A step-by-step workthrough tutorial
Eric
2021/05/29
- Python 3.7.10
- PyTorch 1.7.1 (cpu version) (gpu version works actually well)
- Step 01: Installation
- Step 02: Get Dataset
- Step 03: Arrange File Structure
- Step 04: Have a look of image
- Step 05: Define Custom Dataset Class
- Step 06: Define DataLoader w.r.t Dataset Class
- Step 07: Define LeNet5 - like Structure
- Step 08: Define Loss Function
- Step 09: Training Phase
- Step 10: Plot Accuracy & Loss Curves
- Step 11: Testing Phase
-
Step 01: Installation
-
Anaconda
-
Conda Create Env
# 示範使用 CPU，請使用這個，Python 3.7 和 PyTorch 1.7.X 是好朋友 ^_^ conda create --name "PT_CPU_1.7.1_DEMO" python=3.7 # 確認新環境建置完成 conda env list # 啟動新環境 conda activate PT_CPU_1.7.1_DEMO
-
PyTorch: 1.7.X, and other useful packages (I spent 10 mins on this)
# 核心: 安裝 pytorch 1.7.1 CPU Version conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cpuonly -c pytorch ### Other Useful Packages # 1. GUI 介面，方便寫程式 → jupyter notebook conda install jupyter notebook # 2. 用來讀 Metadata.csv → pandas conda install pandas # 3. 知道 loop 的進度 → tqdm conda install tqdm # 4. 讀影像，其他選擇如 PIL，總之這邊使用 OpenCV → opencv conda install -c conda-forge opencv # 5. 畫圖 → matplotlib conda install matplotlib # 6. 矩陣運算 → numpy conda install numpy
-
示範上考慮到大家不一定都有顯卡，所以使用 CPU，若想要裝 higher version (e.g. 1.8.X) or GPU Version，請參考以下網址。
-
PyTorch Installation Link: https://pytorch.org/get-started/previous-versions/
-
-
Step 02: Get Dataset
-
AIdea AOI Detection: https://aidea-web.tw/topic/285ef3be-44eb-43dd-85cc-f0388bf85ea4
-
Download Page:
-
-
Step 03: Arrange File Structure
-
預期的專案結構如下:
- . / test_images/
- . / train_images/
- . / test.csv
- . / train.csv
- . / LeNet5.ipynb ← 這個 Python File 是要自己新建的
-
train_images
-
train.csv
-
- 在 PyTorch，Train 的時候提供 Batch 需要使用 DataLoader，而要使用 DataLoader 前，會需要先完成自定義的 Dataset Class。
- 下方正式進入程式碼，請大家移動 cd 至專案路徑當中，並開啟 jupyter notebook
jupyter notebook然後點選: LeNet5.ipynb
-
Step 04: Have a look of image
-
import 此步驟需要之套件
import pandas as pd import cv2 as cv import matplotlib.pyplot as plt
-
Mapping table: Python Dictionary is our friend!
label_map_table = { 0: "normal", 1: "void", 2: "Horizontal Defect", 3: "Vertical Defect", 4: "Edge Defect", 5: "Partical" }
-
Visualize it!
root_train = "./train_images/" root_test = "./test_images/" train_csv = "./train.csv" test_csv = "./test.csv" df_train = pd.read_csv(train_csv) # df stands for dataframe id = 137 # change this number to see other outcome png_img = cv.imread(root_train + df_train.ID[id]) label = df_train.Label[id] print(f"[Label] => {df_train.Label[id]}; [Label Actually Means] => {label_map_table[label]}") plt.imshow(png_img) plt.show()
-
Conclusion:
-
-
Step 05: Define Custom Dataset Class
-
import 此步驟需要的套件
from torch.utils.data.dataset import Dataset from torchvision import transforms import numpy as np import pandas as pd import cv2 as cv
-
重頭戲: 動手寫 Custom DataLoader
# [Input Args] # 1. target_csv <string>: It's the metadata file describe the name of image and its label. # 2. root_path <string>: It's the path to the image folder. Combination of this and name is the full path to the image. # 3. height <int>: Use this for elastically resize image to desired shape. # 4. width <int>: Use this for elastically resize image to desired shape. class AOI_Dataset(Dataset): # perform logic operation: think what kind of info I need when loading data def __init__(self, target_csv, root_path, height, width, transform = None): # height, width self.height = height self.width = width # register self self.target_csv = target_csv self.root_path = root_path # 1. Read CSV file through root_path self.df = pd.read_csv(self.target_csv) # 2. Remember the length self.count = len(self.df) # 3. transform self.transforms = transform # input: index # output: pair of (image, lable) def __getitem__(self, index): # Read images img = cv.imread(self.root_path + self.df.ID[index]) # Use resize to a smaller shape when training takes so long. img_resize = cv.resize(img, (self.height, self.width)) # To Tensor img_tensor = self.transforms(np.uint8(img_resize)) # Get label label = self.df.Label[index] return (img_tensor, label) def __len__(self): return self.count
-
REF: PyTorch 官方 Dataset & DataLoader 教學 [Link]
-
-
Step 06: Define DataLoader w.r.t Dataset Class
-
import 此步驟需要的套件
import torch from torchvision import transforms
-
DataLoader
height = 512 width = 512 transform_train = transforms.Compose([ transforms.ToTensor(), # add mroe in the future ]) Train_Dataset = AOI_Dataset(target_csv = train_csv, root_path = root_train, height = height, width = width, transform = transform_train) batch_size = 8 Train_DataLoader = torch.utils.data.DataLoader(Train_Dataset, batch_size = batch_size)
-
-
Step 07: Define LeNet5-like Structure
-
Yann Lecun
- Original Version of LeNet5 in his paper.
-
Structure of Original LeNet5
import torch import torch.nn as nn import torch.nn.functional as F # Use GPU device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") print(device) class LeNet(nn.Module): def __init__(self): super().__init__() self.conv1 = nn.Conv2d( 3, 6, 3, padding = 1) # 加深 channel self.conv2 = nn.Conv2d( 6, 16, 3, padding = 1) # 加深 channel self.conv3 = nn.Conv2d(16, 50, 3, padding = 1) # 加深 channel self.pool = nn.MaxPool2d(2, 2) self.fc1 = nn.Linear(50 * 64 * 64, 120) # Why 64 Why 50 self.fc2 = nn.Linear(120, 84) self.fc3 = nn.Linear(84, 6) def forward(self, x): x = self.pool(F.relu(self.conv1(x))) # 512 x 512 x 3 -> 256 x 256 x 6 x = self.pool(F.relu(self.conv2(x))) # 256 x 256 x 6 -> 128 x 128 x 16 x = self.pool(F.relu(self.conv3(x))) # 128 x 128 x 16 -> 64 x 64 x 50 x = x.view(-1, 50 * 64 * 64) x = F.relu(self.fc1(x)) x = F.relu(self.fc2(x)) x = self.fc3(x) #x = F.softmax(x) return x lenet = LeNet() print(lenet.to(device))
-
-
Step 08: Define Loss Function and setup Hyper Parameters
import torch.optim as optim criterion = nn.CrossEntropyLoss() optimizer = optim.RMSprop(lenet.parameters(), lr=1e-4) # learning rate is also adjustable epoch = 10 # change this to whatever number you'd like
-
Step 09: Training Phase
from tqdm import tqdm import time tic = time.time() train_acc_list = [] val_acc_list = [] loss_list = [] print_probe_num = 10 for epoch in range(epoch): # loop over the dataset multiple times running_loss = 0.0 for i, data in enumerate(Train_DataLoader, 0): # Select input and output pair inputs, labels = data[0].to(device), data[1].to(device) # Clear gradient optimizer.zero_grad() # Forward Propagation outputs = lenet(inputs.float()) # Compute Loss loss = criterion(outputs, labels) # Backward Propagation loss.backward() # Update Weight optimizer.step() # Just want to calculate the running loss 移動平均 loss!! running_loss += loss.item() if i % print_probe_num == (print_probe_num - 1): print('[%d, %5d] loss: %.3f' % (epoch + 1, i + 1, running_loss / print_probe_num)) loss_list.append(running_loss / print_probe_num) running_loss = 0.0 correct = 0 total = 0 # Train with torch.no_grad(): # since we're not training, we don't need to calculate the gradients for our outputs for datum in tqdm(Train_DataLoader): imgs, labs = datum[0].to(device), datum[1].to(device) # calculate outputs by running images through the network outputs = lenet(imgs.float()) # the class with the highest energy is what we choose as prediction _, preds = torch.max(outputs.data, 1) total += labs.size(0) correct += (preds == labs).sum().item() train_acc_list.append(float(correct)/float(total)) print('Accuracy of the network on the train images: %d %%' % (100 * correct / total)) toc = time.time() print(f"Spend {round(toc - tic, 2)} (sec)") print('Finished Training')
- Why zero_grad
-
What step(), backward(), and zero_grad() do [Link]
-
- Why zero_grad
-
Step 10: Plot Accuracy & Loss Curves
import matplotlib.pyplot as plt ## Accuracy plt.figure(figsize = (20, 10)) plt.title("LeNet5: Accuracy Curve", fontsize = 24) plt.xlabel("Epochs" , fontsize = 20) plt.ylabel("Accuracy %", fontsize = 20) plt.plot(train_acc_list, label = "train acc.") plt.legend(loc = 2, fontsize = 20) plt.show() ## Loss plt.figure(figsize=(20, 10)) plt.title("LeNet5: Loss Curve", fontsize = 24) plt.plot(loss_list) plt.xlabel("Probes", fontsize = 20) plt.ylabel("Loss", fontsize = 20) plt.show()
-
Step 11: Testing Phase
transform_test = transforms.Compose([ transforms.ToTensor(), ]) Test_Dataset = AOI_Dataset(target_csv = test_csv, root_path = root_test, height = height, width = width, transform = transform_test) Test_DataLoader = torch.utils.data.DataLoader(dataset = Test_Dataset, batch_size = 1, shuffle = False) Name_of_csv_file = "AI.FREE.SUCCESS.csv" df_test = pd.read_csv(test_csv) df_test_np = df_test.to_numpy() count = -1 with torch.no_grad(): # since we're not training, we don't need to calculate the gradients for our outputs for datum in tqdm(Test_DataLoader): count = count + 1 imgs = datum[0].to(device) # calculate outputs by running images through the network outputs = lenet(imgs.float()) # the class with the highest energy is what we choose as prediction _, preds = torch.max(outputs.data, 1) df_test_np[count][1] = float(preds) df = pd.DataFrame(df_test_np, columns = ['ID','Label']) df.to_csv(Name_of_csv_file, index=False)
- Submit the result to AIdea
- Validation Dataset
- Data Augmentation
- Train on GPU
- Estimate Usage of GPU Memory