-
Notifications
You must be signed in to change notification settings - Fork 26
/
train.py
188 lines (158 loc) · 8.47 KB
/
train.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
import os
import numpy as np
import copy
from tqdm import tqdm
import random
import shutil
from sklearn.metrics import confusion_matrix
import torch
import torch.utils.data as Data
import torch.nn.functional as F
from tensorboardX import SummaryWriter
from network.ulsd import ULSD, lpn_loss_func, loi_loss_func
from network.dataset import Dataset
from config.cfg import parse
from metric.eval_mAPJ import eval_mAPJ
from metric.eval_sAP import eval_sAP
def train(model, loader, cfg, device):
# Option
optimizer = torch.optim.Adam(model.parameters(), lr=cfg.lr, weight_decay=cfg.weight_decay, amsgrad=True)
if cfg.last_epoch != -1:
checkpoint_file = os.path.join(cfg.model_path, f'{cfg.version}-{cfg.last_epoch:03d}.pkl')
checkpoint = torch.load(checkpoint_file, map_location=device)
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=cfg.milestones, last_epoch=cfg.last_epoch)
# Summary
if os.path.exists(cfg.log_path):
shutil.rmtree(cfg.log_path)
writer = SummaryWriter(cfg.log_path)
# Train
step = 1
best_metric = [0.0] * 4
best_state_dict = None
for epoch in range(cfg.last_epoch + 1, cfg.num_epochs):
# Train
model.train()
for images, map_labels, metas in tqdm(loader['train'], desc='train: '):
images = images.to(device)
map_labels = {name: map_labels[name].to(device) for name in map_labels.keys()}
metas = {name: [meta.to(device) for meta in metas[name]] for name in metas.keys()}
map_preds, loi_scores, loi_labels = model(images, metas)
lmap_loss, jmap_loss, joff_loss, eoff, coff_loss, eoff_loss = lpn_loss_func(map_preds, map_labels)
pos_loss, neg_loss = loi_loss_func(loi_scores, loi_labels)
losses = [lmap_loss, jmap_loss, joff_loss, eoff, coff_loss, eoff_loss, pos_loss, neg_loss]
loss = sum([weight * loss for weight, loss in zip(cfg.weights, losses)])
optimizer.zero_grad()
loss.backward()
optimizer.step()
# Visualize
if step % cfg.print_freq == 0:
image = images[0]
lmap_pred = map_preds['lmap'][0]
jmap_pred = map_preds['jmap'][0]
cmap_pred = map_preds['cmap'][0]
lmap_label = map_labels['lmap'][0]
jmap_label = map_labels['jmap'][0]
cmap_label = map_labels['cmap'][0]
image = F.interpolate(image[None, :, :, :], (lmap_label.shape[-2], lmap_label.shape[-1]))[0]
lr = scheduler.get_last_lr()[0]
loi_label = loi_labels[0].detach().cpu().numpy() > 0.5
loi_score = loi_scores[0].detach().cpu().numpy() > 0.5
tn, fp, fn, tp = confusion_matrix(loi_label, loi_score).ravel()
print('epoch: %d/%d | loss: %6f | lmap loss: %6f | jmap loss: %6f | joff loss: %6f | cmap loss: %6f | '
'coff loss: %6f | eoff loss: %6f | pos loss: %6f | neg loss: %6f | lr: %e' % (epoch,
cfg.num_epochs, loss.item(), lmap_loss.item(), jmap_loss.item(), joff_loss.item(), eoff.item(),
coff_loss.item(), eoff_loss.item(), pos_loss.item(), neg_loss.item(), lr))
print('tn', tn, 'fp', fp, 'fn', fn, 'tp', tp)
image_list = [loader['train'].dataset.DeNormalize(image),
lmap_label.repeat((3, 1, 1)), lmap_pred.repeat((3, 1, 1)),
jmap_label.repeat((3, 1, 1)), jmap_pred.repeat((3, 1, 1)),
cmap_label.repeat((3, 1, 1)), cmap_pred.repeat((3, 1, 1))]
writer.add_images('image', image_list, step, dataformats='CHW')
writer.add_scalar('loss', loss, step)
writer.add_scalar('lmap loss', lmap_loss, step)
writer.add_scalar('jmap loss', jmap_loss, step)
writer.add_scalar('joff loss', joff_loss, step)
writer.add_scalar('cmap loss', eoff, step)
writer.add_scalar('coff loss', coff_loss, step)
writer.add_scalar('eoff loss', eoff_loss, step)
writer.add_scalar('pos loss', pos_loss, step)
writer.add_scalar('neg loss', neg_loss, step)
writer.add_scalars('metrics', {'tn': tn, 'fp': fp, 'fn': fn, 'tp': tp}, step)
writer.add_scalar('lr', lr, step)
step += 1
if epoch % cfg.save_freq == 0:
# Save checkpoint
if cfg.save_checkpoint:
checkpoint_file = os.path.join(cfg.model_path, f'{cfg.version}-{epoch:03d}.pkl')
checkpoint = {'model': model.state_dict(), 'optimizer': optimizer.state_dict()}
torch.save(checkpoint, checkpoint_file)
# Val
model.eval()
save_path = os.path.join(cfg.model_path, f'{cfg.version}-{epoch:03d}')
if not os.path.isdir(save_path):
os.mkdir(save_path)
index = 0
for images in tqdm(loader['val'], desc='val: '):
images = images.to(device)
jmaps, joffs, line_preds, line_scores = model(images)
jmaps = jmaps.detach().cpu()
joffs = joffs.detach().cpu()
line_preds = [line_pred.detach().cpu() for line_pred in line_preds]
line_scores = [line_score.detach().cpu() for line_score in line_scores]
for i in range(len(images)):
jmap = jmaps[i].numpy()
joff = joffs[i].numpy()
line_pred = line_preds[i].numpy()
line_score = line_scores[i].numpy()
src_filename = loader['val'].dataset.file_list[index].split()[0]
filename = os.path.split(src_filename)[1]
image_filename = os.path.join(save_path, filename)
npz_filename = image_filename.replace('.png', '.npz')
np.savez_compressed(npz_filename, jmap=jmap, joff=joff,
line_pred=line_pred, line_score=line_score)
index += 1
mAPJ, P, R = eval_mAPJ(cfg.groundtruth_path, save_path)
print(f'APJ: {mAPJ:.1f} | {P:.1f} | {R:.1f}')
msAP, P, R, _ = eval_sAP(cfg.groundtruth_path, save_path)
print(f'sAP: {msAP:.1f} | {P:.1f} | {R:.1f} |')
writer.add_scalar('metric: APJ', mAPJ, step)
writer.add_scalar('metric: sAP', msAP, step)
shutil.rmtree(save_path)
if msAP > best_metric[1]:
best_metric = [mAPJ, msAP, P, R]
best_state_dict = copy.deepcopy(model.state_dict())
print(f'best metric: {best_metric[0]:.1f} | {best_metric[1]:.1f} | {best_metric[2]:.1f} | {best_metric[3]:.1f}')
scheduler.step()
writer.close()
# Save best model
model_filename = os.path.join(cfg.model_path, cfg.model_name)
torch.save(best_state_dict, model_filename)
if __name__ == '__main__':
# Parameter
cfg = parse()
os.makedirs(cfg.model_path, exist_ok=True)
# Use GPU or CPU
use_gpu = cfg.gpu >= 0 and torch.cuda.is_available()
device = torch.device(f'cuda:{cfg.gpu}' if use_gpu else 'cpu')
print('use_gpu: ', use_gpu)
# Seed
random.seed(cfg.seed)
np.random.seed(cfg.seed)
torch.manual_seed(cfg.seed)
torch.backends.cudnn.deterministic = True
if use_gpu:
torch.cuda.manual_seed_all(cfg.seed)
# Load model
model = ULSD(cfg).to(device)
# Load dataset
train_dataset = Dataset(os.path.join(cfg.train_dataset_path, 'train'), cfg, with_label=True)
val_dataset = Dataset(os.path.join(cfg.train_dataset_path, 'test'), cfg, with_label=False)
train_loader = Data.DataLoader(dataset=train_dataset, batch_size=cfg.train_batch_size,
num_workers=cfg.num_workers, shuffle=True, collate_fn=train_dataset.collate)
val_loader = Data.DataLoader(dataset=val_dataset, batch_size=cfg.test_batch_size,
num_workers=cfg.num_workers, shuffle=False)
loader = {'train': train_loader, 'val': val_loader}
# Train network
train(model, loader, cfg, device)