 925,792] # P6/64 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [768, 3, 2]], # 7-P5/32 + [-1, 3, C3, [768]], + [-1, 1, Conv, [1024, 3, 2]], # 9-P6/64 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 11 + ] + +# YOLOv5 v6.0 head +head: + [[-1, 1, Conv, [768, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 8], 1, Concat, [1]], # cat backbone P5 + [-1, 3, C3, [768, False]], # 15 + + [-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 19 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 23 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 20], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 26 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 16], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [768, False]], # 29 (P5/32-large) + + [-1, 1, Conv, [768, 3, 2]], + [[-1, 12], 1, Concat, [1]], # cat head P6 + [-1, 3, C3, [1024, False]], # 32 (P6/64-xlarge) + + [[23, 26, 29, 32], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5, P6) + ] diff --git a/yolov5/models/hub/yolov5n6.yaml b/yolov5/models/hub/yolov5n6.yaml new file mode 100644 index 0000000..0c0c71d --- /dev/null +++ b/yolov5/models/hub/yolov5n6.yaml @@ -0,0 +1,60 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 0.33 # model depth multiple +width_multiple: 0.25 # layer channel multiple +anchors: + - [19,27, 44,40, 38,94] # P3/8 + - [96,68, 86,152, 180,137] # P4/16 + - [140,301, 303,264, 238,542] # P5/32 + - [436,615, 739,380, 925,792] # P6/64 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [768, 3, 2]], # 7-P5/32 + [-1, 3, C3, [768]], + [-1, 1, Conv, [1024, 3, 2]], # 9-P6/64 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 11 + ] + +# YOLOv5 v6.0 head +head: + [[-1, 1, Conv, [768, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 8], 1, Concat, [1]], # cat backbone P5 + [-1, 3, C3, [768, False]], # 15 + + [-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 19 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 23 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 20], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 26 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 16], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [768, False]], # 29 (P5/32-large) + + [-1, 1, Conv, [768, 3, 2]], + [[-1, 12], 1, Concat, [1]], # cat head P6 + [-1, 3, C3, [1024, False]], # 32 (P6/64-xlarge) + + [[23, 26, 29, 32], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5, P6) + ] diff --git a/yolov5/models/hub/yolov5s-ghost.yaml b/yolov5/models/hub/yolov5s-ghost.yaml new file mode 100644 index 0000000..ff9519c --- /dev/null +++ b/yolov5/models/hub/yolov5s-ghost.yaml @@ -0,0 +1,48 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 0.33 # model depth multiple +width_multiple: 0.50 # layer channel multiple +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, GhostConv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3Ghost, [128]], + [-1, 1, GhostConv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3Ghost, [256]], + [-1, 1, GhostConv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3Ghost, [512]], + [-1, 1, GhostConv, [1024, 3, 2]], # 7-P5/32 + [-1, 3, C3Ghost, [1024]], + [-1, 1, SPPF, [1024, 5]], # 9 + ] + +# YOLOv5 v6.0 head +head: + [[-1, 1, GhostConv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3Ghost, [512, False]], # 13 + + [-1, 1, GhostConv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3Ghost, [256, False]], # 17 (P3/8-small) + + [-1, 1, GhostConv, [256, 3, 2]], + [[-1, 14], 1, Concat, [1]], # cat head P4 + [-1, 3, C3Ghost, [512, False]], # 20 (P4/16-medium) + + [-1, 1, GhostConv, [512, 3, 2]], + [[-1, 10], 1, Concat, [1]], # cat head P5 + [-1, 3, C3Ghost, [1024, False]], # 23 (P5/32-large) + + [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/yolov5/models/hub/yolov5s-transformer.yaml b/yolov5/models/hub/yolov5s-transformer.yaml new file mode 100644 index 0000000..100d7c4 --- /dev/null +++ b/yolov5/models/hub/yolov5s-transformer.yaml @@ -0,0 +1,48 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 0.33 # model depth multiple +width_multiple: 0.50 # layer channel multiple +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 + [-1, 3, C3TR, [1024]], # 9 <--- C3TR() Transformer module + [-1, 1, SPPF, [1024, 5]], # 9 + ] + +# YOLOv5 v6.0 head +head: + [[-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 13 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 17 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 14], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 20 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 10], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [1024, False]], # 23 (P5/32-large) + + [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/yolov5/models/hub/yolov5s6.yaml b/yolov5/models/hub/yolov5s6.yaml new file mode 100644 index 0000000..a28fb55 --- /dev/null +++ b/yolov5/models/hub/yolov5s6.yaml @@ -0,0 +1,60 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 0.33 # model depth multiple +width_multiple: 0.50 # layer channel multiple +anchors: + - [19,27, 44,40, 38,94] # P3/8 + - [96,68, 86,152, 180,137] # P4/16 + - [140,301, 303,264, 238,542] # P5/32 + - [436,615, 739,380, 925,792] # P6/64 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [768, 3, 2]], # 7-P5/32 + [-1, 3, C3, [768]], + [-1, 1, Conv, [1024, 3, 2]], # 9-P6/64 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 11 + ] + +# YOLOv5 v6.0 head +head: + [[-1, 1, Conv, [768, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 8], 1, Concat, [1]], # cat backbone P5 + [-1, 3, C3, [768, False]], # 15 + + [-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 19 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 23 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 20], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 26 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 16], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [768, False]], # 29 (P5/32-large) + + [-1, 1, Conv, [768, 3, 2]], + [[-1, 12], 1, Concat, [1]], # cat head P6 + [-1, 3, C3, [1024, False]], # 32 (P6/64-xlarge) + + [[23, 26, 29, 32], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5, P6) + ] diff --git a/yolov5/models/hub/yolov5x6.yaml b/yolov5/models/hub/yolov5x6.yaml new file mode 100644 index 0000000..ba795c4 --- /dev/null +++ b/yolov5/models/hub/yolov5x6.yaml @@ -0,0 +1,60 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 1.33 # model depth multiple +width_multiple: 1.25 # layer channel multiple +anchors: + - [19,27, 44,40, 38,94] # P3/8 + - [96,68, 86,152, 180,137] # P4/16 + - [140,301, 303,264, 238,542] # P5/32 + - [436,615, 739,380, 925,792] # P6/64 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [768, 3, 2]], # 7-P5/32 + [-1, 3, C3, [768]], + [-1, 1, Conv, [1024, 3, 2]], # 9-P6/64 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 11 + ] + +# YOLOv5 v6.0 head +head: + [[-1, 1, Conv, [768, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 8], 1, Concat, [1]], # cat backbone P5 + [-1, 3, C3, [768, False]], # 15 + + [-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 19 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 23 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 20], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 26 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 16], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [768, False]], # 29 (P5/32-large) + + [-1, 1, Conv, [768, 3, 2]], + [[-1, 12], 1, Concat, [1]], # cat head P6 + [-1, 3, C3, [1024, False]], # 32 (P6/64-xlarge) + + [[23, 26, 29, 32], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5, P6) + ] diff --git a/yolov5/models/tf.py b/yolov5/models/tf.py new file mode 100644 index 0000000..b0d98cc --- /dev/null +++ b/yolov5/models/tf.py @@ -0,0 +1,574 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +TensorFlow, Keras and TFLite versions of YOLOv5 +Authored by https://github.com/zldrobit in PR https://github.com/ultralytics/yolov5/pull/1127 + +Usage: + $ python models/tf.py --weights yolov5s.pt + +Export: + $ python path/to/export.py --weights yolov5s.pt --include saved_model pb tflite tfjs +""" + +import argparse +import sys +from copy import deepcopy +from pathlib import Path + +FILE = Path(__file__).resolve() +ROOT = FILE.parents[1] # YOLOv5 root directory +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) # add ROOT to PATH +# ROOT = ROOT.relative_to(Path.cwd()) # relative + +import numpy as np +import tensorflow as tf +import torch +import torch.nn as nn +from tensorflow import keras + +from models.common import (C3, SPP, SPPF, Bottleneck, BottleneckCSP, C3x, Concat, Conv, CrossConv, DWConv, + DWConvTranspose2d, Focus, autopad) +from models.experimental import MixConv2d, attempt_load +from models.yolo import Detect +from utils.activations import SiLU +from utils.general import LOGGER, make_divisible, print_args + + +class TFBN(keras.layers.Layer): + # TensorFlow BatchNormalization wrapper + def __init__(self, w=None): + super().__init__() + self.bn = keras.layers.BatchNormalization( + beta_initializer=keras.initializers.Constant(w.bias.numpy()), + gamma_initializer=keras.initializers.Constant(w.weight.numpy()), + moving_mean_initializer=keras.initializers.Constant(w.running_mean.numpy()), + moving_variance_initializer=keras.initializers.Constant(w.running_var.numpy()), + epsilon=w.eps) + + def call(self, inputs): + return self.bn(inputs) + + +class TFPad(keras.layers.Layer): + # Pad inputs in spatial dimensions 1 and 2 + def __init__(self, pad): + super().__init__() + if isinstance(pad, int): + self.pad = tf.constant([[0, 0], [pad, pad], [pad, pad], [0, 0]]) + else: # tuple/list + self.pad = tf.constant([[0, 0], [pad[0], pad[0]], [pad[1], pad[1]], [0, 0]]) + + def call(self, inputs): + return tf.pad(inputs, self.pad, mode='constant', constant_values=0) + + +class TFConv(keras.layers.Layer): + # Standard convolution + def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True, w=None): + # ch_in, ch_out, weights, kernel, stride, padding, groups + super().__init__() + assert g == 1, "TF v2.2 Conv2D does not support 'groups' argument" + # TensorFlow convolution padding is inconsistent with PyTorch (e.g. k=3 s=2 'SAME' padding) + # see https://stackoverflow.com/questions/52975843/comparing-conv2d-with-padding-between-tensorflow-and-pytorch + conv = keras.layers.Conv2D( + filters=c2, + kernel_size=k, + strides=s, + padding='SAME' if s == 1 else 'VALID', + use_bias=not hasattr(w, 'bn'), + kernel_initializer=keras.initializers.Constant(w.conv.weight.permute(2, 3, 1, 0).numpy()), + bias_initializer='zeros' if hasattr(w, 'bn') else keras.initializers.Constant(w.conv.bias.numpy())) + self.conv = conv if s == 1 else keras.Sequential([TFPad(autopad(k, p)), conv]) + self.bn = TFBN(w.bn) if hasattr(w, 'bn') else tf.identity + self.act = activations(w.act) if act else tf.identity + + def call(self, inputs): + return self.act(self.bn(self.conv(inputs))) + + +class TFDWConv(keras.layers.Layer): + # Depthwise convolution + def __init__(self, c1, c2, k=1, s=1, p=None, act=True, w=None): + # ch_in, ch_out, weights, kernel, stride, padding, groups + super().__init__() + assert c2 % c1 == 0, f'TFDWConv() output={c2} must be a multiple of input={c1} channels' + conv = keras.layers.DepthwiseConv2D( + kernel_size=k, + depth_multiplier=c2 // c1, + strides=s, + padding='SAME' if s == 1 else 'VALID', + use_bias=not hasattr(w, 'bn'), + depthwise_initializer=keras.initializers.Constant(w.conv.weight.permute(2, 3, 1, 0).numpy()), + bias_initializer='zeros' if hasattr(w, 'bn') else keras.initializers.Constant(w.conv.bias.numpy())) + self.conv = conv if s == 1 else keras.Sequential([TFPad(autopad(k, p)), conv]) + self.bn = TFBN(w.bn) if hasattr(w, 'bn') else tf.identity + self.act = activations(w.act) if act else tf.identity + + def call(self, inputs): + return self.act(self.bn(self.conv(inputs))) + + +class TFDWConvTranspose2d(keras.layers.Layer): + # Depthwise ConvTranspose2d + def __init__(self, c1, c2, k=1, s=1, p1=0, p2=0, w=None): + # ch_in, ch_out, weights, kernel, stride, padding, groups + super().__init__() + assert c1 == c2, f'TFDWConv() output={c2} must be equal to input={c1} channels' + assert k == 4 and p1 == 1, 'TFDWConv() only valid for k=4 and p1=1' + weight, bias = w.weight.permute(2, 3, 1, 0).numpy(), w.bias.numpy() + self.c1 = c1 + self.conv = [ + keras.layers.Conv2DTranspose(filters=1, + kernel_size=k, + strides=s, + padding='VALID', + output_padding=p2, + use_bias=True, + kernel_initializer=keras.initializers.Constant(weight[..., i:i + 1]), + bias_initializer=keras.initializers.Constant(bias[i])) for i in range(c1)] + + def call(self, inputs): + return tf.concat([m(x) for m, x in zip(self.conv, tf.split(inputs, self.c1, 3))], 3)[:, 1:-1, 1:-1] + + +class TFFocus(keras.layers.Layer): + # Focus wh information into c-space + def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True, w=None): + # ch_in, ch_out, kernel, stride, padding, groups + super().__init__() + self.conv = TFConv(c1 * 4, c2, k, s, p, g, act, w.conv) + + def call(self, inputs): # x(b,w,h,c) -> y(b,w/2,h/2,4c) + # inputs = inputs / 255 # normalize 0-255 to 0-1 + inputs = [inputs[:, ::2, ::2, :], inputs[:, 1::2, ::2, :], inputs[:, ::2, 1::2, :], inputs[:, 1::2, 1::2, :]] + return self.conv(tf.concat(inputs, 3)) + + +class TFBottleneck(keras.layers.Layer): + # Standard bottleneck + def __init__(self, c1, c2, shortcut=True, g=1, e=0.5, w=None): # ch_in, ch_out, shortcut, groups, expansion + super().__init__() + c_ = int(c2 * e) # hidden channels + self.cv1 = TFConv(c1, c_, 1, 1, w=w.cv1) + self.cv2 = TFConv(c_, c2, 3, 1, g=g, w=w.cv2) + self.add = shortcut and c1 == c2 + + def call(self, inputs): + return inputs + self.cv2(self.cv1(inputs)) if self.add else self.cv2(self.cv1(inputs)) + + +class TFCrossConv(keras.layers.Layer): + # Cross Convolution + def __init__(self, c1, c2, k=3, s=1, g=1, e=1.0, shortcut=False, w=None): + super().__init__() + c_ = int(c2 * e) # hidden channels + self.cv1 = TFConv(c1, c_, (1, k), (1, s), w=w.cv1) + self.cv2 = TFConv(c_, c2, (k, 1), (s, 1), g=g, w=w.cv2) + self.add = shortcut and c1 == c2 + + def call(self, inputs): + return inputs + self.cv2(self.cv1(inputs)) if self.add else self.cv2(self.cv1(inputs)) + + +class TFConv2d(keras.layers.Layer): + # Substitution for PyTorch nn.Conv2D + def __init__(self, c1, c2, k, s=1, g=1, bias=True, w=None): + super().__init__() + assert g == 1, "TF v2.2 Conv2D does not support 'groups' argument" + self.conv = keras.layers.Conv2D(filters=c2, + kernel_size=k, + strides=s, + padding='VALID', + use_bias=bias, + kernel_initializer=keras.initializers.Constant( + w.weight.permute(2, 3, 1, 0).numpy()), + bias_initializer=keras.initializers.Constant(w.bias.numpy()) if bias else None) + + def call(self, inputs): + return self.conv(inputs) + + +class TFBottleneckCSP(keras.layers.Layer): + # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks + def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5, w=None): + # ch_in, ch_out, number, shortcut, groups, expansion + super().__init__() + c_ = int(c2 * e) # hidden channels + self.cv1 = TFConv(c1, c_, 1, 1, w=w.cv1) + self.cv2 = TFConv2d(c1, c_, 1, 1, bias=False, w=w.cv2) + self.cv3 = TFConv2d(c_, c_, 1, 1, bias=False, w=w.cv3) + self.cv4 = TFConv(2 * c_, c2, 1, 1, w=w.cv4) + self.bn = TFBN(w.bn) + self.act = lambda x: keras.activations.swish(x) + self.m = keras.Sequential([TFBottleneck(c_, c_, shortcut, g, e=1.0, w=w.m[j]) for j in range(n)]) + + def call(self, inputs): + y1 = self.cv3(self.m(self.cv1(inputs))) + y2 = self.cv2(inputs) + return self.cv4(self.act(self.bn(tf.concat((y1, y2), axis=3)))) + + +class TFC3(keras.layers.Layer): + # CSP Bottleneck with 3 convolutions + def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5, w=None): + # ch_in, ch_out, number, shortcut, groups, expansion + super().__init__() + c_ = int(c2 * e) # hidden channels + self.cv1 = TFConv(c1, c_, 1, 1, w=w.cv1) + self.cv2 = TFConv(c1, c_, 1, 1, w=w.cv2) + self.cv3 = TFConv(2 * c_, c2, 1, 1, w=w.cv3) + self.m = keras.Sequential([TFBottleneck(c_, c_, shortcut, g, e=1.0, w=w.m[j]) for j in range(n)]) + + def call(self, inputs): + return self.cv3(tf.concat((self.m(self.cv1(inputs)), self.cv2(inputs)), axis=3)) + + +class TFC3x(keras.layers.Layer): + # 3 module with cross-convolutions + def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5, w=None): + # ch_in, ch_out, number, shortcut, groups, expansion + super().__init__() + c_ = int(c2 * e) # hidden channels + self.cv1 = TFConv(c1, c_, 1, 1, w=w.cv1) + self.cv2 = TFConv(c1, c_, 1, 1, w=w.cv2) + self.cv3 = TFConv(2 * c_, c2, 1, 1, w=w.cv3) + self.m = keras.Sequential([ + TFCrossConv(c_, c_, k=3, s=1, g=g, e=1.0, shortcut=shortcut, w=w.m[j]) for j in range(n)]) + + def call(self, inputs): + return self.cv3(tf.concat((self.m(self.cv1(inputs)), self.cv2(inputs)), axis=3)) + + +class TFSPP(keras.layers.Layer): + # Spatial pyramid pooling layer used in YOLOv3-SPP + def __init__(self, c1, c2, k=(5, 9, 13), w=None): + super().__init__() + c_ = c1 // 2 # hidden channels + self.cv1 = TFConv(c1, c_, 1, 1, w=w.cv1) + self.cv2 = TFConv(c_ * (len(k) + 1), c2, 1, 1, w=w.cv2) + self.m = [keras.layers.MaxPool2D(pool_size=x, strides=1, padding='SAME') for x in k] + + def call(self, inputs): + x = self.cv1(inputs) + return self.cv2(tf.concat([x] + [m(x) for m in self.m], 3)) + + +class TFSPPF(keras.layers.Layer): + # Spatial pyramid pooling-Fast layer + def __init__(self, c1, c2, k=5, w=None): + super().__init__() + c_ = c1 // 2 # hidden channels + self.cv1 = TFConv(c1, c_, 1, 1, w=w.cv1) + self.cv2 = TFConv(c_ * 4, c2, 1, 1, w=w.cv2) + self.m = keras.layers.MaxPool2D(pool_size=k, strides=1, padding='SAME') + + def call(self, inputs): + x = self.cv1(inputs) + y1 = self.m(x) + y2 = self.m(y1) + return self.cv2(tf.concat([x, y1, y2, self.m(y2)], 3)) + + +class TFDetect(keras.layers.Layer): + # TF YOLOv5 Detect layer + def __init__(self, nc=80, anchors=(), ch=(), imgsz=(640, 640), w=None): # detection layer + super().__init__() + self.stride = tf.convert_to_tensor(w.stride.numpy(), dtype=tf.float32) + self.nc = nc # number of classes + self.no = nc + 5 # number of outputs per anchor + self.nl = len(anchors) # number of detection layers + self.na = len(anchors[0]) // 2 # number of anchors + self.grid = [tf.zeros(1)] * self.nl # init grid + self.anchors = tf.convert_to_tensor(w.anchors.numpy(), dtype=tf.float32) + self.anchor_grid = tf.reshape(self.anchors * tf.reshape(self.stride, [self.nl, 1, 1]), [self.nl, 1, -1, 1, 2]) + self.m = [TFConv2d(x, self.no * self.na, 1, w=w.m[i]) for i, x in enumerate(ch)] + self.training = False # set to False after building model + self.imgsz = imgsz + for i in range(self.nl): + ny, nx = self.imgsz[0] // self.stride[i], self.imgsz[1] // self.stride[i] + self.grid[i] = self._make_grid(nx, ny) + + def call(self, inputs): + z = [] # inference output + x = [] + for i in range(self.nl): + x.append(self.m[i](inputs[i])) + # x(bs,20,20,255) to x(bs,3,20,20,85) + ny, nx = self.imgsz[0] // self.stride[i], self.imgsz[1] // self.stride[i] + x[i] = tf.reshape(x[i], [-1, ny * nx, self.na, self.no]) + + if not self.training: # inference + y = tf.sigmoid(x[i]) + grid = tf.transpose(self.grid[i], [0, 2, 1, 3]) - 0.5 + anchor_grid = tf.transpose(self.anchor_grid[i], [0, 2, 1, 3]) * 4 + xy = (y[..., 0:2] * 2 + grid) * self.stride[i] # xy + wh = y[..., 2:4] ** 2 * anchor_grid + # Normalize xywh to 0-1 to reduce calibration error + xy /= tf.constant([[self.imgsz[1], self.imgsz[0]]], dtype=tf.float32) + wh /= tf.constant([[self.imgsz[1], self.imgsz[0]]], dtype=tf.float32) + y = tf.concat([xy, wh, y[..., 4:]], -1) + z.append(tf.reshape(y, [-1, self.na * ny * nx, self.no])) + + return tf.transpose(x, [0, 2, 1, 3]) if self.training else (tf.concat(z, 1), x) + + @staticmethod + def _make_grid(nx=20, ny=20): + # yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)]) + # return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float() + xv, yv = tf.meshgrid(tf.range(nx), tf.range(ny)) + return tf.cast(tf.reshape(tf.stack([xv, yv], 2), [1, 1, ny * nx, 2]), dtype=tf.float32) + + +class TFUpsample(keras.layers.Layer): + # TF version of torch.nn.Upsample() + def __init__(self, size, scale_factor, mode, w=None): # warning: all arguments needed including 'w' + super().__init__() + assert scale_factor == 2, "scale_factor must be 2" + self.upsample = lambda x: tf.image.resize(x, (x.shape[1] * 2, x.shape[2] * 2), method=mode) + # self.upsample = keras.layers.UpSampling2D(size=scale_factor, interpolation=mode) + # with default arguments: align_corners=False, half_pixel_centers=False + # self.upsample = lambda x: tf.raw_ops.ResizeNearestNeighbor(images=x, + # size=(x.shape[1] * 2, x.shape[2] * 2)) + + def call(self, inputs): + return self.upsample(inputs) + + +class TFConcat(keras.layers.Layer): + # TF version of torch.concat() + def __init__(self, dimension=1, w=None): + super().__init__() + assert dimension == 1, "convert only NCHW to NHWC concat" + self.d = 3 + + def call(self, inputs): + return tf.concat(inputs, self.d) + + +def parse_model(d, ch, model, imgsz): # model_dict, input_channels(3) + LOGGER.info(f"\n{'':>3}{'from':>18}{'n':>3}{'params':>10} {'module':<40}{'arguments':<30}") + anchors, nc, gd, gw = d['anchors'], d['nc'], d['depth_multiple'], d['width_multiple'] + na = (len(anchors[0]) // 2) if isinstance(anchors, list) else anchors # number of anchors + no = na * (nc + 5) # number of outputs = anchors * (classes + 5) + + layers, save, c2 = [], [], ch[-1] # layers, savelist, ch out + for i, (f, n, m, args) in enumerate(d['backbone'] + d['head']): # from, number, module, args + m_str = m + m = eval(m) if isinstance(m, str) else m # eval strings + for j, a in enumerate(args): + try: + args[j] = eval(a) if isinstance(a, str) else a # eval strings + except NameError: + pass + + n = max(round(n * gd), 1) if n > 1 else n # depth gain + if m in [ + nn.Conv2d, Conv, DWConv, DWConvTranspose2d, Bottleneck, SPP, SPPF, MixConv2d, Focus, CrossConv, + BottleneckCSP, C3, C3x]: + c1, c2 = ch[f], args[0] + c2 = make_divisible(c2 * gw, 8) if c2 != no else c2 + + args = [c1, c2, *args[1:]] + if m in [BottleneckCSP, C3, C3x]: + args.insert(2, n) + n = 1 + elif m is nn.BatchNorm2d: + args = [ch[f]] + elif m is Concat: + c2 = sum(ch[-1 if x == -1 else x + 1] for x in f) + elif m is Detect: + args.append([ch[x + 1] for x in f]) + if isinstance(args[1], int): # number of anchors + args[1] = [list(range(args[1] * 2))] * len(f) + args.append(imgsz) + else: + c2 = ch[f] + + tf_m = eval('TF' + m_str.replace('nn.', '')) + m_ = keras.Sequential([tf_m(*args, w=model.model[i][j]) for j in range(n)]) if n > 1 \ + else tf_m(*args, w=model.model[i]) # module + + torch_m_ = nn.Sequential(*(m(*args) for _ in range(n))) if n > 1 else m(*args) # module + t = str(m)[8:-2].replace('__main__.', '') # module type + np = sum(x.numel() for x in torch_m_.parameters()) # number params + m_.i, m_.f, m_.type, m_.np = i, f, t, np # attach index, 'from' index, type, number params + LOGGER.info(f'{i:>3}{str(f):>18}{str(n):>3}{np:>10} {t:<40}{str(args):<30}') # print + save.extend(x % i for x in ([f] if isinstance(f, int) else f) if x != -1) # append to savelist + layers.append(m_) + ch.append(c2) + return keras.Sequential(layers), sorted(save) + + +class TFModel: + # TF YOLOv5 model + def __init__(self, cfg='yolov5s.yaml', ch=3, nc=None, model=None, imgsz=(640, 640)): # model, channels, classes + super().__init__() + if isinstance(cfg, dict): + self.yaml = cfg # model dict + else: # is *.yaml + import yaml # for torch hub + self.yaml_file = Path(cfg).name + with open(cfg) as f: + self.yaml = yaml.load(f, Loader=yaml.FullLoader) # model dict + + # Define model + if nc and nc != self.yaml['nc']: + LOGGER.info(f"Overriding {cfg} nc={self.yaml['nc']} with nc={nc}") + self.yaml['nc'] = nc # override yaml value + self.model, self.savelist = parse_model(deepcopy(self.yaml), ch=[ch], model=model, imgsz=imgsz) + + def predict(self, + inputs, + tf_nms=False, + agnostic_nms=False, + topk_per_class=100, + topk_all=100, + iou_thres=0.45, + conf_thres=0.25): + y = [] # outputs + x = inputs + for m in self.model.layers: + if m.f != -1: # if not from previous layer + x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f] # from earlier layers + + x = m(x) # run + y.append(x if m.i in self.savelist else None) # save output + + # Add TensorFlow NMS + if tf_nms: + boxes = self._xywh2xyxy(x[0][..., :4]) + probs = x[0][:, :, 4:5] + classes = x[0][:, :, 5:] + scores = probs * classes + if agnostic_nms: + nms = AgnosticNMS()((boxes, classes, scores), topk_all, iou_thres, conf_thres) + else: + boxes = tf.expand_dims(boxes, 2) + nms = tf.image.combined_non_max_suppression(boxes, + scores, + topk_per_class, + topk_all, + iou_thres, + conf_thres, + clip_boxes=False) + return nms, x[1] + return x[0] # output only first tensor [1,6300,85] = [xywh, conf, class0, class1, ...] + # x = x[0][0] # [x(1,6300,85), ...] to x(6300,85) + # xywh = x[..., :4] # x(6300,4) boxes + # conf = x[..., 4:5] # x(6300,1) confidences + # cls = tf.reshape(tf.cast(tf.argmax(x[..., 5:], axis=1), tf.float32), (-1, 1)) # x(6300,1) classes + # return tf.concat([conf, cls, xywh], 1) + + @staticmethod + def _xywh2xyxy(xywh): + # Convert nx4 boxes from [x, y, w, h] to [x1, y1, x2, y2] where xy1=top-left, xy2=bottom-right + x, y, w, h = tf.split(xywh, num_or_size_splits=4, axis=-1) + return tf.concat([x - w / 2, y - h / 2, x + w / 2, y + h / 2], axis=-1) + + +class AgnosticNMS(keras.layers.Layer): + # TF Agnostic NMS + def call(self, input, topk_all, iou_thres, conf_thres): + # wrap map_fn to avoid TypeSpec related error https://stackoverflow.com/a/65809989/3036450 + return tf.map_fn(lambda x: self._nms(x, topk_all, iou_thres, conf_thres), + input, + fn_output_signature=(tf.float32, tf.float32, tf.float32, tf.int32), + name='agnostic_nms') + + @staticmethod + def _nms(x, topk_all=100, iou_thres=0.45, conf_thres=0.25): # agnostic NMS + boxes, classes, scores = x + class_inds = tf.cast(tf.argmax(classes, axis=-1), tf.float32) + scores_inp = tf.reduce_max(scores, -1) + selected_inds = tf.image.non_max_suppression(boxes, + scores_inp, + max_output_size=topk_all, + iou_threshold=iou_thres, + score_threshold=conf_thres) + selected_boxes = tf.gather(boxes, selected_inds) + padded_boxes = tf.pad(selected_boxes, + paddings=[[0, topk_all - tf.shape(selected_boxes)[0]], [0, 0]], + mode="CONSTANT", + constant_values=0.0) + selected_scores = tf.gather(scores_inp, selected_inds) + padded_scores = tf.pad(selected_scores, + paddings=[[0, topk_all - tf.shape(selected_boxes)[0]]], + mode="CONSTANT", + constant_values=-1.0) + selected_classes = tf.gather(class_inds, selected_inds) + padded_classes = tf.pad(selected_classes, + paddings=[[0, topk_all - tf.shape(selected_boxes)[0]]], + mode="CONSTANT", + constant_values=-1.0) + valid_detections = tf.shape(selected_inds)[0] + return padded_boxes, padded_scores, padded_classes, valid_detections + + +def activations(act=nn.SiLU): + # Returns TF activation from input PyTorch activation + if isinstance(act, nn.LeakyReLU): + return lambda x: keras.activations.relu(x, alpha=0.1) + elif isinstance(act, nn.Hardswish): + return lambda x: x * tf.nn.relu6(x + 3) * 0.166666667 + elif isinstance(act, (nn.SiLU, SiLU)): + return lambda x: keras.activations.swish(x) + else: + raise Exception(f'no matching TensorFlow activation found for PyTorch activation {act}') + + +def representative_dataset_gen(dataset, ncalib=100): + # Representative dataset generator for use with converter.representative_dataset, returns a generator of np arrays + for n, (path, img, im0s, vid_cap, string) in enumerate(dataset): + im = np.transpose(img, [1, 2, 0]) + im = np.expand_dims(im, axis=0).astype(np.float32) + im /= 255 + yield [im] + if n >= ncalib: + break + + +def run( + weights=ROOT / 'yolov5s.pt', # weights path + imgsz=(640, 640), # inference size h,w + batch_size=1, # batch size + dynamic=False, # dynamic batch size +): + # PyTorch model + im = torch.zeros((batch_size, 3, *imgsz)) # BCHW image + model = attempt_load(weights, device=torch.device('cpu'), inplace=True, fuse=False) + _ = model(im) # inference + model.info() + + # TensorFlow model + im = tf.zeros((batch_size, *imgsz, 3)) # BHWC image + tf_model = TFModel(cfg=model.yaml, model=model, nc=model.nc, imgsz=imgsz) + _ = tf_model.predict(im) # inference + + # Keras model + im = keras.Input(shape=(*imgsz, 3), batch_size=None if dynamic else batch_size) + keras_model = keras.Model(inputs=im, outputs=tf_model.predict(im)) + keras_model.summary() + + LOGGER.info('PyTorch, TensorFlow and Keras models successfully verified.\nUse export.py for TF model export.') + + +def parse_opt(): + parser = argparse.ArgumentParser() + parser.add_argument('--weights', type=str, default=ROOT / 'yolov5s.pt', help='weights path') + parser.add_argument('--imgsz', '--img', '--img-size', nargs='+', type=int, default=[640], help='inference size h,w') + parser.add_argument('--batch-size', type=int, default=1, help='batch size') + parser.add_argument('--dynamic', action='store_true', help='dynamic batch size') + opt = parser.parse_args() + opt.imgsz *= 2 if len(opt.imgsz) == 1 else 1 # expand + print_args(vars(opt)) + return opt + + +def main(opt): + run(**vars(opt)) + + +if __name__ == "__main__": + opt = parse_opt() + main(opt) diff --git a/yolov5/models/yolo.py b/yolov5/models/yolo.py new file mode 100644 index 0000000..02660e6 --- /dev/null +++ b/yolov5/models/yolo.py @@ -0,0 +1,338 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +YOLO-specific modules + +Usage: + $ python path/to/models/yolo.py --cfg yolov5s.yaml +""" + +import argparse +import os +import platform +import sys +from copy import deepcopy +from pathlib import Path + +FILE = Path(__file__).resolve() +ROOT = FILE.parents[1] # YOLOv5 root directory +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) # add ROOT to PATH +if platform.system() != 'Windows': + ROOT = Path(os.path.relpath(ROOT, Path.cwd())) # relative + +from models.common import * +from models.experimental import * +from utils.autoanchor import check_anchor_order +from utils.general import LOGGER, check_version, check_yaml, make_divisible, print_args +from utils.plots import feature_visualization +from utils.torch_utils import (fuse_conv_and_bn, initialize_weights, model_info, profile, scale_img, select_device, + time_sync) + +try: + import thop # for FLOPs computation +except ImportError: + thop = None + + +class Detect(nn.Module): + stride = None # strides computed during build + onnx_dynamic = False # ONNX export parameter + export = False # export mode + + def __init__(self, nc=80, anchors=(), ch=(), inplace=True): # detection layer + super().__init__() + self.nc = nc # number of classes + self.no = nc + 5 # number of outputs per anchor + self.nl = len(anchors) # number of detection layers + self.na = len(anchors[0]) // 2 # number of anchors + self.grid = [torch.zeros(1)] * self.nl # init grid + self.anchor_grid = [torch.zeros(1)] * self.nl # init anchor grid + self.register_buffer('anchors', torch.tensor(anchors).float().view(self.nl, -1, 2)) # shape(nl,na,2) + self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch) # output conv + self.inplace = inplace # use in-place ops (e.g. slice assignment) + + def forward(self, x): + z = [] # inference output + for i in range(self.nl): + x[i] = self.m[i](x[i]) # conv + bs, _, ny, nx = x[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85) + x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous() + + if not self.training: # inference + if self.onnx_dynamic or self.grid[i].shape[2:4] != x[i].shape[2:4]: + self.grid[i], self.anchor_grid[i] = self._make_grid(nx, ny, i) + + y = x[i].sigmoid() + if self.inplace: + y[..., 0:2] = (y[..., 0:2] * 2 + self.grid[i]) * self.stride[i] # xy + y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh + else: # for YOLOv5 on AWS Inferentia https://github.com/ultralytics/yolov5/pull/2953 + xy, wh, conf = y.split((2, 2, self.nc + 1), 4) # y.tensor_split((2, 4, 5), 4) # torch 1.8.0 + xy = (xy * 2 + self.grid[i]) * self.stride[i] # xy + wh = (wh * 2) ** 2 * self.anchor_grid[i] # wh + y = torch.cat((xy, wh, conf), 4) + z.append(y.view(bs, -1, self.no)) + + return x if self.training else (torch.cat(z, 1),) if self.export else (torch.cat(z, 1), x) + + def _make_grid(self, nx=20, ny=20, i=0): + d = self.anchors[i].device + t = self.anchors[i].dtype + shape = 1, self.na, ny, nx, 2 # grid shape + y, x = torch.arange(ny, device=d, dtype=t), torch.arange(nx, device=d, dtype=t) + if check_version(torch.__version__, '1.10.0'): # torch>=1.10.0 meshgrid workaround for torch>=0.7 compatibility + yv, xv = torch.meshgrid(y, x, indexing='ij') + else: + yv, xv = torch.meshgrid(y, x) + grid = torch.stack((xv, yv), 2).expand(shape) - 0.5 # add grid offset, i.e. y = 2.0 * x - 0.5 + anchor_grid = (self.anchors[i] * self.stride[i]).view((1, self.na, 1, 1, 2)).expand(shape) + return grid, anchor_grid + + +class Model(nn.Module): + # YOLOv5 model + def __init__(self, cfg='yolov5s.yaml', ch=3, nc=None, anchors=None): # model, input channels, number of classes + super().__init__() + if isinstance(cfg, dict): + self.yaml = cfg # model dict + else: # is *.yaml + import yaml # for torch hub + self.yaml_file = Path(cfg).name + with open(cfg, encoding='ascii', errors='ignore') as f: + self.yaml = yaml.safe_load(f) # model dict + + # Define model + ch = self.yaml['ch'] = self.yaml.get('ch', ch) # input channels + if nc and nc != self.yaml['nc']: + LOGGER.info(f"Overriding model.yaml nc={self.yaml['nc']} with nc={nc}") + self.yaml['nc'] = nc # override yaml value + if anchors: + LOGGER.info(f'Overriding model.yaml anchors with anchors={anchors}') + self.yaml['anchors'] = round(anchors) # override yaml value + self.model, self.save = parse_model(deepcopy(self.yaml), ch=[ch]) # model, savelist + self.names = [str(i) for i in range(self.yaml['nc'])] # default names + self.inplace = self.yaml.get('inplace', True) + + # Build strides, anchors + m = self.model[-1] # Detect() + if isinstance(m, Detect): + s = 256 # 2x min stride + m.inplace = self.inplace + m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))]) # forward + check_anchor_order(m) # must be in pixel-space (not grid-space) + m.anchors /= m.stride.view(-1, 1, 1) + self.stride = m.stride + self._initialize_biases() # only run once + + # Init weights, biases + initialize_weights(self) + self.info() + LOGGER.info('') + + def forward(self, x, augment=False, profile=False, visualize=False): + if augment: + return self._forward_augment(x) # augmented inference, None + return self._forward_once(x, profile, visualize) # single-scale inference, train + + def _forward_augment(self, x): + img_size = x.shape[-2:] # height, width + s = [1, 0.83, 0.67] # scales + f = [None, 3, None] # flips (2-ud, 3-lr) + y = [] # outputs + for si, fi in zip(s, f): + xi = scale_img(x.flip(fi) if fi else x, si, gs=int(self.stride.max())) + yi = self._forward_once(xi)[0] # forward + # cv2.imwrite(f'img_{si}.jpg', 255 * xi[0].cpu().numpy().transpose((1, 2, 0))[:, :, ::-1]) # save + yi = self._descale_pred(yi, fi, si, img_size) + y.append(yi) + y = self._clip_augmented(y) # clip augmented tails + return torch.cat(y, 1), None # augmented inference, train + + def _forward_once(self, x, profile=False, visualize=False): + y, dt = [], [] # outputs + for m in self.model: + if m.f != -1: # if not from previous layer + x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f] # from earlier layers + if profile: + self._profile_one_layer(m, x, dt) + x = m(x) # run + y.append(x if m.i in self.save else None) # save output + if visualize: + feature_visualization(x, m.type, m.i, save_dir=visualize) + return x + + def _descale_pred(self, p, flips, scale, img_size): + # de-scale predictions following augmented inference (inverse operation) + if self.inplace: + p[..., :4] /= scale # de-scale + if flips == 2: + p[..., 1] = img_size[0] - p[..., 1] # de-flip ud + elif flips == 3: + p[..., 0] = img_size[1] - p[..., 0] # de-flip lr + else: + x, y, wh = p[..., 0:1] / scale, p[..., 1:2] / scale, p[..., 2:4] / scale # de-scale + if flips == 2: + y = img_size[0] - y # de-flip ud + elif flips == 3: + x = img_size[1] - x # de-flip lr + p = torch.cat((x, y, wh, p[..., 4:]), -1) + return p + + def _clip_augmented(self, y): + # Clip YOLOv5 augmented inference tails + nl = self.model[-1].nl # number of detection layers (P3-P5) + g = sum(4 ** x for x in range(nl)) # grid points + e = 1 # exclude layer count + i = (y[0].shape[1] // g) * sum(4 ** x for x in range(e)) # indices + y[0] = y[0][:, :-i] # large + i = (y[-1].shape[1] // g) * sum(4 ** (nl - 1 - x) for x in range(e)) # indices + y[-1] = y[-1][:, i:] # small + return y + + def _profile_one_layer(self, m, x, dt): + c = isinstance(m, Detect) # is final layer, copy input as inplace fix + o = thop.profile(m, inputs=(x.copy() if c else x,), verbose=False)[0] / 1E9 * 2 if thop else 0 # FLOPs + t = time_sync() + for _ in range(10): + m(x.copy() if c else x) + dt.append((time_sync() - t) * 100) + if m == self.model[0]: + LOGGER.info(f"{'time (ms)':>10s} {'GFLOPs':>10s} {'params':>10s} module") + LOGGER.info(f'{dt[-1]:10.2f} {o:10.2f} {m.np:10.0f} {m.type}') + if c: + LOGGER.info(f"{sum(dt):10.2f} {'-':>10s} {'-':>10s} Total") + + def _initialize_biases(self, cf=None): # initialize biases into Detect(), cf is class frequency + # https://arxiv.org/abs/1708.02002 section 3.3 + # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1. + m = self.model[-1] # Detect() module + for mi, s in zip(m.m, m.stride): # from + b = mi.bias.view(m.na, -1).detach() # conv.bias(255) to (3,85) + b[:, 4] += math.log(8 / (640 / s) ** 2) # obj (8 objects per 640 image) + b[:, 5:] += math.log(0.6 / (m.nc - 0.999999)) if cf is None else torch.log(cf / cf.sum()) # cls + mi.bias = torch.nn.Parameter(b.view(-1), requires_grad=True) + + def _print_biases(self): + m = self.model[-1] # Detect() module + for mi in m.m: # from + b = mi.bias.detach().view(m.na, -1).T # conv.bias(255) to (3,85) + LOGGER.info( + ('%6g Conv2d.bias:' + '%10.3g' * 6) % (mi.weight.shape[1], *b[:5].mean(1).tolist(), b[5:].mean())) + + # def _print_weights(self): + # for m in self.model.modules(): + # if type(m) is Bottleneck: + # LOGGER.info('%10.3g' % (m.w.detach().sigmoid() * 2)) # shortcut weights + + def fuse(self): # fuse model Conv2d() + BatchNorm2d() layers + LOGGER.info('Fusing layers... ') + for m in self.model.modules(): + if isinstance(m, (Conv, DWConv)) and hasattr(m, 'bn'): + m.conv = fuse_conv_and_bn(m.conv, m.bn) # update conv + delattr(m, 'bn') # remove batchnorm + m.forward = m.forward_fuse # update forward + self.info() + return self + + def info(self, verbose=False, img_size=640): # print model information + model_info(self, verbose, img_size) + + def _apply(self, fn): + # Apply to(), cpu(), cuda(), half() to model tensors that are not parameters or registered buffers + self = super()._apply(fn) + m = self.model[-1] # Detect() + if isinstance(m, Detect): + m.stride = fn(m.stride) + m.grid = list(map(fn, m.grid)) + if isinstance(m.anchor_grid, list): + m.anchor_grid = list(map(fn, m.anchor_grid)) + return self + + +def parse_model(d, ch): # model_dict, input_channels(3) + LOGGER.info(f"\n{'':>3}{'from':>18}{'n':>3}{'params':>10} {'module':<40}{'arguments':<30}") + anchors, nc, gd, gw = d['anchors'], d['nc'], d['depth_multiple'], d['width_multiple'] + na = (len(anchors[0]) // 2) if isinstance(anchors, list) else anchors # number of anchors + no = na * (nc + 5) # number of outputs = anchors * (classes + 5) + + layers, save, c2 = [], [], ch[-1] # layers, savelist, ch out + for i, (f, n, m, args) in enumerate(d['backbone'] + d['head']): # from, number, module, args + m = eval(m) if isinstance(m, str) else m # eval strings + for j, a in enumerate(args): + try: + args[j] = eval(a) if isinstance(a, str) else a # eval strings + except NameError: + pass + + n = n_ = max(round(n * gd), 1) if n > 1 else n # depth gain + if m in (Conv, GhostConv, Bottleneck, GhostBottleneck, SPP, SPPF, DWConv, MixConv2d, Focus, CrossConv, + BottleneckCSP, C3, C3TR, C3SPP, C3Ghost, nn.ConvTranspose2d, DWConvTranspose2d, C3x): + c1, c2 = ch[f], args[0] + if c2 != no: # if not output + c2 = make_divisible(c2 * gw, 8) + + args = [c1, c2, *args[1:]] + if m in [BottleneckCSP, C3, C3TR, C3Ghost, C3x]: + args.insert(2, n) # number of repeats + n = 1 + elif m is nn.BatchNorm2d: + args = [ch[f]] + elif m is Concat: + c2 = sum(ch[x] for x in f) + elif m is Detect: + args.append([ch[x] for x in f]) + if isinstance(args[1], int): # number of anchors + args[1] = [list(range(args[1] * 2))] * len(f) + elif m is Contract: + c2 = ch[f] * args[0] ** 2 + elif m is Expand: + c2 = ch[f] // args[0] ** 2 + else: + c2 = ch[f] + + m_ = nn.Sequential(*(m(*args) for _ in range(n))) if n > 1 else m(*args) # module + t = str(m)[8:-2].replace('__main__.', '') # module type + np = sum(x.numel() for x in m_.parameters()) # number params + m_.i, m_.f, m_.type, m_.np = i, f, t, np # attach index, 'from' index, type, number params + LOGGER.info(f'{i:>3}{str(f):>18}{n_:>3}{np:10.0f} {t:<40}{str(args):<30}') # print + save.extend(x % i for x in ([f] if isinstance(f, int) else f) if x != -1) # append to savelist + layers.append(m_) + if i == 0: + ch = [] + ch.append(c2) + return nn.Sequential(*layers), sorted(save) + + +if __name__ == '__main__': + parser = argparse.ArgumentParser() + parser.add_argument('--cfg', type=str, default='yolov5s.yaml', help='model.yaml') + parser.add_argument('--batch-size', type=int, default=1, help='total batch size for all GPUs') + parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') + parser.add_argument('--profile', action='store_true', help='profile model speed') + parser.add_argument('--line-profile', action='store_true', help='profile model speed layer by layer') + parser.add_argument('--test', action='store_true', help='test all yolo*.yaml') + opt = parser.parse_args() + opt.cfg = check_yaml(opt.cfg) # check YAML + print_args(vars(opt)) + device = select_device(opt.device) + + # Create model + im = torch.rand(opt.batch_size, 3, 640, 640).to(device) + model = Model(opt.cfg).to(device) + + # Options + if opt.line_profile: # profile layer by layer + _ = model(im, profile=True) + + elif opt.profile: # profile forward-backward + results = profile(input=im, ops=[model], n=3) + + elif opt.test: # test all models + for cfg in Path(ROOT / 'models').rglob('yolo*.yaml'): + try: + _ = Model(cfg) + except Exception as e: + print(f'Error in {cfg}: {e}') + + else: # report fused model summary + model.fuse() diff --git a/yolov5/models/yolov5l.yaml b/yolov5/models/yolov5l.yaml new file mode 100644 index 0000000..ce8a5de --- /dev/null +++ b/yolov5/models/yolov5l.yaml @@ -0,0 +1,48 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 9 + ] + +# YOLOv5 v6.0 head +head: + [[-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 13 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 17 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 14], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 20 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 10], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [1024, False]], # 23 (P5/32-large) + + [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/yolov5/models/yolov5m.yaml b/yolov5/models/yolov5m.yaml new file mode 100644 index 0000000..ad13ab3 --- /dev/null +++ b/yolov5/models/yolov5m.yaml @@ -0,0 +1,48 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 0.67 # model depth multiple +width_multiple: 0.75 # layer channel multiple +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 9 + ] + +# YOLOv5 v6.0 head +head: + [[-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 13 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 17 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 14], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 20 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 10], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [1024, False]], # 23 (P5/32-large) + + [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/yolov5/models/yolov5n.yaml b/yolov5/models/yolov5n.yaml new file mode 100644 index 0000000..8a28a40 --- /dev/null +++ b/yolov5/models/yolov5n.yaml @@ -0,0 +1,48 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 0.33 # model depth multiple +width_multiple: 0.25 # layer channel multiple +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 9 + ] + +# YOLOv5 v6.0 head +head: + [[-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 13 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 17 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 14], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 20 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 10], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [1024, False]], # 23 (P5/32-large) + + [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/yolov5/models/yolov5s.yaml b/yolov5/models/yolov5s.yaml new file mode 100644 index 0000000..f35beab --- /dev/null +++ b/yolov5/models/yolov5s.yaml @@ -0,0 +1,48 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 0.33 # model depth multiple +width_multiple: 0.50 # layer channel multiple +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 9 + ] + +# YOLOv5 v6.0 head +head: + [[-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 13 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 17 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 14], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 20 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 10], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [1024, False]], # 23 (P5/32-large) + + [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/yolov5/models/yolov5x.yaml b/yolov5/models/yolov5x.yaml new file mode 100644 index 0000000..f617a02 --- /dev/null +++ b/yolov5/models/yolov5x.yaml @@ -0,0 +1,48 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 1.33 # model depth multiple +width_multiple: 1.25 # layer channel multiple +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 9 + ] + +# YOLOv5 v6.0 head +head: + [[-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 13 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 17 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 14], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 20 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 10], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [1024, False]], # 23 (P5/32-large) + + [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/yolov5/requirements.txt b/yolov5/requirements.txt new file mode 100644 index 0000000..1937b93 --- /dev/null +++ b/yolov5/requirements.txt @@ -0,0 +1,40 @@ +# YOLOv5 requirements +# Usage: pip install -r requirements.txt + +# Base ---------------------------------------- +matplotlib>=3.2.2 +numpy>=1.18.5 +opencv-python>=4.1.1 +Pillow>=7.1.2 +PyYAML>=5.3.1 +requests>=2.23.0 +scipy>=1.4.1 # Google Colab version +torch>=1.7.0 +torchvision>=0.8.1 +tqdm>=4.41.0 +protobuf<=3.20.1 # https://github.com/ultralytics/yolov5/issues/8012 + +# Logging ------------------------------------- +tensorboard>=2.4.1 +# wandb + +# Plotting ------------------------------------ +pandas>=1.1.4 +seaborn>=0.11.0 + +# Export -------------------------------------- +# coremltools>=4.1 # CoreML export +# onnx>=1.9.0 # ONNX export +# onnx-simplifier>=0.3.6 # ONNX simplifier +# scikit-learn==0.19.2 # CoreML quantization +# tensorflow>=2.4.1 # TFLite export +# tensorflowjs>=3.9.0 # TF.js export +# openvino-dev # OpenVINO export + +# Extras -------------------------------------- +ipython # interactive notebook +psutil # system utilization +thop # FLOPs computation +# albumentations>=1.0.3 +# pycocotools>=2.0 # COCO mAP +# roboflow diff --git a/yolov5/setup.cfg b/yolov5/setup.cfg new file mode 100644 index 0000000..020a757 --- /dev/null +++ b/yolov5/setup.cfg @@ -0,0 +1,59 @@ +# Project-wide configuration file, can be used for package metadata and other toll configurations +# Example usage: global configuration for PEP8 (via flake8) setting or default pytest arguments +# Local usage: pip install pre-commit, pre-commit run --all-files + +[metadata] +license_file = LICENSE +description_file = README.md + + +[tool:pytest] +norecursedirs = + .git + dist + build +addopts = + --doctest-modules + --durations=25 + --color=yes + + +[flake8] +max-line-length = 120 +exclude = .tox,*.egg,build,temp +select = E,W,F +doctests = True +verbose = 2 +# https://pep8.readthedocs.io/en/latest/intro.html#error-codes +format = pylint +# see: https://www.flake8rules.com/ +ignore = + E731 # Do not assign a lambda expression, use a def + F405 # name may be undefined, or defined from star imports: module + E402 # module level import not at top of file + F401 # module imported but unused + W504 # line break after binary operator + E127 # continuation line over-indented for visual indent + W504 # line break after binary operator + E231 # missing whitespace after ‘,’, ‘;’, or ‘:’ + E501 # line too long + F403 # ‘from module import *’ used; unable to detect undefined names + + +[isort] +# https://pycqa.github.io/isort/docs/configuration/options.html +line_length = 120 +# see: https://pycqa.github.io/isort/docs/configuration/multi_line_output_modes.html +multi_line_output = 0 + + +[yapf] +based_on_style = pep8 +spaces_before_comment = 2 +COLUMN_LIMIT = 120 +COALESCE_BRACKETS = True +SPACES_AROUND_POWER_OPERATOR = True +SPACE_BETWEEN_ENDING_COMMA_AND_CLOSING_BRACKET = False +SPLIT_BEFORE_CLOSING_BRACKET = False +SPLIT_BEFORE_FIRST_ARGUMENT = False +# EACH_DICT_ENTRY_ON_SEPARATE_LINE = False diff --git a/yolov5/train.py b/yolov5/train.py new file mode 100644 index 0000000..a06ad5a --- /dev/null +++ b/yolov5/train.py @@ -0,0 +1,670 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Train a YOLOv5 model on a custom dataset. + +Models and datasets download automatically from the latest YOLOv5 release. +Models: https://github.com/ultralytics/yolov5/tree/master/models +Datasets: https://github.com/ultralytics/yolov5/tree/master/data +Tutorial: https://github.com/ultralytics/yolov5/wiki/Train-Custom-Data + +Usage: + $ python path/to/train.py --data coco128.yaml --weights yolov5s.pt --img 640 # from pretrained (RECOMMENDED) + $ python path/to/train.py --data coco128.yaml --weights '' --cfg yolov5s.yaml --img 640 # from scratch +""" + +import argparse +import math +import os +import random +import sys +import time +from copy import deepcopy +from datetime import datetime +from pathlib import Path + +import numpy as np +import torch +import torch.distributed as dist +import torch.nn as nn +import yaml +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.optim import SGD, Adam, AdamW, lr_scheduler +from tqdm import tqdm + +FILE = Path(__file__).resolve() +ROOT = FILE.parents[0] # YOLOv5 root directory +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) # add ROOT to PATH +ROOT = Path(os.path.relpath(ROOT, Path.cwd())) # relative + +import val # for end-of-epoch mAP +from models.experimental import attempt_load +from models.yolo import Model +from utils.autoanchor import check_anchors +from utils.autobatch import check_train_batch_size +from utils.callbacks import Callbacks +from utils.dataloaders import create_dataloader +from utils.downloads import attempt_download +from utils.general import (LOGGER, check_amp, check_dataset, check_file, check_git_status, check_img_size, + check_requirements, check_suffix, check_version, check_yaml, colorstr, get_latest_run, + increment_path, init_seeds, intersect_dicts, labels_to_class_weights, + labels_to_image_weights, methods, one_cycle, print_args, print_mutation, strip_optimizer) +from utils.loggers import Loggers +from utils.loggers.wandb.wandb_utils import check_wandb_resume +from utils.loss import ComputeLoss +from utils.metrics import fitness +from utils.plots import plot_evolve, plot_labels +from utils.torch_utils import EarlyStopping, ModelEMA, de_parallel, select_device, torch_distributed_zero_first + +LOCAL_RANK = int(os.getenv('LOCAL_RANK', -1)) # https://pytorch.org/docs/stable/elastic/run.html +RANK = int(os.getenv('RANK', -1)) +WORLD_SIZE = int(os.getenv('WORLD_SIZE', 1)) + + +def train(hyp, opt, device, callbacks): # hyp is path/to/hyp.yaml or hyp dictionary + save_dir, epochs, batch_size, weights, single_cls, evolve, data, cfg, resume, noval, nosave, workers, freeze = \ + Path(opt.save_dir), opt.epochs, opt.batch_size, opt.weights, opt.single_cls, opt.evolve, opt.data, opt.cfg, \ + opt.resume, opt.noval, opt.nosave, opt.workers, opt.freeze + callbacks.run('on_pretrain_routine_start') + + # Directories + w = save_dir / 'weights' # weights dir + (w.parent if evolve else w).mkdir(parents=True, exist_ok=True) # make dir + last, best = w / 'last.pt', w / 'best.pt' + + # Hyperparameters + if isinstance(hyp, str): + with open(hyp, errors='ignore') as f: + hyp = yaml.safe_load(f) # load hyps dict + LOGGER.info(colorstr('hyperparameters: ') + ', '.join(f'{k}={v}' for k, v in hyp.items())) + + # Save run settings + if not evolve: + with open(save_dir / 'hyp.yaml', 'w') as f: + yaml.safe_dump(hyp, f, sort_keys=False) + with open(save_dir / 'opt.yaml', 'w') as f: + yaml.safe_dump(vars(opt), f, sort_keys=False) + + # Loggers + data_dict = None + if RANK in {-1, 0}: + loggers = Loggers(save_dir, weights, opt, hyp, LOGGER) # loggers instance + if loggers.wandb: + data_dict = loggers.wandb.data_dict + if resume: + weights, epochs, hyp, batch_size = opt.weights, opt.epochs, opt.hyp, opt.batch_size + + # Register actions + for k in methods(loggers): + callbacks.register_action(k, callback=getattr(loggers, k)) + + # Config + plots = not evolve and not opt.noplots # create plots + cuda = device.type != 'cpu' + init_seeds(1 + RANK) + with torch_distributed_zero_first(LOCAL_RANK): + data_dict = data_dict or check_dataset(data) # check if None + train_path, val_path = data_dict['train'], data_dict['val'] + nc = 1 if single_cls else int(data_dict['nc']) # number of classes + names = ['item'] if single_cls and len(data_dict['names']) != 1 else data_dict['names'] # class names + assert len(names) == nc, f'{len(names)} names found for nc={nc} dataset in {data}' # check + is_coco = isinstance(val_path, str) and val_path.endswith('coco/val2017.txt') # COCO dataset + + # Model + check_suffix(weights, '.pt') # check weights + pretrained = weights.endswith('.pt') + if pretrained: + with torch_distributed_zero_first(LOCAL_RANK): + weights = attempt_download(weights) # download if not found locally + ckpt = torch.load(weights, map_location='cpu') # load checkpoint to CPU to avoid CUDA memory leak + model = Model(cfg or ckpt['model'].yaml, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device) # create + exclude = ['anchor'] if (cfg or hyp.get('anchors')) and not resume else [] # exclude keys + csd = ckpt['model'].float().state_dict() # checkpoint state_dict as FP32 + csd = intersect_dicts(csd, model.state_dict(), exclude=exclude) # intersect + model.load_state_dict(csd, strict=False) # load + LOGGER.info(f'Transferred {len(csd)}/{len(model.state_dict())} items from {weights}') # report + else: + model = Model(cfg, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device) # create + amp = check_amp(model) # check AMP + + # Freeze + freeze = [f'model.{x}.' for x in (freeze if len(freeze) > 1 else range(freeze[0]))] # layers to freeze + for k, v in model.named_parameters(): + v.requires_grad = True # train all layers + if any(x in k for x in freeze): + LOGGER.info(f'freezing {k}') + v.requires_grad = False + + # Image size + gs = max(int(model.stride.max()), 32) # grid size (max stride) + imgsz = check_img_size(opt.imgsz, gs, floor=gs * 2) # verify imgsz is gs-multiple + + # Batch size + if RANK == -1 and batch_size == -1: # single-GPU only, estimate best batch size + batch_size = check_train_batch_size(model, imgsz, amp) + loggers.on_params_update({"batch_size": batch_size}) + + # Optimizer + nbs = 64 # nominal batch size + accumulate = max(round(nbs / batch_size), 1) # accumulate loss before optimizing + hyp['weight_decay'] *= batch_size * accumulate / nbs # scale weight_decay + LOGGER.info(f"Scaled weight_decay = {hyp['weight_decay']}") + + g = [], [], [] # optimizer parameter groups + bn = tuple(v for k, v in nn.__dict__.items() if 'Norm' in k) # normalization layers, i.e. BatchNorm2d() + for v in model.modules(): + if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter): # bias + g[2].append(v.bias) + if isinstance(v, bn): # weight (no decay) + g[1].append(v.weight) + elif hasattr(v, 'weight') and isinstance(v.weight, nn.Parameter): # weight (with decay) + g[0].append(v.weight) + + if opt.optimizer == 'Adam': + optimizer = Adam(g[2], lr=hyp['lr0'], betas=(hyp['momentum'], 0.999)) # adjust beta1 to momentum + elif opt.optimizer == 'AdamW': + optimizer = AdamW(g[2], lr=hyp['lr0'], betas=(hyp['momentum'], 0.999)) # adjust beta1 to momentum + else: + optimizer = SGD(g[2], lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True) + + optimizer.add_param_group({'params': g[0], 'weight_decay': hyp['weight_decay']}) # add g0 with weight_decay + optimizer.add_param_group({'params': g[1]}) # add g1 (BatchNorm2d weights) + LOGGER.info(f"{colorstr('optimizer:')} {type(optimizer).__name__} with parameter groups " + f"{len(g[1])} weight (no decay), {len(g[0])} weight, {len(g[2])} bias") + del g + + # Scheduler + if opt.cos_lr: + lf = one_cycle(1, hyp['lrf'], epochs) # cosine 1->hyp['lrf'] + else: + lf = lambda x: (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf'] # linear + scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf) # plot_lr_scheduler(optimizer, scheduler, epochs) + + # EMA + ema = ModelEMA(model) if RANK in {-1, 0} else None + + # Resume + start_epoch, best_fitness = 0, 0.0 + if pretrained: + # Optimizer + if ckpt['optimizer'] is not None: + optimizer.load_state_dict(ckpt['optimizer']) + best_fitness = ckpt['best_fitness'] + + # EMA + if ema and ckpt.get('ema'): + ema.ema.load_state_dict(ckpt['ema'].float().state_dict()) + ema.updates = ckpt['updates'] + + # Epochs + start_epoch = ckpt['epoch'] + 1 + if resume: + assert start_epoch > 0, f'{weights} training to {epochs} epochs is finished, nothing to resume.' + if epochs < start_epoch: + LOGGER.info(f"{weights} has been trained for {ckpt['epoch']} epochs. Fine-tuning for {epochs} more epochs.") + epochs += ckpt['epoch'] # finetune additional epochs + + del ckpt, csd + + # DP mode + if cuda and RANK == -1 and torch.cuda.device_count() > 1: + LOGGER.warning('WARNING: DP not recommended, use torch.distributed.run for best DDP Multi-GPU results.\n' + 'See Multi-GPU Tutorial at https://github.com/ultralytics/yolov5/issues/475 to get started.') + model = torch.nn.DataParallel(model) + + # SyncBatchNorm + if opt.sync_bn and cuda and RANK != -1: + model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device) + LOGGER.info('Using SyncBatchNorm()') + + # Trainloader + train_loader, dataset = create_dataloader(train_path, + imgsz, + batch_size // WORLD_SIZE, + gs, + single_cls, + hyp=hyp, + augment=True, + cache=None if opt.cache == 'val' else opt.cache, + rect=opt.rect, + rank=LOCAL_RANK, + workers=workers, + image_weights=opt.image_weights, + quad=opt.quad, + prefix=colorstr('train: '), + shuffle=True) + mlc = int(np.concatenate(dataset.labels, 0)[:, 0].max()) # max label class + nb = len(train_loader) # number of batches + assert mlc < nc, f'Label class {mlc} exceeds nc={nc} in {data}. Possible class labels are 0-{nc - 1}' + + # Process 0 + if RANK in {-1, 0}: + val_loader = create_dataloader(val_path, + imgsz, + batch_size // WORLD_SIZE * 2, + gs, + single_cls, + hyp=hyp, + cache=None if noval else opt.cache, + rect=True, + rank=-1, + workers=workers * 2, + pad=0.5, + prefix=colorstr('val: '))[0] + + if not resume: + labels = np.concatenate(dataset.labels, 0) + # c = torch.tensor(labels[:, 0]) # classes + # cf = torch.bincount(c.long(), minlength=nc) + 1. # frequency + # model._initialize_biases(cf.to(device)) + if plots: + plot_labels(labels, names, save_dir) + + # Anchors + if not opt.noautoanchor: + check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz) + model.half().float() # pre-reduce anchor precision + + callbacks.run('on_pretrain_routine_end') + + # DDP mode + if cuda and RANK != -1: + if check_version(torch.__version__, '1.11.0'): + model = DDP(model, device_ids=[LOCAL_RANK], output_device=LOCAL_RANK, static_graph=True) + else: + model = DDP(model, device_ids=[LOCAL_RANK], output_device=LOCAL_RANK) + + # Model attributes + nl = de_parallel(model).model[-1].nl # number of detection layers (to scale hyps) + hyp['box'] *= 3 / nl # scale to layers + hyp['cls'] *= nc / 80 * 3 / nl # scale to classes and layers + hyp['obj'] *= (imgsz / 640) ** 2 * 3 / nl # scale to image size and layers + hyp['label_smoothing'] = opt.label_smoothing + model.nc = nc # attach number of classes to model + model.hyp = hyp # attach hyperparameters to model + model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) * nc # attach class weights + model.names = names + + # Start training + t0 = time.time() + nw = max(round(hyp['warmup_epochs'] * nb), 100) # number of warmup iterations, max(3 epochs, 100 iterations) + # nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training + last_opt_step = -1 + maps = np.zeros(nc) # mAP per class + results = (0, 0, 0, 0, 0, 0, 0) # P, R, mAP@.5, mAP@.5-.95, val_loss(box, obj, cls) + scheduler.last_epoch = start_epoch - 1 # do not move + scaler = torch.cuda.amp.GradScaler(enabled=amp) + stopper = EarlyStopping(patience=opt.patience) + compute_loss = ComputeLoss(model) # init loss class + callbacks.run('on_train_start') + LOGGER.info(f'Image sizes {imgsz} train, {imgsz} val\n' + f'Using {train_loader.num_workers * WORLD_SIZE} dataloader workers\n' + f"Logging results to {colorstr('bold', save_dir)}\n" + f'Starting training for {epochs} epochs...') + for epoch in range(start_epoch, epochs): # epoch ------------------------------------------------------------------ + callbacks.run('on_train_epoch_start') + model.train() + + # Update image weights (optional, single-GPU only) + if opt.image_weights: + cw = model.class_weights.cpu().numpy() * (1 - maps) ** 2 / nc # class weights + iw = labels_to_image_weights(dataset.labels, nc=nc, class_weights=cw) # image weights + dataset.indices = random.choices(range(dataset.n), weights=iw, k=dataset.n) # rand weighted idx + + # Update mosaic border (optional) + # b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs) + # dataset.mosaic_border = [b - imgsz, -b] # height, width borders + + mloss = torch.zeros(3, device=device) # mean losses + if RANK != -1: + train_loader.sampler.set_epoch(epoch) + pbar = enumerate(train_loader) + LOGGER.info(('\n' + '%10s' * 7) % ('Epoch', 'gpu_mem', 'box', 'obj', 'cls', 'labels', 'img_size')) + if RANK in {-1, 0}: + pbar = tqdm(pbar, total=nb, bar_format='{l_bar}{bar:10}{r_bar}{bar:-10b}') # progress bar + optimizer.zero_grad() + for i, (imgs, targets, paths, _) in pbar: # batch ------------------------------------------------------------- + callbacks.run('on_train_batch_start') + ni = i + nb * epoch # number integrated batches (since train start) + imgs = imgs.to(device, non_blocking=True).float() / 255 # uint8 to float32, 0-255 to 0.0-1.0 + + # Warmup + if ni <= nw: + xi = [0, nw] # x interp + # compute_loss.gr = np.interp(ni, xi, [0.0, 1.0]) # iou loss ratio (obj_loss = 1.0 or iou) + accumulate = max(1, np.interp(ni, xi, [1, nbs / batch_size]).round()) + for j, x in enumerate(optimizer.param_groups): + # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0 + x['lr'] = np.interp(ni, xi, [hyp['warmup_bias_lr'] if j == 2 else 0.0, x['initial_lr'] * lf(epoch)]) + if 'momentum' in x: + x['momentum'] = np.interp(ni, xi, [hyp['warmup_momentum'], hyp['momentum']]) + + # Multi-scale + if opt.multi_scale: + sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs # size + sf = sz / max(imgs.shape[2:]) # scale factor + if sf != 1: + ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]] # new shape (stretched to gs-multiple) + imgs = nn.functional.interpolate(imgs, size=ns, mode='bilinear', align_corners=False) + + # Forward + with torch.cuda.amp.autocast(amp): + pred = model(imgs) # forward + loss, loss_items = compute_loss(pred, targets.to(device)) # loss scaled by batch_size + if RANK != -1: + loss *= WORLD_SIZE # gradient averaged between devices in DDP mode + if opt.quad: + loss *= 4. + + # Backward + scaler.scale(loss).backward() + + # Optimize + if ni - last_opt_step >= accumulate: + scaler.step(optimizer) # optimizer.step + scaler.update() + optimizer.zero_grad() + if ema: + ema.update(model) + last_opt_step = ni + + # Log + if RANK in {-1, 0}: + mloss = (mloss * i + loss_items) / (i + 1) # update mean losses + mem = f'{torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0:.3g}G' # (GB) + pbar.set_description(('%10s' * 2 + '%10.4g' * 5) % + (f'{epoch}/{epochs - 1}', mem, *mloss, targets.shape[0], imgs.shape[-1])) + callbacks.run('on_train_batch_end', ni, model, imgs, targets, paths, plots) + if callbacks.stop_training: + return + # end batch ------------------------------------------------------------------------------------------------ + + # Scheduler + lr = [x['lr'] for x in optimizer.param_groups] # for loggers + scheduler.step() + + if RANK in {-1, 0}: + # mAP + callbacks.run('on_train_epoch_end', epoch=epoch) + ema.update_attr(model, include=['yaml', 'nc', 'hyp', 'names', 'stride', 'class_weights']) + final_epoch = (epoch + 1 == epochs) or stopper.possible_stop + if not noval or final_epoch: # Calculate mAP + results, maps, _ = val.run(data_dict, + batch_size=batch_size // WORLD_SIZE * 2, + imgsz=imgsz, + model=ema.ema, + single_cls=single_cls, + dataloader=val_loader, + save_dir=save_dir, + plots=False, + callbacks=callbacks, + compute_loss=compute_loss) + + # Update best mAP + fi = fitness(np.array(results).reshape(1, -1)) # weighted combination of [P, R, mAP@.5, mAP@.5-.95] + if fi > best_fitness: + best_fitness = fi + log_vals = list(mloss) + list(results) + lr + callbacks.run('on_fit_epoch_end', log_vals, epoch, best_fitness, fi) + + # Save model + if (not nosave) or (final_epoch and not evolve): # if save + ckpt = { + 'epoch': epoch, + 'best_fitness': best_fitness, + 'model': deepcopy(de_parallel(model)).half(), + 'ema': deepcopy(ema.ema).half(), + 'updates': ema.updates, + 'optimizer': optimizer.state_dict(), + 'wandb_id': loggers.wandb.wandb_run.id if loggers.wandb else None, + 'date': datetime.now().isoformat()} + + # Save last, best and delete + torch.save(ckpt, last) + if best_fitness == fi: + torch.save(ckpt, best) + if opt.save_period > 0 and epoch % opt.save_period == 0: + torch.save(ckpt, w / f'epoch{epoch}.pt') + del ckpt + callbacks.run('on_model_save', last, epoch, final_epoch, best_fitness, fi) + + # Stop Single-GPU + if RANK == -1 and stopper(epoch=epoch, fitness=fi): + break + + # Stop DDP TODO: known issues shttps://github.com/ultralytics/yolov5/pull/4576 + # stop = stopper(epoch=epoch, fitness=fi) + # if RANK == 0: + # dist.broadcast_object_list([stop], 0) # broadcast 'stop' to all ranks + + # Stop DPP + # with torch_distributed_zero_first(RANK): + # if stop: + # break # must break all DDP ranks + + # end epoch ---------------------------------------------------------------------------------------------------- + # end training ----------------------------------------------------------------------------------------------------- + if RANK in {-1, 0}: + LOGGER.info(f'\n{epoch - start_epoch + 1} epochs completed in {(time.time() - t0) / 3600:.3f} hours.') + for f in last, best: + if f.exists(): + strip_optimizer(f) # strip optimizers + if f is best: + LOGGER.info(f'\nValidating {f}...') + results, _, _ = val.run( + data_dict, + batch_size=batch_size // WORLD_SIZE * 2, + imgsz=imgsz, + model=attempt_load(f, device).half(), + iou_thres=0.65 if is_coco else 0.60, # best pycocotools results at 0.65 + single_cls=single_cls, + dataloader=val_loader, + save_dir=save_dir, + save_json=is_coco, + verbose=True, + plots=plots, + callbacks=callbacks, + compute_loss=compute_loss) # val best model with plots + if is_coco: + callbacks.run('on_fit_epoch_end', list(mloss) + list(results) + lr, epoch, best_fitness, fi) + + callbacks.run('on_train_end', last, best, plots, epoch, results) + + torch.cuda.empty_cache() + return results + + +def parse_opt(known=False): + parser = argparse.ArgumentParser() + parser.add_argument('--weights', type=str, default=ROOT / 'yolov5s.pt', help='initial weights path') + parser.add_argument('--cfg', type=str, default='', help='model.yaml path') + parser.add_argument('--data', type=str, default=ROOT / 'data/coco128.yaml', help='dataset.yaml path') + parser.add_argument('--hyp', type=str, default=ROOT / 'data/hyps/hyp.scratch-low.yaml', help='hyperparameters path') + parser.add_argument('--epochs', type=int, default=300) + parser.add_argument('--batch-size', type=int, default=16, help='total batch size for all GPUs, -1 for autobatch') + parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, help='train, val image size (pixels)') + parser.add_argument('--rect', action='store_true', help='rectangular training') + parser.add_argument('--resume', nargs='?', const=True, default=False, help='resume most recent training') + parser.add_argument('--nosave', action='store_true', help='only save final checkpoint') + parser.add_argument('--noval', action='store_true', help='only validate final epoch') + parser.add_argument('--noautoanchor', action='store_true', help='disable AutoAnchor') + parser.add_argument('--noplots', action='store_true', help='save no plot files') + parser.add_argument('--evolve', type=int, nargs='?', const=300, help='evolve hyperparameters for x generations') + parser.add_argument('--bucket', type=str, default='', help='gsutil bucket') + parser.add_argument('--cache', type=str, nargs='?', const='ram', help='--cache images in "ram" (default) or "disk"') + parser.add_argument('--image-weights', action='store_true', help='use weighted image selection for training') + parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') + parser.add_argument('--multi-scale', action='store_true', help='vary img-size +/- 50%%') + parser.add_argument('--single-cls', action='store_true', help='train multi-class data as single-class') + parser.add_argument('--optimizer', type=str, choices=['SGD', 'Adam', 'AdamW'], default='SGD', help='optimizer') + parser.add_argument('--sync-bn', action='store_true', help='use SyncBatchNorm, only available in DDP mode') + parser.add_argument('--workers', type=int, default=8, help='max dataloader workers (per RANK in DDP mode)') + parser.add_argument('--project', default=ROOT / 'runs/train', help='save to project/name') + parser.add_argument('--name', default='exp', help='save to project/name') + parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment') + parser.add_argument('--quad', action='store_true', help='quad dataloader') + parser.add_argument('--cos-lr', action='store_true', help='cosine LR scheduler') + parser.add_argument('--label-smoothing', type=float, default=0.0, help='Label smoothing epsilon') + parser.add_argument('--patience', type=int, default=100, help='EarlyStopping patience (epochs without improvement)') + parser.add_argument('--freeze', nargs='+', type=int, default=[0], help='Freeze layers: backbone=10, first3=0 1 2') + parser.add_argument('--save-period', type=int, default=-1, help='Save checkpoint every x epochs (disabled if < 1)') + parser.add_argument('--local_rank', type=int, default=-1, help='DDP parameter, do not modify') + + # Weights & Biases arguments + parser.add_argument('--entity', default=None, help='W&B: Entity') + parser.add_argument('--upload_dataset', nargs='?', const=True, default=False, help='W&B: Upload data, "val" option') + parser.add_argument('--bbox_interval', type=int, default=-1, help='W&B: Set bounding-box image logging interval') + parser.add_argument('--artifact_alias', type=str, default='latest', help='W&B: Version of dataset artifact to use') + + opt = parser.parse_known_args()[0] if known else parser.parse_args() + return opt + + +def main(opt, callbacks=Callbacks()): + # Checks + if RANK in {-1, 0}: + print_args(vars(opt)) + check_git_status() + check_requirements(exclude=['thop']) + + # Resume + if opt.resume and not check_wandb_resume(opt) and not opt.evolve: # resume an interrupted run + ckpt = opt.resume if isinstance(opt.resume, str) else get_latest_run() # specified or most recent path + assert os.path.isfile(ckpt), 'ERROR: --resume checkpoint does not exist' + with open(Path(ckpt).parent.parent / 'opt.yaml', errors='ignore') as f: + opt = argparse.Namespace(**yaml.safe_load(f)) # replace + opt.cfg, opt.weights, opt.resume = '', ckpt, True # reinstate + LOGGER.info(f'Resuming training from {ckpt}') + else: + opt.data, opt.cfg, opt.hyp, opt.weights, opt.project = \ + check_file(opt.data), check_yaml(opt.cfg), check_yaml(opt.hyp), str(opt.weights), str(opt.project) # checks + assert len(opt.cfg) or len(opt.weights), 'either --cfg or --weights must be specified' + if opt.evolve: + if opt.project == str(ROOT / 'runs/train'): # if default project name, rename to runs/evolve + opt.project = str(ROOT / 'runs/evolve') + opt.exist_ok, opt.resume = opt.resume, False # pass resume to exist_ok and disable resume + if opt.name == 'cfg': + opt.name = Path(opt.cfg).stem # use model.yaml as name + opt.save_dir = str(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok)) + + # DDP mode + device = select_device(opt.device, batch_size=opt.batch_size) + if LOCAL_RANK != -1: + msg = 'is not compatible with YOLOv5 Multi-GPU DDP training' + assert not opt.image_weights, f'--image-weights {msg}' + assert not opt.evolve, f'--evolve {msg}' + assert opt.batch_size != -1, f'AutoBatch with --batch-size -1 {msg}, please pass a valid --batch-size' + assert opt.batch_size % WORLD_SIZE == 0, f'--batch-size {opt.batch_size} must be multiple of WORLD_SIZE' + assert torch.cuda.device_count() > LOCAL_RANK, 'insufficient CUDA devices for DDP command' + torch.cuda.set_device(LOCAL_RANK) + device = torch.device('cuda', LOCAL_RANK) + dist.init_process_group(backend="nccl" if dist.is_nccl_available() else "gloo") + + # Train + if not opt.evolve: + train(opt.hyp, opt, device, callbacks) + if WORLD_SIZE > 1 and RANK == 0: + LOGGER.info('Destroying process group... ') + dist.destroy_process_group() + + # Evolve hyperparameters (optional) + else: + # Hyperparameter evolution metadata (mutation scale 0-1, lower_limit, upper_limit) + meta = { + 'lr0': (1, 1e-5, 1e-1), # initial learning rate (SGD=1E-2, Adam=1E-3) + 'lrf': (1, 0.01, 1.0), # final OneCycleLR learning rate (lr0 * lrf) + 'momentum': (0.3, 0.6, 0.98), # SGD momentum/Adam beta1 + 'weight_decay': (1, 0.0, 0.001), # optimizer weight decay + 'warmup_epochs': (1, 0.0, 5.0), # warmup epochs (fractions ok) + 'warmup_momentum': (1, 0.0, 0.95), # warmup initial momentum + 'warmup_bias_lr': (1, 0.0, 0.2), # warmup initial bias lr + 'box': (1, 0.02, 0.2), # box loss gain + 'cls': (1, 0.2, 4.0), # cls loss gain + 'cls_pw': (1, 0.5, 2.0), # cls BCELoss positive_weight + 'obj': (1, 0.2, 4.0), # obj loss gain (scale with pixels) + 'obj_pw': (1, 0.5, 2.0), # obj BCELoss positive_weight + 'iou_t': (0, 0.1, 0.7), # IoU training threshold + 'anchor_t': (1, 2.0, 8.0), # anchor-multiple threshold + 'anchors': (2, 2.0, 10.0), # anchors per output grid (0 to ignore) + 'fl_gamma': (0, 0.0, 2.0), # focal loss gamma (efficientDet default gamma=1.5) + 'hsv_h': (1, 0.0, 0.1), # image HSV-Hue augmentation (fraction) + 'hsv_s': (1, 0.0, 0.9), # image HSV-Saturation augmentation (fraction) + 'hsv_v': (1, 0.0, 0.9), # image HSV-Value augmentation (fraction) + 'degrees': (1, 0.0, 45.0), # image rotation (+/- deg) + 'translate': (1, 0.0, 0.9), # image translation (+/- fraction) + 'scale': (1, 0.0, 0.9), # image scale (+/- gain) + 'shear': (1, 0.0, 10.0), # image shear (+/- deg) + 'perspective': (0, 0.0, 0.001), # image perspective (+/- fraction), range 0-0.001 + 'flipud': (1, 0.0, 1.0), # image flip up-down (probability) + 'fliplr': (0, 0.0, 1.0), # image flip left-right (probability) + 'mosaic': (1, 0.0, 1.0), # image mixup (probability) + 'mixup': (1, 0.0, 1.0), # image mixup (probability) + 'copy_paste': (1, 0.0, 1.0)} # segment copy-paste (probability) + + with open(opt.hyp, errors='ignore') as f: + hyp = yaml.safe_load(f) # load hyps dict + if 'anchors' not in hyp: # anchors commented in hyp.yaml + hyp['anchors'] = 3 + opt.noval, opt.nosave, save_dir = True, True, Path(opt.save_dir) # only val/save final epoch + # ei = [isinstance(x, (int, float)) for x in hyp.values()] # evolvable indices + evolve_yaml, evolve_csv = save_dir / 'hyp_evolve.yaml', save_dir / 'evolve.csv' + if opt.bucket: + os.system(f'gsutil cp gs://{opt.bucket}/evolve.csv {evolve_csv}') # download evolve.csv if exists + + for _ in range(opt.evolve): # generations to evolve + if evolve_csv.exists(): # if evolve.csv exists: select best hyps and mutate + # Select parent(s) + parent = 'single' # parent selection method: 'single' or 'weighted' + x = np.loadtxt(evolve_csv, ndmin=2, delimiter=',', skiprows=1) + n = min(5, len(x)) # number of previous results to consider + x = x[np.argsort(-fitness(x))][:n] # top n mutations + w = fitness(x) - fitness(x).min() + 1E-6 # weights (sum > 0) + if parent == 'single' or len(x) == 1: + # x = x[random.randint(0, n - 1)] # random selection + x = x[random.choices(range(n), weights=w)[0]] # weighted selection + elif parent == 'weighted': + x = (x * w.reshape(n, 1)).sum(0) / w.sum() # weighted combination + + # Mutate + mp, s = 0.8, 0.2 # mutation probability, sigma + npr = np.random + npr.seed(int(time.time())) + g = np.array([meta[k][0] for k in hyp.keys()]) # gains 0-1 + ng = len(meta) + v = np.ones(ng) + while all(v == 1): # mutate until a change occurs (prevent duplicates) + v = (g * (npr.random(ng) < mp) * npr.randn(ng) * npr.random() * s + 1).clip(0.3, 3.0) + for i, k in enumerate(hyp.keys()): # plt.hist(v.ravel(), 300) + hyp[k] = float(x[i + 7] * v[i]) # mutate + + # Constrain to limits + for k, v in meta.items(): + hyp[k] = max(hyp[k], v[1]) # lower limit + hyp[k] = min(hyp[k], v[2]) # upper limit + hyp[k] = round(hyp[k], 5) # significant digits + + # Train mutation + results = train(hyp.copy(), opt, device, callbacks) + callbacks = Callbacks() + # Write mutation results + print_mutation(results, hyp.copy(), save_dir, opt.bucket) + + # Plot results + plot_evolve(evolve_csv) + LOGGER.info(f'Hyperparameter evolution finished {opt.evolve} generations\n' + f"Results saved to {colorstr('bold', save_dir)}\n" + f'Usage example: $ python train.py --hyp {evolve_yaml}') + + +def run(**kwargs): + # Usage: import train; train.run(data='coco128.yaml', imgsz=320, weights='yolov5m.pt') + opt = parse_opt(True) + for k, v in kwargs.items(): + setattr(opt, k, v) + main(opt) + return opt + + +if __name__ == "__main__": + opt = parse_opt() + main(opt) diff --git a/yolov5/tutorial.ipynb b/yolov5/tutorial.ipynb new file mode 100644 index 0000000..1f537c5 --- /dev/null +++ b/yolov5/tutorial.ipynb @@ -0,0 +1,1120 @@ +{ + "nbformat": 4, + "nbformat_minor": 0, + "metadata": { + "colab": { + "name": "YOLOv5 Tutorial", + "provenance": [], + "collapsed_sections": [], + "machine_shape": "hm", + "include_colab_link": true + }, + "kernelspec": { + "name": "python3", + "display_name": "Python 3" + }, + "accelerator": "GPU", + "widgets": { + "application/vnd.jupyter.widget-state+json": { + "d90eeb56398f458086e3b2b41dbd9fec": { + "model_module": "@jupyter-widgets/controls", + "model_name": "HBoxModel", + "model_module_version": "1.5.0", + "state": { + "_dom_classes": [], + "_model_module": "@jupyter-widgets/controls", + "_model_module_version": "1.5.0", + "_model_name": "HBoxModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/controls", + "_view_module_version": "1.5.0", + "_view_name": "HBoxView", + "box_style": "", + "children": [ + "IPY_MODEL_d91d8347f17349a4987cea29eac0a49c", + "IPY_MODEL_8f4ffda703ac4348ab7edf1d12a188e1", + "IPY_MODEL_8c2d91f564de45f8a403386eeeccac27" + ], + "layout": "IPY_MODEL_5dd95d3eda8b49f7910620edcdcbdcdc" + } + }, + "d91d8347f17349a4987cea29eac0a49c": { + "model_module": "@jupyter-widgets/controls", + "model_name": "HTMLModel", + "model_module_version": "1.5.0", + "state": { + "_dom_classes": [], + "_model_module": "@jupyter-widgets/controls", + "_model_module_version": "1.5.0", + "_model_name": "HTMLModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/controls", + "_view_module_version": "1.5.0", + "_view_name": "HTMLView", + "description": "", + "description_tooltip": null, + "layout": "IPY_MODEL_520e5b7e80eb450188261cffbc574d25", + "placeholder": "​", + "style": "IPY_MODEL_3cef138c5f7743858bb0f87b65dd3c76", + "value": "100%" + } + }, + "8f4ffda703ac4348ab7edf1d12a188e1": { + "model_module": "@jupyter-widgets/controls", + "model_name": "FloatProgressModel", + "model_module_version": "1.5.0", + "state": { + "_dom_classes": [], + "_model_module": "@jupyter-widgets/controls", + "_model_module_version": "1.5.0", + "_model_name": "FloatProgressModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/controls", + "_view_module_version": "1.5.0", + "_view_name": "ProgressView", + "bar_style": "success", + "description": "", + "description_tooltip": null, + "layout": "IPY_MODEL_c3782c6dda80400ba7f8c5345624bf87", + "max": 818322941, + "min": 0, + "orientation": "horizontal", + "style": "IPY_MODEL_11415bab172a4904b73e29ff60f6fce1", + "value": 818322941 + } + }, + "8c2d91f564de45f8a403386eeeccac27": { + "model_module": "@jupyter-widgets/controls", + "model_name": "HTMLModel", + "model_module_version": "1.5.0", + "state": { + "_dom_classes": [], + "_model_module": "@jupyter-widgets/controls", + "_model_module_version": "1.5.0", + "_model_name": "HTMLModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/controls", + "_view_module_version": "1.5.0", + "_view_name": "HTMLView", + "description": "", + "description_tooltip": null, + "layout": "IPY_MODEL_eac18040908042dbae67a47d23e95b47", + "placeholder": "​", + "style": "IPY_MODEL_e0fc1d6eb478469c9098aa9518d7b358", + "value": " 780M/780M [01:17<00:00, 17.7MB/s]" + } + }, + "5dd95d3eda8b49f7910620edcdcbdcdc": { + "model_module": "@jupyter-widgets/base", + "model_name": "LayoutModel", + "model_module_version": "1.2.0", + "state": { + "_model_module": "@jupyter-widgets/base", + "_model_module_version": "1.2.0", + "_model_name": "LayoutModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/base", + "_view_module_version": "1.2.0", + "_view_name": "LayoutView", + "align_content": null, + "align_items": null, + "align_self": null, + "border": null, + "bottom": null, + "display": null, + "flex": null, + "flex_flow": null, + "grid_area": null, + "grid_auto_columns": null, + "grid_auto_flow": null, + "grid_auto_rows": null, + "grid_column": null, + "grid_gap": null, + "grid_row": null, + "grid_template_areas": null, + "grid_template_columns": null, + "grid_template_rows": null, + "height": null, + "justify_content": null, + "justify_items": null, + "left": null, + "margin": null, + "max_height": null, + "max_width": null, + "min_height": null, + "min_width": null, + "object_fit": null, + "object_position": null, + "order": null, + "overflow": null, + "overflow_x": null, + "overflow_y": null, + "padding": null, + "right": null, + "top": null, + "visibility": null, + "width": null + } + }, + "520e5b7e80eb450188261cffbc574d25": { + "model_module": "@jupyter-widgets/base", + "model_name": "LayoutModel", + "model_module_version": "1.2.0", + "state": { + "_model_module": "@jupyter-widgets/base", + "_model_module_version": "1.2.0", + "_model_name": "LayoutModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/base", + "_view_module_version": "1.2.0", + "_view_name": "LayoutView", + "align_content": null, + "align_items": null, + "align_self": null, + "border": null, + "bottom": null, + "display": null, + "flex": null, + "flex_flow": null, + "grid_area": null, + "grid_auto_columns": null, + "grid_auto_flow": null, + "grid_auto_rows": null, + "grid_column": null, + "grid_gap": null, + "grid_row": null, + "grid_template_areas": null, + "grid_template_columns": null, + "grid_template_rows": null, + "height": null, + "justify_content": null, + "justify_items": null, + "left": null, + "margin": null, + "max_height": null, + "max_width": null, + "min_height": null, + "min_width": null, + "object_fit": null, + "object_position": null, + "order": null, + "overflow": null, + "overflow_x": null, + "overflow_y": null, + "padding": null, + "right": null, + "top": null, + "visibility": null, + "width": null + } + }, + "3cef138c5f7743858bb0f87b65dd3c76": { + "model_module": "@jupyter-widgets/controls", + "model_name": "DescriptionStyleModel", + "model_module_version": "1.5.0", + "state": { + "_model_module": "@jupyter-widgets/controls", + "_model_module_version": "1.5.0", + "_model_name": "DescriptionStyleModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/base", + "_view_module_version": "1.2.0", + "_view_name": "StyleView", + "description_width": "" + } + }, + "c3782c6dda80400ba7f8c5345624bf87": { + "model_module": "@jupyter-widgets/base", + "model_name": "LayoutModel", + "model_module_version": "1.2.0", + "state": { + "_model_module": "@jupyter-widgets/base", + "_model_module_version": "1.2.0", + "_model_name": "LayoutModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/base", + "_view_module_version": "1.2.0", + "_view_name": "LayoutView", + "align_content": null, + "align_items": null, + "align_self": null, + "border": null, + "bottom": null, + "display": null, + "flex": null, + "flex_flow": null, + "grid_area": null, + "grid_auto_columns": null, + "grid_auto_flow": null, + "grid_auto_rows": null, + "grid_column": null, + "grid_gap": null, + "grid_row": null, + "grid_template_areas": null, + "grid_template_columns": null, + "grid_template_rows": null, + "height": null, + "justify_content": null, + "justify_items": null, + "left": null, + "margin": null, + "max_height": null, + "max_width": null, + "min_height": null, + "min_width": null, + "object_fit": null, + "object_position": null, + "order": null, + "overflow": null, + "overflow_x": null, + "overflow_y": null, + "padding": null, + "right": null, + "top": null, + "visibility": null, + "width": null + } + }, + "11415bab172a4904b73e29ff60f6fce1": { + "model_module": "@jupyter-widgets/controls", + "model_name": "ProgressStyleModel", + "model_module_version": "1.5.0", + "state": { + "_model_module": "@jupyter-widgets/controls", + "_model_module_version": "1.5.0", + "_model_name": "ProgressStyleModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/base", + "_view_module_version": "1.2.0", + "_view_name": "StyleView", + "bar_color": null, + "description_width": "" + } + }, + "eac18040908042dbae67a47d23e95b47": { + "model_module": "@jupyter-widgets/base", + "model_name": "LayoutModel", + "model_module_version": "1.2.0", + "state": { + "_model_module": "@jupyter-widgets/base", + "_model_module_version": "1.2.0", + "_model_name": "LayoutModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/base", + "_view_module_version": "1.2.0", + "_view_name": "LayoutView", + "align_content": null, + "align_items": null, + "align_self": null, + "border": null, + "bottom": null, + "display": null, + "flex": null, + "flex_flow": null, + "grid_area": null, + "grid_auto_columns": null, + "grid_auto_flow": null, + "grid_auto_rows": null, + "grid_column": null, + "grid_gap": null, + "grid_row": null, + "grid_template_areas": null, + "grid_template_columns": null, + "grid_template_rows": null, + "height": null, + "justify_content": null, + "justify_items": null, + "left": null, + "margin": null, + "max_height": null, + "max_width": null, + "min_height": null, + "min_width": null, + "object_fit": null, + "object_position": null, + "order": null, + "overflow": null, + "overflow_x": null, + "overflow_y": null, + "padding": null, + "right": null, + "top": null, + "visibility": null, + "width": null + } + }, + "e0fc1d6eb478469c9098aa9518d7b358": { + "model_module": "@jupyter-widgets/controls", + "model_name": "DescriptionStyleModel", + "model_module_version": "1.5.0", + "state": { + "_model_module": "@jupyter-widgets/controls", + "_model_module_version": "1.5.0", + "_model_name": "DescriptionStyleModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/base", + "_view_module_version": "1.2.0", + "_view_name": "StyleView", + "description_width": "" + } + } + } + } + }, + "cells": [ + { + "cell_type": "markdown", + "metadata": { + "id": "view-in-github", + "colab_type": "text" + }, + "source": [ + "\"Open" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "t6MPjfT5NrKQ" + }, + "source": [ + "\n", + "\n", + "\n", + "This is the **official YOLOv5 🚀 notebook** by **Ultralytics**, and is freely available for redistribution under the [GPL-3.0 license](https://choosealicense.com/licenses/gpl-3.0/). \n", + "For more information please visit https://github.com/ultralytics/yolov5 and https://ultralytics.com. Thank you!" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "7mGmQbAO5pQb" + }, + "source": [ + "# Setup\n", + "\n", + "Clone repo, install dependencies and check PyTorch and GPU." + ] + }, + { + "cell_type": "code", + "metadata": { + "id": "wbvMlHd_QwMG", + "colab": { + "base_uri": "https://localhost:8080/" + }, + "outputId": "ebf225bd-e109-4dbd-8561-3b15514ca47c" + }, + "source": [ + "!git clone https://github.com/ultralytics/yolov5 # clone\n", + "%cd yolov5\n", + "%pip install -qr requirements.txt # install\n", + "\n", + "import torch\n", + "import utils\n", + "display = utils.notebook_init() # checks" + ], + "execution_count": 1, + "outputs": [ + { + "output_type": "stream", + "name": "stderr", + "text": [ + "YOLOv5 🚀 v6.1-174-gc4cb7c6 torch 1.11.0+cu113 CUDA:0 (Tesla V100-SXM2-16GB, 16160MiB)\n" + ] + }, + { + "output_type": "stream", + "name": "stdout", + "text": [ + "Setup complete ✅ (8 CPUs, 51.0 GB RAM, 38.2/166.8 GB disk)\n" + ] + } + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "4JnkELT0cIJg" + }, + "source": [ + "# 1. Inference\n", + "\n", + "`detect.py` runs YOLOv5 inference on a variety of sources, downloading models automatically from the [latest YOLOv5 release](https://github.com/ultralytics/yolov5/releases), and saving results to `runs/detect`. Example inference sources are:\n", + "\n", + "```shell\n", + "python detect.py --source 0 # webcam\n", + " img.jpg # image \n", + " vid.mp4 # video\n", + " path/ # directory\n", + " path/*.jpg # glob\n", + " 'https://youtu.be/Zgi9g1ksQHc' # YouTube\n", + " 'rtsp://example.com/media.mp4' # RTSP, RTMP, HTTP stream\n", + "```" + ] + }, + { + "cell_type": "code", + "metadata": { + "id": "zR9ZbuQCH7FX", + "colab": { + "base_uri": "https://localhost:8080/" + }, + "outputId": "2f43338d-f533-4277-ef9f-b37b565e2702" + }, + "source": [ + "!python detect.py --weights yolov5s.pt --img 640 --conf 0.25 --source data/images\n", + "display.Image(filename='runs/detect/exp/zidane.jpg', width=600)" + ], + "execution_count": 2, + "outputs": [ + { + "output_type": "stream", + "name": "stdout", + "text": [ + "\u001b[34m\u001b[1mdetect: \u001b[0mweights=['yolov5s.pt'], source=data/images, data=data/coco128.yaml, imgsz=[640, 640], conf_thres=0.25, iou_thres=0.45, max_det=1000, device=, view_img=False, save_txt=False, save_conf=False, save_crop=False, nosave=False, classes=None, agnostic_nms=False, augment=False, visualize=False, update=False, project=runs/detect, name=exp, exist_ok=False, line_thickness=3, hide_labels=False, hide_conf=False, half=False, dnn=False\n", + "YOLOv5 🚀 v6.1-174-gc4cb7c6 torch 1.11.0+cu113 CUDA:0 (Tesla V100-SXM2-16GB, 16160MiB)\n", + "\n", + "Downloading https://github.com/ultralytics/yolov5/releases/download/v6.1/yolov5s.pt to yolov5s.pt...\n", + "100% 14.1M/14.1M [00:00<00:00, 220MB/s]\n", + "\n", + "Fusing layers... \n", + "YOLOv5s summary: 213 layers, 7225885 parameters, 0 gradients\n", + "image 1/2 /content/yolov5/data/images/bus.jpg: 640x480 4 persons, 1 bus, Done. (0.012s)\n", + "image 2/2 /content/yolov5/data/images/zidane.jpg: 384x640 2 persons, 2 ties, Done. (0.013s)\n", + "Speed: 0.5ms pre-process, 12.5ms inference, 17.3ms NMS per image at shape (1, 3, 640, 640)\n", + "Results saved to \u001b[1mruns/detect/exp\u001b[0m\n" + ] + } + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "hkAzDWJ7cWTr" + }, + "source": [ + "        \n", + "" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "0eq1SMWl6Sfn" + }, + "source": [ + "# 2. Validate\n", + "Validate a model's accuracy on [COCO](https://cocodataset.org/#home) val or test-dev datasets. Models are downloaded automatically from the [latest YOLOv5 release](https://github.com/ultralytics/yolov5/releases). To show results by class use the `--verbose` flag. Note that `pycocotools` metrics may be ~1% better than the equivalent repo metrics, as is visible below, due to slight differences in mAP computation." + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "eyTZYGgRjnMc" + }, + "source": [ + "## COCO val\n", + "Download [COCO val 2017](https://github.com/ultralytics/yolov5/blob/74b34872fdf41941cddcf243951cdb090fbac17b/data/coco.yaml#L14) dataset (1GB - 5000 images), and test model accuracy." + ] + }, + { + "cell_type": "code", + "metadata": { + "id": "WQPtK1QYVaD_", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 49, + "referenced_widgets": [ + "d90eeb56398f458086e3b2b41dbd9fec", + "d91d8347f17349a4987cea29eac0a49c", + "8f4ffda703ac4348ab7edf1d12a188e1", + "8c2d91f564de45f8a403386eeeccac27", + "5dd95d3eda8b49f7910620edcdcbdcdc", + "520e5b7e80eb450188261cffbc574d25", + "3cef138c5f7743858bb0f87b65dd3c76", + "c3782c6dda80400ba7f8c5345624bf87", + "11415bab172a4904b73e29ff60f6fce1", + "eac18040908042dbae67a47d23e95b47", + "e0fc1d6eb478469c9098aa9518d7b358" + ] + }, + "outputId": "26f3c005-cc13-4b7c-8523-844b56a0b0e3" + }, + "source": [ + "# Download COCO val\n", + "torch.hub.download_url_to_file('https://ultralytics.com/assets/coco2017val.zip', 'tmp.zip')\n", + "!unzip -q tmp.zip -d ../datasets && rm tmp.zip" + ], + "execution_count": 3, + "outputs": [ + { + "output_type": "display_data", + "data": { + "text/plain": [ + " 0%| | 0.00/780M [00:00

\n", + "Close the active learning loop by sampling images from your inference conditions with the `roboflow` pip package\n", + "

\n", + "\n", + "Train a YOLOv5s model on the [COCO128](https://www.kaggle.com/ultralytics/coco128) dataset with `--data coco128.yaml`, starting from pretrained `--weights yolov5s.pt`, or from randomly initialized `--weights '' --cfg yolov5s.yaml`.\n", + "\n", + "- **Pretrained [Models](https://github.com/ultralytics/yolov5/tree/master/models)** are downloaded\n", + "automatically from the [latest YOLOv5 release](https://github.com/ultralytics/yolov5/releases)\n", + "- **[Datasets](https://github.com/ultralytics/yolov5/tree/master/data)** available for autodownload include: [COCO](https://github.com/ultralytics/yolov5/blob/master/data/coco.yaml), [COCO128](https://github.com/ultralytics/yolov5/blob/master/data/coco128.yaml), [VOC](https://github.com/ultralytics/yolov5/blob/master/data/VOC.yaml), [Argoverse](https://github.com/ultralytics/yolov5/blob/master/data/Argoverse.yaml), [VisDrone](https://github.com/ultralytics/yolov5/blob/master/data/VisDrone.yaml), [GlobalWheat](https://github.com/ultralytics/yolov5/blob/master/data/GlobalWheat2020.yaml), [xView](https://github.com/ultralytics/yolov5/blob/master/data/xView.yaml), [Objects365](https://github.com/ultralytics/yolov5/blob/master/data/Objects365.yaml), [SKU-110K](https://github.com/ultralytics/yolov5/blob/master/data/SKU-110K.yaml).\n", + "- **Training Results** are saved to `runs/train/` with incrementing run directories, i.e. `runs/train/exp2`, `runs/train/exp3` etc.\n", + "

\n", + "\n", + "## Train on Custom Data with Roboflow 🌟 NEW\n", + "\n", + "[Roboflow](https://roboflow.com/?ref=ultralytics) enables you to easily **organize, label, and prepare** a high quality dataset with your own custom data. Roboflow also makes it easy to establish an active learning pipeline, collaborate with your team on dataset improvement, and integrate directly into your model building workflow with the `roboflow` pip package.\n", + "\n", + "- Custom Training Example: [https://blog.roboflow.com/how-to-train-yolov5-on-a-custom-dataset/](https://blog.roboflow.com/how-to-train-yolov5-on-a-custom-dataset/?ref=ultralytics)\n", + "- Custom Training Notebook: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/roboflow-ai/yolov5-custom-training-tutorial/blob/main/yolov5-custom-training.ipynb)\n", + "
\n", + "\n", + "

Label images lightning fast (including with model-assisted labeling)" + ] + }, + { + "cell_type": "code", + "metadata": { + "id": "bOy5KI2ncnWd" + }, + "source": [ + "# Tensorboard (optional)\n", + "%load_ext tensorboard\n", + "%tensorboard --logdir runs/train" + ], + "execution_count": null, + "outputs": [] + }, + { + "cell_type": "code", + "metadata": { + "id": "2fLAV42oNb7M" + }, + "source": [ + "# Weights & Biases (optional)\n", + "%pip install -q wandb\n", + "import wandb\n", + "wandb.login()" + ], + "execution_count": null, + "outputs": [] + }, + { + "cell_type": "code", + "metadata": { + "id": "1NcFxRcFdJ_O", + "colab": { + "base_uri": "https://localhost:8080/" + }, + "outputId": "6735ae8b-fd75-4ecd-9d32-71d1881e2481" + }, + "source": [ + "# Train YOLOv5s on COCO128 for 3 epochs\n", + "!python train.py --img 640 --batch 16 --epochs 3 --data coco128.yaml --weights yolov5s.pt --cache" + ], + "execution_count": 5, + "outputs": [ + { + "output_type": "stream", + "name": "stdout", + "text": [ + "\u001b[34m\u001b[1mtrain: \u001b[0mweights=yolov5s.pt, cfg=, data=coco128.yaml, hyp=data/hyps/hyp.scratch-low.yaml, epochs=3, batch_size=16, imgsz=640, rect=False, resume=False, nosave=False, noval=False, noautoanchor=False, noplots=False, evolve=None, bucket=, cache=ram, image_weights=False, device=, multi_scale=False, single_cls=False, optimizer=SGD, sync_bn=False, workers=8, project=runs/train, name=exp, exist_ok=False, quad=False, cos_lr=False, label_smoothing=0.0, patience=100, freeze=[0], save_period=-1, local_rank=-1, entity=None, upload_dataset=False, bbox_interval=-1, artifact_alias=latest\n", + "\u001b[34m\u001b[1mgithub: \u001b[0mup to date with https://github.com/ultralytics/yolov5 ✅\n", + "YOLOv5 🚀 v6.1-174-gc4cb7c6 torch 1.11.0+cu113 CUDA:0 (Tesla V100-SXM2-16GB, 16160MiB)\n", + "\n", + "\u001b[34m\u001b[1mhyperparameters: \u001b[0mlr0=0.01, lrf=0.01, momentum=0.937, weight_decay=0.0005, warmup_epochs=3.0, warmup_momentum=0.8, warmup_bias_lr=0.1, box=0.05, cls=0.5, cls_pw=1.0, obj=1.0, obj_pw=1.0, iou_t=0.2, anchor_t=4.0, fl_gamma=0.0, hsv_h=0.015, hsv_s=0.7, hsv_v=0.4, degrees=0.0, translate=0.1, scale=0.5, shear=0.0, perspective=0.0, flipud=0.0, fliplr=0.5, mosaic=1.0, mixup=0.0, copy_paste=0.0\n", + "\u001b[34m\u001b[1mWeights & Biases: \u001b[0mrun 'pip install wandb' to automatically track and visualize YOLOv5 🚀 runs (RECOMMENDED)\n", + "\u001b[34m\u001b[1mTensorBoard: \u001b[0mStart with 'tensorboard --logdir runs/train', view at http://localhost:6006/\n", + "\n", + "Dataset not found ⚠, missing paths ['/content/datasets/coco128/images/train2017']\n", + "Downloading https://ultralytics.com/assets/coco128.zip to coco128.zip...\n", + "100% 6.66M/6.66M [00:00<00:00, 41.0MB/s]\n", + "Dataset download success ✅ (0.9s), saved to \u001b[1m/content/datasets\u001b[0m\n", + "\n", + " from n params module arguments \n", + " 0 -1 1 3520 models.common.Conv [3, 32, 6, 2, 2] \n", + " 1 -1 1 18560 models.common.Conv [32, 64, 3, 2] \n", + " 2 -1 1 18816 models.common.C3 [64, 64, 1] \n", + " 3 -1 1 73984 models.common.Conv [64, 128, 3, 2] \n", + " 4 -1 2 115712 models.common.C3 [128, 128, 2] \n", + " 5 -1 1 295424 models.common.Conv [128, 256, 3, 2] \n", + " 6 -1 3 625152 models.common.C3 [256, 256, 3] \n", + " 7 -1 1 1180672 models.common.Conv [256, 512, 3, 2] \n", + " 8 -1 1 1182720 models.common.C3 [512, 512, 1] \n", + " 9 -1 1 656896 models.common.SPPF [512, 512, 5] \n", + " 10 -1 1 131584 models.common.Conv [512, 256, 1, 1] \n", + " 11 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest'] \n", + " 12 [-1, 6] 1 0 models.common.Concat [1] \n", + " 13 -1 1 361984 models.common.C3 [512, 256, 1, False] \n", + " 14 -1 1 33024 models.common.Conv [256, 128, 1, 1] \n", + " 15 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest'] \n", + " 16 [-1, 4] 1 0 models.common.Concat [1] \n", + " 17 -1 1 90880 models.common.C3 [256, 128, 1, False] \n", + " 18 -1 1 147712 models.common.Conv [128, 128, 3, 2] \n", + " 19 [-1, 14] 1 0 models.common.Concat [1] \n", + " 20 -1 1 296448 models.common.C3 [256, 256, 1, False] \n", + " 21 -1 1 590336 models.common.Conv [256, 256, 3, 2] \n", + " 22 [-1, 10] 1 0 models.common.Concat [1] \n", + " 23 -1 1 1182720 models.common.C3 [512, 512, 1, False] \n", + " 24 [17, 20, 23] 1 229245 models.yolo.Detect [80, [[10, 13, 16, 30, 33, 23], [30, 61, 62, 45, 59, 119], [116, 90, 156, 198, 373, 326]], [128, 256, 512]]\n", + "Model summary: 270 layers, 7235389 parameters, 7235389 gradients, 16.5 GFLOPs\n", + "\n", + "Transferred 349/349 items from yolov5s.pt\n", + "Scaled weight_decay = 0.0005\n", + "\u001b[34m\u001b[1moptimizer:\u001b[0m SGD with parameter groups 57 weight (no decay), 60 weight, 60 bias\n", + "\u001b[34m\u001b[1malbumentations: \u001b[0mversion 1.0.3 required by YOLOv5, but version 0.1.12 is currently installed\n", + "\u001b[34m\u001b[1mtrain: \u001b[0mScanning '/content/datasets/coco128/labels/train2017' images and labels...128 found, 0 missing, 2 empty, 0 corrupt: 100% 128/128 [00:00<00:00, 405.04it/s]\n", + "\u001b[34m\u001b[1mtrain: \u001b[0mNew cache created: /content/datasets/coco128/labels/train2017.cache\n", + "\u001b[34m\u001b[1mtrain: \u001b[0mCaching images (0.1GB ram): 100% 128/128 [00:00<00:00, 977.19it/s]\n", + "\u001b[34m\u001b[1mval: \u001b[0mScanning '/content/datasets/coco128/labels/train2017.cache' images and labels... 128 found, 0 missing, 2 empty, 0 corrupt: 100% 128/128 [00:00\"Weights

" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "-WPvRbS5Swl6" + }, + "source": [ + "## Local Logging\n", + "\n", + "All results are logged by default to `runs/train`, with a new experiment directory created for each new training as `runs/train/exp2`, `runs/train/exp3`, etc. View train and val jpgs to see mosaics, labels, predictions and augmentation effects. Note an Ultralytics **Mosaic Dataloader** is used for training (shown below), which combines 4 images into 1 mosaic during training.\n", + "\n", + "> \n", + "`train_batch0.jpg` shows train batch 0 mosaics and labels\n", + "\n", + "> \n", + "`test_batch0_labels.jpg` shows val batch 0 labels\n", + "\n", + "> \n", + "`test_batch0_pred.jpg` shows val batch 0 _predictions_\n", + "\n", + "Training results are automatically logged to [Tensorboard](https://www.tensorflow.org/tensorboard) and [CSV](https://github.com/ultralytics/yolov5/pull/4148) as `results.csv`, which is plotted as `results.png` (below) after training completes. You can also plot any `results.csv` file manually:\n", + "\n", + "```python\n", + "from utils.plots import plot_results \n", + "plot_results('path/to/results.csv') # plot 'results.csv' as 'results.png'\n", + "```\n", + "\n", + "\"COCO128" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "Zelyeqbyt3GD" + }, + "source": [ + "# Environments\n", + "\n", + "YOLOv5 may be run in any of the following up-to-date verified environments (with all dependencies including [CUDA](https://developer.nvidia.com/cuda)/[CUDNN](https://developer.nvidia.com/cudnn), [Python](https://www.python.org/) and [PyTorch](https://pytorch.org/) preinstalled):\n", + "\n", + "- **Google Colab and Kaggle** notebooks with free GPU: \"Open \"Open\n", + "- **Google Cloud** Deep Learning VM. See [GCP Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/GCP-Quickstart)\n", + "- **Amazon** Deep Learning AMI. See [AWS Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/AWS-Quickstart)\n", + "- **Docker Image**. See [Docker Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/Docker-Quickstart) \"Docker\n" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "6Qu7Iesl0p54" + }, + "source": [ + "# Status\n", + "\n", + "![CI CPU testing](https://github.com/ultralytics/yolov5/workflows/CI%20CPU%20testing/badge.svg)\n", + "\n", + "If this badge is green, all [YOLOv5 GitHub Actions](https://github.com/ultralytics/yolov5/actions) Continuous Integration (CI) tests are currently passing. CI tests verify correct operation of YOLOv5 training ([train.py](https://github.com/ultralytics/yolov5/blob/master/train.py)), testing ([val.py](https://github.com/ultralytics/yolov5/blob/master/val.py)), inference ([detect.py](https://github.com/ultralytics/yolov5/blob/master/detect.py)) and export ([export.py](https://github.com/ultralytics/yolov5/blob/master/export.py)) on macOS, Windows, and Ubuntu every 24 hours and on every commit.\n" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "IEijrePND_2I" + }, + "source": [ + "# Appendix\n", + "\n", + "Optional extras below. Unit tests validate repo functionality and should be run on any PRs submitted.\n" + ] + }, + { + "cell_type": "code", + "metadata": { + "id": "mcKoSIK2WSzj" + }, + "source": [ + "# Reproduce\n", + "for x in 'yolov5n', 'yolov5s', 'yolov5m', 'yolov5l', 'yolov5x':\n", + " !python val.py --weights {x}.pt --data coco.yaml --img 640 --task speed # speed\n", + " !python val.py --weights {x}.pt --data coco.yaml --img 640 --conf 0.001 --iou 0.65 # mAP" + ], + "execution_count": null, + "outputs": [] + }, + { + "cell_type": "code", + "metadata": { + "id": "GMusP4OAxFu6" + }, + "source": [ + "# PyTorch Hub\n", + "import torch\n", + "\n", + "# Model\n", + "model = torch.hub.load('ultralytics/yolov5', 'yolov5s')\n", + "\n", + "# Images\n", + "dir = 'https://ultralytics.com/images/'\n", + "imgs = [dir + f for f in ('zidane.jpg', 'bus.jpg')] # batch of images\n", + "\n", + "# Inference\n", + "results = model(imgs)\n", + "results.print() # or .show(), .save()" + ], + "execution_count": null, + "outputs": [] + }, + { + "cell_type": "code", + "metadata": { + "id": "FGH0ZjkGjejy" + }, + "source": [ + "# CI Checks\n", + "%%shell\n", + "export PYTHONPATH=\"$PWD\" # to run *.py. files in subdirectories\n", + "rm -rf runs # remove runs/\n", + "for m in yolov5n; do # models\n", + " python train.py --img 64 --batch 32 --weights $m.pt --epochs 1 --device 0 # train pretrained\n", + " python train.py --img 64 --batch 32 --weights '' --cfg $m.yaml --epochs 1 --device 0 # train scratch\n", + " for d in 0 cpu; do # devices\n", + " python val.py --weights $m.pt --device $d # val official\n", + " python val.py --weights runs/train/exp/weights/best.pt --device $d # val custom\n", + " python detect.py --weights $m.pt --device $d # detect official\n", + " python detect.py --weights runs/train/exp/weights/best.pt --device $d # detect custom\n", + " done\n", + " python hubconf.py # hub\n", + " python models/yolo.py --cfg $m.yaml # build PyTorch model\n", + " python models/tf.py --weights $m.pt # build TensorFlow model\n", + " python export.py --img 64 --batch 1 --weights $m.pt --include torchscript onnx # export\n", + "done" + ], + "execution_count": null, + "outputs": [] + }, + { + "cell_type": "code", + "metadata": { + "id": "gogI-kwi3Tye" + }, + "source": [ + "# Profile\n", + "from utils.torch_utils import profile\n", + "\n", + "m1 = lambda x: x * torch.sigmoid(x)\n", + "m2 = torch.nn.SiLU()\n", + "results = profile(input=torch.randn(16, 3, 640, 640), ops=[m1, m2], n=100)" + ], + "execution_count": null, + "outputs": [] + }, + { + "cell_type": "code", + "metadata": { + "id": "RVRSOhEvUdb5" + }, + "source": [ + "# Evolve\n", + "!python train.py --img 640 --batch 64 --epochs 100 --data coco128.yaml --weights yolov5s.pt --cache --noautoanchor --evolve\n", + "!d=runs/train/evolve && cp evolve.* $d && zip -r evolve.zip $d && gsutil mv evolve.zip gs://bucket # upload results (optional)" + ], + "execution_count": null, + "outputs": [] + }, + { + "cell_type": "code", + "metadata": { + "id": "BSgFCAcMbk1R" + }, + "source": [ + "# VOC\n", + "for b, m in zip([64, 64, 64, 32, 16], ['yolov5n', 'yolov5s', 'yolov5m', 'yolov5l', 'yolov5x']): # batch, model\n", + " !python train.py --batch {b} --weights {m}.pt --data VOC.yaml --epochs 50 --img 512 --hyp hyp.VOC.yaml --project VOC --name {m} --cache" + ], + "execution_count": null, + "outputs": [] + }, + { + "cell_type": "code", + "metadata": { + "id": "VTRwsvA9u7ln" + }, + "source": [ + "# TensorRT \n", + "# https://docs.nvidia.com/deeplearning/tensorrt/install-guide/index.html#installing-pip\n", + "!pip install -U nvidia-tensorrt --index-url https://pypi.ngc.nvidia.com # install\n", + "!python export.py --weights yolov5s.pt --include engine --imgsz 640 640 --device 0 # export\n", + "!python detect.py --weights yolov5s.engine --imgsz 640 640 --device 0 # inference" + ], + "execution_count": null, + "outputs": [] + } + ] +} diff --git a/yolov5/utils/__init__.py b/yolov5/utils/__init__.py new file mode 100644 index 0000000..da53a4d --- /dev/null +++ b/yolov5/utils/__init__.py @@ -0,0 +1,36 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +utils/initialization +""" + + +def notebook_init(verbose=True): + # Check system software and hardware + print('Checking setup...') + + import os + import shutil + + from utils.general import check_requirements, emojis, is_colab + from utils.torch_utils import select_device # imports + + check_requirements(('psutil', 'IPython')) + import psutil + from IPython import display # to display images and clear console output + + if is_colab(): + shutil.rmtree('/content/sample_data', ignore_errors=True) # remove colab /sample_data directory + + # System info + if verbose: + gb = 1 << 30 # bytes to GiB (1024 ** 3) + ram = psutil.virtual_memory().total + total, used, free = shutil.disk_usage("/") + display.clear_output() + s = f'({os.cpu_count()} CPUs, {ram / gb:.1f} GB RAM, {(total - free) / gb:.1f}/{total / gb:.1f} GB disk)' + else: + s = '' + + select_device(newline=False) + print(emojis(f'Setup complete ✅ {s}')) + return display diff --git a/yolov5/utils/activations.py b/yolov5/utils/activations.py new file mode 100644 index 0000000..084ce8c --- /dev/null +++ b/yolov5/utils/activations.py @@ -0,0 +1,103 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Activation functions +""" + +import torch +import torch.nn as nn +import torch.nn.functional as F + + +class SiLU(nn.Module): + # SiLU activation https://arxiv.org/pdf/1606.08415.pdf + @staticmethod + def forward(x): + return x * torch.sigmoid(x) + + +class Hardswish(nn.Module): + # Hard-SiLU activation + @staticmethod + def forward(x): + # return x * F.hardsigmoid(x) # for TorchScript and CoreML + return x * F.hardtanh(x + 3, 0.0, 6.0) / 6.0 # for TorchScript, CoreML and ONNX + + +class Mish(nn.Module): + # Mish activation https://github.com/digantamisra98/Mish + @staticmethod + def forward(x): + return x * F.softplus(x).tanh() + + +class MemoryEfficientMish(nn.Module): + # Mish activation memory-efficient + class F(torch.autograd.Function): + + @staticmethod + def forward(ctx, x): + ctx.save_for_backward(x) + return x.mul(torch.tanh(F.softplus(x))) # x * tanh(ln(1 + exp(x))) + + @staticmethod + def backward(ctx, grad_output): + x = ctx.saved_tensors[0] + sx = torch.sigmoid(x) + fx = F.softplus(x).tanh() + return grad_output * (fx + x * sx * (1 - fx * fx)) + + def forward(self, x): + return self.F.apply(x) + + +class FReLU(nn.Module): + # FReLU activation https://arxiv.org/abs/2007.11824 + def __init__(self, c1, k=3): # ch_in, kernel + super().__init__() + self.conv = nn.Conv2d(c1, c1, k, 1, 1, groups=c1, bias=False) + self.bn = nn.BatchNorm2d(c1) + + def forward(self, x): + return torch.max(x, self.bn(self.conv(x))) + + +class AconC(nn.Module): + r""" ACON activation (activate or not) + AconC: (p1*x-p2*x) * sigmoid(beta*(p1*x-p2*x)) + p2*x, beta is a learnable parameter + according to "Activate or Not: Learning Customized Activation" . + """ + + def __init__(self, c1): + super().__init__() + self.p1 = nn.Parameter(torch.randn(1, c1, 1, 1)) + self.p2 = nn.Parameter(torch.randn(1, c1, 1, 1)) + self.beta = nn.Parameter(torch.ones(1, c1, 1, 1)) + + def forward(self, x): + dpx = (self.p1 - self.p2) * x + return dpx * torch.sigmoid(self.beta * dpx) + self.p2 * x + + +class MetaAconC(nn.Module): + r""" ACON activation (activate or not) + MetaAconC: (p1*x-p2*x) * sigmoid(beta*(p1*x-p2*x)) + p2*x, beta is generated by a small network + according to "Activate or Not: Learning Customized Activation" . + """ + + def __init__(self, c1, k=1, s=1, r=16): # ch_in, kernel, stride, r + super().__init__() + c2 = max(r, c1 // r) + self.p1 = nn.Parameter(torch.randn(1, c1, 1, 1)) + self.p2 = nn.Parameter(torch.randn(1, c1, 1, 1)) + self.fc1 = nn.Conv2d(c1, c2, k, s, bias=True) + self.fc2 = nn.Conv2d(c2, c1, k, s, bias=True) + # self.bn1 = nn.BatchNorm2d(c2) + # self.bn2 = nn.BatchNorm2d(c1) + + def forward(self, x): + y = x.mean(dim=2, keepdims=True).mean(dim=3, keepdims=True) + # batch-size 1 bug/instabilities https://github.com/ultralytics/yolov5/issues/2891 + # beta = torch.sigmoid(self.bn2(self.fc2(self.bn1(self.fc1(y))))) # bug/unstable + beta = torch.sigmoid(self.fc2(self.fc1(y))) # bug patch BN layers removed + dpx = (self.p1 - self.p2) * x + return dpx * torch.sigmoid(beta * dpx) + self.p2 * x diff --git a/yolov5/utils/augmentations.py b/yolov5/utils/augmentations.py new file mode 100644 index 0000000..3f764c0 --- /dev/null +++ b/yolov5/utils/augmentations.py @@ -0,0 +1,284 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Image augmentation functions +""" + +import math +import random + +import cv2 +import numpy as np + +from utils.general import LOGGER, check_version, colorstr, resample_segments, segment2box +from utils.metrics import bbox_ioa + + +class Albumentations: + # YOLOv5 Albumentations class (optional, only used if package is installed) + def __init__(self): + self.transform = None + try: + import albumentations as A + check_version(A.__version__, '1.0.3', hard=True) # version requirement + + T = [ + A.Blur(p=0.01), + A.MedianBlur(p=0.01), + A.ToGray(p=0.01), + A.CLAHE(p=0.01), + A.RandomBrightnessContrast(p=0.0), + A.RandomGamma(p=0.0), + A.ImageCompression(quality_lower=75, p=0.0)] # transforms + self.transform = A.Compose(T, bbox_params=A.BboxParams(format='yolo', label_fields=['class_labels'])) + + LOGGER.info(colorstr('albumentations: ') + ', '.join(f'{x}' for x in self.transform.transforms if x.p)) + except ImportError: # package not installed, skip + pass + except Exception as e: + LOGGER.info(colorstr('albumentations: ') + f'{e}') + + def __call__(self, im, labels, p=1.0): + if self.transform and random.random() < p: + new = self.transform(image=im, bboxes=labels[:, 1:], class_labels=labels[:, 0]) # transformed + im, labels = new['image'], np.array([[c, *b] for c, b in zip(new['class_labels'], new['bboxes'])]) + return im, labels + + +def augment_hsv(im, hgain=0.5, sgain=0.5, vgain=0.5): + # HSV color-space augmentation + if hgain or sgain or vgain: + r = np.random.uniform(-1, 1, 3) * [hgain, sgain, vgain] + 1 # random gains + hue, sat, val = cv2.split(cv2.cvtColor(im, cv2.COLOR_BGR2HSV)) + dtype = im.dtype # uint8 + + x = np.arange(0, 256, dtype=r.dtype) + lut_hue = ((x * r[0]) % 180).astype(dtype) + lut_sat = np.clip(x * r[1], 0, 255).astype(dtype) + lut_val = np.clip(x * r[2], 0, 255).astype(dtype) + + im_hsv = cv2.merge((cv2.LUT(hue, lut_hue), cv2.LUT(sat, lut_sat), cv2.LUT(val, lut_val))) + cv2.cvtColor(im_hsv, cv2.COLOR_HSV2BGR, dst=im) # no return needed + + +def hist_equalize(im, clahe=True, bgr=False): + # Equalize histogram on BGR image 'im' with im.shape(n,m,3) and range 0-255 + yuv = cv2.cvtColor(im, cv2.COLOR_BGR2YUV if bgr else cv2.COLOR_RGB2YUV) + if clahe: + c = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8)) + yuv[:, :, 0] = c.apply(yuv[:, :, 0]) + else: + yuv[:, :, 0] = cv2.equalizeHist(yuv[:, :, 0]) # equalize Y channel histogram + return cv2.cvtColor(yuv, cv2.COLOR_YUV2BGR if bgr else cv2.COLOR_YUV2RGB) # convert YUV image to RGB + + +def replicate(im, labels): + # Replicate labels + h, w = im.shape[:2] + boxes = labels[:, 1:].astype(int) + x1, y1, x2, y2 = boxes.T + s = ((x2 - x1) + (y2 - y1)) / 2 # side length (pixels) + for i in s.argsort()[:round(s.size * 0.5)]: # smallest indices + x1b, y1b, x2b, y2b = boxes[i] + bh, bw = y2b - y1b, x2b - x1b + yc, xc = int(random.uniform(0, h - bh)), int(random.uniform(0, w - bw)) # offset x, y + x1a, y1a, x2a, y2a = [xc, yc, xc + bw, yc + bh] + im[y1a:y2a, x1a:x2a] = im[y1b:y2b, x1b:x2b] # im4[ymin:ymax, xmin:xmax] + labels = np.append(labels, [[labels[i, 0], x1a, y1a, x2a, y2a]], axis=0) + + return im, labels + + +def letterbox(im, new_shape=(640, 640), color=(114, 114, 114), auto=True, scaleFill=False, scaleup=True, stride=32): + # Resize and pad image while meeting stride-multiple constraints + shape = im.shape[:2] # current shape [height, width] + if isinstance(new_shape, int): + new_shape = (new_shape, new_shape) + + # Scale ratio (new / old) + r = min(new_shape[0] / shape[0], new_shape[1] / shape[1]) + if not scaleup: # only scale down, do not scale up (for better val mAP) + r = min(r, 1.0) + + # Compute padding + ratio = r, r # width, height ratios + new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r)) + dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1] # wh padding + if auto: # minimum rectangle + dw, dh = np.mod(dw, stride), np.mod(dh, stride) # wh padding + elif scaleFill: # stretch + dw, dh = 0.0, 0.0 + new_unpad = (new_shape[1], new_shape[0]) + ratio = new_shape[1] / shape[1], new_shape[0] / shape[0] # width, height ratios + + dw /= 2 # divide padding into 2 sides + dh /= 2 + + if shape[::-1] != new_unpad: # resize + im = cv2.resize(im, new_unpad, interpolation=cv2.INTER_LINEAR) + top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1)) + left, right = int(round(dw - 0.1)), int(round(dw + 0.1)) + im = cv2.copyMakeBorder(im, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color) # add border + return im, ratio, (dw, dh) + + +def random_perspective(im, + targets=(), + segments=(), + degrees=10, + translate=.1, + scale=.1, + shear=10, + perspective=0.0, + border=(0, 0)): + # torchvision.transforms.RandomAffine(degrees=(-10, 10), translate=(0.1, 0.1), scale=(0.9, 1.1), shear=(-10, 10)) + # targets = [cls, xyxy] + + height = im.shape[0] + border[0] * 2 # shape(h,w,c) + width = im.shape[1] + border[1] * 2 + + # Center + C = np.eye(3) + C[0, 2] = -im.shape[1] / 2 # x translation (pixels) + C[1, 2] = -im.shape[0] / 2 # y translation (pixels) + + # Perspective + P = np.eye(3) + P[2, 0] = random.uniform(-perspective, perspective) # x perspective (about y) + P[2, 1] = random.uniform(-perspective, perspective) # y perspective (about x) + + # Rotation and Scale + R = np.eye(3) + a = random.uniform(-degrees, degrees) + # a += random.choice([-180, -90, 0, 90]) # add 90deg rotations to small rotations + s = random.uniform(1 - scale, 1 + scale) + # s = 2 ** random.uniform(-scale, scale) + R[:2] = cv2.getRotationMatrix2D(angle=a, center=(0, 0), scale=s) + + # Shear + S = np.eye(3) + S[0, 1] = math.tan(random.uniform(-shear, shear) * math.pi / 180) # x shear (deg) + S[1, 0] = math.tan(random.uniform(-shear, shear) * math.pi / 180) # y shear (deg) + + # Translation + T = np.eye(3) + T[0, 2] = random.uniform(0.5 - translate, 0.5 + translate) * width # x translation (pixels) + T[1, 2] = random.uniform(0.5 - translate, 0.5 + translate) * height # y translation (pixels) + + # Combined rotation matrix + M = T @ S @ R @ P @ C # order of operations (right to left) is IMPORTANT + if (border[0] != 0) or (border[1] != 0) or (M != np.eye(3)).any(): # image changed + if perspective: + im = cv2.warpPerspective(im, M, dsize=(width, height), borderValue=(114, 114, 114)) + else: # affine + im = cv2.warpAffine(im, M[:2], dsize=(width, height), borderValue=(114, 114, 114)) + + # Visualize + # import matplotlib.pyplot as plt + # ax = plt.subplots(1, 2, figsize=(12, 6))[1].ravel() + # ax[0].imshow(im[:, :, ::-1]) # base + # ax[1].imshow(im2[:, :, ::-1]) # warped + + # Transform label coordinates + n = len(targets) + if n: + use_segments = any(x.any() for x in segments) + new = np.zeros((n, 4)) + if use_segments: # warp segments + segments = resample_segments(segments) # upsample + for i, segment in enumerate(segments): + xy = np.ones((len(segment), 3)) + xy[:, :2] = segment + xy = xy @ M.T # transform + xy = xy[:, :2] / xy[:, 2:3] if perspective else xy[:, :2] # perspective rescale or affine + + # clip + new[i] = segment2box(xy, width, height) + + else: # warp boxes + xy = np.ones((n * 4, 3)) + xy[:, :2] = targets[:, [1, 2, 3, 4, 1, 4, 3, 2]].reshape(n * 4, 2) # x1y1, x2y2, x1y2, x2y1 + xy = xy @ M.T # transform + xy = (xy[:, :2] / xy[:, 2:3] if perspective else xy[:, :2]).reshape(n, 8) # perspective rescale or affine + + # create new boxes + x = xy[:, [0, 2, 4, 6]] + y = xy[:, [1, 3, 5, 7]] + new = np.concatenate((x.min(1), y.min(1), x.max(1), y.max(1))).reshape(4, n).T + + # clip + new[:, [0, 2]] = new[:, [0, 2]].clip(0, width) + new[:, [1, 3]] = new[:, [1, 3]].clip(0, height) + + # filter candidates + i = box_candidates(box1=targets[:, 1:5].T * s, box2=new.T, area_thr=0.01 if use_segments else 0.10) + targets = targets[i] + targets[:, 1:5] = new[i] + + return im, targets + + +def copy_paste(im, labels, segments, p=0.5): + # Implement Copy-Paste augmentation https://arxiv.org/abs/2012.07177, labels as nx5 np.array(cls, xyxy) + n = len(segments) + if p and n: + h, w, c = im.shape # height, width, channels + im_new = np.zeros(im.shape, np.uint8) + for j in random.sample(range(n), k=round(p * n)): + l, s = labels[j], segments[j] + box = w - l[3], l[2], w - l[1], l[4] + ioa = bbox_ioa(box, labels[:, 1:5]) # intersection over area + if (ioa < 0.30).all(): # allow 30% obscuration of existing labels + labels = np.concatenate((labels, [[l[0], *box]]), 0) + segments.append(np.concatenate((w - s[:, 0:1], s[:, 1:2]), 1)) + cv2.drawContours(im_new, [segments[j].astype(np.int32)], -1, (255, 255, 255), cv2.FILLED) + + result = cv2.bitwise_and(src1=im, src2=im_new) + result = cv2.flip(result, 1) # augment segments (flip left-right) + i = result > 0 # pixels to replace + # i[:, :] = result.max(2).reshape(h, w, 1) # act over ch + im[i] = result[i] # cv2.imwrite('debug.jpg', im) # debug + + return im, labels, segments + + +def cutout(im, labels, p=0.5): + # Applies image cutout augmentation https://arxiv.org/abs/1708.04552 + if random.random() < p: + h, w = im.shape[:2] + scales = [0.5] * 1 + [0.25] * 2 + [0.125] * 4 + [0.0625] * 8 + [0.03125] * 16 # image size fraction + for s in scales: + mask_h = random.randint(1, int(h * s)) # create random masks + mask_w = random.randint(1, int(w * s)) + + # box + xmin = max(0, random.randint(0, w) - mask_w // 2) + ymin = max(0, random.randint(0, h) - mask_h // 2) + xmax = min(w, xmin + mask_w) + ymax = min(h, ymin + mask_h) + + # apply random color mask + im[ymin:ymax, xmin:xmax] = [random.randint(64, 191) for _ in range(3)] + + # return unobscured labels + if len(labels) and s > 0.03: + box = np.array([xmin, ymin, xmax, ymax], dtype=np.float32) + ioa = bbox_ioa(box, labels[:, 1:5]) # intersection over area + labels = labels[ioa < 0.60] # remove >60% obscured labels + + return labels + + +def mixup(im, labels, im2, labels2): + # Applies MixUp augmentation https://arxiv.org/pdf/1710.09412.pdf + r = np.random.beta(32.0, 32.0) # mixup ratio, alpha=beta=32.0 + im = (im * r + im2 * (1 - r)).astype(np.uint8) + labels = np.concatenate((labels, labels2), 0) + return im, labels + + +def box_candidates(box1, box2, wh_thr=2, ar_thr=100, area_thr=0.1, eps=1e-16): # box1(4,n), box2(4,n) + # Compute candidate boxes: box1 before augment, box2 after augment, wh_thr (pixels), aspect_ratio_thr, area_ratio + w1, h1 = box1[2] - box1[0], box1[3] - box1[1] + w2, h2 = box2[2] - box2[0], box2[3] - box2[1] + ar = np.maximum(w2 / (h2 + eps), h2 / (w2 + eps)) # aspect ratio + return (w2 > wh_thr) & (h2 > wh_thr) & (w2 * h2 / (w1 * h1 + eps) > area_thr) & (ar < ar_thr) # candidates diff --git a/yolov5/utils/autoanchor.py b/yolov5/utils/autoanchor.py new file mode 100644 index 0000000..1a4c521 --- /dev/null +++ b/yolov5/utils/autoanchor.py @@ -0,0 +1,170 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +AutoAnchor utils +""" + +import random + +import numpy as np +import torch +import yaml +from tqdm import tqdm + +from utils.general import LOGGER, colorstr, emojis + +PREFIX = colorstr('AutoAnchor: ') + + +def check_anchor_order(m): + # Check anchor order against stride order for YOLOv5 Detect() module m, and correct if necessary + a = m.anchors.prod(-1).mean(-1).view(-1) # mean anchor area per output layer + da = a[-1] - a[0] # delta a + ds = m.stride[-1] - m.stride[0] # delta s + if da and (da.sign() != ds.sign()): # same order + LOGGER.info(f'{PREFIX}Reversing anchor order') + m.anchors[:] = m.anchors.flip(0) + + +def check_anchors(dataset, model, thr=4.0, imgsz=640): + # Check anchor fit to data, recompute if necessary + m = model.module.model[-1] if hasattr(model, 'module') else model.model[-1] # Detect() + shapes = imgsz * dataset.shapes / dataset.shapes.max(1, keepdims=True) + scale = np.random.uniform(0.9, 1.1, size=(shapes.shape[0], 1)) # augment scale + wh = torch.tensor(np.concatenate([l[:, 3:5] * s for s, l in zip(shapes * scale, dataset.labels)])).float() # wh + + def metric(k): # compute metric + r = wh[:, None] / k[None] + x = torch.min(r, 1 / r).min(2)[0] # ratio metric + best = x.max(1)[0] # best_x + aat = (x > 1 / thr).float().sum(1).mean() # anchors above threshold + bpr = (best > 1 / thr).float().mean() # best possible recall + return bpr, aat + + stride = m.stride.to(m.anchors.device).view(-1, 1, 1) # model strides + anchors = m.anchors.clone() * stride # current anchors + bpr, aat = metric(anchors.cpu().view(-1, 2)) + s = f'\n{PREFIX}{aat:.2f} anchors/target, {bpr:.3f} Best Possible Recall (BPR). ' + if bpr > 0.98: # threshold to recompute + LOGGER.info(emojis(f'{s}Current anchors are a good fit to dataset ✅')) + else: + LOGGER.info(emojis(f'{s}Anchors are a poor fit to dataset ⚠️, attempting to improve...')) + na = m.anchors.numel() // 2 # number of anchors + try: + anchors = kmean_anchors(dataset, n=na, img_size=imgsz, thr=thr, gen=1000, verbose=False) + except Exception as e: + LOGGER.info(f'{PREFIX}ERROR: {e}') + new_bpr = metric(anchors)[0] + if new_bpr > bpr: # replace anchors + anchors = torch.tensor(anchors, device=m.anchors.device).type_as(m.anchors) + m.anchors[:] = anchors.clone().view_as(m.anchors) + check_anchor_order(m) # must be in pixel-space (not grid-space) + m.anchors /= stride + s = f'{PREFIX}Done ✅ (optional: update model *.yaml to use these anchors in the future)' + else: + s = f'{PREFIX}Done ⚠️ (original anchors better than new anchors, proceeding with original anchors)' + LOGGER.info(emojis(s)) + + +def kmean_anchors(dataset='./data/coco128.yaml', n=9, img_size=640, thr=4.0, gen=1000, verbose=True): + """ Creates kmeans-evolved anchors from training dataset + + Arguments: + dataset: path to data.yaml, or a loaded dataset + n: number of anchors + img_size: image size used for training + thr: anchor-label wh ratio threshold hyperparameter hyp['anchor_t'] used for training, default=4.0 + gen: generations to evolve anchors using genetic algorithm + verbose: print all results + + Return: + k: kmeans evolved anchors + + Usage: + from utils.autoanchor import *; _ = kmean_anchors() + """ + from scipy.cluster.vq import kmeans + + npr = np.random + thr = 1 / thr + + def metric(k, wh): # compute metrics + r = wh[:, None] / k[None] + x = torch.min(r, 1 / r).min(2)[0] # ratio metric + # x = wh_iou(wh, torch.tensor(k)) # iou metric + return x, x.max(1)[0] # x, best_x + + def anchor_fitness(k): # mutation fitness + _, best = metric(torch.tensor(k, dtype=torch.float32), wh) + return (best * (best > thr).float()).mean() # fitness + + def print_results(k, verbose=True): + k = k[np.argsort(k.prod(1))] # sort small to large + x, best = metric(k, wh0) + bpr, aat = (best > thr).float().mean(), (x > thr).float().mean() * n # best possible recall, anch > thr + s = f'{PREFIX}thr={thr:.2f}: {bpr:.4f} best possible recall, {aat:.2f} anchors past thr\n' \ + f'{PREFIX}n={n}, img_size={img_size}, metric_all={x.mean():.3f}/{best.mean():.3f}-mean/best, ' \ + f'past_thr={x[x > thr].mean():.3f}-mean: ' + for x in k: + s += '%i,%i, ' % (round(x[0]), round(x[1])) + if verbose: + LOGGER.info(s[:-2]) + return k + + if isinstance(dataset, str): # *.yaml file + with open(dataset, errors='ignore') as f: + data_dict = yaml.safe_load(f) # model dict + from utils.dataloaders import LoadImagesAndLabels + dataset = LoadImagesAndLabels(data_dict['train'], augment=True, rect=True) + + # Get label wh + shapes = img_size * dataset.shapes / dataset.shapes.max(1, keepdims=True) + wh0 = np.concatenate([l[:, 3:5] * s for s, l in zip(shapes, dataset.labels)]) # wh + + # Filter + i = (wh0 < 3.0).any(1).sum() + if i: + LOGGER.info(f'{PREFIX}WARNING: Extremely small objects found: {i} of {len(wh0)} labels are < 3 pixels in size') + wh = wh0[(wh0 >= 2.0).any(1)] # filter > 2 pixels + # wh = wh * (npr.rand(wh.shape[0], 1) * 0.9 + 0.1) # multiply by random scale 0-1 + + # Kmeans init + try: + LOGGER.info(f'{PREFIX}Running kmeans for {n} anchors on {len(wh)} points...') + assert n <= len(wh) # apply overdetermined constraint + s = wh.std(0) # sigmas for whitening + k = kmeans(wh / s, n, iter=30)[0] * s # points + assert n == len(k) # kmeans may return fewer points than requested if wh is insufficient or too similar + except Exception: + LOGGER.warning(f'{PREFIX}WARNING: switching strategies from kmeans to random init') + k = np.sort(npr.rand(n * 2)).reshape(n, 2) * img_size # random init + wh, wh0 = (torch.tensor(x, dtype=torch.float32) for x in (wh, wh0)) + k = print_results(k, verbose=False) + + # Plot + # k, d = [None] * 20, [None] * 20 + # for i in tqdm(range(1, 21)): + # k[i-1], d[i-1] = kmeans(wh / s, i) # points, mean distance + # fig, ax = plt.subplots(1, 2, figsize=(14, 7), tight_layout=True) + # ax = ax.ravel() + # ax[0].plot(np.arange(1, 21), np.array(d) ** 2, marker='.') + # fig, ax = plt.subplots(1, 2, figsize=(14, 7)) # plot wh + # ax[0].hist(wh[wh[:, 0]<100, 0],400) + # ax[1].hist(wh[wh[:, 1]<100, 1],400) + # fig.savefig('wh.png', dpi=200) + + # Evolve + f, sh, mp, s = anchor_fitness(k), k.shape, 0.9, 0.1 # fitness, generations, mutation prob, sigma + pbar = tqdm(range(gen), bar_format='{l_bar}{bar:10}{r_bar}{bar:-10b}') # progress bar + for _ in pbar: + v = np.ones(sh) + while (v == 1).all(): # mutate until a change occurs (prevent duplicates) + v = ((npr.random(sh) < mp) * random.random() * npr.randn(*sh) * s + 1).clip(0.3, 3.0) + kg = (k.copy() * v).clip(min=2.0) + fg = anchor_fitness(kg) + if fg > f: + f, k = fg, kg.copy() + pbar.desc = f'{PREFIX}Evolving anchors with Genetic Algorithm: fitness = {f:.4f}' + if verbose: + print_results(k, verbose) + + return print_results(k) diff --git a/yolov5/utils/autobatch.py b/yolov5/utils/autobatch.py new file mode 100644 index 0000000..1100945 --- /dev/null +++ b/yolov5/utils/autobatch.py @@ -0,0 +1,57 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Auto-batch utils +""" + +from copy import deepcopy + +import numpy as np +import torch + +from utils.general import LOGGER, colorstr +from utils.torch_utils import profile + + +def check_train_batch_size(model, imgsz=640, amp=True): + # Check YOLOv5 training batch size + with torch.cuda.amp.autocast(amp): + return autobatch(deepcopy(model).train(), imgsz) # compute optimal batch size + + +def autobatch(model, imgsz=640, fraction=0.9, batch_size=16): + # Automatically estimate best batch size to use `fraction` of available CUDA memory + # Usage: + # import torch + # from utils.autobatch import autobatch + # model = torch.hub.load('ultralytics/yolov5', 'yolov5s', autoshape=False) + # print(autobatch(model)) + + prefix = colorstr('AutoBatch: ') + LOGGER.info(f'{prefix}Computing optimal batch size for --imgsz {imgsz}') + device = next(model.parameters()).device # get model device + if device.type == 'cpu': + LOGGER.info(f'{prefix}CUDA not detected, using default CPU batch-size {batch_size}') + return batch_size + + gb = 1 << 30 # bytes to GiB (1024 ** 3) + d = str(device).upper() # 'CUDA:0' + properties = torch.cuda.get_device_properties(device) # device properties + t = properties.total_memory / gb # (GiB) + r = torch.cuda.memory_reserved(device) / gb # (GiB) + a = torch.cuda.memory_allocated(device) / gb # (GiB) + f = t - (r + a) # free inside reserved + LOGGER.info(f'{prefix}{d} ({properties.name}) {t:.2f}G total, {r:.2f}G reserved, {a:.2f}G allocated, {f:.2f}G free') + + batch_sizes = [1, 2, 4, 8, 16] + try: + img = [torch.zeros(b, 3, imgsz, imgsz) for b in batch_sizes] + y = profile(img, model, n=3, device=device) + except Exception as e: + LOGGER.warning(f'{prefix}{e}') + + y = [x[2] for x in y if x] # memory [2] + batch_sizes = batch_sizes[:len(y)] + p = np.polyfit(batch_sizes, y, deg=1) # first degree polynomial fit + b = int((f * fraction - p[1]) / p[0]) # y intercept (optimal batch size) + LOGGER.info(f'{prefix}Using batch-size {b} for {d} {t * fraction:.2f}G/{t:.2f}G ({fraction * 100:.0f}%)') + return b diff --git a/yolov5/utils/aws/__init__.py b/yolov5/utils/aws/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/yolov5/utils/aws/mime.sh b/yolov5/utils/aws/mime.sh new file mode 100644 index 0000000..c319a83 --- /dev/null +++ b/yolov5/utils/aws/mime.sh @@ -0,0 +1,26 @@ +# AWS EC2 instance startup 'MIME' script https://aws.amazon.com/premiumsupport/knowledge-center/execute-user-data-ec2/ +# This script will run on every instance restart, not only on first start +# --- DO NOT COPY ABOVE COMMENTS WHEN PASTING INTO USERDATA --- + +Content-Type: multipart/mixed; boundary="//" +MIME-Version: 1.0 + +--// +Content-Type: text/cloud-config; charset="us-ascii" +MIME-Version: 1.0 +Content-Transfer-Encoding: 7bit +Content-Disposition: attachment; filename="cloud-config.txt" + +#cloud-config +cloud_final_modules: +- [scripts-user, always] + +--// +Content-Type: text/x-shellscript; charset="us-ascii" +MIME-Version: 1.0 +Content-Transfer-Encoding: 7bit +Content-Disposition: attachment; filename="userdata.txt" + +#!/bin/bash +# --- paste contents of userdata.sh here --- +--// diff --git a/yolov5/utils/aws/resume.py b/yolov5/utils/aws/resume.py new file mode 100644 index 0000000..b21731c --- /dev/null +++ b/yolov5/utils/aws/resume.py @@ -0,0 +1,40 @@ +# Resume all interrupted trainings in yolov5/ dir including DDP trainings +# Usage: $ python utils/aws/resume.py + +import os +import sys +from pathlib import Path + +import torch +import yaml + +FILE = Path(__file__).resolve() +ROOT = FILE.parents[2] # YOLOv5 root directory +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) # add ROOT to PATH + +port = 0 # --master_port +path = Path('').resolve() +for last in path.rglob('*/**/last.pt'): + ckpt = torch.load(last) + if ckpt['optimizer'] is None: + continue + + # Load opt.yaml + with open(last.parent.parent / 'opt.yaml', errors='ignore') as f: + opt = yaml.safe_load(f) + + # Get device count + d = opt['device'].split(',') # devices + nd = len(d) # number of devices + ddp = nd > 1 or (nd == 0 and torch.cuda.device_count() > 1) # distributed data parallel + + if ddp: # multi-GPU + port += 1 + cmd = f'python -m torch.distributed.run --nproc_per_node {nd} --master_port {port} train.py --resume {last}' + else: # single-GPU + cmd = f'python train.py --resume {last}' + + cmd += ' > /dev/null 2>&1 &' # redirect output to dev/null and run in daemon thread + print(cmd) + os.system(cmd) diff --git a/yolov5/utils/aws/userdata.sh b/yolov5/utils/aws/userdata.sh new file mode 100644 index 0000000..5fc1332 --- /dev/null +++ b/yolov5/utils/aws/userdata.sh @@ -0,0 +1,27 @@ +#!/bin/bash +# AWS EC2 instance startup script https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/user-data.html +# This script will run only once on first instance start (for a re-start script see mime.sh) +# /home/ubuntu (ubuntu) or /home/ec2-user (amazon-linux) is working dir +# Use >300 GB SSD + +cd home/ubuntu +if [ ! -d yolov5 ]; then + echo "Running first-time script." # install dependencies, download COCO, pull Docker + git clone https://github.com/ultralytics/yolov5 -b master && sudo chmod -R 777 yolov5 + cd yolov5 + bash data/scripts/get_coco.sh && echo "COCO done." & + sudo docker pull ultralytics/yolov5:latest && echo "Docker done." & + python -m pip install --upgrade pip && pip install -r requirements.txt && python detect.py && echo "Requirements done." & + wait && echo "All tasks done." # finish background tasks +else + echo "Running re-start script." # resume interrupted runs + i=0 + list=$(sudo docker ps -qa) # container list i.e. $'one\ntwo\nthree\nfour' + while IFS= read -r id; do + ((i++)) + echo "restarting container $i: $id" + sudo docker start $id + # sudo docker exec -it $id python train.py --resume # single-GPU + sudo docker exec -d $id python utils/aws/resume.py # multi-scenario + done <<<"$list" +fi diff --git a/yolov5/utils/benchmarks.py b/yolov5/utils/benchmarks.py new file mode 100644 index 0000000..d0f2a25 --- /dev/null +++ b/yolov5/utils/benchmarks.py @@ -0,0 +1,148 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Run YOLOv5 benchmarks on all supported export formats + +Format | `export.py --include` | Model +--- | --- | --- +PyTorch | - | yolov5s.pt +TorchScript | `torchscript` | yolov5s.torchscript +ONNX | `onnx` | yolov5s.onnx +OpenVINO | `openvino` | yolov5s_openvino_model/ +TensorRT | `engine` | yolov5s.engine +CoreML | `coreml` | yolov5s.mlmodel +TensorFlow SavedModel | `saved_model` | yolov5s_saved_model/ +TensorFlow GraphDef | `pb` | yolov5s.pb +TensorFlow Lite | `tflite` | yolov5s.tflite +TensorFlow Edge TPU | `edgetpu` | yolov5s_edgetpu.tflite +TensorFlow.js | `tfjs` | yolov5s_web_model/ + +Requirements: + $ pip install -r requirements.txt coremltools onnx onnx-simplifier onnxruntime openvino-dev tensorflow-cpu # CPU + $ pip install -r requirements.txt coremltools onnx onnx-simplifier onnxruntime-gpu openvino-dev tensorflow # GPU + $ pip install -U nvidia-tensorrt --index-url https://pypi.ngc.nvidia.com # TensorRT + +Usage: + $ python utils/benchmarks.py --weights yolov5s.pt --img 640 +""" + +import argparse +import sys +import time +from pathlib import Path + +import pandas as pd + +FILE = Path(__file__).resolve() +ROOT = FILE.parents[1] # YOLOv5 root directory +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) # add ROOT to PATH +# ROOT = ROOT.relative_to(Path.cwd()) # relative + +import export +import val +from utils import notebook_init +from utils.general import LOGGER, check_yaml, print_args +from utils.torch_utils import select_device + + +def run( + weights=ROOT / 'yolov5s.pt', # weights path + imgsz=640, # inference size (pixels) + batch_size=1, # batch size + data=ROOT / 'data/coco128.yaml', # dataset.yaml path + device='', # cuda device, i.e. 0 or 0,1,2,3 or cpu + half=False, # use FP16 half-precision inference + test=False, # test exports only + pt_only=False, # test PyTorch only +): + y, t = [], time.time() + device = select_device(device) + for i, (name, f, suffix, gpu) in export.export_formats().iterrows(): # index, (name, file, suffix, gpu-capable) + try: + assert i != 9, 'Edge TPU not supported' + assert i != 10, 'TF.js not supported' + if device.type != 'cpu': + assert gpu, f'{name} inference not supported on GPU' + + # Export + if f == '-': + w = weights # PyTorch format + else: + w = export.run(weights=weights, imgsz=[imgsz], include=[f], device=device, half=half)[-1] # all others + assert suffix in str(w), 'export failed' + + # Validate + result = val.run(data, w, batch_size, imgsz, plots=False, device=device, task='benchmark', half=half) + metrics = result[0] # metrics (mp, mr, map50, map, *losses(box, obj, cls)) + speeds = result[2] # times (preprocess, inference, postprocess) + y.append([name, round(metrics[3], 4), round(speeds[1], 2)]) # mAP, t_inference + except Exception as e: + LOGGER.warning(f'WARNING: Benchmark failure for {name}: {e}') + y.append([name, None, None]) # mAP, t_inference + if pt_only and i == 0: + break # break after PyTorch + + # Print results + LOGGER.info('\n') + parse_opt() + notebook_init() # print system info + py = pd.DataFrame(y, columns=['Format', 'mAP@0.5:0.95', 'Inference time (ms)'] if map else ['Format', 'Export', '']) + LOGGER.info(f'\nBenchmarks complete ({time.time() - t:.2f}s)') + LOGGER.info(str(py if map else py.iloc[:, :2])) + return py + + +def test( + weights=ROOT / 'yolov5s.pt', # weights path + imgsz=640, # inference size (pixels) + batch_size=1, # batch size + data=ROOT / 'data/coco128.yaml', # dataset.yaml path + device='', # cuda device, i.e. 0 or 0,1,2,3 or cpu + half=False, # use FP16 half-precision inference + test=False, # test exports only + pt_only=False, # test PyTorch only +): + y, t = [], time.time() + device = select_device(device) + for i, (name, f, suffix, gpu) in export.export_formats().iterrows(): # index, (name, file, suffix, gpu-capable) + try: + w = weights if f == '-' else \ + export.run(weights=weights, imgsz=[imgsz], include=[f], device=device, half=half)[-1] # weights + assert suffix in str(w), 'export failed' + y.append([name, True]) + except Exception: + y.append([name, False]) # mAP, t_inference + + # Print results + LOGGER.info('\n') + parse_opt() + notebook_init() # print system info + py = pd.DataFrame(y, columns=['Format', 'Export']) + LOGGER.info(f'\nExports complete ({time.time() - t:.2f}s)') + LOGGER.info(str(py)) + return py + + +def parse_opt(): + parser = argparse.ArgumentParser() + parser.add_argument('--weights', type=str, default=ROOT / 'yolov5s.pt', help='weights path') + parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, help='inference size (pixels)') + parser.add_argument('--batch-size', type=int, default=1, help='batch size') + parser.add_argument('--data', type=str, default=ROOT / 'data/coco128.yaml', help='dataset.yaml path') + parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') + parser.add_argument('--half', action='store_true', help='use FP16 half-precision inference') + parser.add_argument('--test', action='store_true', help='test exports only') + parser.add_argument('--pt-only', action='store_true', help='test PyTorch only') + opt = parser.parse_args() + opt.data = check_yaml(opt.data) # check YAML + print_args(vars(opt)) + return opt + + +def main(opt): + test(**vars(opt)) if opt.test else run(**vars(opt)) + + +if __name__ == "__main__": + opt = parse_opt() + main(opt) diff --git a/yolov5/utils/callbacks.py b/yolov5/utils/callbacks.py new file mode 100644 index 0000000..2b32df0 --- /dev/null +++ b/yolov5/utils/callbacks.py @@ -0,0 +1,71 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Callback utils +""" + + +class Callbacks: + """" + Handles all registered callbacks for YOLOv5 Hooks + """ + + def __init__(self): + # Define the available callbacks + self._callbacks = { + 'on_pretrain_routine_start': [], + 'on_pretrain_routine_end': [], + 'on_train_start': [], + 'on_train_epoch_start': [], + 'on_train_batch_start': [], + 'optimizer_step': [], + 'on_before_zero_grad': [], + 'on_train_batch_end': [], + 'on_train_epoch_end': [], + 'on_val_start': [], + 'on_val_batch_start': [], + 'on_val_image_end': [], + 'on_val_batch_end': [], + 'on_val_end': [], + 'on_fit_epoch_end': [], # fit = train + val + 'on_model_save': [], + 'on_train_end': [], + 'on_params_update': [], + 'teardown': [],} + self.stop_training = False # set True to interrupt training + + def register_action(self, hook, name='', callback=None): + """ + Register a new action to a callback hook + + Args: + hook: The callback hook name to register the action to + name: The name of the action for later reference + callback: The callback to fire + """ + assert hook in self._callbacks, f"hook '{hook}' not found in callbacks {self._callbacks}" + assert callable(callback), f"callback '{callback}' is not callable" + self._callbacks[hook].append({'name': name, 'callback': callback}) + + def get_registered_actions(self, hook=None): + """" + Returns all the registered actions by callback hook + + Args: + hook: The name of the hook to check, defaults to all + """ + return self._callbacks[hook] if hook else self._callbacks + + def run(self, hook, *args, **kwargs): + """ + Loop through the registered actions and fire all callbacks + + Args: + hook: The name of the hook to check, defaults to all + args: Arguments to receive from YOLOv5 + kwargs: Keyword Arguments to receive from YOLOv5 + """ + + assert hook in self._callbacks, f"hook '{hook}' not found in callbacks {self._callbacks}" + + for logger in self._callbacks[hook]: + logger['callback'](*args, **kwargs) diff --git a/yolov5/utils/dataloaders.py b/yolov5/utils/dataloaders.py new file mode 100644 index 0000000..23ee2d5 --- /dev/null +++ b/yolov5/utils/dataloaders.py @@ -0,0 +1,1076 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Dataloaders and dataset utils +""" + +import glob +import hashlib +import json +import math +import os +import random +import shutil +import time +from itertools import repeat +from multiprocessing.pool import Pool, ThreadPool +from pathlib import Path +from threading import Thread +from urllib.parse import urlparse +from zipfile import ZipFile + +import numpy as np +import torch +import torch.nn.functional as F +import yaml +from PIL import ExifTags, Image, ImageOps +from torch.utils.data import DataLoader, Dataset, dataloader, distributed +from tqdm import tqdm + +from utils.augmentations import Albumentations, augment_hsv, copy_paste, letterbox, mixup, random_perspective +from utils.general import (DATASETS_DIR, LOGGER, NUM_THREADS, check_dataset, check_requirements, check_yaml, clean_str, + cv2, segments2boxes, xyn2xy, xywh2xyxy, xywhn2xyxy, xyxy2xywhn) +from utils.torch_utils import torch_distributed_zero_first + +# Parameters +HELP_URL = 'https://github.com/ultralytics/yolov5/wiki/Train-Custom-Data' +IMG_FORMATS = 'bmp', 'dng', 'jpeg', 'jpg', 'mpo', 'png', 'tif', 'tiff', 'webp' # include image suffixes +VID_FORMATS = 'asf', 'avi', 'gif', 'm4v', 'mkv', 'mov', 'mp4', 'mpeg', 'mpg', 'ts', 'wmv' # include video suffixes +BAR_FORMAT = '{l_bar}{bar:10}{r_bar}{bar:-10b}' # tqdm bar format +LOCAL_RANK = int(os.getenv('LOCAL_RANK', -1)) # https://pytorch.org/docs/stable/elastic/run.html + +# Get orientation exif tag +for orientation in ExifTags.TAGS.keys(): + if ExifTags.TAGS[orientation] == 'Orientation': + break + + +def get_hash(paths): + # Returns a single hash value of a list of paths (files or dirs) + size = sum(os.path.getsize(p) for p in paths if os.path.exists(p)) # sizes + h = hashlib.md5(str(size).encode()) # hash sizes + h.update(''.join(paths).encode()) # hash paths + return h.hexdigest() # return hash + + +def exif_size(img): + # Returns exif-corrected PIL size + s = img.size # (width, height) + try: + rotation = dict(img._getexif().items())[orientation] + if rotation in [6, 8]: # rotation 270 or 90 + s = (s[1], s[0]) + except Exception: + pass + + return s + + +def exif_transpose(image): + """ + Transpose a PIL image accordingly if it has an EXIF Orientation tag. + Inplace version of https://github.com/python-pillow/Pillow/blob/master/src/PIL/ImageOps.py exif_transpose() + + :param image: The image to transpose. + :return: An image. + """ + exif = image.getexif() + orientation = exif.get(0x0112, 1) # default 1 + if orientation > 1: + method = { + 2: Image.FLIP_LEFT_RIGHT, + 3: Image.ROTATE_180, + 4: Image.FLIP_TOP_BOTTOM, + 5: Image.TRANSPOSE, + 6: Image.ROTATE_270, + 7: Image.TRANSVERSE, + 8: Image.ROTATE_90,}.get(orientation) + if method is not None: + image = image.transpose(method) + del exif[0x0112] + image.info["exif"] = exif.tobytes() + return image + + +def create_dataloader(path, + imgsz, + batch_size, + stride, + single_cls=False, + hyp=None, + augment=False, + cache=False, + pad=0.0, + rect=False, + rank=-1, + workers=8, + image_weights=False, + quad=False, + prefix='', + shuffle=False): + if rect and shuffle: + LOGGER.warning('WARNING: --rect is incompatible with DataLoader shuffle, setting shuffle=False') + shuffle = False + with torch_distributed_zero_first(rank): # init dataset *.cache only once if DDP + dataset = LoadImagesAndLabels( + path, + imgsz, + batch_size, + augment=augment, # augmentation + hyp=hyp, # hyperparameters + rect=rect, # rectangular batches + cache_images=cache, + single_cls=single_cls, + stride=int(stride), + pad=pad, + image_weights=image_weights, + prefix=prefix) + + batch_size = min(batch_size, len(dataset)) + nd = torch.cuda.device_count() # number of CUDA devices + nw = min([os.cpu_count() // max(nd, 1), batch_size if batch_size > 1 else 0, workers]) # number of workers + sampler = None if rank == -1 else distributed.DistributedSampler(dataset, shuffle=shuffle) + loader = DataLoader if image_weights else InfiniteDataLoader # only DataLoader allows for attribute updates + return loader(dataset, + batch_size=batch_size, + shuffle=shuffle and sampler is None, + num_workers=nw, + sampler=sampler, + pin_memory=True, + collate_fn=LoadImagesAndLabels.collate_fn4 if quad else LoadImagesAndLabels.collate_fn), dataset + + +class InfiniteDataLoader(dataloader.DataLoader): + """ Dataloader that reuses workers + + Uses same syntax as vanilla DataLoader + """ + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + object.__setattr__(self, 'batch_sampler', _RepeatSampler(self.batch_sampler)) + self.iterator = super().__iter__() + + def __len__(self): + return len(self.batch_sampler.sampler) + + def __iter__(self): + for _ in range(len(self)): + yield next(self.iterator) + + +class _RepeatSampler: + """ Sampler that repeats forever + + Args: + sampler (Sampler) + """ + + def __init__(self, sampler): + self.sampler = sampler + + def __iter__(self): + while True: + yield from iter(self.sampler) + + +class LoadImages: + # YOLOv5 image/video dataloader, i.e. `python detect.py --source image.jpg/vid.mp4` + def __init__(self, path, img_size=640, stride=32, auto=True): + p = str(Path(path).resolve()) # os-agnostic absolute path + if '*' in p: + files = sorted(glob.glob(p, recursive=True)) # glob + elif os.path.isdir(p): + files = sorted(glob.glob(os.path.join(p, '*.*'))) # dir + elif os.path.isfile(p): + files = [p] # files + else: + raise Exception(f'ERROR: {p} does not exist') + + images = [x for x in files if x.split('.')[-1].lower() in IMG_FORMATS] + videos = [x for x in files if x.split('.')[-1].lower() in VID_FORMATS] + ni, nv = len(images), len(videos) + + self.img_size = img_size + self.stride = stride + self.files = images + videos + self.nf = ni + nv # number of files + self.video_flag = [False] * ni + [True] * nv + self.mode = 'image' + self.auto = auto + if any(videos): + self.new_video(videos[0]) # new video + else: + self.cap = None + assert self.nf > 0, f'No images or videos found in {p}. ' \ + f'Supported formats are:\nimages: {IMG_FORMATS}\nvideos: {VID_FORMATS}' + + def __iter__(self): + self.count = 0 + return self + + def __next__(self): + if self.count == self.nf: + raise StopIteration + path = self.files[self.count] + + if self.video_flag[self.count]: + # Read video + self.mode = 'video' + ret_val, img0 = self.cap.read() + while not ret_val: + self.count += 1 + self.cap.release() + if self.count == self.nf: # last video + raise StopIteration + path = self.files[self.count] + self.new_video(path) + ret_val, img0 = self.cap.read() + + self.frame += 1 + s = f'video {self.count + 1}/{self.nf} ({self.frame}/{self.frames}) {path}: ' + + else: + # Read image + self.count += 1 + img0 = cv2.imread(path) # BGR + assert img0 is not None, f'Image Not Found {path}' + s = f'image {self.count}/{self.nf} {path}: ' + + # Padded resize + img = letterbox(img0, self.img_size, stride=self.stride, auto=self.auto)[0] + + # Convert + img = img.transpose((2, 0, 1))[::-1] # HWC to CHW, BGR to RGB + img = np.ascontiguousarray(img) + + return path, img, img0, self.cap, s + + def new_video(self, path): + self.frame = 0 + self.cap = cv2.VideoCapture(path) + self.frames = int(self.cap.get(cv2.CAP_PROP_FRAME_COUNT)) + + def __len__(self): + return self.nf # number of files + + +class LoadWebcam: # for inference + # YOLOv5 local webcam dataloader, i.e. `python detect.py --source 0` + def __init__(self, pipe='0', img_size=640, stride=32): + self.img_size = img_size + self.stride = stride + self.pipe = eval(pipe) if pipe.isnumeric() else pipe + self.cap = cv2.VideoCapture(self.pipe) # video capture object + self.cap.set(cv2.CAP_PROP_BUFFERSIZE, 3) # set buffer size + + def __iter__(self): + self.count = -1 + return self + + def __next__(self): + self.count += 1 + if cv2.waitKey(1) == ord('q'): # q to quit + self.cap.release() + cv2.destroyAllWindows() + raise StopIteration + + # Read frame + ret_val, img0 = self.cap.read() + img0 = cv2.flip(img0, 1) # flip left-right + + # Print + assert ret_val, f'Camera Error {self.pipe}' + img_path = 'webcam.jpg' + s = f'webcam {self.count}: ' + + # Padded resize + img = letterbox(img0, self.img_size, stride=self.stride)[0] + + # Convert + img = img.transpose((2, 0, 1))[::-1] # HWC to CHW, BGR to RGB + img = np.ascontiguousarray(img) + + return img_path, img, img0, None, s + + def __len__(self): + return 0 + + +class LoadStreams: + # YOLOv5 streamloader, i.e. `python detect.py --source 'rtsp://example.com/media.mp4' # RTSP, RTMP, HTTP streams` + def __init__(self, sources='streams.txt', img_size=640, stride=32, auto=True): + self.mode = 'stream' + self.img_size = img_size + self.stride = stride + + if os.path.isfile(sources): + with open(sources) as f: + sources = [x.strip() for x in f.read().strip().splitlines() if len(x.strip())] + else: + sources = [sources] + + n = len(sources) + self.imgs, self.fps, self.frames, self.threads = [None] * n, [0] * n, [0] * n, [None] * n + self.sources = [clean_str(x) for x in sources] # clean source names for later + self.auto = auto + for i, s in enumerate(sources): # index, source + # Start thread to read frames from video stream + st = f'{i + 1}/{n}: {s}... ' + if urlparse(s).hostname in ('www.youtube.com', 'youtube.com', 'youtu.be'): # if source is YouTube video + check_requirements(('pafy', 'youtube_dl==2020.12.2')) + import pafy + s = pafy.new(s).getbest(preftype="mp4").url # YouTube URL + s = eval(s) if s.isnumeric() else s # i.e. s = '0' local webcam + cap = cv2.VideoCapture(s) + assert cap.isOpened(), f'{st}Failed to open {s}' + w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH)) + h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT)) + fps = cap.get(cv2.CAP_PROP_FPS) # warning: may return 0 or nan + self.frames[i] = max(int(cap.get(cv2.CAP_PROP_FRAME_COUNT)), 0) or float('inf') # infinite stream fallback + self.fps[i] = max((fps if math.isfinite(fps) else 0) % 100, 0) or 30 # 30 FPS fallback + + _, self.imgs[i] = cap.read() # guarantee first frame + self.threads[i] = Thread(target=self.update, args=([i, cap, s]), daemon=True) + LOGGER.info(f"{st} Success ({self.frames[i]} frames {w}x{h} at {self.fps[i]:.2f} FPS)") + self.threads[i].start() + LOGGER.info('') # newline + + # check for common shapes + s = np.stack([letterbox(x, self.img_size, stride=self.stride, auto=self.auto)[0].shape for x in self.imgs]) + self.rect = np.unique(s, axis=0).shape[0] == 1 # rect inference if all shapes equal + if not self.rect: + LOGGER.warning('WARNING: Stream shapes differ. For optimal performance supply similarly-shaped streams.') + + def update(self, i, cap, stream): + # Read stream `i` frames in daemon thread + n, f, read = 0, self.frames[i], 1 # frame number, frame array, inference every 'read' frame + while cap.isOpened() and n < f: + n += 1 + # _, self.imgs[index] = cap.read() + cap.grab() + if n % read == 0: + success, im = cap.retrieve() + if success: + self.imgs[i] = im + else: + LOGGER.warning('WARNING: Video stream unresponsive, please check your IP camera connection.') + self.imgs[i] = np.zeros_like(self.imgs[i]) + cap.open(stream) # re-open stream if signal was lost + time.sleep(1 / self.fps[i]) # wait time + + def __iter__(self): + self.count = -1 + return self + + def __next__(self): + self.count += 1 + if not all(x.is_alive() for x in self.threads) or cv2.waitKey(1) == ord('q'): # q to quit + cv2.destroyAllWindows() + raise StopIteration + + # Letterbox + img0 = self.imgs.copy() + img = [letterbox(x, self.img_size, stride=self.stride, auto=self.rect and self.auto)[0] for x in img0] + + # Stack + img = np.stack(img, 0) + + # Convert + img = img[..., ::-1].transpose((0, 3, 1, 2)) # BGR to RGB, BHWC to BCHW + img = np.ascontiguousarray(img) + + return self.sources, img, img0, None, '' + + def __len__(self): + return len(self.sources) # 1E12 frames = 32 streams at 30 FPS for 30 years + + +def img2label_paths(img_paths): + # Define label paths as a function of image paths + sa, sb = f'{os.sep}images{os.sep}', f'{os.sep}labels{os.sep}' # /images/, /labels/ substrings + return [sb.join(x.rsplit(sa, 1)).rsplit('.', 1)[0] + '.txt' for x in img_paths] + + +class LoadImagesAndLabels(Dataset): + # YOLOv5 train_loader/val_loader, loads images and labels for training and validation + cache_version = 0.6 # dataset labels *.cache version + rand_interp_methods = [cv2.INTER_NEAREST, cv2.INTER_LINEAR, cv2.INTER_CUBIC, cv2.INTER_AREA, cv2.INTER_LANCZOS4] + + def __init__(self, + path, + img_size=640, + batch_size=16, + augment=False, + hyp=None, + rect=False, + image_weights=False, + cache_images=False, + single_cls=False, + stride=32, + pad=0.0, + prefix=''): + self.img_size = img_size + self.augment = augment + self.hyp = hyp + self.image_weights = image_weights + self.rect = False if image_weights else rect + self.mosaic = self.augment and not self.rect # load 4 images at a time into a mosaic (only during training) + self.mosaic_border = [-img_size // 2, -img_size // 2] + self.stride = stride + self.path = path + self.albumentations = Albumentations() if augment else None + + try: + f = [] # image files + for p in path if isinstance(path, list) else [path]: + p = Path(p) # os-agnostic + if p.is_dir(): # dir + f += glob.glob(str(p / '**' / '*.*'), recursive=True) + # f = list(p.rglob('*.*')) # pathlib + elif p.is_file(): # file + with open(p) as t: + t = t.read().strip().splitlines() + parent = str(p.parent) + os.sep + f += [x.replace('./', parent) if x.startswith('./') else x for x in t] # local to global path + # f += [p.parent / x.lstrip(os.sep) for x in t] # local to global path (pathlib) + else: + raise Exception(f'{prefix}{p} does not exist') + self.im_files = sorted(x.replace('/', os.sep) for x in f if x.split('.')[-1].lower() in IMG_FORMATS) + # self.img_files = sorted([x for x in f if x.suffix[1:].lower() in IMG_FORMATS]) # pathlib + assert self.im_files, f'{prefix}No images found' + except Exception as e: + raise Exception(f'{prefix}Error loading data from {path}: {e}\nSee {HELP_URL}') + + # Check cache + self.label_files = img2label_paths(self.im_files) # labels + cache_path = (p if p.is_file() else Path(self.label_files[0]).parent).with_suffix('.cache') + try: + cache, exists = np.load(cache_path, allow_pickle=True).item(), True # load dict + assert cache['version'] == self.cache_version # same version + assert cache['hash'] == get_hash(self.label_files + self.im_files) # same hash + except Exception: + cache, exists = self.cache_labels(cache_path, prefix), False # cache + + # Display cache + nf, nm, ne, nc, n = cache.pop('results') # found, missing, empty, corrupt, total + if exists and LOCAL_RANK in {-1, 0}: + d = f"Scanning '{cache_path}' images and labels... {nf} found, {nm} missing, {ne} empty, {nc} corrupt" + tqdm(None, desc=prefix + d, total=n, initial=n, bar_format=BAR_FORMAT) # display cache results + if cache['msgs']: + LOGGER.info('\n'.join(cache['msgs'])) # display warnings + assert nf > 0 or not augment, f'{prefix}No labels in {cache_path}. Can not train without labels. See {HELP_URL}' + + # Read cache + [cache.pop(k) for k in ('hash', 'version', 'msgs')] # remove items + labels, shapes, self.segments = zip(*cache.values()) + self.labels = list(labels) + self.shapes = np.array(shapes, dtype=np.float64) + self.im_files = list(cache.keys()) # update + self.label_files = img2label_paths(cache.keys()) # update + n = len(shapes) # number of images + bi = np.floor(np.arange(n) / batch_size).astype(np.int) # batch index + nb = bi[-1] + 1 # number of batches + self.batch = bi # batch index of image + self.n = n + self.indices = range(n) + + # Update labels + include_class = [] # filter labels to include only these classes (optional) + include_class_array = np.array(include_class).reshape(1, -1) + for i, (label, segment) in enumerate(zip(self.labels, self.segments)): + if include_class: + j = (label[:, 0:1] == include_class_array).any(1) + self.labels[i] = label[j] + if segment: + self.segments[i] = segment[j] + if single_cls: # single-class training, merge all classes into 0 + self.labels[i][:, 0] = 0 + if segment: + self.segments[i][:, 0] = 0 + + # Rectangular Training + if self.rect: + # Sort by aspect ratio + s = self.shapes # wh + ar = s[:, 1] / s[:, 0] # aspect ratio + irect = ar.argsort() + self.im_files = [self.im_files[i] for i in irect] + self.label_files = [self.label_files[i] for i in irect] + self.labels = [self.labels[i] for i in irect] + self.shapes = s[irect] # wh + ar = ar[irect] + + # Set training image shapes + shapes = [[1, 1]] * nb + for i in range(nb): + ari = ar[bi == i] + mini, maxi = ari.min(), ari.max() + if maxi < 1: + shapes[i] = [maxi, 1] + elif mini > 1: + shapes[i] = [1, 1 / mini] + + self.batch_shapes = np.ceil(np.array(shapes) * img_size / stride + pad).astype(np.int) * stride + + # Cache images into RAM/disk for faster training (WARNING: large datasets may exceed system resources) + self.ims = [None] * n + self.npy_files = [Path(f).with_suffix('.npy') for f in self.im_files] + if cache_images: + gb = 0 # Gigabytes of cached images + self.im_hw0, self.im_hw = [None] * n, [None] * n + fcn = self.cache_images_to_disk if cache_images == 'disk' else self.load_image + results = ThreadPool(NUM_THREADS).imap(fcn, range(n)) + pbar = tqdm(enumerate(results), total=n, bar_format=BAR_FORMAT, disable=LOCAL_RANK > 0) + for i, x in pbar: + if cache_images == 'disk': + gb += self.npy_files[i].stat().st_size + else: # 'ram' + self.ims[i], self.im_hw0[i], self.im_hw[i] = x # im, hw_orig, hw_resized = load_image(self, i) + gb += self.ims[i].nbytes + pbar.desc = f'{prefix}Caching images ({gb / 1E9:.1f}GB {cache_images})' + pbar.close() + + def cache_labels(self, path=Path('./labels.cache'), prefix=''): + # Cache dataset labels, check images and read shapes + x = {} # dict + nm, nf, ne, nc, msgs = 0, 0, 0, 0, [] # number missing, found, empty, corrupt, messages + desc = f"{prefix}Scanning '{path.parent / path.stem}' images and labels..." + with Pool(NUM_THREADS) as pool: + pbar = tqdm(pool.imap(verify_image_label, zip(self.im_files, self.label_files, repeat(prefix))), + desc=desc, + total=len(self.im_files), + bar_format=BAR_FORMAT) + for im_file, lb, shape, segments, nm_f, nf_f, ne_f, nc_f, msg in pbar: + nm += nm_f + nf += nf_f + ne += ne_f + nc += nc_f + if im_file: + x[im_file] = [lb, shape, segments] + if msg: + msgs.append(msg) + pbar.desc = f"{desc}{nf} found, {nm} missing, {ne} empty, {nc} corrupt" + + pbar.close() + if msgs: + LOGGER.info('\n'.join(msgs)) + if nf == 0: + LOGGER.warning(f'{prefix}WARNING: No labels found in {path}. See {HELP_URL}') + x['hash'] = get_hash(self.label_files + self.im_files) + x['results'] = nf, nm, ne, nc, len(self.im_files) + x['msgs'] = msgs # warnings + x['version'] = self.cache_version # cache version + try: + np.save(path, x) # save cache for next time + path.with_suffix('.cache.npy').rename(path) # remove .npy suffix + LOGGER.info(f'{prefix}New cache created: {path}') + except Exception as e: + LOGGER.warning(f'{prefix}WARNING: Cache directory {path.parent} is not writeable: {e}') # not writeable + return x + + def __len__(self): + return len(self.im_files) + + # def __iter__(self): + # self.count = -1 + # print('ran dataset iter') + # #self.shuffled_vector = np.random.permutation(self.nF) if self.augment else np.arange(self.nF) + # return self + + def __getitem__(self, index): + index = self.indices[index] # linear, shuffled, or image_weights + + hyp = self.hyp + mosaic = self.mosaic and random.random() < hyp['mosaic'] + if mosaic: + # Load mosaic + img, labels = self.load_mosaic(index) + shapes = None + + # MixUp augmentation + if random.random() < hyp['mixup']: + img, labels = mixup(img, labels, *self.load_mosaic(random.randint(0, self.n - 1))) + + else: + # Load image + img, (h0, w0), (h, w) = self.load_image(index) + + # Letterbox + shape = self.batch_shapes[self.batch[index]] if self.rect else self.img_size # final letterboxed shape + img, ratio, pad = letterbox(img, shape, auto=False, scaleup=self.augment) + shapes = (h0, w0), ((h / h0, w / w0), pad) # for COCO mAP rescaling + + labels = self.labels[index].copy() + if labels.size: # normalized xywh to pixel xyxy format + labels[:, 1:] = xywhn2xyxy(labels[:, 1:], ratio[0] * w, ratio[1] * h, padw=pad[0], padh=pad[1]) + + if self.augment: + img, labels = random_perspective(img, + labels, + degrees=hyp['degrees'], + translate=hyp['translate'], + scale=hyp['scale'], + shear=hyp['shear'], + perspective=hyp['perspective']) + + nl = len(labels) # number of labels + if nl: + labels[:, 1:5] = xyxy2xywhn(labels[:, 1:5], w=img.shape[1], h=img.shape[0], clip=True, eps=1E-3) + + if self.augment: + # Albumentations + img, labels = self.albumentations(img, labels) + nl = len(labels) # update after albumentations + + # HSV color-space + augment_hsv(img, hgain=hyp['hsv_h'], sgain=hyp['hsv_s'], vgain=hyp['hsv_v']) + + # Flip up-down + if random.random() < hyp['flipud']: + img = np.flipud(img) + if nl: + labels[:, 2] = 1 - labels[:, 2] + + # Flip left-right + if random.random() < hyp['fliplr']: + img = np.fliplr(img) + if nl: + labels[:, 1] = 1 - labels[:, 1] + + # Cutouts + # labels = cutout(img, labels, p=0.5) + # nl = len(labels) # update after cutout + + labels_out = torch.zeros((nl, 6)) + if nl: + labels_out[:, 1:] = torch.from_numpy(labels) + + # Convert + img = img.transpose((2, 0, 1))[::-1] # HWC to CHW, BGR to RGB + img = np.ascontiguousarray(img) + + return torch.from_numpy(img), labels_out, self.im_files[index], shapes + + def load_image(self, i): + # Loads 1 image from dataset index 'i', returns (im, original hw, resized hw) + im, f, fn = self.ims[i], self.im_files[i], self.npy_files[i], + if im is None: # not cached in RAM + if fn.exists(): # load npy + im = np.load(fn) + else: # read image + im = cv2.imread(f) # BGR + assert im is not None, f'Image Not Found {f}' + h0, w0 = im.shape[:2] # orig hw + r = self.img_size / max(h0, w0) # ratio + if r != 1: # if sizes are not equal + interp = cv2.INTER_LINEAR if (self.augment or r > 1) else cv2.INTER_AREA + im = cv2.resize(im, (int(w0 * r), int(h0 * r)), interpolation=interp) + return im, (h0, w0), im.shape[:2] # im, hw_original, hw_resized + else: + return self.ims[i], self.im_hw0[i], self.im_hw[i] # im, hw_original, hw_resized + + def cache_images_to_disk(self, i): + # Saves an image as an *.npy file for faster loading + f = self.npy_files[i] + if not f.exists(): + np.save(f.as_posix(), cv2.imread(self.im_files[i])) + + def load_mosaic(self, index): + # YOLOv5 4-mosaic loader. Loads 1 image + 3 random images into a 4-image mosaic + labels4, segments4 = [], [] + s = self.img_size + yc, xc = (int(random.uniform(-x, 2 * s + x)) for x in self.mosaic_border) # mosaic center x, y + indices = [index] + random.choices(self.indices, k=3) # 3 additional image indices + random.shuffle(indices) + for i, index in enumerate(indices): + # Load image + img, _, (h, w) = self.load_image(index) + + # place img in img4 + if i == 0: # top left + img4 = np.full((s * 2, s * 2, img.shape[2]), 114, dtype=np.uint8) # base image with 4 tiles + x1a, y1a, x2a, y2a = max(xc - w, 0), max(yc - h, 0), xc, yc # xmin, ymin, xmax, ymax (large image) + x1b, y1b, x2b, y2b = w - (x2a - x1a), h - (y2a - y1a), w, h # xmin, ymin, xmax, ymax (small image) + elif i == 1: # top right + x1a, y1a, x2a, y2a = xc, max(yc - h, 0), min(xc + w, s * 2), yc + x1b, y1b, x2b, y2b = 0, h - (y2a - y1a), min(w, x2a - x1a), h + elif i == 2: # bottom left + x1a, y1a, x2a, y2a = max(xc - w, 0), yc, xc, min(s * 2, yc + h) + x1b, y1b, x2b, y2b = w - (x2a - x1a), 0, w, min(y2a - y1a, h) + elif i == 3: # bottom right + x1a, y1a, x2a, y2a = xc, yc, min(xc + w, s * 2), min(s * 2, yc + h) + x1b, y1b, x2b, y2b = 0, 0, min(w, x2a - x1a), min(y2a - y1a, h) + + img4[y1a:y2a, x1a:x2a] = img[y1b:y2b, x1b:x2b] # img4[ymin:ymax, xmin:xmax] + padw = x1a - x1b + padh = y1a - y1b + + # Labels + labels, segments = self.labels[index].copy(), self.segments[index].copy() + if labels.size: + labels[:, 1:] = xywhn2xyxy(labels[:, 1:], w, h, padw, padh) # normalized xywh to pixel xyxy format + segments = [xyn2xy(x, w, h, padw, padh) for x in segments] + labels4.append(labels) + segments4.extend(segments) + + # Concat/clip labels + labels4 = np.concatenate(labels4, 0) + for x in (labels4[:, 1:], *segments4): + np.clip(x, 0, 2 * s, out=x) # clip when using random_perspective() + # img4, labels4 = replicate(img4, labels4) # replicate + + # Augment + img4, labels4, segments4 = copy_paste(img4, labels4, segments4, p=self.hyp['copy_paste']) + img4, labels4 = random_perspective(img4, + labels4, + segments4, + degrees=self.hyp['degrees'], + translate=self.hyp['translate'], + scale=self.hyp['scale'], + shear=self.hyp['shear'], + perspective=self.hyp['perspective'], + border=self.mosaic_border) # border to remove + + return img4, labels4 + + def load_mosaic9(self, index): + # YOLOv5 9-mosaic loader. Loads 1 image + 8 random images into a 9-image mosaic + labels9, segments9 = [], [] + s = self.img_size + indices = [index] + random.choices(self.indices, k=8) # 8 additional image indices + random.shuffle(indices) + hp, wp = -1, -1 # height, width previous + for i, index in enumerate(indices): + # Load image + img, _, (h, w) = self.load_image(index) + + # place img in img9 + if i == 0: # center + img9 = np.full((s * 3, s * 3, img.shape[2]), 114, dtype=np.uint8) # base image with 4 tiles + h0, w0 = h, w + c = s, s, s + w, s + h # xmin, ymin, xmax, ymax (base) coordinates + elif i == 1: # top + c = s, s - h, s + w, s + elif i == 2: # top right + c = s + wp, s - h, s + wp + w, s + elif i == 3: # right + c = s + w0, s, s + w0 + w, s + h + elif i == 4: # bottom right + c = s + w0, s + hp, s + w0 + w, s + hp + h + elif i == 5: # bottom + c = s + w0 - w, s + h0, s + w0, s + h0 + h + elif i == 6: # bottom left + c = s + w0 - wp - w, s + h0, s + w0 - wp, s + h0 + h + elif i == 7: # left + c = s - w, s + h0 - h, s, s + h0 + elif i == 8: # top left + c = s - w, s + h0 - hp - h, s, s + h0 - hp + + padx, pady = c[:2] + x1, y1, x2, y2 = (max(x, 0) for x in c) # allocate coords + + # Labels + labels, segments = self.labels[index].copy(), self.segments[index].copy() + if labels.size: + labels[:, 1:] = xywhn2xyxy(labels[:, 1:], w, h, padx, pady) # normalized xywh to pixel xyxy format + segments = [xyn2xy(x, w, h, padx, pady) for x in segments] + labels9.append(labels) + segments9.extend(segments) + + # Image + img9[y1:y2, x1:x2] = img[y1 - pady:, x1 - padx:] # img9[ymin:ymax, xmin:xmax] + hp, wp = h, w # height, width previous + + # Offset + yc, xc = (int(random.uniform(0, s)) for _ in self.mosaic_border) # mosaic center x, y + img9 = img9[yc:yc + 2 * s, xc:xc + 2 * s] + + # Concat/clip labels + labels9 = np.concatenate(labels9, 0) + labels9[:, [1, 3]] -= xc + labels9[:, [2, 4]] -= yc + c = np.array([xc, yc]) # centers + segments9 = [x - c for x in segments9] + + for x in (labels9[:, 1:], *segments9): + np.clip(x, 0, 2 * s, out=x) # clip when using random_perspective() + # img9, labels9 = replicate(img9, labels9) # replicate + + # Augment + img9, labels9 = random_perspective(img9, + labels9, + segments9, + degrees=self.hyp['degrees'], + translate=self.hyp['translate'], + scale=self.hyp['scale'], + shear=self.hyp['shear'], + perspective=self.hyp['perspective'], + border=self.mosaic_border) # border to remove + + return img9, labels9 + + @staticmethod + def collate_fn(batch): + im, label, path, shapes = zip(*batch) # transposed + for i, lb in enumerate(label): + lb[:, 0] = i # add target image index for build_targets() + return torch.stack(im, 0), torch.cat(label, 0), path, shapes + + @staticmethod + def collate_fn4(batch): + img, label, path, shapes = zip(*batch) # transposed + n = len(shapes) // 4 + im4, label4, path4, shapes4 = [], [], path[:n], shapes[:n] + + ho = torch.tensor([[0.0, 0, 0, 1, 0, 0]]) + wo = torch.tensor([[0.0, 0, 1, 0, 0, 0]]) + s = torch.tensor([[1, 1, 0.5, 0.5, 0.5, 0.5]]) # scale + for i in range(n): # zidane torch.zeros(16,3,720,1280) # BCHW + i *= 4 + if random.random() < 0.5: + im = F.interpolate(img[i].unsqueeze(0).float(), scale_factor=2.0, mode='bilinear', + align_corners=False)[0].type(img[i].type()) + lb = label[i] + else: + im = torch.cat((torch.cat((img[i], img[i + 1]), 1), torch.cat((img[i + 2], img[i + 3]), 1)), 2) + lb = torch.cat((label[i], label[i + 1] + ho, label[i + 2] + wo, label[i + 3] + ho + wo), 0) * s + im4.append(im) + label4.append(lb) + + for i, lb in enumerate(label4): + lb[:, 0] = i # add target image index for build_targets() + + return torch.stack(im4, 0), torch.cat(label4, 0), path4, shapes4 + + +# Ancillary functions -------------------------------------------------------------------------------------------------- +def create_folder(path='./new'): + # Create folder + if os.path.exists(path): + shutil.rmtree(path) # delete output folder + os.makedirs(path) # make new output folder + + +def flatten_recursive(path=DATASETS_DIR / 'coco128'): + # Flatten a recursive directory by bringing all files to top level + new_path = Path(str(path) + '_flat') + create_folder(new_path) + for file in tqdm(glob.glob(str(Path(path)) + '/**/*.*', recursive=True)): + shutil.copyfile(file, new_path / Path(file).name) + + +def extract_boxes(path=DATASETS_DIR / 'coco128'): # from utils.datasets import *; extract_boxes() + # Convert detection dataset into classification dataset, with one directory per class + path = Path(path) # images dir + shutil.rmtree(path / 'classifier') if (path / 'classifier').is_dir() else None # remove existing + files = list(path.rglob('*.*')) + n = len(files) # number of files + for im_file in tqdm(files, total=n): + if im_file.suffix[1:] in IMG_FORMATS: + # image + im = cv2.imread(str(im_file))[..., ::-1] # BGR to RGB + h, w = im.shape[:2] + + # labels + lb_file = Path(img2label_paths([str(im_file)])[0]) + if Path(lb_file).exists(): + with open(lb_file) as f: + lb = np.array([x.split() for x in f.read().strip().splitlines()], dtype=np.float32) # labels + + for j, x in enumerate(lb): + c = int(x[0]) # class + f = (path / 'classifier') / f'{c}' / f'{path.stem}_{im_file.stem}_{j}.jpg' # new filename + if not f.parent.is_dir(): + f.parent.mkdir(parents=True) + + b = x[1:] * [w, h, w, h] # box + # b[2:] = b[2:].max() # rectangle to square + b[2:] = b[2:] * 1.2 + 3 # pad + b = xywh2xyxy(b.reshape(-1, 4)).ravel().astype(np.int) + + b[[0, 2]] = np.clip(b[[0, 2]], 0, w) # clip boxes outside of image + b[[1, 3]] = np.clip(b[[1, 3]], 0, h) + assert cv2.imwrite(str(f), im[b[1]:b[3], b[0]:b[2]]), f'box failure in {f}' + + +def autosplit(path=DATASETS_DIR / 'coco128/images', weights=(0.9, 0.1, 0.0), annotated_only=False): + """ Autosplit a dataset into train/val/test splits and save path/autosplit_*.txt files + Usage: from utils.datasets import *; autosplit() + Arguments + path: Path to images directory + weights: Train, val, test weights (list, tuple) + annotated_only: Only use images with an annotated txt file + """ + path = Path(path) # images dir + files = sorted(x for x in path.rglob('*.*') if x.suffix[1:].lower() in IMG_FORMATS) # image files only + n = len(files) # number of files + random.seed(0) # for reproducibility + indices = random.choices([0, 1, 2], weights=weights, k=n) # assign each image to a split + + txt = ['autosplit_train.txt', 'autosplit_val.txt', 'autosplit_test.txt'] # 3 txt files + [(path.parent / x).unlink(missing_ok=True) for x in txt] # remove existing + + print(f'Autosplitting images from {path}' + ', using *.txt labeled images only' * annotated_only) + for i, img in tqdm(zip(indices, files), total=n): + if not annotated_only or Path(img2label_paths([str(img)])[0]).exists(): # check label + with open(path.parent / txt[i], 'a') as f: + f.write('./' + img.relative_to(path.parent).as_posix() + '\n') # add image to txt file + + +def verify_image_label(args): + # Verify one image-label pair + im_file, lb_file, prefix = args + nm, nf, ne, nc, msg, segments = 0, 0, 0, 0, '', [] # number (missing, found, empty, corrupt), message, segments + try: + # verify images + im = Image.open(im_file) + im.verify() # PIL verify + shape = exif_size(im) # image size + assert (shape[0] > 9) & (shape[1] > 9), f'image size {shape} <10 pixels' + assert im.format.lower() in IMG_FORMATS, f'invalid image format {im.format}' + if im.format.lower() in ('jpg', 'jpeg'): + with open(im_file, 'rb') as f: + f.seek(-2, 2) + if f.read() != b'\xff\xd9': # corrupt JPEG + ImageOps.exif_transpose(Image.open(im_file)).save(im_file, 'JPEG', subsampling=0, quality=100) + msg = f'{prefix}WARNING: {im_file}: corrupt JPEG restored and saved' + + # verify labels + if os.path.isfile(lb_file): + nf = 1 # label found + with open(lb_file) as f: + lb = [x.split() for x in f.read().strip().splitlines() if len(x)] + if any(len(x) > 6 for x in lb): # is segment + classes = np.array([x[0] for x in lb], dtype=np.float32) + segments = [np.array(x[1:], dtype=np.float32).reshape(-1, 2) for x in lb] # (cls, xy1...) + lb = np.concatenate((classes.reshape(-1, 1), segments2boxes(segments)), 1) # (cls, xywh) + lb = np.array(lb, dtype=np.float32) + nl = len(lb) + if nl: + assert lb.shape[1] == 5, f'labels require 5 columns, {lb.shape[1]} columns detected' + assert (lb >= 0).all(), f'negative label values {lb[lb < 0]}' + assert (lb[:, 1:] <= 1).all(), f'non-normalized or out of bounds coordinates {lb[:, 1:][lb[:, 1:] > 1]}' + _, i = np.unique(lb, axis=0, return_index=True) + if len(i) < nl: # duplicate row check + lb = lb[i] # remove duplicates + if segments: + segments = segments[i] + msg = f'{prefix}WARNING: {im_file}: {nl - len(i)} duplicate labels removed' + else: + ne = 1 # label empty + lb = np.zeros((0, 5), dtype=np.float32) + else: + nm = 1 # label missing + lb = np.zeros((0, 5), dtype=np.float32) + return im_file, lb, shape, segments, nm, nf, ne, nc, msg + except Exception as e: + nc = 1 + msg = f'{prefix}WARNING: {im_file}: ignoring corrupt image/label: {e}' + return [None, None, None, None, nm, nf, ne, nc, msg] + + +def dataset_stats(path='coco128.yaml', autodownload=False, verbose=False, profile=False, hub=False): + """ Return dataset statistics dictionary with images and instances counts per split per class + To run in parent directory: export PYTHONPATH="$PWD/yolov5" + Usage1: from utils.datasets import *; dataset_stats('coco128.yaml', autodownload=True) + Usage2: from utils.datasets import *; dataset_stats('path/to/coco128_with_yaml.zip') + Arguments + path: Path to data.yaml or data.zip (with data.yaml inside data.zip) + autodownload: Attempt to download dataset if not found locally + verbose: Print stats dictionary + """ + + def round_labels(labels): + # Update labels to integer class and 6 decimal place floats + return [[int(c), *(round(x, 4) for x in points)] for c, *points in labels] + + def unzip(path): + # Unzip data.zip TODO: CONSTRAINT: path/to/abc.zip MUST unzip to 'path/to/abc/' + if str(path).endswith('.zip'): # path is data.zip + assert Path(path).is_file(), f'Error unzipping {path}, file not found' + ZipFile(path).extractall(path=path.parent) # unzip + dir = path.with_suffix('') # dataset directory == zip name + return True, str(dir), next(dir.rglob('*.yaml')) # zipped, data_dir, yaml_path + else: # path is data.yaml + return False, None, path + + def hub_ops(f, max_dim=1920): + # HUB ops for 1 image 'f': resize and save at reduced quality in /dataset-hub for web/app viewing + f_new = im_dir / Path(f).name # dataset-hub image filename + try: # use PIL + im = Image.open(f) + r = max_dim / max(im.height, im.width) # ratio + if r < 1.0: # image too large + im = im.resize((int(im.width * r), int(im.height * r))) + im.save(f_new, 'JPEG', quality=75, optimize=True) # save + except Exception as e: # use OpenCV + print(f'WARNING: HUB ops PIL failure {f}: {e}') + im = cv2.imread(f) + im_height, im_width = im.shape[:2] + r = max_dim / max(im_height, im_width) # ratio + if r < 1.0: # image too large + im = cv2.resize(im, (int(im_width * r), int(im_height * r)), interpolation=cv2.INTER_AREA) + cv2.imwrite(str(f_new), im) + + zipped, data_dir, yaml_path = unzip(Path(path)) + with open(check_yaml(yaml_path), errors='ignore') as f: + data = yaml.safe_load(f) # data dict + if zipped: + data['path'] = data_dir # TODO: should this be dir.resolve()? + check_dataset(data, autodownload) # download dataset if missing + hub_dir = Path(data['path'] + ('-hub' if hub else '')) + stats = {'nc': data['nc'], 'names': data['names']} # statistics dictionary + for split in 'train', 'val', 'test': + if data.get(split) is None: + stats[split] = None # i.e. no test set + continue + x = [] + dataset = LoadImagesAndLabels(data[split]) # load dataset + for label in tqdm(dataset.labels, total=dataset.n, desc='Statistics'): + x.append(np.bincount(label[:, 0].astype(int), minlength=data['nc'])) + x = np.array(x) # shape(128x80) + stats[split] = { + 'instance_stats': { + 'total': int(x.sum()), + 'per_class': x.sum(0).tolist()}, + 'image_stats': { + 'total': dataset.n, + 'unlabelled': int(np.all(x == 0, 1).sum()), + 'per_class': (x > 0).sum(0).tolist()}, + 'labels': [{ + str(Path(k).name): round_labels(v.tolist())} for k, v in zip(dataset.im_files, dataset.labels)]} + + if hub: + im_dir = hub_dir / 'images' + im_dir.mkdir(parents=True, exist_ok=True) + for _ in tqdm(ThreadPool(NUM_THREADS).imap(hub_ops, dataset.im_files), total=dataset.n, desc='HUB Ops'): + pass + + # Profile + stats_path = hub_dir / 'stats.json' + if profile: + for _ in range(1): + file = stats_path.with_suffix('.npy') + t1 = time.time() + np.save(file, stats) + t2 = time.time() + x = np.load(file, allow_pickle=True) + print(f'stats.npy times: {time.time() - t2:.3f}s read, {t2 - t1:.3f}s write') + + file = stats_path.with_suffix('.json') + t1 = time.time() + with open(file, 'w') as f: + json.dump(stats, f) # save stats *.json + t2 = time.time() + with open(file) as f: + x = json.load(f) # load hyps dict + print(f'stats.json times: {time.time() - t2:.3f}s read, {t2 - t1:.3f}s write') + + # Save, print and return + if hub: + print(f'Saving {stats_path.resolve()}...') + with open(stats_path, 'w') as f: + json.dump(stats, f) # save stats.json + if verbose: + print(json.dumps(stats, indent=2, sort_keys=False)) + return stats diff --git a/yolov5/utils/docker/.dockerignore b/yolov5/utils/docker/.dockerignore new file mode 100644 index 0000000..af51ccc --- /dev/null +++ b/yolov5/utils/docker/.dockerignore @@ -0,0 +1,222 @@ +# Repo-specific DockerIgnore ------------------------------------------------------------------------------------------- +#.git +.cache +.idea +runs +output +coco +storage.googleapis.com + +data/samples/* +**/results*.csv +*.jpg + +# Neural Network weights ----------------------------------------------------------------------------------------------- +**/*.pt +**/*.pth +**/*.onnx +**/*.engine +**/*.mlmodel +**/*.torchscript +**/*.torchscript.pt +**/*.tflite +**/*.h5 +**/*.pb +*_saved_model/ +*_web_model/ +*_openvino_model/ + +# Below Copied From .gitignore ----------------------------------------------------------------------------------------- +# Below Copied From .gitignore ----------------------------------------------------------------------------------------- + + +# GitHub Python GitIgnore ---------------------------------------------------------------------------------------------- +# Byte-compiled / optimized / DLL files +__pycache__/ +*.py[cod] +*$py.class + +# C extensions +*.so + +# Distribution / packaging +.Python +env/ +build/ +develop-eggs/ +dist/ +downloads/ +eggs/ +.eggs/ +lib/ +lib64/ +parts/ +sdist/ +var/ +wheels/ +*.egg-info/ +wandb/ +.installed.cfg +*.egg + +# PyInstaller +# Usually these files are written by a python script from a template +# before PyInstaller builds the exe, so as to inject date/other infos into it. +*.manifest +*.spec + +# Installer logs +pip-log.txt +pip-delete-this-directory.txt + +# Unit test / coverage reports +htmlcov/ +.tox/ +.coverage +.coverage.* +.cache +nosetests.xml +coverage.xml +*.cover +.hypothesis/ + +# Translations +*.mo +*.pot + +# Django stuff: +*.log +local_settings.py + +# Flask stuff: +instance/ +.webassets-cache + +# Scrapy stuff: +.scrapy + +# Sphinx documentation +docs/_build/ + +# PyBuilder +target/ + +# Jupyter Notebook +.ipynb_checkpoints + +# pyenv +.python-version + +# celery beat schedule file +celerybeat-schedule + +# SageMath parsed files +*.sage.py + +# dotenv +.env + +# virtualenv +.venv* +venv*/ +ENV*/ + +# Spyder project settings +.spyderproject +.spyproject + +# Rope project settings +.ropeproject + +# mkdocs documentation +/site + +# mypy +.mypy_cache/ + + +# https://github.com/github/gitignore/blob/master/Global/macOS.gitignore ----------------------------------------------- + +# General +.DS_Store +.AppleDouble +.LSOverride + +# Icon must end with two \r +Icon +Icon? + +# Thumbnails +._* + +# Files that might appear in the root of a volume +.DocumentRevisions-V100 +.fseventsd +.Spotlight-V100 +.TemporaryItems +.Trashes +.VolumeIcon.icns +.com.apple.timemachine.donotpresent + +# Directories potentially created on remote AFP share +.AppleDB +.AppleDesktop +Network Trash Folder +Temporary Items +.apdisk + + +# https://github.com/github/gitignore/blob/master/Global/JetBrains.gitignore +# Covers JetBrains IDEs: IntelliJ, RubyMine, PhpStorm, AppCode, PyCharm, CLion, Android Studio and WebStorm +# Reference: https://intellij-support.jetbrains.com/hc/en-us/articles/206544839 + +# User-specific stuff: +.idea/* +.idea/**/workspace.xml +.idea/**/tasks.xml +.idea/dictionaries +.html # Bokeh Plots +.pg # TensorFlow Frozen Graphs +.avi # videos + +# Sensitive or high-churn files: +.idea/**/dataSources/ +.idea/**/dataSources.ids +.idea/**/dataSources.local.xml +.idea/**/sqlDataSources.xml +.idea/**/dynamic.xml +.idea/**/uiDesigner.xml + +# Gradle: +.idea/**/gradle.xml +.idea/**/libraries + +# CMake +cmake-build-debug/ +cmake-build-release/ + +# Mongo Explorer plugin: +.idea/**/mongoSettings.xml + +## File-based project format: +*.iws + +## Plugin-specific files: + +# IntelliJ +out/ + +# mpeltonen/sbt-idea plugin +.idea_modules/ + +# JIRA plugin +atlassian-ide-plugin.xml + +# Cursive Clojure plugin +.idea/replstate.xml + +# Crashlytics plugin (for Android Studio and IntelliJ) +com_crashlytics_export_strings.xml +crashlytics.properties +crashlytics-build.properties +fabric.properties diff --git a/yolov5/utils/docker/Dockerfile b/yolov5/utils/docker/Dockerfile new file mode 100644 index 0000000..284a4fc --- /dev/null +++ b/yolov5/utils/docker/Dockerfile @@ -0,0 +1,67 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# Builds ultralytics/yolov5:latest image on DockerHub https://hub.docker.com/r/ultralytics/yolov5 +# Image is CUDA-optimized for YOLOv5 single/multi-GPU training and inference + +# Start FROM NVIDIA PyTorch image https://ngc.nvidia.com/catalog/containers/nvidia:pytorch +FROM nvcr.io/nvidia/pytorch:22.04-py3 +RUN rm -rf /opt/pytorch # remove 1.2GB dir + +# Downloads to user config dir +ADD https://ultralytics.com/assets/Arial.ttf https://ultralytics.com/assets/Arial.Unicode.ttf /root/.config/Ultralytics/ + +# Install linux packages +RUN apt update && apt install --no-install-recommends -y zip htop screen libgl1-mesa-glx + +# Install pip packages +COPY requirements.txt . +RUN python -m pip install --upgrade pip +RUN pip uninstall -y torch torchvision torchtext Pillow +RUN pip install --no-cache -r requirements.txt albumentations wandb gsutil notebook Pillow>=9.1.0 \ + --extra-index-url https://download.pytorch.org/whl/cu113 + +# Create working directory +RUN mkdir -p /usr/src/app +WORKDIR /usr/src/app + +# Copy contents +COPY . /usr/src/app +RUN git clone https://github.com/ultralytics/yolov5 /usr/src/yolov5 + +# Set environment variables +ENV OMP_NUM_THREADS=8 + + +# Usage Examples ------------------------------------------------------------------------------------------------------- + +# Build and Push +# t=ultralytics/yolov5:latest && sudo docker build -f utils/docker/Dockerfile -t $t . && sudo docker push $t + +# Pull and Run +# t=ultralytics/yolov5:latest && sudo docker pull $t && sudo docker run -it --ipc=host --gpus all $t + +# Pull and Run with local directory access +# t=ultralytics/yolov5:latest && sudo docker pull $t && sudo docker run -it --ipc=host --gpus all -v "$(pwd)"/datasets:/usr/src/datasets $t + +# Kill all +# sudo docker kill $(sudo docker ps -q) + +# Kill all image-based +# sudo docker kill $(sudo docker ps -qa --filter ancestor=ultralytics/yolov5:latest) + +# Bash into running container +# sudo docker exec -it 5a9b5863d93d bash + +# Bash into stopped container +# id=$(sudo docker ps -qa) && sudo docker start $id && sudo docker exec -it $id bash + +# Clean up +# docker system prune -a --volumes + +# Update Ubuntu drivers +# https://www.maketecheasier.com/install-nvidia-drivers-ubuntu/ + +# DDP test +# python -m torch.distributed.run --nproc_per_node 2 --master_port 1 train.py --epochs 3 + +# GCP VM from Image +# docker.io/ultralytics/yolov5:latest diff --git a/yolov5/utils/docker/Dockerfile-arm64 b/yolov5/utils/docker/Dockerfile-arm64 new file mode 100644 index 0000000..2e26105 --- /dev/null +++ b/yolov5/utils/docker/Dockerfile-arm64 @@ -0,0 +1,42 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# Builds ultralytics/yolov5:latest-arm64 image on DockerHub https://hub.docker.com/r/ultralytics/yolov5 +# Image is aarch64-compatible for Apple M1 and other ARM architectures i.e. Jetson Nano and Raspberry Pi + +# Start FROM Ubuntu image https://hub.docker.com/_/ubuntu +FROM arm64v8/ubuntu:20.04 + +# Downloads to user config dir +ADD https://ultralytics.com/assets/Arial.ttf https://ultralytics.com/assets/Arial.Unicode.ttf /root/.config/Ultralytics/ + +# Install linux packages +RUN apt update +RUN DEBIAN_FRONTEND=noninteractive TZ=Etc/UTC apt install -y tzdata +RUN apt install --no-install-recommends -y python3-pip git zip curl htop gcc \ + libgl1-mesa-glx libglib2.0-0 libpython3.8-dev +# RUN alias python=python3 + +# Install pip packages +COPY requirements.txt . +RUN python3 -m pip install --upgrade pip +RUN pip install --no-cache -r requirements.txt gsutil notebook \ + tensorflow-aarch64 + # tensorflowjs \ + # onnx onnx-simplifier onnxruntime \ + # coremltools openvino-dev \ + +# Create working directory +RUN mkdir -p /usr/src/app +WORKDIR /usr/src/app + +# Copy contents +COPY . /usr/src/app +RUN git clone https://github.com/ultralytics/yolov5 /usr/src/yolov5 + + +# Usage Examples ------------------------------------------------------------------------------------------------------- + +# Build and Push +# t=ultralytics/yolov5:latest-M1 && sudo docker build --platform linux/arm64 -f utils/docker/Dockerfile-arm64 -t $t . && sudo docker push $t + +# Pull and Run +# t=ultralytics/yolov5:latest-M1 && sudo docker pull $t && sudo docker run -it --ipc=host -v "$(pwd)"/datasets:/usr/src/datasets $t diff --git a/yolov5/utils/docker/Dockerfile-cpu b/yolov5/utils/docker/Dockerfile-cpu new file mode 100644 index 0000000..c8aa8c6 --- /dev/null +++ b/yolov5/utils/docker/Dockerfile-cpu @@ -0,0 +1,39 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# Builds ultralytics/yolov5:latest-cpu image on DockerHub https://hub.docker.com/r/ultralytics/yolov5 +# Image is CPU-optimized for ONNX, OpenVINO and PyTorch YOLOv5 deployments + +# Start FROM Ubuntu image https://hub.docker.com/_/ubuntu +FROM ubuntu:20.04 + +# Downloads to user config dir +ADD https://ultralytics.com/assets/Arial.ttf https://ultralytics.com/assets/Arial.Unicode.ttf /root/.config/Ultralytics/ + +# Install linux packages +RUN apt update +RUN DEBIAN_FRONTEND=noninteractive TZ=Etc/UTC apt install -y tzdata +RUN apt install --no-install-recommends -y python3-pip git zip curl htop libgl1-mesa-glx libglib2.0-0 libpython3.8-dev +# RUN alias python=python3 + +# Install pip packages +COPY requirements.txt . +RUN python3 -m pip install --upgrade pip +RUN pip install --no-cache -r requirements.txt albumentations gsutil notebook \ + coremltools onnx onnx-simplifier onnxruntime openvino-dev tensorflow-cpu tensorflowjs \ + --extra-index-url https://download.pytorch.org/whl/cpu + +# Create working directory +RUN mkdir -p /usr/src/app +WORKDIR /usr/src/app + +# Copy contents +COPY . /usr/src/app +RUN git clone https://github.com/ultralytics/yolov5 /usr/src/yolov5 + + +# Usage Examples ------------------------------------------------------------------------------------------------------- + +# Build and Push +# t=ultralytics/yolov5:latest-cpu && sudo docker build -f utils/docker/Dockerfile-cpu -t $t . && sudo docker push $t + +# Pull and Run +# t=ultralytics/yolov5:latest-cpu && sudo docker pull $t && sudo docker run -it --ipc=host -v "$(pwd)"/datasets:/usr/src/datasets $t diff --git a/yolov5/utils/downloads.py b/yolov5/utils/downloads.py new file mode 100644 index 0000000..ebe5bd3 --- /dev/null +++ b/yolov5/utils/downloads.py @@ -0,0 +1,178 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Download utils +""" + +import logging +import os +import platform +import subprocess +import time +import urllib +from pathlib import Path +from zipfile import ZipFile + +import requests +import torch + + +def is_url(url): + # Check if online file exists + try: + r = urllib.request.urlopen(url) # response + return r.getcode() == 200 + except urllib.request.HTTPError: + return False + + +def gsutil_getsize(url=''): + # gs://bucket/file size https://cloud.google.com/storage/docs/gsutil/commands/du + s = subprocess.check_output(f'gsutil du {url}', shell=True).decode('utf-8') + return eval(s.split(' ')[0]) if len(s) else 0 # bytes + + +def safe_download(file, url, url2=None, min_bytes=1E0, error_msg=''): + # Attempts to download file from url or url2, checks and removes incomplete downloads < min_bytes + from utils.general import LOGGER + + file = Path(file) + assert_msg = f"Downloaded file '{file}' does not exist or size is < min_bytes={min_bytes}" + try: # url1 + LOGGER.info(f'Downloading {url} to {file}...') + torch.hub.download_url_to_file(url, str(file), progress=LOGGER.level <= logging.INFO) + assert file.exists() and file.stat().st_size > min_bytes, assert_msg # check + except Exception as e: # url2 + file.unlink(missing_ok=True) # remove partial downloads + LOGGER.info(f'ERROR: {e}\nRe-attempting {url2 or url} to {file}...') + os.system(f"curl -L '{url2 or url}' -o '{file}' --retry 3 -C -") # curl download, retry and resume on fail + finally: + if not file.exists() or file.stat().st_size < min_bytes: # check + file.unlink(missing_ok=True) # remove partial downloads + LOGGER.info(f"ERROR: {assert_msg}\n{error_msg}") + LOGGER.info('') + + +def attempt_download(file, repo='ultralytics/yolov5', release='v6.1'): + # Attempt file download from GitHub release assets if not found locally. release = 'latest', 'v6.1', etc. + from utils.general import LOGGER + + def github_assets(repository, version='latest'): + # Return GitHub repo tag (i.e. 'v6.1') and assets (i.e. ['yolov5s.pt', 'yolov5m.pt', ...]) + if version != 'latest': + version = f'tags/{version}' # i.e. tags/v6.1 + response = requests.get(f'https://api.github.com/repos/{repository}/releases/{version}').json() # github api + return response['tag_name'], [x['name'] for x in response['assets']] # tag, assets + + file = Path(str(file).strip().replace("'", '')) + if not file.exists(): + # URL specified + name = Path(urllib.parse.unquote(str(file))).name # decode '%2F' to '/' etc. + if str(file).startswith(('http:/', 'https:/')): # download + url = str(file).replace(':/', '://') # Pathlib turns :// -> :/ + file = name.split('?')[0] # parse authentication https://url.com/file.txt?auth... + if Path(file).is_file(): + LOGGER.info(f'Found {url} locally at {file}') # file already exists + else: + safe_download(file=file, url=url, min_bytes=1E5) + return file + + # GitHub assets + assets = [ + 'yolov5n.pt', 'yolov5s.pt', 'yolov5m.pt', 'yolov5l.pt', 'yolov5x.pt', 'yolov5n6.pt', 'yolov5s6.pt', + 'yolov5m6.pt', 'yolov5l6.pt', 'yolov5x6.pt'] + try: + tag, assets = github_assets(repo, release) + except Exception: + try: + tag, assets = github_assets(repo) # latest release + except Exception: + try: + tag = subprocess.check_output('git tag', shell=True, stderr=subprocess.STDOUT).decode().split()[-1] + except Exception: + tag = release + + file.parent.mkdir(parents=True, exist_ok=True) # make parent dir (if required) + if name in assets: + url3 = 'https://drive.google.com/drive/folders/1EFQTEUeXWSFww0luse2jB9M1QNZQGwNl' # backup gdrive mirror + safe_download( + file, + url=f'https://github.com/{repo}/releases/download/{tag}/{name}', + url2=f'https://storage.googleapis.com/{repo}/{tag}/{name}', # backup url (optional) + min_bytes=1E5, + error_msg=f'{file} missing, try downloading from https://github.com/{repo}/releases/{tag} or {url3}') + + return str(file) + + +def gdrive_download(id='16TiPfZj7htmTyhntwcZyEEAejOUxuT6m', file='tmp.zip'): + # Downloads a file from Google Drive. from yolov5.utils.downloads import *; gdrive_download() + t = time.time() + file = Path(file) + cookie = Path('cookie') # gdrive cookie + print(f'Downloading https://drive.google.com/uc?export=download&id={id} as {file}... ', end='') + file.unlink(missing_ok=True) # remove existing file + cookie.unlink(missing_ok=True) # remove existing cookie + + # Attempt file download + out = "NUL" if platform.system() == "Windows" else "/dev/null" + os.system(f'curl -c ./cookie -s -L "drive.google.com/uc?export=download&id={id}" > {out}') + if os.path.exists('cookie'): # large file + s = f'curl -Lb ./cookie "drive.google.com/uc?export=download&confirm={get_token()}&id={id}" -o {file}' + else: # small file + s = f'curl -s -L -o {file} "drive.google.com/uc?export=download&id={id}"' + r = os.system(s) # execute, capture return + cookie.unlink(missing_ok=True) # remove existing cookie + + # Error check + if r != 0: + file.unlink(missing_ok=True) # remove partial + print('Download error ') # raise Exception('Download error') + return r + + # Unzip if archive + if file.suffix == '.zip': + print('unzipping... ', end='') + ZipFile(file).extractall(path=file.parent) # unzip + file.unlink() # remove zip + + print(f'Done ({time.time() - t:.1f}s)') + return r + + +def get_token(cookie="./cookie"): + with open(cookie) as f: + for line in f: + if "download" in line: + return line.split()[-1] + return "" + + +# Google utils: https://cloud.google.com/storage/docs/reference/libraries ---------------------------------------------- +# +# +# def upload_blob(bucket_name, source_file_name, destination_blob_name): +# # Uploads a file to a bucket +# # https://cloud.google.com/storage/docs/uploading-objects#storage-upload-object-python +# +# storage_client = storage.Client() +# bucket = storage_client.get_bucket(bucket_name) +# blob = bucket.blob(destination_blob_name) +# +# blob.upload_from_filename(source_file_name) +# +# print('File {} uploaded to {}.'.format( +# source_file_name, +# destination_blob_name)) +# +# +# def download_blob(bucket_name, source_blob_name, destination_file_name): +# # Uploads a blob from a bucket +# storage_client = storage.Client() +# bucket = storage_client.get_bucket(bucket_name) +# blob = bucket.blob(source_blob_name) +# +# blob.download_to_filename(destination_file_name) +# +# print('Blob {} downloaded to {}.'.format( +# source_blob_name, +# destination_file_name)) diff --git a/yolov5/utils/flask_rest_api/README.md b/yolov5/utils/flask_rest_api/README.md new file mode 100644 index 0000000..a726acb --- /dev/null +++ b/yolov5/utils/flask_rest_api/README.md @@ -0,0 +1,73 @@ +# Flask REST API + +[REST](https://en.wikipedia.org/wiki/Representational_state_transfer) [API](https://en.wikipedia.org/wiki/API)s are +commonly used to expose Machine Learning (ML) models to other services. This folder contains an example REST API +created using Flask to expose the YOLOv5s model from [PyTorch Hub](https://pytorch.org/hub/ultralytics_yolov5/). + +## Requirements + +[Flask](https://palletsprojects.com/p/flask/) is required. Install with: + +```shell +$ pip install Flask +``` + +## Run + +After Flask installation run: + +```shell +$ python3 restapi.py --port 5000 +``` + +Then use [curl](https://curl.se/) to perform a request: + +```shell +$ curl -X POST -F image=@zidane.jpg 'http://localhost:5000/v1/object-detection/yolov5s' +``` + +The model inference results are returned as a JSON response: + +```json +[ + { + "class": 0, + "confidence": 0.8900438547, + "height": 0.9318675399, + "name": "person", + "width": 0.3264600933, + "xcenter": 0.7438579798, + "ycenter": 0.5207948685 + }, + { + "class": 0, + "confidence": 0.8440024257, + "height": 0.7155083418, + "name": "person", + "width": 0.6546785235, + "xcenter": 0.427829951, + "ycenter": 0.6334488392 + }, + { + "class": 27, + "confidence": 0.3771208823, + "height": 0.3902671337, + "name": "tie", + "width": 0.0696444362, + "xcenter": 0.3675483763, + "ycenter": 0.7991207838 + }, + { + "class": 27, + "confidence": 0.3527112305, + "height": 0.1540903747, + "name": "tie", + "width": 0.0336618312, + "xcenter": 0.7814827561, + "ycenter": 0.5065554976 + } +] +``` + +An example python script to perform inference using [requests](https://docs.python-requests.org/en/master/) is given +in `example_request.py` diff --git a/yolov5/utils/flask_rest_api/example_request.py b/yolov5/utils/flask_rest_api/example_request.py new file mode 100644 index 0000000..773ad89 --- /dev/null +++ b/yolov5/utils/flask_rest_api/example_request.py @@ -0,0 +1,19 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Perform test request +""" + +import pprint + +import requests + +DETECTION_URL = "http://localhost:5000/v1/object-detection/yolov5s" +IMAGE = "zidane.jpg" + +# Read image +with open(IMAGE, "rb") as f: + image_data = f.read() + +response = requests.post(DETECTION_URL, files={"image": image_data}).json() + +pprint.pprint(response) diff --git a/yolov5/utils/flask_rest_api/restapi.py b/yolov5/utils/flask_rest_api/restapi.py new file mode 100644 index 0000000..08036dd --- /dev/null +++ b/yolov5/utils/flask_rest_api/restapi.py @@ -0,0 +1,46 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Run a Flask REST API exposing a YOLOv5s model +""" + +import argparse +import io + +import torch +from flask import Flask, request +from PIL import Image + +app = Flask(__name__) + +DETECTION_URL = "/v1/object-detection/yolov5s" + + +@app.route(DETECTION_URL, methods=["POST"]) +def predict(): + if request.method != "POST": + return + + if request.files.get("image"): + # Method 1 + # with request.files["image"] as f: + # im = Image.open(io.BytesIO(f.read())) + + # Method 2 + im_file = request.files["image"] + im_bytes = im_file.read() + im = Image.open(io.BytesIO(im_bytes)) + + results = model(im, size=640) # reduce size=320 for faster inference + return results.pandas().xyxy[0].to_json(orient="records") + + +if __name__ == "__main__": + parser = argparse.ArgumentParser(description="Flask API exposing YOLOv5 model") + parser.add_argument("--port", default=5000, type=int, help="port number") + opt = parser.parse_args() + + # Fix known issue urllib.error.HTTPError 403: rate limit exceeded https://github.com/ultralytics/yolov5/pull/7210 + torch.hub._validate_not_a_forked_repo = lambda a, b, c: True + + model = torch.hub.load("ultralytics/yolov5", "yolov5s", force_reload=True) # force_reload to recache + app.run(host="0.0.0.0", port=opt.port) # debug=True causes Restarting with stat diff --git a/yolov5/utils/general.py b/yolov5/utils/general.py new file mode 100644 index 0000000..7905e72 --- /dev/null +++ b/yolov5/utils/general.py @@ -0,0 +1,1017 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +General utils +""" + +import contextlib +import glob +import inspect +import logging +import math +import os +import platform +import random +import re +import shutil +import signal +import threading +import time +import urllib +from datetime import datetime +from itertools import repeat +from multiprocessing.pool import ThreadPool +from pathlib import Path +from subprocess import check_output +from typing import Optional +from zipfile import ZipFile + +import cv2 +import numpy as np +import pandas as pd +import pkg_resources as pkg +import torch +import torchvision +import yaml + +from utils.downloads import gsutil_getsize +from utils.metrics import box_iou, fitness + +FILE = Path(__file__).resolve() +ROOT = FILE.parents[1] # YOLOv5 root directory +RANK = int(os.getenv('RANK', -1)) + +# Settings +DATASETS_DIR = ROOT.parent / 'datasets' # YOLOv5 datasets directory +NUM_THREADS = min(8, max(1, os.cpu_count() - 1)) # number of YOLOv5 multiprocessing threads +AUTOINSTALL = str(os.getenv('YOLOv5_AUTOINSTALL', True)).lower() == 'true' # global auto-install mode +VERBOSE = str(os.getenv('YOLOv5_VERBOSE', True)).lower() == 'true' # global verbose mode +FONT = 'Arial.ttf' # https://ultralytics.com/assets/Arial.ttf + +torch.set_printoptions(linewidth=320, precision=5, profile='long') +np.set_printoptions(linewidth=320, formatter={'float_kind': '{:11.5g}'.format}) # format short g, %precision=5 +pd.options.display.max_columns = 10 +cv2.setNumThreads(0) # prevent OpenCV from multithreading (incompatible with PyTorch DataLoader) +os.environ['NUMEXPR_MAX_THREADS'] = str(NUM_THREADS) # NumExpr max threads +os.environ['OMP_NUM_THREADS'] = str(NUM_THREADS) # OpenMP max threads (PyTorch and SciPy) + + +def is_kaggle(): + # Is environment a Kaggle Notebook? + try: + assert os.environ.get('PWD') == '/kaggle/working' + assert os.environ.get('KAGGLE_URL_BASE') == 'https://www.kaggle.com' + return True + except AssertionError: + return False + + +def is_writeable(dir, test=False): + # Return True if directory has write permissions, test opening a file with write permissions if test=True + if not test: + return os.access(dir, os.R_OK) # possible issues on Windows + file = Path(dir) / 'tmp.txt' + try: + with open(file, 'w'): # open file with write permissions + pass + file.unlink() # remove file + return True + except OSError: + return False + + +def set_logging(name=None, verbose=VERBOSE): + # Sets level and returns logger + if is_kaggle(): + for h in logging.root.handlers: + logging.root.removeHandler(h) # remove all handlers associated with the root logger object + rank = int(os.getenv('RANK', -1)) # rank in world for Multi-GPU trainings + level = logging.INFO if verbose and rank in {-1, 0} else logging.WARNING + log = logging.getLogger(name) + log.setLevel(level) + handler = logging.StreamHandler() + handler.setFormatter(logging.Formatter("%(message)s")) + handler.setLevel(level) + log.addHandler(handler) + + +set_logging() # run before defining LOGGER +LOGGER = logging.getLogger("yolov5") # define globally (used in train.py, val.py, detect.py, etc.) + + +def user_config_dir(dir='Ultralytics', env_var='YOLOV5_CONFIG_DIR'): + # Return path of user configuration directory. Prefer environment variable if exists. Make dir if required. + env = os.getenv(env_var) + if env: + path = Path(env) # use environment variable + else: + cfg = {'Windows': 'AppData/Roaming', 'Linux': '.config', 'Darwin': 'Library/Application Support'} # 3 OS dirs + path = Path.home() / cfg.get(platform.system(), '') # OS-specific config dir + path = (path if is_writeable(path) else Path('/tmp')) / dir # GCP and AWS lambda fix, only /tmp is writeable + path.mkdir(exist_ok=True) # make if required + return path + + +CONFIG_DIR = user_config_dir() # Ultralytics settings dir + + +class Profile(contextlib.ContextDecorator): + # Usage: @Profile() decorator or 'with Profile():' context manager + def __enter__(self): + self.start = time.time() + + def __exit__(self, type, value, traceback): + print(f'Profile results: {time.time() - self.start:.5f}s') + + +class Timeout(contextlib.ContextDecorator): + # Usage: @Timeout(seconds) decorator or 'with Timeout(seconds):' context manager + def __init__(self, seconds, *, timeout_msg='', suppress_timeout_errors=True): + self.seconds = int(seconds) + self.timeout_message = timeout_msg + self.suppress = bool(suppress_timeout_errors) + + def _timeout_handler(self, signum, frame): + raise TimeoutError(self.timeout_message) + + def __enter__(self): + if platform.system() != 'Windows': # not supported on Windows + signal.signal(signal.SIGALRM, self._timeout_handler) # Set handler for SIGALRM + signal.alarm(self.seconds) # start countdown for SIGALRM to be raised + + def __exit__(self, exc_type, exc_val, exc_tb): + if platform.system() != 'Windows': + signal.alarm(0) # Cancel SIGALRM if it's scheduled + if self.suppress and exc_type is TimeoutError: # Suppress TimeoutError + return True + + +class WorkingDirectory(contextlib.ContextDecorator): + # Usage: @WorkingDirectory(dir) decorator or 'with WorkingDirectory(dir):' context manager + def __init__(self, new_dir): + self.dir = new_dir # new dir + self.cwd = Path.cwd().resolve() # current dir + + def __enter__(self): + os.chdir(self.dir) + + def __exit__(self, exc_type, exc_val, exc_tb): + os.chdir(self.cwd) + + +def try_except(func): + # try-except function. Usage: @try_except decorator + def handler(*args, **kwargs): + try: + func(*args, **kwargs) + except Exception as e: + print(e) + + return handler + + +def threaded(func): + # Multi-threads a target function and returns thread. Usage: @threaded decorator + def wrapper(*args, **kwargs): + thread = threading.Thread(target=func, args=args, kwargs=kwargs, daemon=True) + thread.start() + return thread + + return wrapper + + +def methods(instance): + # Get class/instance methods + return [f for f in dir(instance) if callable(getattr(instance, f)) and not f.startswith("__")] + + +def print_args(args: Optional[dict] = None, show_file=True, show_fcn=False): + # Print function arguments (optional args dict) + x = inspect.currentframe().f_back # previous frame + file, _, fcn, _, _ = inspect.getframeinfo(x) + if args is None: # get args automatically + args, _, _, frm = inspect.getargvalues(x) + args = {k: v for k, v in frm.items() if k in args} + s = (f'{Path(file).stem}: ' if show_file else '') + (f'{fcn}: ' if show_fcn else '') + LOGGER.info(colorstr(s) + ', '.join(f'{k}={v}' for k, v in args.items())) + + +def init_seeds(seed=0): + # Initialize random number generator (RNG) seeds https://pytorch.org/docs/stable/notes/randomness.html + # cudnn seed 0 settings are slower and more reproducible, else faster and less reproducible + import torch.backends.cudnn as cudnn + random.seed(seed) + np.random.seed(seed) + torch.manual_seed(seed) + cudnn.benchmark, cudnn.deterministic = (False, True) if seed == 0 else (True, False) + + +def intersect_dicts(da, db, exclude=()): + # Dictionary intersection of matching keys and shapes, omitting 'exclude' keys, using da values + return {k: v for k, v in da.items() if k in db and not any(x in k for x in exclude) and v.shape == db[k].shape} + + +def get_latest_run(search_dir='.'): + # Return path to most recent 'last.pt' in /runs (i.e. to --resume from) + last_list = glob.glob(f'{search_dir}/**/last*.pt', recursive=True) + return max(last_list, key=os.path.getctime) if last_list else '' + + +def is_docker(): + # Is environment a Docker container? + return Path('/workspace').exists() # or Path('/.dockerenv').exists() + + +def is_colab(): + # Is environment a Google Colab instance? + try: + import google.colab + return True + except ImportError: + return False + + +def is_pip(): + # Is file in a pip package? + return 'site-packages' in Path(__file__).resolve().parts + + +def is_ascii(s=''): + # Is string composed of all ASCII (no UTF) characters? (note str().isascii() introduced in python 3.7) + s = str(s) # convert list, tuple, None, etc. to str + return len(s.encode().decode('ascii', 'ignore')) == len(s) + + +def is_chinese(s='人工智能'): + # Is string composed of any Chinese characters? + return bool(re.search('[\u4e00-\u9fff]', str(s))) + + +def emojis(str=''): + # Return platform-dependent emoji-safe version of string + return str.encode().decode('ascii', 'ignore') if platform.system() == 'Windows' else str + + +def file_age(path=__file__): + # Return days since last file update + dt = (datetime.now() - datetime.fromtimestamp(Path(path).stat().st_mtime)) # delta + return dt.days # + dt.seconds / 86400 # fractional days + + +def file_date(path=__file__): + # Return human-readable file modification date, i.e. '2021-3-26' + t = datetime.fromtimestamp(Path(path).stat().st_mtime) + return f'{t.year}-{t.month}-{t.day}' + + +def file_size(path): + # Return file/dir size (MB) + mb = 1 << 20 # bytes to MiB (1024 ** 2) + path = Path(path) + if path.is_file(): + return path.stat().st_size / mb + elif path.is_dir(): + return sum(f.stat().st_size for f in path.glob('**/*') if f.is_file()) / mb + else: + return 0.0 + + +def check_online(): + # Check internet connectivity + import socket + try: + socket.create_connection(("1.1.1.1", 443), 5) # check host accessibility + return True + except OSError: + return False + + +def git_describe(path=ROOT): # path must be a directory + # Return human-readable git description, i.e. v5.0-5-g3e25f1e https://git-scm.com/docs/git-describe + try: + assert (Path(path) / '.git').is_dir() + return check_output(f'git -C {path} describe --tags --long --always', shell=True).decode()[:-1] + except Exception: + return '' + + +@try_except +@WorkingDirectory(ROOT) +def check_git_status(): + # Recommend 'git pull' if code is out of date + msg = ', for updates see https://github.com/ultralytics/yolov5' + s = colorstr('github: ') # string + assert Path('.git').exists(), s + 'skipping check (not a git repository)' + msg + assert not is_docker(), s + 'skipping check (Docker image)' + msg + assert check_online(), s + 'skipping check (offline)' + msg + + cmd = 'git fetch && git config --get remote.origin.url' + url = check_output(cmd, shell=True, timeout=5).decode().strip().rstrip('.git') # git fetch + branch = check_output('git rev-parse --abbrev-ref HEAD', shell=True).decode().strip() # checked out + n = int(check_output(f'git rev-list {branch}..origin/master --count', shell=True)) # commits behind + if n > 0: + s += f"⚠️ YOLOv5 is out of date by {n} commit{'s' * (n > 1)}. Use `git pull` or `git clone {url}` to update." + else: + s += f'up to date with {url} ✅' + LOGGER.info(emojis(s)) # emoji-safe + + +def check_python(minimum='3.7.0'): + # Check current python version vs. required python version + check_version(platform.python_version(), minimum, name='Python ', hard=True) + + +def check_version(current='0.0.0', minimum='0.0.0', name='version ', pinned=False, hard=False, verbose=False): + # Check version vs. required version + current, minimum = (pkg.parse_version(x) for x in (current, minimum)) + result = (current == minimum) if pinned else (current >= minimum) # bool + s = f'{name}{minimum} required by YOLOv5, but {name}{current} is currently installed' # string + if hard: + assert result, s # assert min requirements met + if verbose and not result: + LOGGER.warning(s) + return result + + +@try_except +def check_requirements(requirements=ROOT / 'requirements.txt', exclude=(), install=True, cmds=()): + # Check installed dependencies meet requirements (pass *.txt file or list of packages) + prefix = colorstr('red', 'bold', 'requirements:') + check_python() # check python version + if isinstance(requirements, (str, Path)): # requirements.txt file + file = Path(requirements) + assert file.exists(), f"{prefix} {file.resolve()} not found, check failed." + with file.open() as f: + requirements = [f'{x.name}{x.specifier}' for x in pkg.parse_requirements(f) if x.name not in exclude] + else: # list or tuple of packages + requirements = [x for x in requirements if x not in exclude] + + n = 0 # number of packages updates + for i, r in enumerate(requirements): + try: + pkg.require(r) + except Exception: # DistributionNotFound or VersionConflict if requirements not met + s = f"{prefix} {r} not found and is required by YOLOv5" + if install and AUTOINSTALL: # check environment variable + LOGGER.info(f"{s}, attempting auto-update...") + try: + assert check_online(), f"'pip install {r}' skipped (offline)" + LOGGER.info(check_output(f'pip install "{r}" {cmds[i] if cmds else ""}', shell=True).decode()) + n += 1 + except Exception as e: + LOGGER.warning(f'{prefix} {e}') + else: + LOGGER.info(f'{s}. Please install and rerun your command.') + + if n: # if packages updated + source = file.resolve() if 'file' in locals() else requirements + s = f"{prefix} {n} package{'s' * (n > 1)} updated per {source}\n" \ + f"{prefix} ⚠️ {colorstr('bold', 'Restart runtime or rerun command for updates to take effect')}\n" + LOGGER.info(emojis(s)) + + +def check_img_size(imgsz, s=32, floor=0): + # Verify image size is a multiple of stride s in each dimension + if isinstance(imgsz, int): # integer i.e. img_size=640 + new_size = max(make_divisible(imgsz, int(s)), floor) + else: # list i.e. img_size=[640, 480] + imgsz = list(imgsz) # convert to list if tuple + new_size = [max(make_divisible(x, int(s)), floor) for x in imgsz] + if new_size != imgsz: + LOGGER.warning(f'WARNING: --img-size {imgsz} must be multiple of max stride {s}, updating to {new_size}') + return new_size + + +def check_imshow(): + # Check if environment supports image displays + try: + assert not is_docker(), 'cv2.imshow() is disabled in Docker environments' + assert not is_colab(), 'cv2.imshow() is disabled in Google Colab environments' + cv2.imshow('test', np.zeros((1, 1, 3))) + cv2.waitKey(1) + cv2.destroyAllWindows() + cv2.waitKey(1) + return True + except Exception as e: + LOGGER.warning(f'WARNING: Environment does not support cv2.imshow() or PIL Image.show() image displays\n{e}') + return False + + +def check_suffix(file='yolov5s.pt', suffix=('.pt',), msg=''): + # Check file(s) for acceptable suffix + if file and suffix: + if isinstance(suffix, str): + suffix = [suffix] + for f in file if isinstance(file, (list, tuple)) else [file]: + s = Path(f).suffix.lower() # file suffix + if len(s): + assert s in suffix, f"{msg}{f} acceptable suffix is {suffix}" + + +def check_yaml(file, suffix=('.yaml', '.yml')): + # Search/download YAML file (if necessary) and return path, checking suffix + return check_file(file, suffix) + + +def check_file(file, suffix=''): + # Search/download file (if necessary) and return path + check_suffix(file, suffix) # optional + file = str(file) # convert to str() + if Path(file).is_file() or not file: # exists + return file + elif file.startswith(('http:/', 'https:/')): # download + url = file # warning: Pathlib turns :// -> :/ + file = Path(urllib.parse.unquote(file).split('?')[0]).name # '%2F' to '/', split https://url.com/file.txt?auth + if Path(file).is_file(): + LOGGER.info(f'Found {url} locally at {file}') # file already exists + else: + LOGGER.info(f'Downloading {url} to {file}...') + torch.hub.download_url_to_file(url, file) + assert Path(file).exists() and Path(file).stat().st_size > 0, f'File download failed: {url}' # check + return file + else: # search + files = [] + for d in 'data', 'models', 'utils': # search directories + files.extend(glob.glob(str(ROOT / d / '**' / file), recursive=True)) # find file + assert len(files), f'File not found: {file}' # assert file was found + assert len(files) == 1, f"Multiple files match '{file}', specify exact path: {files}" # assert unique + return files[0] # return file + + +def check_font(font=FONT, progress=False): + # Download font to CONFIG_DIR if necessary + font = Path(font) + file = CONFIG_DIR / font.name + if not font.exists() and not file.exists(): + url = "https://ultralytics.com/assets/" + font.name + LOGGER.info(f'Downloading {url} to {file}...') + torch.hub.download_url_to_file(url, str(file), progress=progress) + + +def check_dataset(data, autodownload=True): + # Download, check and/or unzip dataset if not found locally + + # Download (optional) + extract_dir = '' + if isinstance(data, (str, Path)) and str(data).endswith('.zip'): # i.e. gs://bucket/dir/coco128.zip + download(data, dir=DATASETS_DIR, unzip=True, delete=False, curl=False, threads=1) + data = next((DATASETS_DIR / Path(data).stem).rglob('*.yaml')) + extract_dir, autodownload = data.parent, False + + # Read yaml (optional) + if isinstance(data, (str, Path)): + with open(data, errors='ignore') as f: + data = yaml.safe_load(f) # dictionary + + # Checks + for k in 'train', 'val', 'nc': + assert k in data, emojis(f"data.yaml '{k}:' field missing ❌") + if 'names' not in data: + LOGGER.warning(emojis("data.yaml 'names:' field missing ⚠, assigning default names 'class0', 'class1', etc.")) + data['names'] = [f'class{i}' for i in range(data['nc'])] # default names + + # Resolve paths + path = Path(extract_dir or data.get('path') or '') # optional 'path' default to '.' + if not path.is_absolute(): + path = (ROOT / path).resolve() + for k in 'train', 'val', 'test': + if data.get(k): # prepend path + data[k] = str(path / data[k]) if isinstance(data[k], str) else [str(path / x) for x in data[k]] + + # Parse yaml + train, val, test, s = (data.get(x) for x in ('train', 'val', 'test', 'download')) + if val: + val = [Path(x).resolve() for x in (val if isinstance(val, list) else [val])] # val path + if not all(x.exists() for x in val): + LOGGER.info(emojis('\nDataset not found ⚠, missing paths %s' % [str(x) for x in val if not x.exists()])) + if not s or not autodownload: + raise Exception(emojis('Dataset not found ❌')) + t = time.time() + root = path.parent if 'path' in data else '..' # unzip directory i.e. '../' + if s.startswith('http') and s.endswith('.zip'): # URL + f = Path(s).name # filename + LOGGER.info(f'Downloading {s} to {f}...') + torch.hub.download_url_to_file(s, f) + Path(root).mkdir(parents=True, exist_ok=True) # create root + ZipFile(f).extractall(path=root) # unzip + Path(f).unlink() # remove zip + r = None # success + elif s.startswith('bash '): # bash script + LOGGER.info(f'Running {s} ...') + r = os.system(s) + else: # python script + r = exec(s, {'yaml': data}) # return None + dt = f'({round(time.time() - t, 1)}s)' + s = f"success ✅ {dt}, saved to {colorstr('bold', root)}" if r in (0, None) else f"failure {dt} ❌" + LOGGER.info(emojis(f"Dataset download {s}")) + check_font('Arial.ttf' if is_ascii(data['names']) else 'Arial.Unicode.ttf', progress=True) # download fonts + return data # dictionary + + +def check_amp(model): + # Check PyTorch Automatic Mixed Precision (AMP) functionality. Return True on correct operation + from models.common import AutoShape, DetectMultiBackend + + def amp_allclose(model, im): + # All close FP32 vs AMP results + m = AutoShape(model, verbose=False) # model + a = m(im).xywhn[0] # FP32 inference + m.amp = True + b = m(im).xywhn[0] # AMP inference + return a.shape == b.shape and torch.allclose(a, b, atol=0.1) # close to 10% absolute tolerance + + prefix = colorstr('AMP: ') + device = next(model.parameters()).device # get model device + if device.type == 'cpu': + return False # AMP disabled on CPU + f = ROOT / 'data' / 'images' / 'bus.jpg' # image to check + im = f if f.exists() else 'https://ultralytics.com/images/bus.jpg' if check_online() else np.ones((640, 640, 3)) + try: + assert amp_allclose(model, im) or amp_allclose(DetectMultiBackend('yolov5n.pt', device), im) + LOGGER.info(emojis(f'{prefix}checks passed ✅')) + return True + except Exception: + help_url = 'https://github.com/ultralytics/yolov5/issues/7908' + LOGGER.warning(emojis(f'{prefix}checks failed ❌, disabling Automatic Mixed Precision. See {help_url}')) + return False + + +def url2file(url): + # Convert URL to filename, i.e. https://url.com/file.txt?auth -> file.txt + url = str(Path(url)).replace(':/', '://') # Pathlib turns :// -> :/ + return Path(urllib.parse.unquote(url)).name.split('?')[0] # '%2F' to '/', split https://url.com/file.txt?auth + + +def download(url, dir='.', unzip=True, delete=True, curl=False, threads=1, retry=3): + # Multi-threaded file download and unzip function, used in data.yaml for autodownload + def download_one(url, dir): + # Download 1 file + success = True + f = dir / Path(url).name # filename + if Path(url).is_file(): # exists in current path + Path(url).rename(f) # move to dir + elif not f.exists(): + LOGGER.info(f'Downloading {url} to {f}...') + for i in range(retry + 1): + if curl: + s = 'sS' if threads > 1 else '' # silent + r = os.system(f'curl -{s}L "{url}" -o "{f}" --retry 9 -C -') # curl download with retry, continue + success = r == 0 + else: + torch.hub.download_url_to_file(url, f, progress=threads == 1) # torch download + success = f.is_file() + if success: + break + elif i < retry: + LOGGER.warning(f'Download failure, retrying {i + 1}/{retry} {url}...') + else: + LOGGER.warning(f'Failed to download {url}...') + + if unzip and success and f.suffix in ('.zip', '.gz'): + LOGGER.info(f'Unzipping {f}...') + if f.suffix == '.zip': + ZipFile(f).extractall(path=dir) # unzip + elif f.suffix == '.gz': + os.system(f'tar xfz {f} --directory {f.parent}') # unzip + if delete: + f.unlink() # remove zip + + dir = Path(dir) + dir.mkdir(parents=True, exist_ok=True) # make directory + if threads > 1: + pool = ThreadPool(threads) + pool.imap(lambda x: download_one(*x), zip(url, repeat(dir))) # multi-threaded + pool.close() + pool.join() + else: + for u in [url] if isinstance(url, (str, Path)) else url: + download_one(u, dir) + + +def make_divisible(x, divisor): + # Returns nearest x divisible by divisor + if isinstance(divisor, torch.Tensor): + divisor = int(divisor.max()) # to int + return math.ceil(x / divisor) * divisor + + +def clean_str(s): + # Cleans a string by replacing special characters with underscore _ + return re.sub(pattern="[|@#!¡·$€%&()=?¿^*;:,¨´><+]", repl="_", string=s) + + +def one_cycle(y1=0.0, y2=1.0, steps=100): + # lambda function for sinusoidal ramp from y1 to y2 https://arxiv.org/pdf/1812.01187.pdf + return lambda x: ((1 - math.cos(x * math.pi / steps)) / 2) * (y2 - y1) + y1 + + +def colorstr(*input): + # Colors a string https://en.wikipedia.org/wiki/ANSI_escape_code, i.e. colorstr('blue', 'hello world') + *args, string = input if len(input) > 1 else ('blue', 'bold', input[0]) # color arguments, string + colors = { + 'black': '\033[30m', # basic colors + 'red': '\033[31m', + 'green': '\033[32m', + 'yellow': '\033[33m', + 'blue': '\033[34m', + 'magenta': '\033[35m', + 'cyan': '\033[36m', + 'white': '\033[37m', + 'bright_black': '\033[90m', # bright colors + 'bright_red': '\033[91m', + 'bright_green': '\033[92m', + 'bright_yellow': '\033[93m', + 'bright_blue': '\033[94m', + 'bright_magenta': '\033[95m', + 'bright_cyan': '\033[96m', + 'bright_white': '\033[97m', + 'end': '\033[0m', # misc + 'bold': '\033[1m', + 'underline': '\033[4m'} + return ''.join(colors[x] for x in args) + f'{string}' + colors['end'] + + +def labels_to_class_weights(labels, nc=80): + # Get class weights (inverse frequency) from training labels + if labels[0] is None: # no labels loaded + return torch.Tensor() + + labels = np.concatenate(labels, 0) # labels.shape = (866643, 5) for COCO + classes = labels[:, 0].astype(np.int) # labels = [class xywh] + weights = np.bincount(classes, minlength=nc) # occurrences per class + + # Prepend gridpoint count (for uCE training) + # gpi = ((320 / 32 * np.array([1, 2, 4])) ** 2 * 3).sum() # gridpoints per image + # weights = np.hstack([gpi * len(labels) - weights.sum() * 9, weights * 9]) ** 0.5 # prepend gridpoints to start + + weights[weights == 0] = 1 # replace empty bins with 1 + weights = 1 / weights # number of targets per class + weights /= weights.sum() # normalize + return torch.from_numpy(weights) + + +def labels_to_image_weights(labels, nc=80, class_weights=np.ones(80)): + # Produces image weights based on class_weights and image contents + # Usage: index = random.choices(range(n), weights=image_weights, k=1) # weighted image sample + class_counts = np.array([np.bincount(x[:, 0].astype(np.int), minlength=nc) for x in labels]) + return (class_weights.reshape(1, nc) * class_counts).sum(1) + + +def coco80_to_coco91_class(): # converts 80-index (val2014) to 91-index (paper) + # https://tech.amikelive.com/node-718/what-object-categories-labels-are-in-coco-dataset/ + # a = np.loadtxt('data/coco.names', dtype='str', delimiter='\n') + # b = np.loadtxt('data/coco_paper.names', dtype='str', delimiter='\n') + # x1 = [list(a[i] == b).index(True) + 1 for i in range(80)] # darknet to coco + # x2 = [list(b[i] == a).index(True) if any(b[i] == a) else None for i in range(91)] # coco to darknet + return [ + 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 27, 28, 31, 32, 33, 34, + 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, + 64, 65, 67, 70, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 84, 85, 86, 87, 88, 89, 90] + + +def xyxy2xywh(x): + # Convert nx4 boxes from [x1, y1, x2, y2] to [x, y, w, h] where xy1=top-left, xy2=bottom-right + y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x) + y[:, 0] = (x[:, 0] + x[:, 2]) / 2 # x center + y[:, 1] = (x[:, 1] + x[:, 3]) / 2 # y center + y[:, 2] = x[:, 2] - x[:, 0] # width + y[:, 3] = x[:, 3] - x[:, 1] # height + return y + + +def xywh2xyxy(x): + # Convert nx4 boxes from [x, y, w, h] to [x1, y1, x2, y2] where xy1=top-left, xy2=bottom-right + y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x) + y[:, 0] = x[:, 0] - x[:, 2] / 2 # top left x + y[:, 1] = x[:, 1] - x[:, 3] / 2 # top left y + y[:, 2] = x[:, 0] + x[:, 2] / 2 # bottom right x + y[:, 3] = x[:, 1] + x[:, 3] / 2 # bottom right y + return y + + +def xywhn2xyxy(x, w=640, h=640, padw=0, padh=0): + # Convert nx4 boxes from [x, y, w, h] normalized to [x1, y1, x2, y2] where xy1=top-left, xy2=bottom-right + y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x) + y[:, 0] = w * (x[:, 0] - x[:, 2] / 2) + padw # top left x + y[:, 1] = h * (x[:, 1] - x[:, 3] / 2) + padh # top left y + y[:, 2] = w * (x[:, 0] + x[:, 2] / 2) + padw # bottom right x + y[:, 3] = h * (x[:, 1] + x[:, 3] / 2) + padh # bottom right y + return y + + +def xyxy2xywhn(x, w=640, h=640, clip=False, eps=0.0): + # Convert nx4 boxes from [x1, y1, x2, y2] to [x, y, w, h] normalized where xy1=top-left, xy2=bottom-right + if clip: + clip_coords(x, (h - eps, w - eps)) # warning: inplace clip + y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x) + y[:, 0] = ((x[:, 0] + x[:, 2]) / 2) / w # x center + y[:, 1] = ((x[:, 1] + x[:, 3]) / 2) / h # y center + y[:, 2] = (x[:, 2] - x[:, 0]) / w # width + y[:, 3] = (x[:, 3] - x[:, 1]) / h # height + return y + + +def xyn2xy(x, w=640, h=640, padw=0, padh=0): + # Convert normalized segments into pixel segments, shape (n,2) + y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x) + y[:, 0] = w * x[:, 0] + padw # top left x + y[:, 1] = h * x[:, 1] + padh # top left y + return y + + +def segment2box(segment, width=640, height=640): + # Convert 1 segment label to 1 box label, applying inside-image constraint, i.e. (xy1, xy2, ...) to (xyxy) + x, y = segment.T # segment xy + inside = (x >= 0) & (y >= 0) & (x <= width) & (y <= height) + x, y, = x[inside], y[inside] + return np.array([x.min(), y.min(), x.max(), y.max()]) if any(x) else np.zeros((1, 4)) # xyxy + + +def segments2boxes(segments): + # Convert segment labels to box labels, i.e. (cls, xy1, xy2, ...) to (cls, xywh) + boxes = [] + for s in segments: + x, y = s.T # segment xy + boxes.append([x.min(), y.min(), x.max(), y.max()]) # cls, xyxy + return xyxy2xywh(np.array(boxes)) # cls, xywh + + +def resample_segments(segments, n=1000): + # Up-sample an (n,2) segment + for i, s in enumerate(segments): + x = np.linspace(0, len(s) - 1, n) + xp = np.arange(len(s)) + segments[i] = np.concatenate([np.interp(x, xp, s[:, i]) for i in range(2)]).reshape(2, -1).T # segment xy + return segments + + +def scale_coords(img1_shape, coords, img0_shape, ratio_pad=None): + # Rescale coords (xyxy) from img1_shape to img0_shape + if ratio_pad is None: # calculate from img0_shape + gain = min(img1_shape[0] / img0_shape[0], img1_shape[1] / img0_shape[1]) # gain = old / new + pad = (img1_shape[1] - img0_shape[1] * gain) / 2, (img1_shape[0] - img0_shape[0] * gain) / 2 # wh padding + else: + gain = ratio_pad[0][0] + pad = ratio_pad[1] + + coords[:, [0, 2]] -= pad[0] # x padding + coords[:, [1, 3]] -= pad[1] # y padding + coords[:, :4] /= gain + clip_coords(coords, img0_shape) + return coords + + +def clip_coords(boxes, shape): + # Clip bounding xyxy bounding boxes to image shape (height, width) + if isinstance(boxes, torch.Tensor): # faster individually + boxes[:, 0].clamp_(0, shape[1]) # x1 + boxes[:, 1].clamp_(0, shape[0]) # y1 + boxes[:, 2].clamp_(0, shape[1]) # x2 + boxes[:, 3].clamp_(0, shape[0]) # y2 + else: # np.array (faster grouped) + boxes[:, [0, 2]] = boxes[:, [0, 2]].clip(0, shape[1]) # x1, x2 + boxes[:, [1, 3]] = boxes[:, [1, 3]].clip(0, shape[0]) # y1, y2 + + +def non_max_suppression(prediction, + conf_thres=0.25, + iou_thres=0.45, + classes=None, + agnostic=False, + multi_label=False, + labels=(), + max_det=300): + """Non-Maximum Suppression (NMS) on inference results to reject overlapping bounding boxes + + Returns: + list of detections, on (n,6) tensor per image [xyxy, conf, cls] + """ + + bs = prediction.shape[0] # batch size + nc = prediction.shape[2] - 5 # number of classes + xc = prediction[..., 4] > conf_thres # candidates + + # Checks + assert 0 <= conf_thres <= 1, f'Invalid Confidence threshold {conf_thres}, valid values are between 0.0 and 1.0' + assert 0 <= iou_thres <= 1, f'Invalid IoU {iou_thres}, valid values are between 0.0 and 1.0' + + # Settings + # min_wh = 2 # (pixels) minimum box width and height + max_wh = 7680 # (pixels) maximum box width and height + max_nms = 30000 # maximum number of boxes into torchvision.ops.nms() + time_limit = 0.3 + 0.03 * bs # seconds to quit after + redundant = True # require redundant detections + multi_label &= nc > 1 # multiple labels per box (adds 0.5ms/img) + merge = False # use merge-NMS + + t = time.time() + output = [torch.zeros((0, 6), device=prediction.device)] * bs + for xi, x in enumerate(prediction): # image index, image inference + # Apply constraints + # x[((x[..., 2:4] < min_wh) | (x[..., 2:4] > max_wh)).any(1), 4] = 0 # width-height + x = x[xc[xi]] # confidence + + # Cat apriori labels if autolabelling + if labels and len(labels[xi]): + lb = labels[xi] + v = torch.zeros((len(lb), nc + 5), device=x.device) + v[:, :4] = lb[:, 1:5] # box + v[:, 4] = 1.0 # conf + v[range(len(lb)), lb[:, 0].long() + 5] = 1.0 # cls + x = torch.cat((x, v), 0) + + # If none remain process next image + if not x.shape[0]: + continue + + # Compute conf + x[:, 5:] *= x[:, 4:5] # conf = obj_conf * cls_conf + + # Box (center x, center y, width, height) to (x1, y1, x2, y2) + box = xywh2xyxy(x[:, :4]) + + # Detections matrix nx6 (xyxy, conf, cls) + if multi_label: + i, j = (x[:, 5:] > conf_thres).nonzero(as_tuple=False).T + x = torch.cat((box[i], x[i, j + 5, None], j[:, None].float()), 1) + else: # best class only + conf, j = x[:, 5:].max(1, keepdim=True) + x = torch.cat((box, conf, j.float()), 1)[conf.view(-1) > conf_thres] + + # Filter by class + if classes is not None: + x = x[(x[:, 5:6] == torch.tensor(classes, device=x.device)).any(1)] + + # Apply finite constraint + # if not torch.isfinite(x).all(): + # x = x[torch.isfinite(x).all(1)] + + # Check shape + n = x.shape[0] # number of boxes + if not n: # no boxes + continue + elif n > max_nms: # excess boxes + x = x[x[:, 4].argsort(descending=True)[:max_nms]] # sort by confidence + + # Batched NMS + c = x[:, 5:6] * (0 if agnostic else max_wh) # classes + boxes, scores = x[:, :4] + c, x[:, 4] # boxes (offset by class), scores + i = torchvision.ops.nms(boxes, scores, iou_thres) # NMS + if i.shape[0] > max_det: # limit detections + i = i[:max_det] + if merge and (1 < n < 3E3): # Merge NMS (boxes merged using weighted mean) + # update boxes as boxes(i,4) = weights(i,n) * boxes(n,4) + iou = box_iou(boxes[i], boxes) > iou_thres # iou matrix + weights = iou * scores[None] # box weights + x[i, :4] = torch.mm(weights, x[:, :4]).float() / weights.sum(1, keepdim=True) # merged boxes + if redundant: + i = i[iou.sum(1) > 1] # require redundancy + + output[xi] = x[i] + if (time.time() - t) > time_limit: + LOGGER.warning(f'WARNING: NMS time limit {time_limit:.3f}s exceeded') + break # time limit exceeded + + return output + + +def strip_optimizer(f='best.pt', s=''): # from utils.general import *; strip_optimizer() + # Strip optimizer from 'f' to finalize training, optionally save as 's' + x = torch.load(f, map_location=torch.device('cpu')) + if x.get('ema'): + x['model'] = x['ema'] # replace model with ema + for k in 'optimizer', 'best_fitness', 'wandb_id', 'ema', 'updates': # keys + x[k] = None + x['epoch'] = -1 + x['model'].half() # to FP16 + for p in x['model'].parameters(): + p.requires_grad = False + torch.save(x, s or f) + mb = os.path.getsize(s or f) / 1E6 # filesize + LOGGER.info(f"Optimizer stripped from {f},{f' saved as {s},' if s else ''} {mb:.1f}MB") + + +def print_mutation(results, hyp, save_dir, bucket, prefix=colorstr('evolve: ')): + evolve_csv = save_dir / 'evolve.csv' + evolve_yaml = save_dir / 'hyp_evolve.yaml' + keys = ('metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/mAP_0.5:0.95', 'val/box_loss', + 'val/obj_loss', 'val/cls_loss') + tuple(hyp.keys()) # [results + hyps] + keys = tuple(x.strip() for x in keys) + vals = results + tuple(hyp.values()) + n = len(keys) + + # Download (optional) + if bucket: + url = f'gs://{bucket}/evolve.csv' + if gsutil_getsize(url) > (evolve_csv.stat().st_size if evolve_csv.exists() else 0): + os.system(f'gsutil cp {url} {save_dir}') # download evolve.csv if larger than local + + # Log to evolve.csv + s = '' if evolve_csv.exists() else (('%20s,' * n % keys).rstrip(',') + '\n') # add header + with open(evolve_csv, 'a') as f: + f.write(s + ('%20.5g,' * n % vals).rstrip(',') + '\n') + + # Save yaml + with open(evolve_yaml, 'w') as f: + data = pd.read_csv(evolve_csv) + data = data.rename(columns=lambda x: x.strip()) # strip keys + i = np.argmax(fitness(data.values[:, :4])) # + generations = len(data) + f.write('# YOLOv5 Hyperparameter Evolution Results\n' + f'# Best generation: {i}\n' + + f'# Last generation: {generations - 1}\n' + '# ' + ', '.join(f'{x.strip():>20s}' for x in keys[:7]) + + '\n' + '# ' + ', '.join(f'{x:>20.5g}' for x in data.values[i, :7]) + '\n\n') + yaml.safe_dump(data.loc[i][7:].to_dict(), f, sort_keys=False) + + # Print to screen + LOGGER.info(prefix + f'{generations} generations finished, current result:\n' + prefix + + ', '.join(f'{x.strip():>20s}' for x in keys) + '\n' + prefix + ', '.join(f'{x:20.5g}' + for x in vals) + '\n\n') + + if bucket: + os.system(f'gsutil cp {evolve_csv} {evolve_yaml} gs://{bucket}') # upload + + +def apply_classifier(x, model, img, im0): + # Apply a second stage classifier to YOLO outputs + # Example model = torchvision.models.__dict__['efficientnet_b0'](pretrained=True).to(device).eval() + im0 = [im0] if isinstance(im0, np.ndarray) else im0 + for i, d in enumerate(x): # per image + if d is not None and len(d): + d = d.clone() + + # Reshape and pad cutouts + b = xyxy2xywh(d[:, :4]) # boxes + b[:, 2:] = b[:, 2:].max(1)[0].unsqueeze(1) # rectangle to square + b[:, 2:] = b[:, 2:] * 1.3 + 30 # pad + d[:, :4] = xywh2xyxy(b).long() + + # Rescale boxes from img_size to im0 size + scale_coords(img.shape[2:], d[:, :4], im0[i].shape) + + # Classes + pred_cls1 = d[:, 5].long() + ims = [] + for a in d: + cutout = im0[i][int(a[1]):int(a[3]), int(a[0]):int(a[2])] + im = cv2.resize(cutout, (224, 224)) # BGR + + im = im[:, :, ::-1].transpose(2, 0, 1) # BGR to RGB, to 3x416x416 + im = np.ascontiguousarray(im, dtype=np.float32) # uint8 to float32 + im /= 255 # 0 - 255 to 0.0 - 1.0 + ims.append(im) + + pred_cls2 = model(torch.Tensor(ims).to(d.device)).argmax(1) # classifier prediction + x[i] = x[i][pred_cls1 == pred_cls2] # retain matching class detections + + return x + + +def increment_path(path, exist_ok=False, sep='', mkdir=False): + # Increment file or directory path, i.e. runs/exp --> runs/exp{sep}2, runs/exp{sep}3, ... etc. + path = Path(path) # os-agnostic + if path.exists() and not exist_ok: + path, suffix = (path.with_suffix(''), path.suffix) if path.is_file() else (path, '') + + # Method 1 + for n in range(2, 9999): + p = f'{path}{sep}{n}{suffix}' # increment path + if not os.path.exists(p): # + break + path = Path(p) + + # Method 2 (deprecated) + # dirs = glob.glob(f"{path}{sep}*") # similar paths + # matches = [re.search(rf"{path.stem}{sep}(\d+)", d) for d in dirs] + # i = [int(m.groups()[0]) for m in matches if m] # indices + # n = max(i) + 1 if i else 2 # increment number + # path = Path(f"{path}{sep}{n}{suffix}") # increment path + + if mkdir: + path.mkdir(parents=True, exist_ok=True) # make directory + + return path + + +# OpenCV Chinese-friendly functions ------------------------------------------------------------------------------------ +imshow_ = cv2.imshow # copy to avoid recursion errors + + +def imread(path, flags=cv2.IMREAD_COLOR): + return cv2.imdecode(np.fromfile(path, np.uint8), flags) + + +def imwrite(path, im): + try: + cv2.imencode(Path(path).suffix, im)[1].tofile(path) + return True + except Exception: + return False + + +def imshow(path, im): + imshow_(path.encode('unicode_escape').decode(), im) + + +cv2.imread, cv2.imwrite, cv2.imshow = imread, imwrite, imshow # redefine + +# Variables ------------------------------------------------------------------------------------------------------------ +NCOLS = 0 if is_docker() else shutil.get_terminal_size().columns # terminal window size for tqdm diff --git a/yolov5/utils/google_app_engine/Dockerfile b/yolov5/utils/google_app_engine/Dockerfile new file mode 100644 index 0000000..0155618 --- /dev/null +++ b/yolov5/utils/google_app_engine/Dockerfile @@ -0,0 +1,25 @@ +FROM gcr.io/google-appengine/python + +# Create a virtualenv for dependencies. This isolates these packages from +# system-level packages. +# Use -p python3 or -p python3.7 to select python version. Default is version 2. +RUN virtualenv /env -p python3 + +# Setting these environment variables are the same as running +# source /env/bin/activate. +ENV VIRTUAL_ENV /env +ENV PATH /env/bin:$PATH + +RUN apt-get update && apt-get install -y python-opencv + +# Copy the application's requirements.txt and run pip to install all +# dependencies into the virtualenv. +ADD requirements.txt /app/requirements.txt +RUN pip install -r /app/requirements.txt + +# Add the application source code. +ADD . /app + +# Run a WSGI server to serve the application. gunicorn must be declared as +# a dependency in requirements.txt. +CMD gunicorn -b :$PORT main:app diff --git a/yolov5/utils/google_app_engine/additional_requirements.txt b/yolov5/utils/google_app_engine/additional_requirements.txt new file mode 100644 index 0000000..42d7ffc --- /dev/null +++ b/yolov5/utils/google_app_engine/additional_requirements.txt @@ -0,0 +1,4 @@ +# add these requirements in your app on top of the existing ones +pip==21.1 +Flask==1.0.2 +gunicorn==19.9.0 diff --git a/yolov5/utils/google_app_engine/app.yaml b/yolov5/utils/google_app_engine/app.yaml new file mode 100644 index 0000000..5056b7c --- /dev/null +++ b/yolov5/utils/google_app_engine/app.yaml @@ -0,0 +1,14 @@ +runtime: custom +env: flex + +service: yolov5app + +liveness_check: + initial_delay_sec: 600 + +manual_scaling: + instances: 1 +resources: + cpu: 1 + memory_gb: 4 + disk_size_gb: 20 diff --git a/yolov5/utils/loggers/__init__.py b/yolov5/utils/loggers/__init__.py new file mode 100644 index 0000000..42b696b --- /dev/null +++ b/yolov5/utils/loggers/__init__.py @@ -0,0 +1,187 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Logging utils +""" + +import os +import warnings + +import pkg_resources as pkg +import torch +from torch.utils.tensorboard import SummaryWriter + +from utils.general import colorstr, cv2, emojis +from utils.loggers.wandb.wandb_utils import WandbLogger +from utils.plots import plot_images, plot_results +from utils.torch_utils import de_parallel + +LOGGERS = ('csv', 'tb', 'wandb') # text-file, TensorBoard, Weights & Biases +RANK = int(os.getenv('RANK', -1)) + +try: + import wandb + + assert hasattr(wandb, '__version__') # verify package import not local dir + if pkg.parse_version(wandb.__version__) >= pkg.parse_version('0.12.2') and RANK in {0, -1}: + try: + wandb_login_success = wandb.login(timeout=30) + except wandb.errors.UsageError: # known non-TTY terminal issue + wandb_login_success = False + if not wandb_login_success: + wandb = None +except (ImportError, AssertionError): + wandb = None + + +class Loggers(): + # YOLOv5 Loggers class + def __init__(self, save_dir=None, weights=None, opt=None, hyp=None, logger=None, include=LOGGERS): + self.save_dir = save_dir + self.weights = weights + self.opt = opt + self.hyp = hyp + self.logger = logger # for printing results to console + self.include = include + self.keys = [ + 'train/box_loss', + 'train/obj_loss', + 'train/cls_loss', # train loss + 'metrics/precision', + 'metrics/recall', + 'metrics/mAP_0.5', + 'metrics/mAP_0.5:0.95', # metrics + 'val/box_loss', + 'val/obj_loss', + 'val/cls_loss', # val loss + 'x/lr0', + 'x/lr1', + 'x/lr2'] # params + self.best_keys = ['best/epoch', 'best/precision', 'best/recall', 'best/mAP_0.5', 'best/mAP_0.5:0.95'] + for k in LOGGERS: + setattr(self, k, None) # init empty logger dictionary + self.csv = True # always log to csv + + # Message + if not wandb: + prefix = colorstr('Weights & Biases: ') + s = f"{prefix}run 'pip install wandb' to automatically track and visualize YOLOv5 🚀 runs (RECOMMENDED)" + self.logger.info(emojis(s)) + + # TensorBoard + s = self.save_dir + if 'tb' in self.include and not self.opt.evolve: + prefix = colorstr('TensorBoard: ') + self.logger.info(f"{prefix}Start with 'tensorboard --logdir {s.parent}', view at http://localhost:6006/") + self.tb = SummaryWriter(str(s)) + + # W&B + if wandb and 'wandb' in self.include: + wandb_artifact_resume = isinstance(self.opt.resume, str) and self.opt.resume.startswith('wandb-artifact://') + run_id = torch.load(self.weights).get('wandb_id') if self.opt.resume and not wandb_artifact_resume else None + self.opt.hyp = self.hyp # add hyperparameters + self.wandb = WandbLogger(self.opt, run_id) + # temp warn. because nested artifacts not supported after 0.12.10 + if pkg.parse_version(wandb.__version__) >= pkg.parse_version('0.12.11'): + self.logger.warning( + "YOLOv5 temporarily requires wandb version 0.12.10 or below. Some features may not work as expected." + ) + else: + self.wandb = None + + def on_train_start(self): + # Callback runs on train start + pass + + def on_pretrain_routine_end(self): + # Callback runs on pre-train routine end + paths = self.save_dir.glob('*labels*.jpg') # training labels + if self.wandb: + self.wandb.log({"Labels": [wandb.Image(str(x), caption=x.name) for x in paths]}) + + def on_train_batch_end(self, ni, model, imgs, targets, paths, plots): + # Callback runs on train batch end + if plots: + if ni == 0: + if not self.opt.sync_bn: # --sync known issue https://github.com/ultralytics/yolov5/issues/3754 + with warnings.catch_warnings(): + warnings.simplefilter('ignore') # suppress jit trace warning + self.tb.add_graph(torch.jit.trace(de_parallel(model), imgs[0:1], strict=False), []) + if ni < 3: + f = self.save_dir / f'train_batch{ni}.jpg' # filename + plot_images(imgs, targets, paths, f) + if self.wandb and ni == 10: + files = sorted(self.save_dir.glob('train*.jpg')) + self.wandb.log({'Mosaics': [wandb.Image(str(f), caption=f.name) for f in files if f.exists()]}) + + def on_train_epoch_end(self, epoch): + # Callback runs on train epoch end + if self.wandb: + self.wandb.current_epoch = epoch + 1 + + def on_val_image_end(self, pred, predn, path, names, im): + # Callback runs on val image end + if self.wandb: + self.wandb.val_one_image(pred, predn, path, names, im) + + def on_val_end(self): + # Callback runs on val end + if self.wandb: + files = sorted(self.save_dir.glob('val*.jpg')) + self.wandb.log({"Validation": [wandb.Image(str(f), caption=f.name) for f in files]}) + + def on_fit_epoch_end(self, vals, epoch, best_fitness, fi): + # Callback runs at the end of each fit (train+val) epoch + x = dict(zip(self.keys, vals)) + if self.csv: + file = self.save_dir / 'results.csv' + n = len(x) + 1 # number of cols + s = '' if file.exists() else (('%20s,' * n % tuple(['epoch'] + self.keys)).rstrip(',') + '\n') # add header + with open(file, 'a') as f: + f.write(s + ('%20.5g,' * n % tuple([epoch] + vals)).rstrip(',') + '\n') + + if self.tb: + for k, v in x.items(): + self.tb.add_scalar(k, v, epoch) + + if self.wandb: + if best_fitness == fi: + best_results = [epoch] + vals[3:7] + for i, name in enumerate(self.best_keys): + self.wandb.wandb_run.summary[name] = best_results[i] # log best results in the summary + self.wandb.log(x) + self.wandb.end_epoch(best_result=best_fitness == fi) + + def on_model_save(self, last, epoch, final_epoch, best_fitness, fi): + # Callback runs on model save event + if self.wandb: + if ((epoch + 1) % self.opt.save_period == 0 and not final_epoch) and self.opt.save_period != -1: + self.wandb.log_model(last.parent, self.opt, epoch, fi, best_model=best_fitness == fi) + + def on_train_end(self, last, best, plots, epoch, results): + # Callback runs on training end + if plots: + plot_results(file=self.save_dir / 'results.csv') # save results.png + files = ['results.png', 'confusion_matrix.png', *(f'{x}_curve.png' for x in ('F1', 'PR', 'P', 'R'))] + files = [(self.save_dir / f) for f in files if (self.save_dir / f).exists()] # filter + self.logger.info(f"Results saved to {colorstr('bold', self.save_dir)}") + + if self.tb: + for f in files: + self.tb.add_image(f.stem, cv2.imread(str(f))[..., ::-1], epoch, dataformats='HWC') + + if self.wandb: + self.wandb.log(dict(zip(self.keys[3:10], results))) + self.wandb.log({"Results": [wandb.Image(str(f), caption=f.name) for f in files]}) + # Calling wandb.log. TODO: Refactor this into WandbLogger.log_model + if not self.opt.evolve: + wandb.log_artifact(str(best if best.exists() else last), + type='model', + name=f'run_{self.wandb.wandb_run.id}_model', + aliases=['latest', 'best', 'stripped']) + self.wandb.finish_run() + + def on_params_update(self, params): + # Update hyperparams or configs of the experiment + # params: A dict containing {param: value} pairs + if self.wandb: + self.wandb.wandb_run.config.update(params, allow_val_change=True) diff --git a/yolov5/utils/loggers/wandb/README.md b/yolov5/utils/loggers/wandb/README.md new file mode 100644 index 0000000..d78324b --- /dev/null +++ b/yolov5/utils/loggers/wandb/README.md @@ -0,0 +1,162 @@ +📚 This guide explains how to use **Weights & Biases** (W&B) with YOLOv5 🚀. UPDATED 29 September 2021. + +- [About Weights & Biases](#about-weights-&-biases) +- [First-Time Setup](#first-time-setup) +- [Viewing runs](#viewing-runs) +- [Disabling wandb](#disabling-wandb) +- [Advanced Usage: Dataset Versioning and Evaluation](#advanced-usage) +- [Reports: Share your work with the world!](#reports) + +## About Weights & Biases + +Think of [W&B](https://wandb.ai/site?utm_campaign=repo_yolo_wandbtutorial) like GitHub for machine learning models. With a few lines of code, save everything you need to debug, compare and reproduce your models — architecture, hyperparameters, git commits, model weights, GPU usage, and even datasets and predictions. + +Used by top researchers including teams at OpenAI, Lyft, Github, and MILA, W&B is part of the new standard of best practices for machine learning. How W&B can help you optimize your machine learning workflows: + +- [Debug](https://wandb.ai/wandb/getting-started/reports/Visualize-Debug-Machine-Learning-Models--VmlldzoyNzY5MDk#Free-2) model performance in real time +- [GPU usage](https://wandb.ai/wandb/getting-started/reports/Visualize-Debug-Machine-Learning-Models--VmlldzoyNzY5MDk#System-4) visualized automatically +- [Custom charts](https://wandb.ai/wandb/customizable-charts/reports/Powerful-Custom-Charts-To-Debug-Model-Peformance--VmlldzoyNzY4ODI) for powerful, extensible visualization +- [Share insights](https://wandb.ai/wandb/getting-started/reports/Visualize-Debug-Machine-Learning-Models--VmlldzoyNzY5MDk#Share-8) interactively with collaborators +- [Optimize hyperparameters](https://docs.wandb.com/sweeps) efficiently +- [Track](https://docs.wandb.com/artifacts) datasets, pipelines, and production models + +## First-Time Setup + +
+ Toggle Details +When you first train, W&B will prompt you to create a new account and will generate an **API key** for you. If you are an existing user you can retrieve your key from https://wandb.ai/authorize. This key is used to tell W&B where to log your data. You only need to supply your key once, and then it is remembered on the same device. + +W&B will create a cloud **project** (default is 'YOLOv5') for your training runs, and each new training run will be provided a unique run **name** within that project as project/name. You can also manually set your project and run name as: + +```shell +$ python train.py --project ... --name ... +``` + +YOLOv5 notebook example: Open In Colab Open In Kaggle +Screen Shot 2021-09-29 at 10 23 13 PM + +
+ +## Viewing Runs + +
+ Toggle Details +Run information streams from your environment to the W&B cloud console as you train. This allows you to monitor and even cancel runs in realtime . All important information is logged: + +- Training & Validation losses +- Metrics: Precision, Recall, mAP@0.5, mAP@0.5:0.95 +- Learning Rate over time +- A bounding box debugging panel, showing the training progress over time +- GPU: Type, **GPU Utilization**, power, temperature, **CUDA memory usage** +- System: Disk I/0, CPU utilization, RAM memory usage +- Your trained model as W&B Artifact +- Environment: OS and Python types, Git repository and state, **training command** + +

Weights & Biases dashboard

+
+ +## Disabling wandb + +- training after running `wandb disabled` inside that directory creates no wandb run + ![Screenshot (84)](https://user-images.githubusercontent.com/15766192/143441777-c780bdd7-7cb4-4404-9559-b4316030a985.png) + +- To enable wandb again, run `wandb online` + ![Screenshot (85)](https://user-images.githubusercontent.com/15766192/143441866-7191b2cb-22f0-4e0f-ae64-2dc47dc13078.png) + +## Advanced Usage + +You can leverage W&B artifacts and Tables integration to easily visualize and manage your datasets, models and training evaluations. Here are some quick examples to get you started. + +
+

1: Train and Log Evaluation simultaneousy

+ This is an extension of the previous section, but it'll also training after uploading the dataset. This also evaluation Table + Evaluation table compares your predictions and ground truths across the validation set for each epoch. It uses the references to the already uploaded datasets, + so no images will be uploaded from your system more than once. +
+ Usage + Code $ python train.py --upload_data val + +![Screenshot from 2021-11-21 17-40-06](https://user-images.githubusercontent.com/15766192/142761183-c1696d8c-3f38-45ab-991a-bb0dfd98ae7d.png) + +
+ +

2. Visualize and Version Datasets

+ Log, visualize, dynamically query, and understand your data with W&B Tables. You can use the following command to log your dataset as a W&B Table. This will generate a {dataset}_wandb.yaml file which can be used to train from dataset artifact. +
+ Usage + Code $ python utils/logger/wandb/log_dataset.py --project ... --name ... --data .. + +![Screenshot (64)](https://user-images.githubusercontent.com/15766192/128486078-d8433890-98a3-4d12-8986-b6c0e3fc64b9.png) + +
+ +

3: Train using dataset artifact

+ When you upload a dataset as described in the first section, you get a new config file with an added `_wandb` to its name. This file contains the information that + can be used to train a model directly from the dataset artifact. This also logs evaluation +
+ Usage + Code $ python train.py --data {data}_wandb.yaml + +![Screenshot (72)](https://user-images.githubusercontent.com/15766192/128979739-4cf63aeb-a76f-483f-8861-1c0100b938a5.png) + +
+ +

4: Save model checkpoints as artifacts

+ To enable saving and versioning checkpoints of your experiment, pass `--save_period n` with the base cammand, where `n` represents checkpoint interval. + You can also log both the dataset and model checkpoints simultaneously. If not passed, only the final model will be logged + +
+ Usage + Code $ python train.py --save_period 1 + +![Screenshot (68)](https://user-images.githubusercontent.com/15766192/128726138-ec6c1f60-639d-437d-b4ee-3acd9de47ef3.png) + +
+ +
+ +

5: Resume runs from checkpoint artifacts.

+Any run can be resumed using artifacts if the --resume argument starts with wandb-artifact:// prefix followed by the run path, i.e, wandb-artifact://username/project/runid . This doesn't require the model checkpoint to be present on the local system. + +
+ Usage + Code $ python train.py --resume wandb-artifact://{run_path} + +![Screenshot (70)](https://user-images.githubusercontent.com/15766192/128728988-4e84b355-6c87-41ae-a591-14aecf45343e.png) + +
+ +

6: Resume runs from dataset artifact & checkpoint artifacts.

+ Local dataset or model checkpoints are not required. This can be used to resume runs directly on a different device + The syntax is same as the previous section, but you'll need to lof both the dataset and model checkpoints as artifacts, i.e, set bot --upload_dataset or + train from _wandb.yaml file and set --save_period + +
+ Usage + Code $ python train.py --resume wandb-artifact://{run_path} + +![Screenshot (70)](https://user-images.githubusercontent.com/15766192/128728988-4e84b355-6c87-41ae-a591-14aecf45343e.png) + +
+ + + +

Reports

+W&B Reports can be created from your saved runs for sharing online. Once a report is created you will receive a link you can use to publically share your results. Here is an example report created from the COCO128 tutorial trainings of all four YOLOv5 models ([link](https://wandb.ai/glenn-jocher/yolov5_tutorial/reports/YOLOv5-COCO128-Tutorial-Results--VmlldzozMDI5OTY)). + +Weights & Biases Reports + +## Environments + +YOLOv5 may be run in any of the following up-to-date verified environments (with all dependencies including [CUDA](https://developer.nvidia.com/cuda)/[CUDNN](https://developer.nvidia.com/cudnn), [Python](https://www.python.org/) and [PyTorch](https://pytorch.org/) preinstalled): + +- **Google Colab and Kaggle** notebooks with free GPU: Open In Colab Open In Kaggle +- **Google Cloud** Deep Learning VM. See [GCP Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/GCP-Quickstart) +- **Amazon** Deep Learning AMI. See [AWS Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/AWS-Quickstart) +- **Docker Image**. See [Docker Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/Docker-Quickstart) Docker Pulls + +## Status + +![CI CPU testing](https://github.com/ultralytics/yolov5/workflows/CI%20CPU%20testing/badge.svg) + +If this badge is green, all [YOLOv5 GitHub Actions](https://github.com/ultralytics/yolov5/actions) Continuous Integration (CI) tests are currently passing. CI tests verify correct operation of YOLOv5 training ([train.py](https://github.com/ultralytics/yolov5/blob/master/train.py)), validation ([val.py](https://github.com/ultralytics/yolov5/blob/master/val.py)), inference ([detect.py](https://github.com/ultralytics/yolov5/blob/master/detect.py)) and export ([export.py](https://github.com/ultralytics/yolov5/blob/master/export.py)) on macOS, Windows, and Ubuntu every 24 hours and on every commit. diff --git a/yolov5/utils/loggers/wandb/__init__.py b/yolov5/utils/loggers/wandb/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/yolov5/utils/loggers/wandb/log_dataset.py b/yolov5/utils/loggers/wandb/log_dataset.py new file mode 100644 index 0000000..06e81fb --- /dev/null +++ b/yolov5/utils/loggers/wandb/log_dataset.py @@ -0,0 +1,27 @@ +import argparse + +from wandb_utils import WandbLogger + +from utils.general import LOGGER + +WANDB_ARTIFACT_PREFIX = 'wandb-artifact://' + + +def create_dataset_artifact(opt): + logger = WandbLogger(opt, None, job_type='Dataset Creation') # TODO: return value unused + if not logger.wandb: + LOGGER.info("install wandb using `pip install wandb` to log the dataset") + + +if __name__ == '__main__': + parser = argparse.ArgumentParser() + parser.add_argument('--data', type=str, default='data/coco128.yaml', help='data.yaml path') + parser.add_argument('--single-cls', action='store_true', help='train as single-class dataset') + parser.add_argument('--project', type=str, default='YOLOv5', help='name of W&B Project') + parser.add_argument('--entity', default=None, help='W&B entity') + parser.add_argument('--name', type=str, default='log dataset', help='name of W&B run') + + opt = parser.parse_args() + opt.resume = False # Explicitly disallow resume check for dataset upload job + + create_dataset_artifact(opt) diff --git a/yolov5/utils/loggers/wandb/sweep.py b/yolov5/utils/loggers/wandb/sweep.py new file mode 100644 index 0000000..d49ea6f --- /dev/null +++ b/yolov5/utils/loggers/wandb/sweep.py @@ -0,0 +1,41 @@ +import sys +from pathlib import Path + +import wandb + +FILE = Path(__file__).resolve() +ROOT = FILE.parents[3] # YOLOv5 root directory +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) # add ROOT to PATH + +from train import parse_opt, train +from utils.callbacks import Callbacks +from utils.general import increment_path +from utils.torch_utils import select_device + + +def sweep(): + wandb.init() + # Get hyp dict from sweep agent. Copy because train() modifies parameters which confused wandb. + hyp_dict = vars(wandb.config).get("_items").copy() + + # Workaround: get necessary opt args + opt = parse_opt(known=True) + opt.batch_size = hyp_dict.get("batch_size") + opt.save_dir = str(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok or opt.evolve)) + opt.epochs = hyp_dict.get("epochs") + opt.nosave = True + opt.data = hyp_dict.get("data") + opt.weights = str(opt.weights) + opt.cfg = str(opt.cfg) + opt.data = str(opt.data) + opt.hyp = str(opt.hyp) + opt.project = str(opt.project) + device = select_device(opt.device, batch_size=opt.batch_size) + + # train + train(hyp_dict, opt, device, callbacks=Callbacks()) + + +if __name__ == "__main__": + sweep() diff --git a/yolov5/utils/loggers/wandb/sweep.yaml b/yolov5/utils/loggers/wandb/sweep.yaml new file mode 100644 index 0000000..688b1ea --- /dev/null +++ b/yolov5/utils/loggers/wandb/sweep.yaml @@ -0,0 +1,143 @@ +# Hyperparameters for training +# To set range- +# Provide min and max values as: +# parameter: +# +# min: scalar +# max: scalar +# OR +# +# Set a specific list of search space- +# parameter: +# values: [scalar1, scalar2, scalar3...] +# +# You can use grid, bayesian and hyperopt search strategy +# For more info on configuring sweeps visit - https://docs.wandb.ai/guides/sweeps/configuration + +program: utils/loggers/wandb/sweep.py +method: random +metric: + name: metrics/mAP_0.5 + goal: maximize + +parameters: + # hyperparameters: set either min, max range or values list + data: + value: "data/coco128.yaml" + batch_size: + values: [64] + epochs: + values: [10] + + lr0: + distribution: uniform + min: 1e-5 + max: 1e-1 + lrf: + distribution: uniform + min: 0.01 + max: 1.0 + momentum: + distribution: uniform + min: 0.6 + max: 0.98 + weight_decay: + distribution: uniform + min: 0.0 + max: 0.001 + warmup_epochs: + distribution: uniform + min: 0.0 + max: 5.0 + warmup_momentum: + distribution: uniform + min: 0.0 + max: 0.95 + warmup_bias_lr: + distribution: uniform + min: 0.0 + max: 0.2 + box: + distribution: uniform + min: 0.02 + max: 0.2 + cls: + distribution: uniform + min: 0.2 + max: 4.0 + cls_pw: + distribution: uniform + min: 0.5 + max: 2.0 + obj: + distribution: uniform + min: 0.2 + max: 4.0 + obj_pw: + distribution: uniform + min: 0.5 + max: 2.0 + iou_t: + distribution: uniform + min: 0.1 + max: 0.7 + anchor_t: + distribution: uniform + min: 2.0 + max: 8.0 + fl_gamma: + distribution: uniform + min: 0.0 + max: 4.0 + hsv_h: + distribution: uniform + min: 0.0 + max: 0.1 + hsv_s: + distribution: uniform + min: 0.0 + max: 0.9 + hsv_v: + distribution: uniform + min: 0.0 + max: 0.9 + degrees: + distribution: uniform + min: 0.0 + max: 45.0 + translate: + distribution: uniform + min: 0.0 + max: 0.9 + scale: + distribution: uniform + min: 0.0 + max: 0.9 + shear: + distribution: uniform + min: 0.0 + max: 10.0 + perspective: + distribution: uniform + min: 0.0 + max: 0.001 + flipud: + distribution: uniform + min: 0.0 + max: 1.0 + fliplr: + distribution: uniform + min: 0.0 + max: 1.0 + mosaic: + distribution: uniform + min: 0.0 + max: 1.0 + mixup: + distribution: uniform + min: 0.0 + max: 1.0 + copy_paste: + distribution: uniform + min: 0.0 + max: 1.0 diff --git a/yolov5/utils/loggers/wandb/wandb_utils.py b/yolov5/utils/loggers/wandb/wandb_utils.py new file mode 100644 index 0000000..04521bf --- /dev/null +++ b/yolov5/utils/loggers/wandb/wandb_utils.py @@ -0,0 +1,577 @@ +"""Utilities and tools for tracking runs with Weights & Biases.""" + +import logging +import os +import sys +from contextlib import contextmanager +from pathlib import Path +from typing import Dict + +import yaml +from tqdm import tqdm + +FILE = Path(__file__).resolve() +ROOT = FILE.parents[3] # YOLOv5 root directory +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) # add ROOT to PATH + +from utils.dataloaders import LoadImagesAndLabels, img2label_paths +from utils.general import LOGGER, check_dataset, check_file + +try: + import wandb + + assert hasattr(wandb, '__version__') # verify package import not local dir +except (ImportError, AssertionError): + wandb = None + +RANK = int(os.getenv('RANK', -1)) +WANDB_ARTIFACT_PREFIX = 'wandb-artifact://' + + +def remove_prefix(from_string, prefix=WANDB_ARTIFACT_PREFIX): + return from_string[len(prefix):] + + +def check_wandb_config_file(data_config_file): + wandb_config = '_wandb.'.join(data_config_file.rsplit('.', 1)) # updated data.yaml path + if Path(wandb_config).is_file(): + return wandb_config + return data_config_file + + +def check_wandb_dataset(data_file): + is_trainset_wandb_artifact = False + is_valset_wandb_artifact = False + if check_file(data_file) and data_file.endswith('.yaml'): + with open(data_file, errors='ignore') as f: + data_dict = yaml.safe_load(f) + is_trainset_wandb_artifact = isinstance(data_dict['train'], + str) and data_dict['train'].startswith(WANDB_ARTIFACT_PREFIX) + is_valset_wandb_artifact = isinstance(data_dict['val'], + str) and data_dict['val'].startswith(WANDB_ARTIFACT_PREFIX) + if is_trainset_wandb_artifact or is_valset_wandb_artifact: + return data_dict + else: + return check_dataset(data_file) + + +def get_run_info(run_path): + run_path = Path(remove_prefix(run_path, WANDB_ARTIFACT_PREFIX)) + run_id = run_path.stem + project = run_path.parent.stem + entity = run_path.parent.parent.stem + model_artifact_name = 'run_' + run_id + '_model' + return entity, project, run_id, model_artifact_name + + +def check_wandb_resume(opt): + process_wandb_config_ddp_mode(opt) if RANK not in [-1, 0] else None + if isinstance(opt.resume, str): + if opt.resume.startswith(WANDB_ARTIFACT_PREFIX): + if RANK not in [-1, 0]: # For resuming DDP runs + entity, project, run_id, model_artifact_name = get_run_info(opt.resume) + api = wandb.Api() + artifact = api.artifact(entity + '/' + project + '/' + model_artifact_name + ':latest') + modeldir = artifact.download() + opt.weights = str(Path(modeldir) / "last.pt") + return True + return None + + +def process_wandb_config_ddp_mode(opt): + with open(check_file(opt.data), errors='ignore') as f: + data_dict = yaml.safe_load(f) # data dict + train_dir, val_dir = None, None + if isinstance(data_dict['train'], str) and data_dict['train'].startswith(WANDB_ARTIFACT_PREFIX): + api = wandb.Api() + train_artifact = api.artifact(remove_prefix(data_dict['train']) + ':' + opt.artifact_alias) + train_dir = train_artifact.download() + train_path = Path(train_dir) / 'data/images/' + data_dict['train'] = str(train_path) + + if isinstance(data_dict['val'], str) and data_dict['val'].startswith(WANDB_ARTIFACT_PREFIX): + api = wandb.Api() + val_artifact = api.artifact(remove_prefix(data_dict['val']) + ':' + opt.artifact_alias) + val_dir = val_artifact.download() + val_path = Path(val_dir) / 'data/images/' + data_dict['val'] = str(val_path) + if train_dir or val_dir: + ddp_data_path = str(Path(val_dir) / 'wandb_local_data.yaml') + with open(ddp_data_path, 'w') as f: + yaml.safe_dump(data_dict, f) + opt.data = ddp_data_path + + +class WandbLogger(): + """Log training runs, datasets, models, and predictions to Weights & Biases. + + This logger sends information to W&B at wandb.ai. By default, this information + includes hyperparameters, system configuration and metrics, model metrics, + and basic data metrics and analyses. + + By providing additional command line arguments to train.py, datasets, + models and predictions can also be logged. + + For more on how this logger is used, see the Weights & Biases documentation: + https://docs.wandb.com/guides/integrations/yolov5 + """ + + def __init__(self, opt, run_id=None, job_type='Training'): + """ + - Initialize WandbLogger instance + - Upload dataset if opt.upload_dataset is True + - Setup trainig processes if job_type is 'Training' + + arguments: + opt (namespace) -- Commandline arguments for this run + run_id (str) -- Run ID of W&B run to be resumed + job_type (str) -- To set the job_type for this run + + """ + # Pre-training routine -- + self.job_type = job_type + self.wandb, self.wandb_run = wandb, None if not wandb else wandb.run + self.val_artifact, self.train_artifact = None, None + self.train_artifact_path, self.val_artifact_path = None, None + self.result_artifact = None + self.val_table, self.result_table = None, None + self.bbox_media_panel_images = [] + self.val_table_path_map = None + self.max_imgs_to_log = 16 + self.wandb_artifact_data_dict = None + self.data_dict = None + # It's more elegant to stick to 1 wandb.init call, + # but useful config data is overwritten in the WandbLogger's wandb.init call + if isinstance(opt.resume, str): # checks resume from artifact + if opt.resume.startswith(WANDB_ARTIFACT_PREFIX): + entity, project, run_id, model_artifact_name = get_run_info(opt.resume) + model_artifact_name = WANDB_ARTIFACT_PREFIX + model_artifact_name + assert wandb, 'install wandb to resume wandb runs' + # Resume wandb-artifact:// runs here| workaround for not overwriting wandb.config + self.wandb_run = wandb.init(id=run_id, + project=project, + entity=entity, + resume='allow', + allow_val_change=True) + opt.resume = model_artifact_name + elif self.wandb: + self.wandb_run = wandb.init(config=opt, + resume="allow", + project='YOLOv5' if opt.project == 'runs/train' else Path(opt.project).stem, + entity=opt.entity, + name=opt.name if opt.name != 'exp' else None, + job_type=job_type, + id=run_id, + allow_val_change=True) if not wandb.run else wandb.run + if self.wandb_run: + if self.job_type == 'Training': + if opt.upload_dataset: + if not opt.resume: + self.wandb_artifact_data_dict = self.check_and_upload_dataset(opt) + + if opt.resume: + # resume from artifact + if isinstance(opt.resume, str) and opt.resume.startswith(WANDB_ARTIFACT_PREFIX): + self.data_dict = dict(self.wandb_run.config.data_dict) + else: # local resume + self.data_dict = check_wandb_dataset(opt.data) + else: + self.data_dict = check_wandb_dataset(opt.data) + self.wandb_artifact_data_dict = self.wandb_artifact_data_dict or self.data_dict + + # write data_dict to config. useful for resuming from artifacts. Do this only when not resuming. + self.wandb_run.config.update({'data_dict': self.wandb_artifact_data_dict}, allow_val_change=True) + self.setup_training(opt) + + if self.job_type == 'Dataset Creation': + self.wandb_run.config.update({"upload_dataset": True}) + self.data_dict = self.check_and_upload_dataset(opt) + + def check_and_upload_dataset(self, opt): + """ + Check if the dataset format is compatible and upload it as W&B artifact + + arguments: + opt (namespace)-- Commandline arguments for current run + + returns: + Updated dataset info dictionary where local dataset paths are replaced by WAND_ARFACT_PREFIX links. + """ + assert wandb, 'Install wandb to upload dataset' + config_path = self.log_dataset_artifact(opt.data, opt.single_cls, + 'YOLOv5' if opt.project == 'runs/train' else Path(opt.project).stem) + with open(config_path, errors='ignore') as f: + wandb_data_dict = yaml.safe_load(f) + return wandb_data_dict + + def setup_training(self, opt): + """ + Setup the necessary processes for training YOLO models: + - Attempt to download model checkpoint and dataset artifacts if opt.resume stats with WANDB_ARTIFACT_PREFIX + - Update data_dict, to contain info of previous run if resumed and the paths of dataset artifact if downloaded + - Setup log_dict, initialize bbox_interval + + arguments: + opt (namespace) -- commandline arguments for this run + + """ + self.log_dict, self.current_epoch = {}, 0 + self.bbox_interval = opt.bbox_interval + if isinstance(opt.resume, str): + modeldir, _ = self.download_model_artifact(opt) + if modeldir: + self.weights = Path(modeldir) / "last.pt" + config = self.wandb_run.config + opt.weights, opt.save_period, opt.batch_size, opt.bbox_interval, opt.epochs, opt.hyp, opt.imgsz = str( + self.weights), config.save_period, config.batch_size, config.bbox_interval, config.epochs,\ + config.hyp, config.imgsz + data_dict = self.data_dict + if self.val_artifact is None: # If --upload_dataset is set, use the existing artifact, don't download + self.train_artifact_path, self.train_artifact = self.download_dataset_artifact( + data_dict.get('train'), opt.artifact_alias) + self.val_artifact_path, self.val_artifact = self.download_dataset_artifact( + data_dict.get('val'), opt.artifact_alias) + + if self.train_artifact_path is not None: + train_path = Path(self.train_artifact_path) / 'data/images/' + data_dict['train'] = str(train_path) + if self.val_artifact_path is not None: + val_path = Path(self.val_artifact_path) / 'data/images/' + data_dict['val'] = str(val_path) + + if self.val_artifact is not None: + self.result_artifact = wandb.Artifact("run_" + wandb.run.id + "_progress", "evaluation") + columns = ["epoch", "id", "ground truth", "prediction"] + columns.extend(self.data_dict['names']) + self.result_table = wandb.Table(columns) + self.val_table = self.val_artifact.get("val") + if self.val_table_path_map is None: + self.map_val_table_path() + if opt.bbox_interval == -1: + self.bbox_interval = opt.bbox_interval = (opt.epochs // 10) if opt.epochs > 10 else 1 + if opt.evolve or opt.noplots: + self.bbox_interval = opt.bbox_interval = opt.epochs + 1 # disable bbox_interval + train_from_artifact = self.train_artifact_path is not None and self.val_artifact_path is not None + # Update the the data_dict to point to local artifacts dir + if train_from_artifact: + self.data_dict = data_dict + + def download_dataset_artifact(self, path, alias): + """ + download the model checkpoint artifact if the path starts with WANDB_ARTIFACT_PREFIX + + arguments: + path -- path of the dataset to be used for training + alias (str)-- alias of the artifact to be download/used for training + + returns: + (str, wandb.Artifact) -- path of the downladed dataset and it's corresponding artifact object if dataset + is found otherwise returns (None, None) + """ + if isinstance(path, str) and path.startswith(WANDB_ARTIFACT_PREFIX): + artifact_path = Path(remove_prefix(path, WANDB_ARTIFACT_PREFIX) + ":" + alias) + dataset_artifact = wandb.use_artifact(artifact_path.as_posix().replace("\\", "/")) + assert dataset_artifact is not None, "'Error: W&B dataset artifact doesn\'t exist'" + datadir = dataset_artifact.download() + return datadir, dataset_artifact + return None, None + + def download_model_artifact(self, opt): + """ + download the model checkpoint artifact if the resume path starts with WANDB_ARTIFACT_PREFIX + + arguments: + opt (namespace) -- Commandline arguments for this run + """ + if opt.resume.startswith(WANDB_ARTIFACT_PREFIX): + model_artifact = wandb.use_artifact(remove_prefix(opt.resume, WANDB_ARTIFACT_PREFIX) + ":latest") + assert model_artifact is not None, 'Error: W&B model artifact doesn\'t exist' + modeldir = model_artifact.download() + # epochs_trained = model_artifact.metadata.get('epochs_trained') + total_epochs = model_artifact.metadata.get('total_epochs') + is_finished = total_epochs is None + assert not is_finished, 'training is finished, can only resume incomplete runs.' + return modeldir, model_artifact + return None, None + + def log_model(self, path, opt, epoch, fitness_score, best_model=False): + """ + Log the model checkpoint as W&B artifact + + arguments: + path (Path) -- Path of directory containing the checkpoints + opt (namespace) -- Command line arguments for this run + epoch (int) -- Current epoch number + fitness_score (float) -- fitness score for current epoch + best_model (boolean) -- Boolean representing if the current checkpoint is the best yet. + """ + model_artifact = wandb.Artifact('run_' + wandb.run.id + '_model', + type='model', + metadata={ + 'original_url': str(path), + 'epochs_trained': epoch + 1, + 'save period': opt.save_period, + 'project': opt.project, + 'total_epochs': opt.epochs, + 'fitness_score': fitness_score}) + model_artifact.add_file(str(path / 'last.pt'), name='last.pt') + wandb.log_artifact(model_artifact, + aliases=['latest', 'last', 'epoch ' + str(self.current_epoch), 'best' if best_model else '']) + LOGGER.info(f"Saving model artifact on epoch {epoch + 1}") + + def log_dataset_artifact(self, data_file, single_cls, project, overwrite_config=False): + """ + Log the dataset as W&B artifact and return the new data file with W&B links + + arguments: + data_file (str) -- the .yaml file with information about the dataset like - path, classes etc. + single_class (boolean) -- train multi-class data as single-class + project (str) -- project name. Used to construct the artifact path + overwrite_config (boolean) -- overwrites the data.yaml file if set to true otherwise creates a new + file with _wandb postfix. Eg -> data_wandb.yaml + + returns: + the new .yaml file with artifact links. it can be used to start training directly from artifacts + """ + upload_dataset = self.wandb_run.config.upload_dataset + log_val_only = isinstance(upload_dataset, str) and upload_dataset == 'val' + self.data_dict = check_dataset(data_file) # parse and check + data = dict(self.data_dict) + nc, names = (1, ['item']) if single_cls else (int(data['nc']), data['names']) + names = {k: v for k, v in enumerate(names)} # to index dictionary + + # log train set + if not log_val_only: + self.train_artifact = self.create_dataset_table(LoadImagesAndLabels(data['train'], rect=True, batch_size=1), + names, + name='train') if data.get('train') else None + if data.get('train'): + data['train'] = WANDB_ARTIFACT_PREFIX + str(Path(project) / 'train') + + self.val_artifact = self.create_dataset_table( + LoadImagesAndLabels(data['val'], rect=True, batch_size=1), names, name='val') if data.get('val') else None + if data.get('val'): + data['val'] = WANDB_ARTIFACT_PREFIX + str(Path(project) / 'val') + + path = Path(data_file) + # create a _wandb.yaml file with artifacts links if both train and test set are logged + if not log_val_only: + path = (path.stem if overwrite_config else path.stem + '_wandb') + '.yaml' # updated data.yaml path + path = ROOT / 'data' / path + data.pop('download', None) + data.pop('path', None) + with open(path, 'w') as f: + yaml.safe_dump(data, f) + LOGGER.info(f"Created dataset config file {path}") + + if self.job_type == 'Training': # builds correct artifact pipeline graph + if not log_val_only: + self.wandb_run.log_artifact( + self.train_artifact) # calling use_artifact downloads the dataset. NOT NEEDED! + self.wandb_run.use_artifact(self.val_artifact) + self.val_artifact.wait() + self.val_table = self.val_artifact.get('val') + self.map_val_table_path() + else: + self.wandb_run.log_artifact(self.train_artifact) + self.wandb_run.log_artifact(self.val_artifact) + return path + + def map_val_table_path(self): + """ + Map the validation dataset Table like name of file -> it's id in the W&B Table. + Useful for - referencing artifacts for evaluation. + """ + self.val_table_path_map = {} + LOGGER.info("Mapping dataset") + for i, data in enumerate(tqdm(self.val_table.data)): + self.val_table_path_map[data[3]] = data[0] + + def create_dataset_table(self, dataset: LoadImagesAndLabels, class_to_id: Dict[int, str], name: str = 'dataset'): + """ + Create and return W&B artifact containing W&B Table of the dataset. + + arguments: + dataset -- instance of LoadImagesAndLabels class used to iterate over the data to build Table + class_to_id -- hash map that maps class ids to labels + name -- name of the artifact + + returns: + dataset artifact to be logged or used + """ + # TODO: Explore multiprocessing to slpit this loop parallely| This is essential for speeding up the the logging + artifact = wandb.Artifact(name=name, type="dataset") + img_files = tqdm([dataset.path]) if isinstance(dataset.path, str) and Path(dataset.path).is_dir() else None + img_files = tqdm(dataset.im_files) if not img_files else img_files + for img_file in img_files: + if Path(img_file).is_dir(): + artifact.add_dir(img_file, name='data/images') + labels_path = 'labels'.join(dataset.path.rsplit('images', 1)) + artifact.add_dir(labels_path, name='data/labels') + else: + artifact.add_file(img_file, name='data/images/' + Path(img_file).name) + label_file = Path(img2label_paths([img_file])[0]) + artifact.add_file(str(label_file), name='data/labels/' + + label_file.name) if label_file.exists() else None + table = wandb.Table(columns=["id", "train_image", "Classes", "name"]) + class_set = wandb.Classes([{'id': id, 'name': name} for id, name in class_to_id.items()]) + for si, (img, labels, paths, shapes) in enumerate(tqdm(dataset)): + box_data, img_classes = [], {} + for cls, *xywh in labels[:, 1:].tolist(): + cls = int(cls) + box_data.append({ + "position": { + "middle": [xywh[0], xywh[1]], + "width": xywh[2], + "height": xywh[3]}, + "class_id": cls, + "box_caption": "%s" % (class_to_id[cls])}) + img_classes[cls] = class_to_id[cls] + boxes = {"ground_truth": {"box_data": box_data, "class_labels": class_to_id}} # inference-space + table.add_data(si, wandb.Image(paths, classes=class_set, boxes=boxes), list(img_classes.values()), + Path(paths).name) + artifact.add(table, name) + return artifact + + def log_training_progress(self, predn, path, names): + """ + Build evaluation Table. Uses reference from validation dataset table. + + arguments: + predn (list): list of predictions in the native space in the format - [xmin, ymin, xmax, ymax, confidence, class] + path (str): local path of the current evaluation image + names (dict(int, str)): hash map that maps class ids to labels + """ + class_set = wandb.Classes([{'id': id, 'name': name} for id, name in names.items()]) + box_data = [] + avg_conf_per_class = [0] * len(self.data_dict['names']) + pred_class_count = {} + for *xyxy, conf, cls in predn.tolist(): + if conf >= 0.25: + cls = int(cls) + box_data.append({ + "position": { + "minX": xyxy[0], + "minY": xyxy[1], + "maxX": xyxy[2], + "maxY": xyxy[3]}, + "class_id": cls, + "box_caption": f"{names[cls]} {conf:.3f}", + "scores": { + "class_score": conf}, + "domain": "pixel"}) + avg_conf_per_class[cls] += conf + + if cls in pred_class_count: + pred_class_count[cls] += 1 + else: + pred_class_count[cls] = 1 + + for pred_class in pred_class_count.keys(): + avg_conf_per_class[pred_class] = avg_conf_per_class[pred_class] / pred_class_count[pred_class] + + boxes = {"predictions": {"box_data": box_data, "class_labels": names}} # inference-space + id = self.val_table_path_map[Path(path).name] + self.result_table.add_data(self.current_epoch, id, self.val_table.data[id][1], + wandb.Image(self.val_table.data[id][1], boxes=boxes, classes=class_set), + *avg_conf_per_class) + + def val_one_image(self, pred, predn, path, names, im): + """ + Log validation data for one image. updates the result Table if validation dataset is uploaded and log bbox media panel + + arguments: + pred (list): list of scaled predictions in the format - [xmin, ymin, xmax, ymax, confidence, class] + predn (list): list of predictions in the native space - [xmin, ymin, xmax, ymax, confidence, class] + path (str): local path of the current evaluation image + """ + if self.val_table and self.result_table: # Log Table if Val dataset is uploaded as artifact + self.log_training_progress(predn, path, names) + + if len(self.bbox_media_panel_images) < self.max_imgs_to_log and self.current_epoch > 0: + if self.current_epoch % self.bbox_interval == 0: + box_data = [{ + "position": { + "minX": xyxy[0], + "minY": xyxy[1], + "maxX": xyxy[2], + "maxY": xyxy[3]}, + "class_id": int(cls), + "box_caption": f"{names[int(cls)]} {conf:.3f}", + "scores": { + "class_score": conf}, + "domain": "pixel"} for *xyxy, conf, cls in pred.tolist()] + boxes = {"predictions": {"box_data": box_data, "class_labels": names}} # inference-space + self.bbox_media_panel_images.append(wandb.Image(im, boxes=boxes, caption=path.name)) + + def log(self, log_dict): + """ + save the metrics to the logging dictionary + + arguments: + log_dict (Dict) -- metrics/media to be logged in current step + """ + if self.wandb_run: + for key, value in log_dict.items(): + self.log_dict[key] = value + + def end_epoch(self, best_result=False): + """ + commit the log_dict, model artifacts and Tables to W&B and flush the log_dict. + + arguments: + best_result (boolean): Boolean representing if the result of this evaluation is best or not + """ + if self.wandb_run: + with all_logging_disabled(): + if self.bbox_media_panel_images: + self.log_dict["BoundingBoxDebugger"] = self.bbox_media_panel_images + try: + wandb.log(self.log_dict) + except BaseException as e: + LOGGER.info( + f"An error occurred in wandb logger. The training will proceed without interruption. More info\n{e}" + ) + self.wandb_run.finish() + self.wandb_run = None + + self.log_dict = {} + self.bbox_media_panel_images = [] + if self.result_artifact: + self.result_artifact.add(self.result_table, 'result') + wandb.log_artifact(self.result_artifact, + aliases=[ + 'latest', 'last', 'epoch ' + str(self.current_epoch), + ('best' if best_result else '')]) + + wandb.log({"evaluation": self.result_table}) + columns = ["epoch", "id", "ground truth", "prediction"] + columns.extend(self.data_dict['names']) + self.result_table = wandb.Table(columns) + self.result_artifact = wandb.Artifact("run_" + wandb.run.id + "_progress", "evaluation") + + def finish_run(self): + """ + Log metrics if any and finish the current W&B run + """ + if self.wandb_run: + if self.log_dict: + with all_logging_disabled(): + wandb.log(self.log_dict) + wandb.run.finish() + + +@contextmanager +def all_logging_disabled(highest_level=logging.CRITICAL): + """ source - https://gist.github.com/simon-weber/7853144 + A context manager that will prevent any logging messages triggered during the body from being processed. + :param highest_level: the maximum logging level in use. + This would only need to be changed if a custom level greater than CRITICAL is defined. + """ + previous_level = logging.root.manager.disable + logging.disable(highest_level) + try: + yield + finally: + logging.disable(previous_level) diff --git a/yolov5/utils/loss.py b/yolov5/utils/loss.py new file mode 100644 index 0000000..9b9c3d9 --- /dev/null +++ b/yolov5/utils/loss.py @@ -0,0 +1,234 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Loss functions +""" + +import torch +import torch.nn as nn + +from utils.metrics import bbox_iou +from utils.torch_utils import de_parallel + + +def smooth_BCE(eps=0.1): # https://github.com/ultralytics/yolov3/issues/238#issuecomment-598028441 + # return positive, negative label smoothing BCE targets + return 1.0 - 0.5 * eps, 0.5 * eps + + +class BCEBlurWithLogitsLoss(nn.Module): + # BCEwithLogitLoss() with reduced missing label effects. + def __init__(self, alpha=0.05): + super().__init__() + self.loss_fcn = nn.BCEWithLogitsLoss(reduction='none') # must be nn.BCEWithLogitsLoss() + self.alpha = alpha + + def forward(self, pred, true): + loss = self.loss_fcn(pred, true) + pred = torch.sigmoid(pred) # prob from logits + dx = pred - true # reduce only missing label effects + # dx = (pred - true).abs() # reduce missing label and false label effects + alpha_factor = 1 - torch.exp((dx - 1) / (self.alpha + 1e-4)) + loss *= alpha_factor + return loss.mean() + + +class FocalLoss(nn.Module): + # Wraps focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5) + def __init__(self, loss_fcn, gamma=1.5, alpha=0.25): + super().__init__() + self.loss_fcn = loss_fcn # must be nn.BCEWithLogitsLoss() + self.gamma = gamma + self.alpha = alpha + self.reduction = loss_fcn.reduction + self.loss_fcn.reduction = 'none' # required to apply FL to each element + + def forward(self, pred, true): + loss = self.loss_fcn(pred, true) + # p_t = torch.exp(-loss) + # loss *= self.alpha * (1.000001 - p_t) ** self.gamma # non-zero power for gradient stability + + # TF implementation https://github.com/tensorflow/addons/blob/v0.7.1/tensorflow_addons/losses/focal_loss.py + pred_prob = torch.sigmoid(pred) # prob from logits + p_t = true * pred_prob + (1 - true) * (1 - pred_prob) + alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha) + modulating_factor = (1.0 - p_t) ** self.gamma + loss *= alpha_factor * modulating_factor + + if self.reduction == 'mean': + return loss.mean() + elif self.reduction == 'sum': + return loss.sum() + else: # 'none' + return loss + + +class QFocalLoss(nn.Module): + # Wraps Quality focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5) + def __init__(self, loss_fcn, gamma=1.5, alpha=0.25): + super().__init__() + self.loss_fcn = loss_fcn # must be nn.BCEWithLogitsLoss() + self.gamma = gamma + self.alpha = alpha + self.reduction = loss_fcn.reduction + self.loss_fcn.reduction = 'none' # required to apply FL to each element + + def forward(self, pred, true): + loss = self.loss_fcn(pred, true) + + pred_prob = torch.sigmoid(pred) # prob from logits + alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha) + modulating_factor = torch.abs(true - pred_prob) ** self.gamma + loss *= alpha_factor * modulating_factor + + if self.reduction == 'mean': + return loss.mean() + elif self.reduction == 'sum': + return loss.sum() + else: # 'none' + return loss + + +class ComputeLoss: + sort_obj_iou = False + + # Compute losses + def __init__(self, model, autobalance=False): + device = next(model.parameters()).device # get model device + h = model.hyp # hyperparameters + + # Define criteria + BCEcls = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['cls_pw']], device=device)) + BCEobj = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['obj_pw']], device=device)) + + # Class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3 + self.cp, self.cn = smooth_BCE(eps=h.get('label_smoothing', 0.0)) # positive, negative BCE targets + + # Focal loss + g = h['fl_gamma'] # focal loss gamma + if g > 0: + BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g) + + m = de_parallel(model).model[-1] # Detect() module + self.balance = {3: [4.0, 1.0, 0.4]}.get(m.nl, [4.0, 1.0, 0.25, 0.06, 0.02]) # P3-P7 + self.ssi = list(m.stride).index(16) if autobalance else 0 # stride 16 index + self.BCEcls, self.BCEobj, self.gr, self.hyp, self.autobalance = BCEcls, BCEobj, 1.0, h, autobalance + self.na = m.na # number of anchors + self.nc = m.nc # number of classes + self.nl = m.nl # number of layers + self.anchors = m.anchors + self.device = device + + def __call__(self, p, targets): # predictions, targets + lcls = torch.zeros(1, device=self.device) # class loss + lbox = torch.zeros(1, device=self.device) # box loss + lobj = torch.zeros(1, device=self.device) # object loss + tcls, tbox, indices, anchors = self.build_targets(p, targets) # targets + + # Losses + for i, pi in enumerate(p): # layer index, layer predictions + b, a, gj, gi = indices[i] # image, anchor, gridy, gridx + tobj = torch.zeros(pi.shape[:4], dtype=pi.dtype, device=self.device) # target obj + + n = b.shape[0] # number of targets + if n: + # pxy, pwh, _, pcls = pi[b, a, gj, gi].tensor_split((2, 4, 5), dim=1) # faster, requires torch 1.8.0 + pxy, pwh, _, pcls = pi[b, a, gj, gi].split((2, 2, 1, self.nc), 1) # target-subset of predictions + + # Regression + pxy = pxy.sigmoid() * 2 - 0.5 + pwh = (pwh.sigmoid() * 2) ** 2 * anchors[i] + pbox = torch.cat((pxy, pwh), 1) # predicted box + iou = bbox_iou(pbox, tbox[i], CIoU=True).squeeze() # iou(prediction, target) + lbox += (1.0 - iou).mean() # iou loss + + # Objectness + iou = iou.detach().clamp(0).type(tobj.dtype) + if self.sort_obj_iou: + j = iou.argsort() + b, a, gj, gi, iou = b[j], a[j], gj[j], gi[j], iou[j] + if self.gr < 1: + iou = (1.0 - self.gr) + self.gr * iou + tobj[b, a, gj, gi] = iou # iou ratio + + # Classification + if self.nc > 1: # cls loss (only if multiple classes) + t = torch.full_like(pcls, self.cn, device=self.device) # targets + t[range(n), tcls[i]] = self.cp + lcls += self.BCEcls(pcls, t) # BCE + + # Append targets to text file + # with open('targets.txt', 'a') as file: + # [file.write('%11.5g ' * 4 % tuple(x) + '\n') for x in torch.cat((txy[i], twh[i]), 1)] + + obji = self.BCEobj(pi[..., 4], tobj) + lobj += obji * self.balance[i] # obj loss + if self.autobalance: + self.balance[i] = self.balance[i] * 0.9999 + 0.0001 / obji.detach().item() + + if self.autobalance: + self.balance = [x / self.balance[self.ssi] for x in self.balance] + lbox *= self.hyp['box'] + lobj *= self.hyp['obj'] + lcls *= self.hyp['cls'] + bs = tobj.shape[0] # batch size + + return (lbox + lobj + lcls) * bs, torch.cat((lbox, lobj, lcls)).detach() + + def build_targets(self, p, targets): + # Build targets for compute_loss(), input targets(image,class,x,y,w,h) + na, nt = self.na, targets.shape[0] # number of anchors, targets + tcls, tbox, indices, anch = [], [], [], [] + gain = torch.ones(7, device=self.device) # normalized to gridspace gain + ai = torch.arange(na, device=self.device).float().view(na, 1).repeat(1, nt) # same as .repeat_interleave(nt) + targets = torch.cat((targets.repeat(na, 1, 1), ai[..., None]), 2) # append anchor indices + + g = 0.5 # bias + off = torch.tensor( + [ + [0, 0], + [1, 0], + [0, 1], + [-1, 0], + [0, -1], # j,k,l,m + # [1, 1], [1, -1], [-1, 1], [-1, -1], # jk,jm,lk,lm + ], + device=self.device).float() * g # offsets + + for i in range(self.nl): + anchors, shape = self.anchors[i], p[i].shape + gain[2:6] = torch.tensor(shape)[[3, 2, 3, 2]] # xyxy gain + + # Match targets to anchors + t = targets * gain # shape(3,n,7) + if nt: + # Matches + r = t[..., 4:6] / anchors[:, None] # wh ratio + j = torch.max(r, 1 / r).max(2)[0] < self.hyp['anchor_t'] # compare + # j = wh_iou(anchors, t[:, 4:6]) > model.hyp['iou_t'] # iou(3,n)=wh_iou(anchors(3,2), gwh(n,2)) + t = t[j] # filter + + # Offsets + gxy = t[:, 2:4] # grid xy + gxi = gain[[2, 3]] - gxy # inverse + j, k = ((gxy % 1 < g) & (gxy > 1)).T + l, m = ((gxi % 1 < g) & (gxi > 1)).T + j = torch.stack((torch.ones_like(j), j, k, l, m)) + t = t.repeat((5, 1, 1))[j] + offsets = (torch.zeros_like(gxy)[None] + off[:, None])[j] + else: + t = targets[0] + offsets = 0 + + # Define + bc, gxy, gwh, a = t.chunk(4, 1) # (image, class), grid xy, grid wh, anchors + a, (b, c) = a.long().view(-1), bc.long().T # anchors, image, class + gij = (gxy - offsets).long() + gi, gj = gij.T # grid indices + + # Append + indices.append((b, a, gj.clamp_(0, shape[2] - 1), gi.clamp_(0, shape[3] - 1))) # image, anchor, grid + tbox.append(torch.cat((gxy - gij, gwh), 1)) # box + anch.append(anchors[a]) # anchors + tcls.append(c) # class + + return tcls, tbox, indices, anch diff --git a/yolov5/utils/metrics.py b/yolov5/utils/metrics.py new file mode 100644 index 0000000..e17747b --- /dev/null +++ b/yolov5/utils/metrics.py @@ -0,0 +1,355 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Model validation metrics +""" + +import math +import warnings +from pathlib import Path + +import matplotlib.pyplot as plt +import numpy as np +import torch + + +def fitness(x): + # Model fitness as a weighted combination of metrics + w = [0.0, 0.0, 0.1, 0.9] # weights for [P, R, mAP@0.5, mAP@0.5:0.95] + return (x[:, :4] * w).sum(1) + + +def smooth(y, f=0.05): + # Box filter of fraction f + nf = round(len(y) * f * 2) // 2 + 1 # number of filter elements (must be odd) + p = np.ones(nf // 2) # ones padding + yp = np.concatenate((p * y[0], y, p * y[-1]), 0) # y padded + return np.convolve(yp, np.ones(nf) / nf, mode='valid') # y-smoothed + + +def ap_per_class(tp, conf, pred_cls, target_cls, plot=False, save_dir='.', names=(), eps=1e-16): + """ Compute the average precision, given the recall and precision curves. + Source: https://github.com/rafaelpadilla/Object-Detection-Metrics. + # Arguments + tp: True positives (nparray, nx1 or nx10). + conf: Objectness value from 0-1 (nparray). + pred_cls: Predicted object classes (nparray). + target_cls: True object classes (nparray). + plot: Plot precision-recall curve at mAP@0.5 + save_dir: Plot save directory + # Returns + The average precision as computed in py-faster-rcnn. + """ + + # Sort by objectness + i = np.argsort(-conf) + tp, conf, pred_cls = tp[i], conf[i], pred_cls[i] + + # Find unique classes + unique_classes, nt = np.unique(target_cls, return_counts=True) + nc = unique_classes.shape[0] # number of classes, number of detections + + # Create Precision-Recall curve and compute AP for each class + px, py = np.linspace(0, 1, 1000), [] # for plotting + ap, p, r = np.zeros((nc, tp.shape[1])), np.zeros((nc, 1000)), np.zeros((nc, 1000)) + for ci, c in enumerate(unique_classes): + i = pred_cls == c + n_l = nt[ci] # number of labels + n_p = i.sum() # number of predictions + if n_p == 0 or n_l == 0: + continue + + # Accumulate FPs and TPs + fpc = (1 - tp[i]).cumsum(0) + tpc = tp[i].cumsum(0) + + # Recall + recall = tpc / (n_l + eps) # recall curve + r[ci] = np.interp(-px, -conf[i], recall[:, 0], left=0) # negative x, xp because xp decreases + + # Precision + precision = tpc / (tpc + fpc) # precision curve + p[ci] = np.interp(-px, -conf[i], precision[:, 0], left=1) # p at pr_score + + # AP from recall-precision curve + for j in range(tp.shape[1]): + ap[ci, j], mpre, mrec = compute_ap(recall[:, j], precision[:, j]) + if plot and j == 0: + py.append(np.interp(px, mrec, mpre)) # precision at mAP@0.5 + + # Compute F1 (harmonic mean of precision and recall) + f1 = 2 * p * r / (p + r + eps) + names = [v for k, v in names.items() if k in unique_classes] # list: only classes that have data + names = dict(enumerate(names)) # to dict + if plot: + plot_pr_curve(px, py, ap, Path(save_dir) / 'PR_curve.png', names) + plot_mc_curve(px, f1, Path(save_dir) / 'F1_curve.png', names, ylabel='F1') + plot_mc_curve(px, p, Path(save_dir) / 'P_curve.png', names, ylabel='Precision') + plot_mc_curve(px, r, Path(save_dir) / 'R_curve.png', names, ylabel='Recall') + + i = smooth(f1.mean(0), 0.1).argmax() # max F1 index + p, r, f1 = p[:, i], r[:, i], f1[:, i] + tp = (r * nt).round() # true positives + fp = (tp / (p + eps) - tp).round() # false positives + return tp, fp, p, r, f1, ap, unique_classes.astype(int) + + +def compute_ap(recall, precision): + """ Compute the average precision, given the recall and precision curves + # Arguments + recall: The recall curve (list) + precision: The precision curve (list) + # Returns + Average precision, precision curve, recall curve + """ + + # Append sentinel values to beginning and end + mrec = np.concatenate(([0.0], recall, [1.0])) + mpre = np.concatenate(([1.0], precision, [0.0])) + + # Compute the precision envelope + mpre = np.flip(np.maximum.accumulate(np.flip(mpre))) + + # Integrate area under curve + method = 'interp' # methods: 'continuous', 'interp' + if method == 'interp': + x = np.linspace(0, 1, 101) # 101-point interp (COCO) + ap = np.trapz(np.interp(x, mrec, mpre), x) # integrate + else: # 'continuous' + i = np.where(mrec[1:] != mrec[:-1])[0] # points where x axis (recall) changes + ap = np.sum((mrec[i + 1] - mrec[i]) * mpre[i + 1]) # area under curve + + return ap, mpre, mrec + + +class ConfusionMatrix: + # Updated version of https://github.com/kaanakan/object_detection_confusion_matrix + def __init__(self, nc, conf=0.25, iou_thres=0.45): + self.matrix = np.zeros((nc + 1, nc + 1)) + self.nc = nc # number of classes + self.conf = conf + self.iou_thres = iou_thres + + def process_batch(self, detections, labels): + """ + Return intersection-over-union (Jaccard index) of boxes. + Both sets of boxes are expected to be in (x1, y1, x2, y2) format. + Arguments: + detections (Array[N, 6]), x1, y1, x2, y2, conf, class + labels (Array[M, 5]), class, x1, y1, x2, y2 + Returns: + None, updates confusion matrix accordingly + """ + detections = detections[detections[:, 4] > self.conf] + gt_classes = labels[:, 0].int() + detection_classes = detections[:, 5].int() + iou = box_iou(labels[:, 1:], detections[:, :4]) + + x = torch.where(iou > self.iou_thres) + if x[0].shape[0]: + matches = torch.cat((torch.stack(x, 1), iou[x[0], x[1]][:, None]), 1).cpu().numpy() + if x[0].shape[0] > 1: + matches = matches[matches[:, 2].argsort()[::-1]] + matches = matches[np.unique(matches[:, 1], return_index=True)[1]] + matches = matches[matches[:, 2].argsort()[::-1]] + matches = matches[np.unique(matches[:, 0], return_index=True)[1]] + else: + matches = np.zeros((0, 3)) + + n = matches.shape[0] > 0 + m0, m1, _ = matches.transpose().astype(int) + for i, gc in enumerate(gt_classes): + j = m0 == i + if n and sum(j) == 1: + self.matrix[detection_classes[m1[j]], gc] += 1 # correct + else: + self.matrix[self.nc, gc] += 1 # background FP + + if n: + for i, dc in enumerate(detection_classes): + if not any(m1 == i): + self.matrix[dc, self.nc] += 1 # background FN + + def matrix(self): + return self.matrix + + def tp_fp(self): + tp = self.matrix.diagonal() # true positives + fp = self.matrix.sum(1) - tp # false positives + # fn = self.matrix.sum(0) - tp # false negatives (missed detections) + return tp[:-1], fp[:-1] # remove background class + + def plot(self, normalize=True, save_dir='', names=()): + try: + import seaborn as sn + + array = self.matrix / ((self.matrix.sum(0).reshape(1, -1) + 1E-9) if normalize else 1) # normalize columns + array[array < 0.005] = np.nan # don't annotate (would appear as 0.00) + + fig = plt.figure(figsize=(12, 9), tight_layout=True) + nc, nn = self.nc, len(names) # number of classes, names + sn.set(font_scale=1.0 if nc < 50 else 0.8) # for label size + labels = (0 < nn < 99) and (nn == nc) # apply names to ticklabels + with warnings.catch_warnings(): + warnings.simplefilter('ignore') # suppress empty matrix RuntimeWarning: All-NaN slice encountered + sn.heatmap(array, + annot=nc < 30, + annot_kws={ + "size": 8}, + cmap='Blues', + fmt='.2f', + square=True, + vmin=0.0, + xticklabels=names + ['background FP'] if labels else "auto", + yticklabels=names + ['background FN'] if labels else "auto").set_facecolor((1, 1, 1)) + fig.axes[0].set_xlabel('True') + fig.axes[0].set_ylabel('Predicted') + fig.savefig(Path(save_dir) / 'confusion_matrix.png', dpi=250) + plt.close() + except Exception as e: + print(f'WARNING: ConfusionMatrix plot failure: {e}') + + def print(self): + for i in range(self.nc + 1): + print(' '.join(map(str, self.matrix[i]))) + + +def bbox_iou(box1, box2, xywh=True, GIoU=False, DIoU=False, CIoU=False, eps=1e-7): + # Returns Intersection over Union (IoU) of box1(1,4) to box2(n,4) + + # Get the coordinates of bounding boxes + if xywh: # transform from xywh to xyxy + (x1, y1, w1, h1), (x2, y2, w2, h2) = box1.chunk(4, 1), box2.chunk(4, 1) + w1_, h1_, w2_, h2_ = w1 / 2, h1 / 2, w2 / 2, h2 / 2 + b1_x1, b1_x2, b1_y1, b1_y2 = x1 - w1_, x1 + w1_, y1 - h1_, y1 + h1_ + b2_x1, b2_x2, b2_y1, b2_y2 = x2 - w2_, x2 + w2_, y2 - h2_, y2 + h2_ + else: # x1, y1, x2, y2 = box1 + b1_x1, b1_y1, b1_x2, b1_y2 = box1.chunk(4, 1) + b2_x1, b2_y1, b2_x2, b2_y2 = box2.chunk(4, 1) + w1, h1 = b1_x2 - b1_x1, b1_y2 - b1_y1 + eps + w2, h2 = b2_x2 - b2_x1, b2_y2 - b2_y1 + eps + + # Intersection area + inter = (torch.min(b1_x2, b2_x2) - torch.max(b1_x1, b2_x1)).clamp(0) * \ + (torch.min(b1_y2, b2_y2) - torch.max(b1_y1, b2_y1)).clamp(0) + + # Union Area + union = w1 * h1 + w2 * h2 - inter + eps + + # IoU + iou = inter / union + if CIoU or DIoU or GIoU: + cw = torch.max(b1_x2, b2_x2) - torch.min(b1_x1, b2_x1) # convex (smallest enclosing box) width + ch = torch.max(b1_y2, b2_y2) - torch.min(b1_y1, b2_y1) # convex height + if CIoU or DIoU: # Distance or Complete IoU https://arxiv.org/abs/1911.08287v1 + c2 = cw ** 2 + ch ** 2 + eps # convex diagonal squared + rho2 = ((b2_x1 + b2_x2 - b1_x1 - b1_x2) ** 2 + (b2_y1 + b2_y2 - b1_y1 - b1_y2) ** 2) / 4 # center dist ** 2 + if CIoU: # https://github.com/Zzh-tju/DIoU-SSD-pytorch/blob/master/utils/box/box_utils.py#L47 + v = (4 / math.pi ** 2) * torch.pow(torch.atan(w2 / h2) - torch.atan(w1 / h1), 2) + with torch.no_grad(): + alpha = v / (v - iou + (1 + eps)) + return iou - (rho2 / c2 + v * alpha) # CIoU + return iou - rho2 / c2 # DIoU + c_area = cw * ch + eps # convex area + return iou - (c_area - union) / c_area # GIoU https://arxiv.org/pdf/1902.09630.pdf + return iou # IoU + + +def box_area(box): + # box = xyxy(4,n) + return (box[2] - box[0]) * (box[3] - box[1]) + + +def box_iou(box1, box2): + # https://github.com/pytorch/vision/blob/master/torchvision/ops/boxes.py + """ + Return intersection-over-union (Jaccard index) of boxes. + Both sets of boxes are expected to be in (x1, y1, x2, y2) format. + Arguments: + box1 (Tensor[N, 4]) + box2 (Tensor[M, 4]) + Returns: + iou (Tensor[N, M]): the NxM matrix containing the pairwise + IoU values for every element in boxes1 and boxes2 + """ + + # inter(N,M) = (rb(N,M,2) - lt(N,M,2)).clamp(0).prod(2) + (a1, a2), (b1, b2) = box1[:, None].chunk(2, 2), box2.chunk(2, 1) + inter = (torch.min(a2, b2) - torch.max(a1, b1)).clamp(0).prod(2) + + # IoU = inter / (area1 + area2 - inter) + return inter / (box_area(box1.T)[:, None] + box_area(box2.T) - inter) + + +def bbox_ioa(box1, box2, eps=1E-7): + """ Returns the intersection over box2 area given box1, box2. Boxes are x1y1x2y2 + box1: np.array of shape(4) + box2: np.array of shape(nx4) + returns: np.array of shape(n) + """ + + # Get the coordinates of bounding boxes + b1_x1, b1_y1, b1_x2, b1_y2 = box1 + b2_x1, b2_y1, b2_x2, b2_y2 = box2.T + + # Intersection area + inter_area = (np.minimum(b1_x2, b2_x2) - np.maximum(b1_x1, b2_x1)).clip(0) * \ + (np.minimum(b1_y2, b2_y2) - np.maximum(b1_y1, b2_y1)).clip(0) + + # box2 area + box2_area = (b2_x2 - b2_x1) * (b2_y2 - b2_y1) + eps + + # Intersection over box2 area + return inter_area / box2_area + + +def wh_iou(wh1, wh2): + # Returns the nxm IoU matrix. wh1 is nx2, wh2 is mx2 + wh1 = wh1[:, None] # [N,1,2] + wh2 = wh2[None] # [1,M,2] + inter = torch.min(wh1, wh2).prod(2) # [N,M] + return inter / (wh1.prod(2) + wh2.prod(2) - inter) # iou = inter / (area1 + area2 - inter) + + +# Plots ---------------------------------------------------------------------------------------------------------------- + + +def plot_pr_curve(px, py, ap, save_dir=Path('pr_curve.png'), names=()): + # Precision-recall curve + fig, ax = plt.subplots(1, 1, figsize=(9, 6), tight_layout=True) + py = np.stack(py, axis=1) + + if 0 < len(names) < 21: # display per-class legend if < 21 classes + for i, y in enumerate(py.T): + ax.plot(px, y, linewidth=1, label=f'{names[i]} {ap[i, 0]:.3f}') # plot(recall, precision) + else: + ax.plot(px, py, linewidth=1, color='grey') # plot(recall, precision) + + ax.plot(px, py.mean(1), linewidth=3, color='blue', label='all classes %.3f mAP@0.5' % ap[:, 0].mean()) + ax.set_xlabel('Recall') + ax.set_ylabel('Precision') + ax.set_xlim(0, 1) + ax.set_ylim(0, 1) + plt.legend(bbox_to_anchor=(1.04, 1), loc="upper left") + fig.savefig(save_dir, dpi=250) + plt.close() + + +def plot_mc_curve(px, py, save_dir=Path('mc_curve.png'), names=(), xlabel='Confidence', ylabel='Metric'): + # Metric-confidence curve + fig, ax = plt.subplots(1, 1, figsize=(9, 6), tight_layout=True) + + if 0 < len(names) < 21: # display per-class legend if < 21 classes + for i, y in enumerate(py): + ax.plot(px, y, linewidth=1, label=f'{names[i]}') # plot(confidence, metric) + else: + ax.plot(px, py.T, linewidth=1, color='grey') # plot(confidence, metric) + + y = smooth(py.mean(0), 0.05) + ax.plot(px, y, linewidth=3, color='blue', label=f'all classes {y.max():.2f} at {px[y.argmax()]:.3f}') + ax.set_xlabel(xlabel) + ax.set_ylabel(ylabel) + ax.set_xlim(0, 1) + ax.set_ylim(0, 1) + plt.legend(bbox_to_anchor=(1.04, 1), loc="upper left") + fig.savefig(save_dir, dpi=250) + plt.close() diff --git a/yolov5/utils/plots.py b/yolov5/utils/plots.py new file mode 100644 index 0000000..1bbb9c0 --- /dev/null +++ b/yolov5/utils/plots.py @@ -0,0 +1,489 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Plotting utils +""" + +import math +import os +from copy import copy +from pathlib import Path +from urllib.error import URLError + +import cv2 +import matplotlib +import matplotlib.pyplot as plt +import numpy as np +import pandas as pd +import seaborn as sn +import torch +from PIL import Image, ImageDraw, ImageFont + +from utils.general import (CONFIG_DIR, FONT, LOGGER, Timeout, check_font, check_requirements, clip_coords, + increment_path, is_ascii, threaded, try_except, xywh2xyxy, xyxy2xywh) +from utils.metrics import fitness + +# Settings +RANK = int(os.getenv('RANK', -1)) +matplotlib.rc('font', **{'size': 11}) +matplotlib.use('Agg') # for writing to files only + + +class Colors: + # Ultralytics color palette https://ultralytics.com/ + def __init__(self): + # hex = matplotlib.colors.TABLEAU_COLORS.values() + hexs = ('FF3838', 'FF9D97', 'FF701F', 'FFB21D', 'CFD231', '48F90A', '92CC17', '3DDB86', '1A9334', '00D4BB', + '2C99A8', '00C2FF', '344593', '6473FF', '0018EC', '8438FF', '520085', 'CB38FF', 'FF95C8', 'FF37C7') + self.palette = [self.hex2rgb(f'#{c}') for c in hexs] + self.n = len(self.palette) + + def __call__(self, i, bgr=False): + c = self.palette[int(i) % self.n] + return (c[2], c[1], c[0]) if bgr else c + + @staticmethod + def hex2rgb(h): # rgb order (PIL) + return tuple(int(h[1 + i:1 + i + 2], 16) for i in (0, 2, 4)) + + +colors = Colors() # create instance for 'from utils.plots import colors' + + +def check_pil_font(font=FONT, size=10): + # Return a PIL TrueType Font, downloading to CONFIG_DIR if necessary + font = Path(font) + font = font if font.exists() else (CONFIG_DIR / font.name) + try: + return ImageFont.truetype(str(font) if font.exists() else font.name, size) + except Exception: # download if missing + try: + check_font(font) + return ImageFont.truetype(str(font), size) + except TypeError: + check_requirements('Pillow>=8.4.0') # known issue https://github.com/ultralytics/yolov5/issues/5374 + except URLError: # not online + return ImageFont.load_default() + + +class Annotator: + # YOLOv5 Annotator for train/val mosaics and jpgs and detect/hub inference annotations + def __init__(self, im, line_width=None, font_size=None, font='Arial.ttf', pil=False, example='abc'): + assert im.data.contiguous, 'Image not contiguous. Apply np.ascontiguousarray(im) to Annotator() input images.' + non_ascii = not is_ascii(example) # non-latin labels, i.e. asian, arabic, cyrillic + self.pil = pil or non_ascii + if self.pil: # use PIL + self.im = im if isinstance(im, Image.Image) else Image.fromarray(im) + self.draw = ImageDraw.Draw(self.im) + self.font = check_pil_font(font='Arial.Unicode.ttf' if non_ascii else font, + size=font_size or max(round(sum(self.im.size) / 2 * 0.035), 12)) + else: # use cv2 + self.im = im + self.lw = line_width or max(round(sum(im.shape) / 2 * 0.003), 2) # line width + + def box_label(self, box, label='', color=(128, 128, 128), txt_color=(255, 255, 255)): + # Add one xyxy box to image with label + if self.pil or not is_ascii(label): + self.draw.rectangle(box, width=self.lw, outline=color) # box + if label: + w, h = self.font.getsize(label) # text width, height + outside = box[1] - h >= 0 # label fits outside box + self.draw.rectangle( + (box[0], box[1] - h if outside else box[1], box[0] + w + 1, + box[1] + 1 if outside else box[1] + h + 1), + fill=color, + ) + # self.draw.text((box[0], box[1]), label, fill=txt_color, font=self.font, anchor='ls') # for PIL>8.0 + self.draw.text((box[0], box[1] - h if outside else box[1]), label, fill=txt_color, font=self.font) + else: # cv2 + p1, p2 = (int(box[0]), int(box[1])), (int(box[2]), int(box[3])) + cv2.rectangle(self.im, p1, p2, color, thickness=self.lw, lineType=cv2.LINE_AA) + if label: + tf = max(self.lw - 1, 1) # font thickness + w, h = cv2.getTextSize(label, 0, fontScale=self.lw / 3, thickness=tf)[0] # text width, height + outside = p1[1] - h >= 3 + p2 = p1[0] + w, p1[1] - h - 3 if outside else p1[1] + h + 3 + cv2.rectangle(self.im, p1, p2, color, -1, cv2.LINE_AA) # filled + cv2.putText(self.im, + label, (p1[0], p1[1] - 2 if outside else p1[1] + h + 2), + 0, + self.lw / 3, + txt_color, + thickness=tf, + lineType=cv2.LINE_AA) + + def rectangle(self, xy, fill=None, outline=None, width=1): + # Add rectangle to image (PIL-only) + self.draw.rectangle(xy, fill, outline, width) + + def text(self, xy, text, txt_color=(255, 255, 255)): + # Add text to image (PIL-only) + w, h = self.font.getsize(text) # text width, height + self.draw.text((xy[0], xy[1] - h + 1), text, fill=txt_color, font=self.font) + + def result(self): + # Return annotated image as array + return np.asarray(self.im) + + +def feature_visualization(x, module_type, stage, n=32, save_dir=Path('runs/detect/exp')): + """ + x: Features to be visualized + module_type: Module type + stage: Module stage within model + n: Maximum number of feature maps to plot + save_dir: Directory to save results + """ + if 'Detect' not in module_type: + batch, channels, height, width = x.shape # batch, channels, height, width + if height > 1 and width > 1: + f = save_dir / f"stage{stage}_{module_type.split('.')[-1]}_features.png" # filename + + blocks = torch.chunk(x[0].cpu(), channels, dim=0) # select batch index 0, block by channels + n = min(n, channels) # number of plots + fig, ax = plt.subplots(math.ceil(n / 8), 8, tight_layout=True) # 8 rows x n/8 cols + ax = ax.ravel() + plt.subplots_adjust(wspace=0.05, hspace=0.05) + for i in range(n): + ax[i].imshow(blocks[i].squeeze()) # cmap='gray' + ax[i].axis('off') + + LOGGER.info(f'Saving {f}... ({n}/{channels})') + plt.savefig(f, dpi=300, bbox_inches='tight') + plt.close() + np.save(str(f.with_suffix('.npy')), x[0].cpu().numpy()) # npy save + + +def hist2d(x, y, n=100): + # 2d histogram used in labels.png and evolve.png + xedges, yedges = np.linspace(x.min(), x.max(), n), np.linspace(y.min(), y.max(), n) + hist, xedges, yedges = np.histogram2d(x, y, (xedges, yedges)) + xidx = np.clip(np.digitize(x, xedges) - 1, 0, hist.shape[0] - 1) + yidx = np.clip(np.digitize(y, yedges) - 1, 0, hist.shape[1] - 1) + return np.log(hist[xidx, yidx]) + + +def butter_lowpass_filtfilt(data, cutoff=1500, fs=50000, order=5): + from scipy.signal import butter, filtfilt + + # https://stackoverflow.com/questions/28536191/how-to-filter-smooth-with-scipy-numpy + def butter_lowpass(cutoff, fs, order): + nyq = 0.5 * fs + normal_cutoff = cutoff / nyq + return butter(order, normal_cutoff, btype='low', analog=False) + + b, a = butter_lowpass(cutoff, fs, order=order) + return filtfilt(b, a, data) # forward-backward filter + + +def output_to_target(output): + # Convert model output to target format [batch_id, class_id, x, y, w, h, conf] + targets = [] + for i, o in enumerate(output): + for *box, conf, cls in o.cpu().numpy(): + targets.append([i, cls, *list(*xyxy2xywh(np.array(box)[None])), conf]) + return np.array(targets) + + +@threaded +def plot_images(images, targets, paths=None, fname='images.jpg', names=None, max_size=1920, max_subplots=16): + # Plot image grid with labels + if isinstance(images, torch.Tensor): + images = images.cpu().float().numpy() + if isinstance(targets, torch.Tensor): + targets = targets.cpu().numpy() + if np.max(images[0]) <= 1: + images *= 255 # de-normalise (optional) + bs, _, h, w = images.shape # batch size, _, height, width + bs = min(bs, max_subplots) # limit plot images + ns = np.ceil(bs ** 0.5) # number of subplots (square) + + # Build Image + mosaic = np.full((int(ns * h), int(ns * w), 3), 255, dtype=np.uint8) # init + for i, im in enumerate(images): + if i == max_subplots: # if last batch has fewer images than we expect + break + x, y = int(w * (i // ns)), int(h * (i % ns)) # block origin + im = im.transpose(1, 2, 0) + mosaic[y:y + h, x:x + w, :] = im + + # Resize (optional) + scale = max_size / ns / max(h, w) + if scale < 1: + h = math.ceil(scale * h) + w = math.ceil(scale * w) + mosaic = cv2.resize(mosaic, tuple(int(x * ns) for x in (w, h))) + + # Annotate + fs = int((h + w) * ns * 0.01) # font size + annotator = Annotator(mosaic, line_width=round(fs / 10), font_size=fs, pil=True, example=names) + for i in range(i + 1): + x, y = int(w * (i // ns)), int(h * (i % ns)) # block origin + annotator.rectangle([x, y, x + w, y + h], None, (255, 255, 255), width=2) # borders + if paths: + annotator.text((x + 5, y + 5 + h), text=Path(paths[i]).name[:40], txt_color=(220, 220, 220)) # filenames + if len(targets) > 0: + ti = targets[targets[:, 0] == i] # image targets + boxes = xywh2xyxy(ti[:, 2:6]).T + classes = ti[:, 1].astype('int') + labels = ti.shape[1] == 6 # labels if no conf column + conf = None if labels else ti[:, 6] # check for confidence presence (label vs pred) + + if boxes.shape[1]: + if boxes.max() <= 1.01: # if normalized with tolerance 0.01 + boxes[[0, 2]] *= w # scale to pixels + boxes[[1, 3]] *= h + elif scale < 1: # absolute coords need scale if image scales + boxes *= scale + boxes[[0, 2]] += x + boxes[[1, 3]] += y + for j, box in enumerate(boxes.T.tolist()): + cls = classes[j] + color = colors(cls) + cls = names[cls] if names else cls + if labels or conf[j] > 0.25: # 0.25 conf thresh + label = f'{cls}' if labels else f'{cls} {conf[j]:.1f}' + annotator.box_label(box, label, color=color) + annotator.im.save(fname) # save + + +def plot_lr_scheduler(optimizer, scheduler, epochs=300, save_dir=''): + # Plot LR simulating training for full epochs + optimizer, scheduler = copy(optimizer), copy(scheduler) # do not modify originals + y = [] + for _ in range(epochs): + scheduler.step() + y.append(optimizer.param_groups[0]['lr']) + plt.plot(y, '.-', label='LR') + plt.xlabel('epoch') + plt.ylabel('LR') + plt.grid() + plt.xlim(0, epochs) + plt.ylim(0) + plt.savefig(Path(save_dir) / 'LR.png', dpi=200) + plt.close() + + +def plot_val_txt(): # from utils.plots import *; plot_val() + # Plot val.txt histograms + x = np.loadtxt('val.txt', dtype=np.float32) + box = xyxy2xywh(x[:, :4]) + cx, cy = box[:, 0], box[:, 1] + + fig, ax = plt.subplots(1, 1, figsize=(6, 6), tight_layout=True) + ax.hist2d(cx, cy, bins=600, cmax=10, cmin=0) + ax.set_aspect('equal') + plt.savefig('hist2d.png', dpi=300) + + fig, ax = plt.subplots(1, 2, figsize=(12, 6), tight_layout=True) + ax[0].hist(cx, bins=600) + ax[1].hist(cy, bins=600) + plt.savefig('hist1d.png', dpi=200) + + +def plot_targets_txt(): # from utils.plots import *; plot_targets_txt() + # Plot targets.txt histograms + x = np.loadtxt('targets.txt', dtype=np.float32).T + s = ['x targets', 'y targets', 'width targets', 'height targets'] + fig, ax = plt.subplots(2, 2, figsize=(8, 8), tight_layout=True) + ax = ax.ravel() + for i in range(4): + ax[i].hist(x[i], bins=100, label=f'{x[i].mean():.3g} +/- {x[i].std():.3g}') + ax[i].legend() + ax[i].set_title(s[i]) + plt.savefig('targets.jpg', dpi=200) + + +def plot_val_study(file='', dir='', x=None): # from utils.plots import *; plot_val_study() + # Plot file=study.txt generated by val.py (or plot all study*.txt in dir) + save_dir = Path(file).parent if file else Path(dir) + plot2 = False # plot additional results + if plot2: + ax = plt.subplots(2, 4, figsize=(10, 6), tight_layout=True)[1].ravel() + + fig2, ax2 = plt.subplots(1, 1, figsize=(8, 4), tight_layout=True) + # for f in [save_dir / f'study_coco_{x}.txt' for x in ['yolov5n6', 'yolov5s6', 'yolov5m6', 'yolov5l6', 'yolov5x6']]: + for f in sorted(save_dir.glob('study*.txt')): + y = np.loadtxt(f, dtype=np.float32, usecols=[0, 1, 2, 3, 7, 8, 9], ndmin=2).T + x = np.arange(y.shape[1]) if x is None else np.array(x) + if plot2: + s = ['P', 'R', 'mAP@.5', 'mAP@.5:.95', 't_preprocess (ms/img)', 't_inference (ms/img)', 't_NMS (ms/img)'] + for i in range(7): + ax[i].plot(x, y[i], '.-', linewidth=2, markersize=8) + ax[i].set_title(s[i]) + + j = y[3].argmax() + 1 + ax2.plot(y[5, 1:j], + y[3, 1:j] * 1E2, + '.-', + linewidth=2, + markersize=8, + label=f.stem.replace('study_coco_', '').replace('yolo', 'YOLO')) + + ax2.plot(1E3 / np.array([209, 140, 97, 58, 35, 18]), [34.6, 40.5, 43.0, 47.5, 49.7, 51.5], + 'k.-', + linewidth=2, + markersize=8, + alpha=.25, + label='EfficientDet') + + ax2.grid(alpha=0.2) + ax2.set_yticks(np.arange(20, 60, 5)) + ax2.set_xlim(0, 57) + ax2.set_ylim(25, 55) + ax2.set_xlabel('GPU Speed (ms/img)') + ax2.set_ylabel('COCO AP val') + ax2.legend(loc='lower right') + f = save_dir / 'study.png' + print(f'Saving {f}...') + plt.savefig(f, dpi=300) + + +@try_except # known issue https://github.com/ultralytics/yolov5/issues/5395 +@Timeout(30) # known issue https://github.com/ultralytics/yolov5/issues/5611 +def plot_labels(labels, names=(), save_dir=Path('')): + # plot dataset labels + LOGGER.info(f"Plotting labels to {save_dir / 'labels.jpg'}... ") + c, b = labels[:, 0], labels[:, 1:].transpose() # classes, boxes + nc = int(c.max() + 1) # number of classes + x = pd.DataFrame(b.transpose(), columns=['x', 'y', 'width', 'height']) + + # seaborn correlogram + sn.pairplot(x, corner=True, diag_kind='auto', kind='hist', diag_kws=dict(bins=50), plot_kws=dict(pmax=0.9)) + plt.savefig(save_dir / 'labels_correlogram.jpg', dpi=200) + plt.close() + + # matplotlib labels + matplotlib.use('svg') # faster + ax = plt.subplots(2, 2, figsize=(8, 8), tight_layout=True)[1].ravel() + y = ax[0].hist(c, bins=np.linspace(0, nc, nc + 1) - 0.5, rwidth=0.8) + try: # color histogram bars by class + [y[2].patches[i].set_color([x / 255 for x in colors(i)]) for i in range(nc)] # known issue #3195 + except Exception: + pass + ax[0].set_ylabel('instances') + if 0 < len(names) < 30: + ax[0].set_xticks(range(len(names))) + ax[0].set_xticklabels(names, rotation=90, fontsize=10) + else: + ax[0].set_xlabel('classes') + sn.histplot(x, x='x', y='y', ax=ax[2], bins=50, pmax=0.9) + sn.histplot(x, x='width', y='height', ax=ax[3], bins=50, pmax=0.9) + + # rectangles + labels[:, 1:3] = 0.5 # center + labels[:, 1:] = xywh2xyxy(labels[:, 1:]) * 2000 + img = Image.fromarray(np.ones((2000, 2000, 3), dtype=np.uint8) * 255) + for cls, *box in labels[:1000]: + ImageDraw.Draw(img).rectangle(box, width=1, outline=colors(cls)) # plot + ax[1].imshow(img) + ax[1].axis('off') + + for a in [0, 1, 2, 3]: + for s in ['top', 'right', 'left', 'bottom']: + ax[a].spines[s].set_visible(False) + + plt.savefig(save_dir / 'labels.jpg', dpi=200) + matplotlib.use('Agg') + plt.close() + + +def plot_evolve(evolve_csv='path/to/evolve.csv'): # from utils.plots import *; plot_evolve() + # Plot evolve.csv hyp evolution results + evolve_csv = Path(evolve_csv) + data = pd.read_csv(evolve_csv) + keys = [x.strip() for x in data.columns] + x = data.values + f = fitness(x) + j = np.argmax(f) # max fitness index + plt.figure(figsize=(10, 12), tight_layout=True) + matplotlib.rc('font', **{'size': 8}) + print(f'Best results from row {j} of {evolve_csv}:') + for i, k in enumerate(keys[7:]): + v = x[:, 7 + i] + mu = v[j] # best single result + plt.subplot(6, 5, i + 1) + plt.scatter(v, f, c=hist2d(v, f, 20), cmap='viridis', alpha=.8, edgecolors='none') + plt.plot(mu, f.max(), 'k+', markersize=15) + plt.title(f'{k} = {mu:.3g}', fontdict={'size': 9}) # limit to 40 characters + if i % 5 != 0: + plt.yticks([]) + print(f'{k:>15}: {mu:.3g}') + f = evolve_csv.with_suffix('.png') # filename + plt.savefig(f, dpi=200) + plt.close() + print(f'Saved {f}') + + +def plot_results(file='path/to/results.csv', dir=''): + # Plot training results.csv. Usage: from utils.plots import *; plot_results('path/to/results.csv') + save_dir = Path(file).parent if file else Path(dir) + fig, ax = plt.subplots(2, 5, figsize=(12, 6), tight_layout=True) + ax = ax.ravel() + files = list(save_dir.glob('results*.csv')) + assert len(files), f'No results.csv files found in {save_dir.resolve()}, nothing to plot.' + for f in files: + try: + data = pd.read_csv(f) + s = [x.strip() for x in data.columns] + x = data.values[:, 0] + for i, j in enumerate([1, 2, 3, 4, 5, 8, 9, 10, 6, 7]): + y = data.values[:, j].astype('float') + # y[y == 0] = np.nan # don't show zero values + ax[i].plot(x, y, marker='.', label=f.stem, linewidth=2, markersize=8) + ax[i].set_title(s[j], fontsize=12) + # if j in [8, 9, 10]: # share train and val loss y axes + # ax[i].get_shared_y_axes().join(ax[i], ax[i - 5]) + except Exception as e: + LOGGER.info(f'Warning: Plotting error for {f}: {e}') + ax[1].legend() + fig.savefig(save_dir / 'results.png', dpi=200) + plt.close() + + +def profile_idetection(start=0, stop=0, labels=(), save_dir=''): + # Plot iDetection '*.txt' per-image logs. from utils.plots import *; profile_idetection() + ax = plt.subplots(2, 4, figsize=(12, 6), tight_layout=True)[1].ravel() + s = ['Images', 'Free Storage (GB)', 'RAM Usage (GB)', 'Battery', 'dt_raw (ms)', 'dt_smooth (ms)', 'real-world FPS'] + files = list(Path(save_dir).glob('frames*.txt')) + for fi, f in enumerate(files): + try: + results = np.loadtxt(f, ndmin=2).T[:, 90:-30] # clip first and last rows + n = results.shape[1] # number of rows + x = np.arange(start, min(stop, n) if stop else n) + results = results[:, x] + t = (results[0] - results[0].min()) # set t0=0s + results[0] = x + for i, a in enumerate(ax): + if i < len(results): + label = labels[fi] if len(labels) else f.stem.replace('frames_', '') + a.plot(t, results[i], marker='.', label=label, linewidth=1, markersize=5) + a.set_title(s[i]) + a.set_xlabel('time (s)') + # if fi == len(files) - 1: + # a.set_ylim(bottom=0) + for side in ['top', 'right']: + a.spines[side].set_visible(False) + else: + a.remove() + except Exception as e: + print(f'Warning: Plotting error for {f}; {e}') + ax[1].legend() + plt.savefig(Path(save_dir) / 'idetection_profile.png', dpi=200) + + +def save_one_box(xyxy, im, file=Path('im.jpg'), gain=1.02, pad=10, square=False, BGR=False, save=True): + # Save image crop as {file} with crop size multiple {gain} and {pad} pixels. Save and/or return crop + xyxy = torch.tensor(xyxy).view(-1, 4) + b = xyxy2xywh(xyxy) # boxes + if square: + b[:, 2:] = b[:, 2:].max(1)[0].unsqueeze(1) # attempt rectangle to square + b[:, 2:] = b[:, 2:] * gain + pad # box wh * gain + pad + xyxy = xywh2xyxy(b).long() + clip_coords(xyxy, im.shape) + crop = im[int(xyxy[0, 1]):int(xyxy[0, 3]), int(xyxy[0, 0]):int(xyxy[0, 2]), ::(1 if BGR else -1)] + if save: + file.parent.mkdir(parents=True, exist_ok=True) # make directory + f = str(increment_path(file).with_suffix('.jpg')) + # cv2.imwrite(f, crop) # https://github.com/ultralytics/yolov5/issues/7007 chroma subsampling issue + Image.fromarray(cv2.cvtColor(crop, cv2.COLOR_BGR2RGB)).save(f, quality=95, subsampling=0) + return crop diff --git a/yolov5/utils/torch_utils.py b/yolov5/utils/torch_utils.py new file mode 100644 index 0000000..d11df83 --- /dev/null +++ b/yolov5/utils/torch_utils.py @@ -0,0 +1,314 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +PyTorch utils +""" + +import math +import os +import platform +import subprocess +import time +import warnings +from contextlib import contextmanager +from copy import deepcopy +from pathlib import Path + +import torch +import torch.distributed as dist +import torch.nn as nn +import torch.nn.functional as F + +from utils.general import LOGGER, file_date, git_describe + +try: + import thop # for FLOPs computation +except ImportError: + thop = None + +# Suppress PyTorch warnings +warnings.filterwarnings('ignore', message='User provided device_type of \'cuda\', but CUDA is not available. Disabling') + + +@contextmanager +def torch_distributed_zero_first(local_rank: int): + # Decorator to make all processes in distributed training wait for each local_master to do something + if local_rank not in [-1, 0]: + dist.barrier(device_ids=[local_rank]) + yield + if local_rank == 0: + dist.barrier(device_ids=[0]) + + +def device_count(): + # Returns number of CUDA devices available. Safe version of torch.cuda.device_count(). Supports Linux and Windows + assert platform.system() in ('Linux', 'Windows'), 'device_count() only supported on Linux or Windows' + try: + cmd = 'nvidia-smi -L | wc -l' if platform.system() == 'Linux' else 'nvidia-smi -L | find /c /v ""' # Windows + return int(subprocess.run(cmd, shell=True, capture_output=True, check=True).stdout.decode().split()[-1]) + except Exception: + return 0 + + +def select_device(device='', batch_size=0, newline=True): + # device = None or 'cpu' or 0 or '0' or '0,1,2,3' + s = f'YOLOv5 🚀 {git_describe() or file_date()} Python-{platform.python_version()} torch-{torch.__version__} ' + device = str(device).strip().lower().replace('cuda:', '').replace('none', '') # to string, 'cuda:0' to '0' + cpu = device == 'cpu' + mps = device == 'mps' # Apple Metal Performance Shaders (MPS) + if cpu or mps: + os.environ['CUDA_VISIBLE_DEVICES'] = '-1' # force torch.cuda.is_available() = False + elif device: # non-cpu device requested + os.environ['CUDA_VISIBLE_DEVICES'] = device # set environment variable - must be before assert is_available() + assert torch.cuda.is_available() and torch.cuda.device_count() >= len(device.replace(',', '')), \ + f"Invalid CUDA '--device {device}' requested, use '--device cpu' or pass valid CUDA device(s)" + + cuda = not cpu and torch.cuda.is_available() + if cuda: + devices = device.split(',') if device else '0' # range(torch.cuda.device_count()) # i.e. 0,1,6,7 + n = len(devices) # device count + if n > 1 and batch_size > 0: # check batch_size is divisible by device_count + assert batch_size % n == 0, f'batch-size {batch_size} not multiple of GPU count {n}' + space = ' ' * (len(s) + 1) + for i, d in enumerate(devices): + p = torch.cuda.get_device_properties(i) + s += f"{'' if i == 0 else space}CUDA:{d} ({p.name}, {p.total_memory / (1 << 20):.0f}MiB)\n" # bytes to MB + elif mps: + s += 'MPS\n' + else: + s += 'CPU\n' + + if not newline: + s = s.rstrip() + LOGGER.info(s.encode().decode('ascii', 'ignore') if platform.system() == 'Windows' else s) # emoji-safe + return torch.device('cuda:0' if cuda else 'mps' if mps else 'cpu') + + +def time_sync(): + # PyTorch-accurate time + if torch.cuda.is_available(): + torch.cuda.synchronize() + return time.time() + + +def profile(input, ops, n=10, device=None): + # YOLOv5 speed/memory/FLOPs profiler + # + # Usage: + # input = torch.randn(16, 3, 640, 640) + # m1 = lambda x: x * torch.sigmoid(x) + # m2 = nn.SiLU() + # profile(input, [m1, m2], n=100) # profile over 100 iterations + + results = [] + if not isinstance(device, torch.device): + device = select_device(device) + print(f"{'Params':>12s}{'GFLOPs':>12s}{'GPU_mem (GB)':>14s}{'forward (ms)':>14s}{'backward (ms)':>14s}" + f"{'input':>24s}{'output':>24s}") + + for x in input if isinstance(input, list) else [input]: + x = x.to(device) + x.requires_grad = True + for m in ops if isinstance(ops, list) else [ops]: + m = m.to(device) if hasattr(m, 'to') else m # device + m = m.half() if hasattr(m, 'half') and isinstance(x, torch.Tensor) and x.dtype is torch.float16 else m + tf, tb, t = 0, 0, [0, 0, 0] # dt forward, backward + try: + flops = thop.profile(m, inputs=(x,), verbose=False)[0] / 1E9 * 2 # GFLOPs + except Exception: + flops = 0 + + try: + for _ in range(n): + t[0] = time_sync() + y = m(x) + t[1] = time_sync() + try: + _ = (sum(yi.sum() for yi in y) if isinstance(y, list) else y).sum().backward() + t[2] = time_sync() + except Exception: # no backward method + # print(e) # for debug + t[2] = float('nan') + tf += (t[1] - t[0]) * 1000 / n # ms per op forward + tb += (t[2] - t[1]) * 1000 / n # ms per op backward + mem = torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0 # (GB) + s_in, s_out = (tuple(x.shape) if isinstance(x, torch.Tensor) else 'list' for x in (x, y)) # shapes + p = sum(x.numel() for x in m.parameters()) if isinstance(m, nn.Module) else 0 # parameters + print(f'{p:12}{flops:12.4g}{mem:>14.3f}{tf:14.4g}{tb:14.4g}{str(s_in):>24s}{str(s_out):>24s}') + results.append([p, flops, mem, tf, tb, s_in, s_out]) + except Exception as e: + print(e) + results.append(None) + torch.cuda.empty_cache() + return results + + +def is_parallel(model): + # Returns True if model is of type DP or DDP + return type(model) in (nn.parallel.DataParallel, nn.parallel.DistributedDataParallel) + + +def de_parallel(model): + # De-parallelize a model: returns single-GPU model if model is of type DP or DDP + return model.module if is_parallel(model) else model + + +def initialize_weights(model): + for m in model.modules(): + t = type(m) + if t is nn.Conv2d: + pass # nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') + elif t is nn.BatchNorm2d: + m.eps = 1e-3 + m.momentum = 0.03 + elif t in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6, nn.SiLU]: + m.inplace = True + + +def find_modules(model, mclass=nn.Conv2d): + # Finds layer indices matching module class 'mclass' + return [i for i, m in enumerate(model.module_list) if isinstance(m, mclass)] + + +def sparsity(model): + # Return global model sparsity + a, b = 0, 0 + for p in model.parameters(): + a += p.numel() + b += (p == 0).sum() + return b / a + + +def prune(model, amount=0.3): + # Prune model to requested global sparsity + import torch.nn.utils.prune as prune + print('Pruning model... ', end='') + for name, m in model.named_modules(): + if isinstance(m, nn.Conv2d): + prune.l1_unstructured(m, name='weight', amount=amount) # prune + prune.remove(m, 'weight') # make permanent + print(' %.3g global sparsity' % sparsity(model)) + + +def fuse_conv_and_bn(conv, bn): + # Fuse Conv2d() and BatchNorm2d() layers https://tehnokv.com/posts/fusing-batchnorm-and-conv/ + fusedconv = nn.Conv2d(conv.in_channels, + conv.out_channels, + kernel_size=conv.kernel_size, + stride=conv.stride, + padding=conv.padding, + groups=conv.groups, + bias=True).requires_grad_(False).to(conv.weight.device) + + # Prepare filters + w_conv = conv.weight.clone().view(conv.out_channels, -1) + w_bn = torch.diag(bn.weight.div(torch.sqrt(bn.eps + bn.running_var))) + fusedconv.weight.copy_(torch.mm(w_bn, w_conv).view(fusedconv.weight.shape)) + + # Prepare spatial bias + b_conv = torch.zeros(conv.weight.size(0), device=conv.weight.device) if conv.bias is None else conv.bias + b_bn = bn.bias - bn.weight.mul(bn.running_mean).div(torch.sqrt(bn.running_var + bn.eps)) + fusedconv.bias.copy_(torch.mm(w_bn, b_conv.reshape(-1, 1)).reshape(-1) + b_bn) + + return fusedconv + + +def model_info(model, verbose=False, img_size=640): + # Model information. img_size may be int or list, i.e. img_size=640 or img_size=[640, 320] + n_p = sum(x.numel() for x in model.parameters()) # number parameters + n_g = sum(x.numel() for x in model.parameters() if x.requires_grad) # number gradients + if verbose: + print(f"{'layer':>5} {'name':>40} {'gradient':>9} {'parameters':>12} {'shape':>20} {'mu':>10} {'sigma':>10}") + for i, (name, p) in enumerate(model.named_parameters()): + name = name.replace('module_list.', '') + print('%5g %40s %9s %12g %20s %10.3g %10.3g' % + (i, name, p.requires_grad, p.numel(), list(p.shape), p.mean(), p.std())) + + try: # FLOPs + from thop import profile + stride = max(int(model.stride.max()), 32) if hasattr(model, 'stride') else 32 + img = torch.zeros((1, model.yaml.get('ch', 3), stride, stride), device=next(model.parameters()).device) # input + flops = profile(deepcopy(model), inputs=(img,), verbose=False)[0] / 1E9 * 2 # stride GFLOPs + img_size = img_size if isinstance(img_size, list) else [img_size, img_size] # expand if int/float + fs = ', %.1f GFLOPs' % (flops * img_size[0] / stride * img_size[1] / stride) # 640x640 GFLOPs + except Exception: + fs = '' + + name = Path(model.yaml_file).stem.replace('yolov5', 'YOLOv5') if hasattr(model, 'yaml_file') else 'Model' + LOGGER.info(f"{name} summary: {len(list(model.modules()))} layers, {n_p} parameters, {n_g} gradients{fs}") + + +def scale_img(img, ratio=1.0, same_shape=False, gs=32): # img(16,3,256,416) + # Scales img(bs,3,y,x) by ratio constrained to gs-multiple + if ratio == 1.0: + return img + h, w = img.shape[2:] + s = (int(h * ratio), int(w * ratio)) # new size + img = F.interpolate(img, size=s, mode='bilinear', align_corners=False) # resize + if not same_shape: # pad/crop img + h, w = (math.ceil(x * ratio / gs) * gs for x in (h, w)) + return F.pad(img, [0, w - s[1], 0, h - s[0]], value=0.447) # value = imagenet mean + + +def copy_attr(a, b, include=(), exclude=()): + # Copy attributes from b to a, options to only include [...] and to exclude [...] + for k, v in b.__dict__.items(): + if (len(include) and k not in include) or k.startswith('_') or k in exclude: + continue + else: + setattr(a, k, v) + + +class EarlyStopping: + # YOLOv5 simple early stopper + def __init__(self, patience=30): + self.best_fitness = 0.0 # i.e. mAP + self.best_epoch = 0 + self.patience = patience or float('inf') # epochs to wait after fitness stops improving to stop + self.possible_stop = False # possible stop may occur next epoch + + def __call__(self, epoch, fitness): + if fitness >= self.best_fitness: # >= 0 to allow for early zero-fitness stage of training + self.best_epoch = epoch + self.best_fitness = fitness + delta = epoch - self.best_epoch # epochs without improvement + self.possible_stop = delta >= (self.patience - 1) # possible stop may occur next epoch + stop = delta >= self.patience # stop training if patience exceeded + if stop: + LOGGER.info(f'Stopping training early as no improvement observed in last {self.patience} epochs. ' + f'Best results observed at epoch {self.best_epoch}, best model saved as best.pt.\n' + f'To update EarlyStopping(patience={self.patience}) pass a new patience value, ' + f'i.e. `python train.py --patience 300` or use `--patience 0` to disable EarlyStopping.') + return stop + + +class ModelEMA: + """ Updated Exponential Moving Average (EMA) from https://github.com/rwightman/pytorch-image-models + Keeps a moving average of everything in the model state_dict (parameters and buffers) + For EMA details see https://www.tensorflow.org/api_docs/python/tf/train/ExponentialMovingAverage + """ + + def __init__(self, model, decay=0.9999, tau=2000, updates=0): + # Create EMA + self.ema = deepcopy(de_parallel(model)).eval() # FP32 EMA + # if next(model.parameters()).device.type != 'cpu': + # self.ema.half() # FP16 EMA + self.updates = updates # number of EMA updates + self.decay = lambda x: decay * (1 - math.exp(-x / tau)) # decay exponential ramp (to help early epochs) + for p in self.ema.parameters(): + p.requires_grad_(False) + + def update(self, model): + # Update EMA parameters + with torch.no_grad(): + self.updates += 1 + d = self.decay(self.updates) + + msd = de_parallel(model).state_dict() # model state_dict + for k, v in self.ema.state_dict().items(): + if v.dtype.is_floating_point: + v *= d + v += (1 - d) * msd[k].detach() + + def update_attr(self, model, include=(), exclude=('process_group', 'reducer')): + # Update EMA attributes + copy_attr(self.ema, model, include, exclude) diff --git a/yolov5/val.py b/yolov5/val.py new file mode 100644 index 0000000..dc7f28f --- /dev/null +++ b/yolov5/val.py @@ -0,0 +1,394 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Validate a trained YOLOv5 model accuracy on a custom dataset + +Usage: + $ python path/to/val.py --weights yolov5s.pt --data coco128.yaml --img 640 + +Usage - formats: + $ python path/to/val.py --weights yolov5s.pt # PyTorch + yolov5s.torchscript # TorchScript + yolov5s.onnx # ONNX Runtime or OpenCV DNN with --dnn + yolov5s.xml # OpenVINO + yolov5s.engine # TensorRT + yolov5s.mlmodel # CoreML (macOS-only) + yolov5s_saved_model # TensorFlow SavedModel + yolov5s.pb # TensorFlow GraphDef + yolov5s.tflite # TensorFlow Lite + yolov5s_edgetpu.tflite # TensorFlow Edge TPU +""" + +import argparse +import json +import os +import sys +from pathlib import Path + +import numpy as np +import torch +from tqdm import tqdm + +FILE = Path(__file__).resolve() +ROOT = FILE.parents[0] # YOLOv5 root directory +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) # add ROOT to PATH +ROOT = Path(os.path.relpath(ROOT, Path.cwd())) # relative + +from models.common import DetectMultiBackend +from utils.callbacks import Callbacks +from utils.dataloaders import create_dataloader +from utils.general import (LOGGER, check_dataset, check_img_size, check_requirements, check_yaml, + coco80_to_coco91_class, colorstr, emojis, increment_path, non_max_suppression, print_args, + scale_coords, xywh2xyxy, xyxy2xywh) +from utils.metrics import ConfusionMatrix, ap_per_class, box_iou +from utils.plots import output_to_target, plot_images, plot_val_study +from utils.torch_utils import select_device, time_sync + + +def save_one_txt(predn, save_conf, shape, file): + # Save one txt result + gn = torch.tensor(shape)[[1, 0, 1, 0]] # normalization gain whwh + for *xyxy, conf, cls in predn.tolist(): + xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist() # normalized xywh + line = (cls, *xywh, conf) if save_conf else (cls, *xywh) # label format + with open(file, 'a') as f: + f.write(('%g ' * len(line)).rstrip() % line + '\n') + + +def save_one_json(predn, jdict, path, class_map): + # Save one JSON result {"image_id": 42, "category_id": 18, "bbox": [258.15, 41.29, 348.26, 243.78], "score": 0.236} + image_id = int(path.stem) if path.stem.isnumeric() else path.stem + box = xyxy2xywh(predn[:, :4]) # xywh + box[:, :2] -= box[:, 2:] / 2 # xy center to top-left corner + for p, b in zip(predn.tolist(), box.tolist()): + jdict.append({ + 'image_id': image_id, + 'category_id': class_map[int(p[5])], + 'bbox': [round(x, 3) for x in b], + 'score': round(p[4], 5)}) + + +def process_batch(detections, labels, iouv): + """ + Return correct predictions matrix. Both sets of boxes are in (x1, y1, x2, y2) format. + Arguments: + detections (Array[N, 6]), x1, y1, x2, y2, conf, class + labels (Array[M, 5]), class, x1, y1, x2, y2 + Returns: + correct (Array[N, 10]), for 10 IoU levels + """ + correct = torch.zeros(detections.shape[0], iouv.shape[0], dtype=torch.bool, device=iouv.device) + iou = box_iou(labels[:, 1:], detections[:, :4]) + correct_class = labels[:, 0:1] == detections[:, 5] + for i in range(len(iouv)): + x = torch.where((iou >= iouv[i]) & correct_class) # IoU > threshold and classes match + if x[0].shape[0]: + matches = torch.cat((torch.stack(x, 1), iou[x[0], x[1]][:, None]), 1).cpu().numpy() # [label, detect, iou] + if x[0].shape[0] > 1: + matches = matches[matches[:, 2].argsort()[::-1]] + matches = matches[np.unique(matches[:, 1], return_index=True)[1]] + # matches = matches[matches[:, 2].argsort()[::-1]] + matches = matches[np.unique(matches[:, 0], return_index=True)[1]] + correct[matches[:, 1].astype(int), i] = True + return correct + + +@torch.no_grad() +def run( + data, + weights=None, # model.pt path(s) + batch_size=32, # batch size + imgsz=640, # inference size (pixels) + conf_thres=0.001, # confidence threshold + iou_thres=0.6, # NMS IoU threshold + task='val', # train, val, test, speed or study + device='', # cuda device, i.e. 0 or 0,1,2,3 or cpu + workers=8, # max dataloader workers (per RANK in DDP mode) + single_cls=False, # treat as single-class dataset + augment=False, # augmented inference + verbose=False, # verbose output + save_txt=False, # save results to *.txt + save_hybrid=False, # save label+prediction hybrid results to *.txt + save_conf=False, # save confidences in --save-txt labels + save_json=False, # save a COCO-JSON results file + project=ROOT / 'runs/val', # save to project/name + name='exp', # save to project/name + exist_ok=False, # existing project/name ok, do not increment + half=True, # use FP16 half-precision inference + dnn=False, # use OpenCV DNN for ONNX inference + model=None, + dataloader=None, + save_dir=Path(''), + plots=True, + callbacks=Callbacks(), + compute_loss=None, +): + # Initialize/load model and set device + training = model is not None + if training: # called by train.py + device, pt, jit, engine = next(model.parameters()).device, True, False, False # get model device, PyTorch model + half &= device.type != 'cpu' # half precision only supported on CUDA + model.half() if half else model.float() + else: # called directly + device = select_device(device, batch_size=batch_size) + + # Directories + save_dir = increment_path(Path(project) / name, exist_ok=exist_ok) # increment run + (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True) # make dir + + # Load model + model = DetectMultiBackend(weights, device=device, dnn=dnn, data=data, fp16=half) + stride, pt, jit, engine = model.stride, model.pt, model.jit, model.engine + imgsz = check_img_size(imgsz, s=stride) # check image size + half = model.fp16 # FP16 supported on limited backends with CUDA + if engine: + batch_size = model.batch_size + else: + device = model.device + if not (pt or jit): + batch_size = 1 # export.py models default to batch-size 1 + LOGGER.info(f'Forcing --batch-size 1 square inference (1,3,{imgsz},{imgsz}) for non-PyTorch models') + + # Data + data = check_dataset(data) # check + + # Configure + model.eval() + cuda = device.type != 'cpu' + is_coco = isinstance(data.get('val'), str) and data['val'].endswith(f'coco{os.sep}val2017.txt') # COCO dataset + nc = 1 if single_cls else int(data['nc']) # number of classes + iouv = torch.linspace(0.5, 0.95, 10, device=device) # iou vector for mAP@0.5:0.95 + niou = iouv.numel() + + # Dataloader + if not training: + if pt and not single_cls: # check --weights are trained on --data + ncm = model.model.nc + assert ncm == nc, f'{weights[0]} ({ncm} classes) trained on different --data than what you passed ({nc} ' \ + f'classes). Pass correct combination of --weights and --data that are trained together.' + model.warmup(imgsz=(1 if pt else batch_size, 3, imgsz, imgsz)) # warmup + pad = 0.0 if task in ('speed', 'benchmark') else 0.5 + rect = False if task == 'benchmark' else pt # square inference for benchmarks + task = task if task in ('train', 'val', 'test') else 'val' # path to train/val/test images + dataloader = create_dataloader(data[task], + imgsz, + batch_size, + stride, + single_cls, + pad=pad, + rect=rect, + workers=workers, + prefix=colorstr(f'{task}: '))[0] + + seen = 0 + confusion_matrix = ConfusionMatrix(nc=nc) + names = {k: v for k, v in enumerate(model.names if hasattr(model, 'names') else model.module.names)} + class_map = coco80_to_coco91_class() if is_coco else list(range(1000)) + s = ('%20s' + '%11s' * 6) % ('Class', 'Images', 'Labels', 'P', 'R', 'mAP@.5', 'mAP@.5:.95') + dt, p, r, f1, mp, mr, map50, map = [0.0, 0.0, 0.0], 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 + loss = torch.zeros(3, device=device) + jdict, stats, ap, ap_class = [], [], [], [] + callbacks.run('on_val_start') + pbar = tqdm(dataloader, desc=s, bar_format='{l_bar}{bar:10}{r_bar}{bar:-10b}') # progress bar + for batch_i, (im, targets, paths, shapes) in enumerate(pbar): + callbacks.run('on_val_batch_start') + t1 = time_sync() + if cuda: + im = im.to(device, non_blocking=True) + targets = targets.to(device) + im = im.half() if half else im.float() # uint8 to fp16/32 + im /= 255 # 0 - 255 to 0.0 - 1.0 + nb, _, height, width = im.shape # batch size, channels, height, width + t2 = time_sync() + dt[0] += t2 - t1 + + # Inference + out, train_out = model(im) if training else model(im, augment=augment, val=True) # inference, loss outputs + dt[1] += time_sync() - t2 + + # Loss + if compute_loss: + loss += compute_loss([x.float() for x in train_out], targets)[1] # box, obj, cls + + # NMS + targets[:, 2:] *= torch.tensor((width, height, width, height), device=device) # to pixels + lb = [targets[targets[:, 0] == i, 1:] for i in range(nb)] if save_hybrid else [] # for autolabelling + t3 = time_sync() + out = non_max_suppression(out, conf_thres, iou_thres, labels=lb, multi_label=True, agnostic=single_cls) + dt[2] += time_sync() - t3 + + # Metrics + for si, pred in enumerate(out): + labels = targets[targets[:, 0] == si, 1:] + nl, npr = labels.shape[0], pred.shape[0] # number of labels, predictions + path, shape = Path(paths[si]), shapes[si][0] + correct = torch.zeros(npr, niou, dtype=torch.bool, device=device) # init + seen += 1 + + if npr == 0: + if nl: + stats.append((correct, *torch.zeros((3, 0), device=device))) + continue + + # Predictions + if single_cls: + pred[:, 5] = 0 + predn = pred.clone() + scale_coords(im[si].shape[1:], predn[:, :4], shape, shapes[si][1]) # native-space pred + + # Evaluate + if nl: + tbox = xywh2xyxy(labels[:, 1:5]) # target boxes + scale_coords(im[si].shape[1:], tbox, shape, shapes[si][1]) # native-space labels + labelsn = torch.cat((labels[:, 0:1], tbox), 1) # native-space labels + correct = process_batch(predn, labelsn, iouv) + if plots: + confusion_matrix.process_batch(predn, labelsn) + stats.append((correct, pred[:, 4], pred[:, 5], labels[:, 0])) # (correct, conf, pcls, tcls) + + # Save/log + if save_txt: + save_one_txt(predn, save_conf, shape, file=save_dir / 'labels' / (path.stem + '.txt')) + if save_json: + save_one_json(predn, jdict, path, class_map) # append to COCO-JSON dictionary + callbacks.run('on_val_image_end', pred, predn, path, names, im[si]) + + # Plot images + if plots and batch_i < 3: + plot_images(im, targets, paths, save_dir / f'val_batch{batch_i}_labels.jpg', names) # labels + plot_images(im, output_to_target(out), paths, save_dir / f'val_batch{batch_i}_pred.jpg', names) # pred + + callbacks.run('on_val_batch_end') + + # Compute metrics + stats = [torch.cat(x, 0).cpu().numpy() for x in zip(*stats)] # to numpy + if len(stats) and stats[0].any(): + tp, fp, p, r, f1, ap, ap_class = ap_per_class(*stats, plot=plots, save_dir=save_dir, names=names) + ap50, ap = ap[:, 0], ap.mean(1) # AP@0.5, AP@0.5:0.95 + mp, mr, map50, map = p.mean(), r.mean(), ap50.mean(), ap.mean() + nt = np.bincount(stats[3].astype(int), minlength=nc) # number of targets per class + else: + nt = torch.zeros(1) + + # Print results + pf = '%20s' + '%11i' * 2 + '%11.3g' * 4 # print format + LOGGER.info(pf % ('all', seen, nt.sum(), mp, mr, map50, map)) + + # Print results per class + if (verbose or (nc < 50 and not training)) and nc > 1 and len(stats): + for i, c in enumerate(ap_class): + LOGGER.info(pf % (names[c], seen, nt[c], p[i], r[i], ap50[i], ap[i])) + + # Print speeds + t = tuple(x / seen * 1E3 for x in dt) # speeds per image + if not training: + shape = (batch_size, 3, imgsz, imgsz) + LOGGER.info(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {shape}' % t) + + # Plots + if plots: + confusion_matrix.plot(save_dir=save_dir, names=list(names.values())) + callbacks.run('on_val_end') + + # Save JSON + if save_json and len(jdict): + w = Path(weights[0] if isinstance(weights, list) else weights).stem if weights is not None else '' # weights + anno_json = str(Path(data.get('path', '../coco')) / 'annotations/instances_val2017.json') # annotations json + pred_json = str(save_dir / f"{w}_predictions.json") # predictions json + LOGGER.info(f'\nEvaluating pycocotools mAP... saving {pred_json}...') + with open(pred_json, 'w') as f: + json.dump(jdict, f) + + try: # https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb + check_requirements(['pycocotools']) + from pycocotools.coco import COCO + from pycocotools.cocoeval import COCOeval + + anno = COCO(anno_json) # init annotations api + pred = anno.loadRes(pred_json) # init predictions api + eval = COCOeval(anno, pred, 'bbox') + if is_coco: + eval.params.imgIds = [int(Path(x).stem) for x in dataloader.dataset.im_files] # image IDs to evaluate + eval.evaluate() + eval.accumulate() + eval.summarize() + map, map50 = eval.stats[:2] # update results (mAP@0.5:0.95, mAP@0.5) + except Exception as e: + LOGGER.info(f'pycocotools unable to run: {e}') + + # Return results + model.float() # for training + if not training: + s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else '' + LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}{s}") + maps = np.zeros(nc) + map + for i, c in enumerate(ap_class): + maps[c] = ap[i] + return (mp, mr, map50, map, *(loss.cpu() / len(dataloader)).tolist()), maps, t + + +def parse_opt(): + parser = argparse.ArgumentParser() + parser.add_argument('--data', type=str, default=ROOT / 'data/coco128.yaml', help='dataset.yaml path') + parser.add_argument('--weights', nargs='+', type=str, default=ROOT / 'yolov5s.pt', help='model.pt path(s)') + parser.add_argument('--batch-size', type=int, default=32, help='batch size') + parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, help='inference size (pixels)') + parser.add_argument('--conf-thres', type=float, default=0.001, help='confidence threshold') + parser.add_argument('--iou-thres', type=float, default=0.6, help='NMS IoU threshold') + parser.add_argument('--task', default='val', help='train, val, test, speed or study') + parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') + parser.add_argument('--workers', type=int, default=8, help='max dataloader workers (per RANK in DDP mode)') + parser.add_argument('--single-cls', action='store_true', help='treat as single-class dataset') + parser.add_argument('--augment', action='store_true', help='augmented inference') + parser.add_argument('--verbose', action='store_true', help='report mAP by class') + parser.add_argument('--save-txt', action='store_true', help='save results to *.txt') + parser.add_argument('--save-hybrid', action='store_true', help='save label+prediction hybrid results to *.txt') + parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels') + parser.add_argument('--save-json', action='store_true', help='save a COCO-JSON results file') + parser.add_argument('--project', default=ROOT / 'runs/val', help='save to project/name') + parser.add_argument('--name', default='exp', help='save to project/name') + parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment') + parser.add_argument('--half', action='store_true', help='use FP16 half-precision inference') + parser.add_argument('--dnn', action='store_true', help='use OpenCV DNN for ONNX inference') + opt = parser.parse_args() + opt.data = check_yaml(opt.data) # check YAML + opt.save_json |= opt.data.endswith('coco.yaml') + opt.save_txt |= opt.save_hybrid + print_args(vars(opt)) + return opt + + +def main(opt): + check_requirements(requirements=ROOT / 'requirements.txt', exclude=('tensorboard', 'thop')) + + if opt.task in ('train', 'val', 'test'): # run normally + if opt.conf_thres > 0.001: # https://github.com/ultralytics/yolov5/issues/1466 + LOGGER.info(emojis(f'WARNING: confidence threshold {opt.conf_thres} > 0.001 produces invalid results ⚠️')) + run(**vars(opt)) + + else: + weights = opt.weights if isinstance(opt.weights, list) else [opt.weights] + opt.half = True # FP16 for fastest results + if opt.task == 'speed': # speed benchmarks + # python val.py --task speed --data coco.yaml --batch 1 --weights yolov5n.pt yolov5s.pt... + opt.conf_thres, opt.iou_thres, opt.save_json = 0.25, 0.45, False + for opt.weights in weights: + run(**vars(opt), plots=False) + + elif opt.task == 'study': # speed vs mAP benchmarks + # python val.py --task study --data coco.yaml --iou 0.7 --weights yolov5n.pt yolov5s.pt... + for opt.weights in weights: + f = f'study_{Path(opt.data).stem}_{Path(opt.weights).stem}.txt' # filename to save to + x, y = list(range(256, 1536 + 128, 128)), [] # x axis (image sizes), y axis + for opt.imgsz in x: # img-size + LOGGER.info(f'\nRunning {f} --imgsz {opt.imgsz}...') + r, _, t = run(**vars(opt), plots=False) + y.append(r + t) # results and times + np.savetxt(f, y, fmt='%10.4g') # save + os.system('zip -r study.zip study_*.txt') + plot_val_study(x=x) # plot + + +if __name__ == "__main__": + opt = parse_opt() + main(opt)