-
Notifications
You must be signed in to change notification settings - Fork 5
/
train.py
259 lines (226 loc) · 10.9 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
import argparse
import datetime
import logging
import math
import random
import time
import torch
from os import path as osp
#from basicsr.data import create_dataloader, create_dataset
from basicsr.data.data_sampler import EnlargedSampler
from basicsr.data.prefetch_dataloader import CPUPrefetcher, CUDAPrefetcher
#from basicsr.models import create_model
from basicsr.utils import (MessageLogger, check_resume, get_env_info,
get_root_logger, get_time_str, init_tb_logger,
init_wandb_logger, make_exp_dirs, mkdir_and_rename,
set_random_seed)
from basicsr.utils.dist_util import get_dist_info, init_dist
from basicsr.utils.options import dict2str, parse
from data import create_dataloader, create_dataset
from model import create_model
import os
#os.environ["CUDA_VISIBLE_DEVICES"] = "0"
#
def parse_options(is_train=True):
parser = argparse.ArgumentParser()
parser.add_argument(
'-opt', type=str, required=True, help='Path to option YAML file.')
parser.add_argument(
'--launcher',
choices=['none', 'pytorch', 'slurm'],
default='none',
help='job launcher')
parser.add_argument('--local_rank', type=int, default=0)
args = parser.parse_args()
opt = parse(args.opt, is_train=is_train) # args.opt: yaml文件的路径;parse:打开opt并且写入新的路径等
# distributed settings
if args.launcher == 'none':
opt['dist'] = False
print('Disable distributed.', flush=True)
else:
opt['dist'] = True
if args.launcher == 'slurm' and 'dist_params' in opt:
init_dist(args.launcher, **opt['dist_params'])
else:
init_dist(args.launcher)
opt['rank'], opt['world_size'] = get_dist_info() # 0, 1
# random seed
seed = opt.get('manual_seed')
if seed is None:
seed = random.randint(1, 10000)
opt['manual_seed'] = seed
set_random_seed(seed + opt['rank'])
return opt
# 初始化logger 和 tensorboard logger 配置
def init_loggers(opt):
log_file = osp.join(opt['path']['log'], f"train_{opt['name']}_{get_time_str()}.log") # opt['path']['log'] = experiments_root -- log文件在experiment_root下面, f {} 就和 '{}'.format()类似
logger = get_root_logger(
logger_name='BVQA360', log_level=logging.INFO, log_file=log_file) # logger 初始化
#logger.info(get_env_info())
logger.info(dict2str(opt)) # 在logger文件里打印出
# initialize tensorboard logger and wandb logger
tb_logger = None
if opt['logger'].get('use_tb_logger') and 'debug' not in opt['name']:
tb_logger = init_tb_logger(log_dir=osp.join(opt['tb_pth'], 'tb_logger', opt['name'])) # /media/yl/yl_8t/tb_logger360/tb_logger/07015_yl_train_BVQA360_Basic_ODV-VQA240/ 中存放tensorboard信息
if (opt['logger'].get('wandb')
is not None) and (opt['logger']['wandb'].get('project')
is not None) and ('debug' not in opt['name']):
assert opt['logger'].get('use_tb_logger') is True, (
'should turn on tensorboard when using wandb')
init_wandb_logger(opt)
return logger, tb_logger
# 构建train和validation的dataloader
def create_train_val_dataloader(opt, logger):
# create train and val dataloaders
train_loader, val_loader = None, None
for phase, dataset_opt in opt['datasets'].items():
if phase == 'train':
dataset_enlarge_ratio = dataset_opt.get('dataset_enlarge_ratio', 1)
# dict.get(key, value=None)
# 当value的值存在时返回其本身,当key的值不存在时返回None(即默认参数)
train_set = create_dataset(dataset_opt)
train_sampler = EnlargedSampler(train_set, opt['world_size'],
opt['rank'], dataset_enlarge_ratio)
train_loader = create_dataloader(
train_set,
dataset_opt,
num_gpu=opt['num_gpu'],
dist=opt['dist'],
sampler=train_sampler,
seed=opt['manual_seed'])
num_iter_per_epoch = math.ceil(
len(train_set) * dataset_enlarge_ratio /
(dataset_opt['batch_size_per_gpu'] * opt['world_size']))
total_iters = int(opt['train']['total_iter'])
total_epochs = math.ceil(total_iters / (num_iter_per_epoch))
logger.info(
'Training statistics:'
f'\n\tNumber of train images: {len(train_set)}'
f'\n\tDataset enlarge ratio: {dataset_enlarge_ratio}'
f'\n\tBatch size per gpu: {dataset_opt["batch_size_per_gpu"]}'
f'\n\tWorld size (gpu number): {opt["world_size"]}'
f'\n\tRequire iter number per epoch: {num_iter_per_epoch}'
f'\n\tTotal epochs: {total_epochs}; iters: {total_iters}.')
# elif phase == 'val':
# val_set = create_dataset(dataset_opt)
# val_loader = create_dataloader(
# val_set,
# dataset_opt,
# num_gpu=opt['num_gpu'],
# dist=opt['dist'],
# sampler=None,
# seed=opt['manual_seed'])
# logger.info(
# f'Number of val images/folders in {dataset_opt["name"]}: '
# f'{len(val_set)}')
else:
raise ValueError(f'Dataset phase {phase} is not recognized.')
return train_loader, train_sampler, val_loader, total_epochs, total_iters
def main():
# parse options, set distributed setting, set ramdom seed
opt = parse_options(is_train=True)
torch.backends.cudnn.benchmark = True
# torch.backends.cudnn.deterministic = True
# load resume states if necessary
# 加载当前训练的状态,包括学习率等优化参数
if opt['path'].get('resume_state'): #Python 字典(Dictionary) get() 函数返回指定键的值。
device_id = torch.cuda.current_device()
resume_state = torch.load(
opt['path']['resume_state'],
map_location=lambda storage, loc: storage.cuda(device_id))
else:
resume_state = None
# mkdir for experiments and logger
if resume_state is None: # 新建experiment_root以及子目录:model,log,visual等;旧的expriment_root目录改名为当前时间戳
make_exp_dirs(opt) # 将当前训练的experiment目录改成旧的时间戳,新的无时间戳
if opt['logger'].get('use_tb_logger') and 'debug' not in opt[
'name'] and opt['rank'] == 0:
mkdir_and_rename(osp.join(opt['tb_pth'], 'tb_logger', opt['name']))
# initialize loggers
logger, tb_logger = init_loggers(opt)
# create train and validation dataloaders
result = create_train_val_dataloader(opt, logger)
train_loader, train_sampler, val_loader, total_epochs, total_iters = result
# create model
if resume_state: # resume training
check_resume(opt, resume_state['iter'])
model = create_model(opt)
model.resume_training(resume_state) # handle optimizers and schedulers
logger.info(f"Resuming training from epoch: {resume_state['epoch']}, "
f"iter: {resume_state['iter']}.")
start_epoch = resume_state['epoch']
current_iter = resume_state['iter']
else:
model = create_model(opt)
start_epoch = 0
current_iter = 0
# create message logger (formatted outputs)
msg_logger = MessageLogger(opt, current_iter, tb_logger) ##########记录loss的tensor board
# dataloader prefetcher
prefetch_mode = opt['datasets']['train'].get('prefetch_mode')
if prefetch_mode is None or prefetch_mode == 'cpu':
prefetcher = CPUPrefetcher(train_loader)
elif prefetch_mode == 'cuda':
prefetcher = CUDAPrefetcher(train_loader, opt)
logger.info(f'Use {prefetch_mode} prefetch dataloader')
if opt['datasets']['train'].get('pin_memory') is not True:
raise ValueError('Please set pin_memory=True for CUDAPrefetcher.')
else:
raise ValueError(f'Wrong prefetch_mode {prefetch_mode}.'
"Supported ones are: None, 'cuda', 'cpu'.")
# training
logger.info(
f'Start training from epoch: {start_epoch}, iter: {current_iter}')
data_time, iter_time = time.time(), time.time()
start_time = time.time()
for epoch in range(start_epoch, total_epochs + 1):
train_sampler.set_epoch(epoch)
prefetcher.reset()
train_data = prefetcher.next()
#print(train_data)
while train_data is not None:
data_time = time.time() - data_time
current_iter += 1
print('>>>>>>>>>>current iter', current_iter)
if current_iter > total_iters:
break
# update learning rate
model.update_learning_rate(
current_iter, warmup_iter=opt['train'].get('warmup_iter', -1))
# training
model.feed_data(train_data)
model.optimize_parameters(current_iter, tb_logger)
iter_time = time.time() - iter_time
# log
if current_iter % opt['logger']['print_freq'] == 0:
log_vars = {'epoch': epoch, 'iter': current_iter}
log_vars.update({'lrs': model.get_current_learning_rate()})
log_vars.update({'time': iter_time, 'data_time': data_time})
log_vars.update(model.get_current_log())
msg_logger(log_vars) #
# save models and training states
if current_iter % opt['logger']['save_checkpoint_freq'] == 0:
logger.info('Saving models and training states.')
model.save(epoch, current_iter)
# validation
if opt.get('val') is not None and (current_iter %
opt['val']['val_freq'] == 0):
model.validation(val_loader, current_iter, tb_logger,
opt['val']['save_img'])
data_time = time.time()
iter_time = time.time()
train_data = prefetcher.next()
# end of iter
# end of epoch
consumed_time = str(
datetime.timedelta(seconds=int(time.time() - start_time)))
logger.info(f'End of training. Time consumed: {consumed_time}')
logger.info('Save the latest model.')
model.save(epoch=-1, current_iter=-1) # -1 stands for the latest
if opt.get('val') is not None:
model.validation(val_loader, current_iter, tb_logger,
opt['val']['save_img'])
if tb_logger:
tb_logger.close()
if __name__ == '__main__':
main()