-
Notifications
You must be signed in to change notification settings - Fork 8
/
train.py
282 lines (237 loc) · 10.2 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
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
import argparse
import math
from tqdm import tqdm
import numpy as np
import os
import os.path as osp
import torch
from torch.optim import SGD, Adam, AdamW, lr_scheduler
from torch.cuda import amp
from model import build_unet3plus, UNet3Plus
from torch.utils.data import DataLoader
from datasets import build_data_loader
from config.config import cfg
from utils.loss import build_u3p_loss
from utils.logger import AverageMeter, SummaryLogger
from utils.metrics import StreamSegMetrics
def one_cycle(y1=0.0, y2=1.0, steps=100):
return lambda x: ((1 - math.cos(x * math.pi / steps)) / 2) * (y2 - y1) + y1
class Trainer:
global_iter = 0
start_epoch = 0
epoch = 0 # current epoch
loss_dict = dict()
val_loss_dict = dict()
val_score_dict = None
best_val_score_dict = None
def __init__(self, cfg, model, train_loader, val_loader):
self.cfg_all = cfg
# build metrics
self.metrics = StreamSegMetrics(cfg.data.num_classes)
cfg = self.cfg = cfg.train
save_dir = osp.join(cfg.logger.log_dir, cfg.save_name)
os.makedirs(save_dir, exist_ok=True)
hyp_path = osp.join(save_dir, cfg.save_name+'.yaml')
with open(hyp_path, "w") as f:
f.write(cfg.dump())
self.model: UNet3Plus = model
self.train_loader: DataLoader = train_loader
self.val_loader: DataLoader = val_loader
# build loss
self.criterion = build_u3p_loss(cfg.loss_type, cfg.aux_weight)
self.scaler = amp.GradScaler(enabled=cfg.device == 'cuda') # mixed precision training
# build optimizer
if cfg.optimizer == 'sgd':
self.optimizer = SGD(self.model.parameters(), lr=cfg.lr, weight_decay=cfg.weight_decay, momentum=cfg.momentum, nesterov=cfg.nesterov)
elif cfg.optimizer == 'adam':
self.optimizer = Adam(self.model.parameters(), lr=cfg.lr, weight_decay=cfg.weight_decay)
elif cfg.optimizer == 'adamw':
self.optimizer = AdamW(self.model.parameters(), lr=cfg.lr, weight_decay=cfg.weight_decay)
else:
raise ValueError('Unknown optimizer')
if cfg.scheduler == 'linear':
self.lr_func = lambda x: (1 - x / (cfg.epochs - 1)) * (1.0 - cfg.lrf) + cfg.lrf # linear
elif cfg.scheduler == 'cyclic':
self.lr_func = one_cycle(1, cfg.lrf, cfg.epochs)
else:
raise ValueError('Unknown scheduler')
# build scheduler
self.scheduler = lr_scheduler.LambdaLR(self.optimizer, lr_lambda=self.lr_func)
self.logger = SummaryLogger(self.cfg_all)
self.model.to(cfg.device)
if cfg.resume:
self.resume(cfg.resume)
def resume(self, resume_path):
print('resuming from {}'.format(resume_path))
saved = torch.load(resume_path, map_location=self.cfg.device)
self.model.load_state_dict(saved['state_dict'])
self.optimizer.load_state_dict(saved['optimizer'])
self.scheduler.load_state_dict(saved['scheduler'])
self.scheduler.step()
self.epoch = saved['epoch'] + 1
self.start_epoch = saved['epoch'] + 1
self.global_iter = saved['global_iter']
def train(self):
for epoch in range(self.start_epoch, self.cfg.epochs):
self.logger.info(f'start training {epoch+1}/{self.cfg.epochs}')
self.train_one_epoch()
self.end_train_epoch()
def train_one_epoch(self):
model = self.model
model.train()
device = self.cfg.device
pbar = enumerate(self.train_loader)
num_batches = len(self.train_loader)
batch_size = self.train_loader.batch_size
accum_steps = self.cfg.accum_steps
pbar = tqdm(pbar, total=num_batches, bar_format='{l_bar}{bar:10}{r_bar}{bar:-10b} epoch: ' \
+ f'{self.epoch + 1}/{self.cfg.epochs}') # progress bar
for i, batch in pbar:
self.warmup()
imgs, masks = batch[0].to(device), batch[1].to(device, dtype=torch.long)
self.global_iter += batch_size
with amp.autocast():
preds = model(imgs)
loss, batch_loss_dict = self.criterion(preds, masks)
self.update_loss_dict(self.loss_dict, batch_loss_dict)
self.scaler.scale(loss).backward()
if (i+1) % accum_steps == 0 or i == num_batches - 1:
self.scaler.step(self.optimizer)
self.scaler.update()
self.optimizer.zero_grad()
pbar.close()
def end_train_epoch(self):
self.epoch += 1
if self.epoch % self.cfg.val_interval == 0 or self.epoch == self.cfg.epochs:
val_dict = self.val_score_dict = self.validate()
miou = val_dict['Mean IoU']
if self.best_val_score_dict is None or miou > self.best_val_score_dict['Mean IoU']:
self.best_val_score_dict = val_dict
self.save_checkpoint(self.cfg.save_name + '_best.ckpt')
self.log_results()
self.save_checkpoint(self.cfg.save_name + '_last.ckpt')
self.scheduler.step()
def save_checkpoint(self, save_name):
state = {
'epoch': self.epoch,
'global_iter': self.global_iter,
'state_dict': self.model.state_dict(),
'optimizer': self.optimizer.state_dict(),
'scheduler': self.scheduler.state_dict(),
}
torch.save(state, osp.join(self.cfg.logger.log_dir, self.cfg.save_name, save_name))
def warmup(self):
ni = self.global_iter
warmup_iters = max(self.cfg.warmup_iters, len(self.train_loader.dataset) * 3)
if ni <= warmup_iters:
xi = [0, warmup_iters] # x interp
for j, x in enumerate(self.optimizer.param_groups):
x['lr'] = np.interp(ni, xi, [0.1 if j == 2 else 0.0, x['initial_lr'] * self.lr_func(self.epoch)])
if 'momentum' in x:
x['momentum'] = np.interp(ni, xi, [0.8, self.cfg.momentum])
def update_loss_dict(self, loss_dict, batch_loss_dict=None):
if batch_loss_dict is None:
if loss_dict is None:
return
for k in loss_dict:
loss_dict[k].reset()
elif len(loss_dict) == 0:
for k, v in batch_loss_dict.items():
loss_dict[k] = AverageMeter(val=v)
else:
for k, v in batch_loss_dict.items():
loss_dict[k].update(v)
def log_results(self):
log_dict = {
'Train': {},
'Val': {}
}
for k, v in self.loss_dict.items():
log_dict['Train'][k] = v.avg
self.update_loss_dict(self.loss_dict, None)
log_dict['Train']['lr'] = self.optimizer.param_groups[0]['lr']
for k, v in self.val_loss_dict.items():
log_dict['Val'][k] = v.avg
self.update_loss_dict(self.val_loss_dict, None)
for k, v in self.val_score_dict.items():
if k == 'Class IoU':
print(v)
# self.logger.cmd_logger.info(v)
continue
log_dict['Val'][k] = v
self.logger.summary(log_dict, self.global_iter)
def validate(self):
"""Do validation and return specified samples"""
self.metrics.reset()
self.model.eval()
device = self.cfg.device
pbar = enumerate(self.val_loader)
pbar = tqdm(pbar, total=len(self.val_loader), bar_format='{l_bar}{bar:10}{r_bar}{bar:-10b}') # progress bar
with torch.no_grad():
for i, (images, labels) in pbar:
images = images.to(device)
labels = labels.to(device, dtype=torch.long)
outputs = self.model(images)
preds = outputs.detach().max(dim=1)[1].cpu().numpy()
targets = labels.cpu().numpy()
self.metrics.update(targets, preds)
_, batch_loss_dict = self.criterion(outputs, labels)
self.update_loss_dict(self.val_loss_dict, batch_loss_dict)
score = self.metrics.get_results()
pbar.close()
return score
def main(args):
cfg.merge_from_file(args.cfg)
if args.seed is not None:
cfg.train.seed = int(args.seed)
if args.resume:
cfg.train.resume = args.resume
if args.data_dir:
cfg.data.data_dir = args.data_dir
if args.use_tensorboard is not None:
cfg.train.logger.use_tensorboard = args.use_tensorboard == 1
elif args.use_wandb is not None:
cfg.train.logger.use_wandb = args.use_wandb == 1
cfg.freeze()
print(cfg)
import torch
import random
import numpy as np
seed = 42
torch.cuda.manual_seed_all(seed)
torch.manual_seed(seed)
random.seed(seed)
np.random.seed(seed)
model, data = cfg.model, cfg.data
model = build_unet3plus(data.num_classes, model.encoder, model.skip_ch, model.aux_losses, model.use_cgm, model.pretrained, model.dropout)
# model = UNet_3Plus_DeepSup()
if data.type in ['voc2012', 'voc2012_aug']:
train_loader, val_loader = build_data_loader(data.data_dir, data.batch_size, data.num_workers, data.max_training_samples, data.crop_size)
else:
raise NotImplementedError
trainer = Trainer(cfg, model, train_loader, val_loader)
trainer.train()
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Train segmentation network')
parser.add_argument('--cfg',
help='experiment configure file name',
default="config/u3p_resnet34_voc.yaml",
type=str)
parser.add_argument('--seed',
help='random seed',
default=None)
parser.add_argument('--resume',
help='resume from checkpoint',
default=None,
type=str)
parser.add_argument('--data_dir',
default=None,
type=str)
parser.add_argument('--use_wandb',
default=None,
type=int)
parser.add_argument('--use_tensorboard',
default=None,
type=int)
args = parser.parse_args()
main(args)