-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathtrain.py
180 lines (150 loc) · 7.1 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
import os
import numpy as np
import copy
import random
import json
import shutil
import tqdm
from sklearn.metrics import confusion_matrix
import torch
import torch.utils.data as Data
from tensorboardX import SummaryWriter
from network.build import build_model
from network.dataset import Dataset
from config.cfg import parse
from metric.eval_mAPJ import eval_mAPJ
from metric.eval_sAP import eval_sAP
import warnings
warnings.filterwarnings('ignore')
def to_device(data, device):
if isinstance(data, torch.Tensor):
return data.to(device)
if isinstance(data, dict):
for key in data:
if isinstance(data[key], torch.Tensor):
data[key] = data[key].to(device)
return data
if isinstance(data, list):
return [to_device(d, device) for d in data]
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:
print('Load pretrained model...')
checkpoint_file = os.path.join(cfg.model_path, cfg.model_name)
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.StepLR(optimizer, step_size=cfg.step_size, last_epoch=cfg.last_epoch)
# Summary
writer = SummaryWriter(cfg.log_path)
# Train
step = (cfg.last_epoch + 1) * len(loader['train'].dataset) // cfg.train_batch_size + 1
step_ = (cfg.last_epoch + 1) * len(loader['train'].dataset) + cfg.train_batch_size
best_sAP = [0 for _ in range(5)]
best_state_dict = None
for epoch in range(cfg.last_epoch + 1, cfg.num_epochs):
# Train
model.train()
for images, annotations in tqdm.tqdm(loader['train'], desc='train: '):
images, annotations = images.to(device), to_device(annotations, device)
loss_dict, labels, scores = model(images, annotations)
loss = sum([cfg.loss_weights[k] * loss_dict[k] for k in cfg.loss_weights.keys()])
optimizer.zero_grad()
loss.backward()
optimizer.step()
# Visualize
step_ = step * cfg.train_batch_size
if step % cfg.print_freq == 0:
lr = scheduler.get_last_lr()[0]
score = scores.detach().cpu().numpy() > 0.5
label = labels.detach().cpu().numpy() > 0.5
tn, fp, fn, tp = confusion_matrix(label, score).ravel()
msg = f'epoch: {epoch}/{cfg.num_epochs} | lr: {lr:e} | loss: {loss.item():6f} |'
for key, value in loss_dict.items():
msg += f' {key}: {value.item():6f} |'
print(msg)
print(f'tp: {tp} tn: {tn} fp: {fp} fn: {fn}')
writer.add_scalar('lr', lr, step_)
writer.add_scalar('loss', loss, step_)
for key, value in loss_dict.items():
writer.add_scalar(key, value, step_)
step += 1
scheduler.step()
if epoch % cfg.save_freq == 0:
# Save model
save_path = os.path.join(cfg.model_path, f'{os.path.splitext(cfg.model_name)[0]}-{epoch:03d}')
os.makedirs(save_path, exist_ok=True)
checkpoint_file = os.path.join(cfg.model_path, cfg.model_name)
checkpoint = {'model': model.state_dict(), 'optimizer': optimizer.state_dict()}
torch.save(checkpoint, checkpoint_file)
# Val
model.eval()
results = []
for images, annotations in tqdm.tqdm(loader['val'], desc='val: '):
images, annotations = images.to(device), to_device(annotations, device)
outputs = model(images, annotations)
for output in outputs:
for k in output.keys():
if isinstance(output[k], torch.Tensor):
output[k] = output[k].tolist()
results.append(output)
with open(os.path.join(save_path, 'result.json'), 'w') as f:
json.dump(results, f)
gt_file = os.path.join(cfg.dataset_path, 'test.json')
pred_file = os.path.join(save_path, 'result.json')
mAPJ, PJ, RJ = eval_mAPJ(gt_file, pred_file)
print(f'mAPJ: {mAPJ:.1f} | {PJ:.1f} | {RJ:.1f}')
msAP, P, R, sAP = eval_sAP(gt_file, pred_file)
print(f'msAP: {msAP:.1f} | {P:.1f} | {R:.1f} | {sAP[0]:.1f} | {sAP[1]:.1f} | {sAP[2]:.1f}')
writer.add_scalar('mAPJ', mAPJ, step_)
writer.add_scalar('msAP', msAP, step_)
shutil.rmtree(save_path)
if msAP > best_sAP[3]:
best_sAP = [mAPJ, PJ, RJ, msAP, P, R, *sAP]
best_state_dict = copy.deepcopy(model.state_dict())
msg = f'best msAP: {best_sAP[0]:.1f} | {best_sAP[1]:.1f} | {best_sAP[2]:.1f} | ' \
f'{best_sAP[3]:.1f} | {best_sAP[4]:.1f} | {best_sAP[5]:.1f} | ' \
f'{best_sAP[6]:.1f} | {best_sAP[7]:.1f} | {best_sAP[8]:.1f}'
print(msg)
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
os.environ['CUDA_VISIBLE_DEVICES'] = str(cfg.gpu)
use_gpu = cfg.gpu >= 0 and torch.cuda.is_available()
device = torch.device(f'cuda:0' 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 = build_model(cfg).to(device)
if cfg.pretrained_model_name != '':
pretrained_model_filename = os.path.join(cfg.model_path, cfg.pretrained_model_name)
print(f'Loading pretrained model: {pretrained_model_filename}')
checkpoint = torch.load(pretrained_model_filename, map_location=device)
if 'model' in checkpoint.keys():
state_dict = checkpoint['model']
else:
state_dict = checkpoint
model.load_state_dict(state_dict)
# Load dataset
train_dataset = Dataset(cfg, split='train')
val_dataset = Dataset(cfg, split='test')
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, collate_fn=train_dataset.collate)
loader = {'train': train_loader, 'val': val_loader}
# Train network
train(model, loader, cfg, device)