-
Notifications
You must be signed in to change notification settings - Fork 5
/
train.py
245 lines (209 loc) · 11.1 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
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
import datetime
import logging
import math
import os.path as osp
import time
from os import path as osp
import torch
from basicsr.archs import build_network
from basicsr.data import build_dataloader, build_dataset
from basicsr.data.data_sampler import EnlargedSampler
from basicsr.data.prefetch_dataloader import CPUPrefetcher, CUDAPrefetcher
from basicsr.models import build_model
from basicsr.utils import (AvgTimer, MessageLogger, check_resume, get_env_info, get_root_logger, get_time_str,
init_tb_logger, init_wandb_logger, scandir)
from basicsr.utils.options import copy_opt_file, dict2str
from torch.utils.data import ConcatDataset
import archs # noqa
import data # noqa
import models # noqa
from utils import parse_options, make_exp_dirs, mkdir_and_rename
def init_tb_loggers(opt):
# initialize wandb logger before tensorboard logger to allow proper sync
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)
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['root_path'], 'tb_logger', opt['name']))
return tb_logger
def create_train_val_dataloader(opt, logger):
# create train and val dataloaders
train_loader, val_loaders = None, []
for phase, dataset_opt in opt['datasets'].items():
dataset_opt['bit'] = opt['bit']
dataset_opt['scale'] = opt['scale']
dataset_opt['gt_size'] = opt['train']['gt_size']
dataset_opt['batch_size_per_gpu'] = opt['train']['batch_size_per_gpu']
if phase == 'train':
dataset_enlarge_ratio = dataset_opt.get('dataset_enlarge_ratio', 1)#如果有就返回该值,没有就返回1
train_sets = [build_dataset(dataset_opt)]
datasets_used = f"\n\t\t{dataset_opt['name']}: {len(train_sets[0])} images" #得到一个字符串
extra_datasets = dataset_opt.get('extra_datasets', None)
if extra_datasets is not None:
for _, extra_dataset in extra_datasets.items(): #对字典迭代
_dataset_opt = dataset_opt
_dataset_opt['name'] = extra_dataset['name']
_dataset_opt['dataroot_gt'] = extra_dataset['dataroot_gt']
_dataset_opt['dataroot_lq'] = extra_dataset['dataroot_lq']
_dataset_opt['meta_info_file'] = extra_dataset.get('meta_info_file', 'None')
_dataset_opt['filename_tmpl'] = extra_dataset.get('filename_tmpl', '{}')
_dataset_opt['io_backend'] = extra_dataset['io_backend']
_extra_dataset = build_dataset(_dataset_opt)
train_sets.append(_extra_dataset)
datasets_used = f"{datasets_used}\n\t\t{_dataset_opt['name']}: {len(_extra_dataset)} images"
train_set = ConcatDataset(train_sets)
train_sampler = EnlargedSampler(train_set, opt['world_size'], opt['rank'], dataset_enlarge_ratio)
train_loader = build_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:' #在log中打印
f'\n\tTraining set(s) used: {datasets_used}'
f'\n\tNumber of total 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.split('_')[0] == 'val': #在_位置分割,0代表前面的
val_set = build_dataset(dataset_opt)
val_loader = build_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"]}: {len(val_set)}')
val_loaders.append(val_loader)
else:
raise ValueError(f'Dataset phase {phase} is not recognized.')
return train_loader, train_sampler, val_loaders, total_epochs, total_iters
def load_resume_state(opt):
resume_state_path = None
if opt['auto_resume']:
state_path = osp.join('experiments', opt['name'], 'training_states')
if osp.isdir(state_path):
states = list(scandir(state_path, suffix='state', recursive=False, full_path=False))
if len(states) != 0:
states = [float(v.split('.state')[0]) for v in states]
resume_state_path = osp.join(state_path, f'{max(states):.0f}.state')
opt['path']['resume_state'] = resume_state_path
else:
if opt['path'].get('resume_state'):
resume_state_path = opt['path']['resume_state']
if resume_state_path is None:
resume_state = None
else:
device_id = torch.cuda.current_device()
resume_state = torch.load(resume_state_path, map_location=lambda storage, loc: storage.cuda(device_id))
check_resume(opt, resume_state['iter'])
return resume_state
def train_pipeline(root_path):
# parse options, set distributed setting, set random seed
opt, args = parse_options(root_path, is_train=True)
opt['root_path'] = root_path
torch.backends.cudnn.benchmark = True
# torch.backends.cudnn.deterministic = True
# load resume states if necessary
resume_state = load_resume_state(opt) #选择最大的.state
# mkdir for experiments and logger
if resume_state is None:
make_exp_dirs(opt)
if opt['logger'].get('use_tb_logger') and 'debug' not in opt['name'] and opt['rank'] == 0:
mkdir_and_rename(osp.join(opt['root_path'], 'tb_logger', opt['name']))
# copy the yml file to the experiment root
copy_opt_file(args.expe, opt['path']['experiments_root'])
copy_opt_file(args.task, opt['path']['experiments_root'])
# WARNING: should not use get_root_logger in the above codes, including the called functions
# Otherwise the logger will not be properly initialized
log_file = osp.join(opt['path']['log'], f"train_{opt['name']}_{get_time_str()}.log")
logger = get_root_logger(logger_name='basicsr', log_level=logging.INFO, log_file=log_file)
logger.info(get_env_info())
# initialize wandb and tb loggers
tb_logger = init_tb_loggers(opt)
# create train and validation dataloaders
result = create_train_val_dataloader(opt, logger)
train_loader, train_sampler, val_loaders, total_epochs, total_iters = result
# create model
model = build_model(opt)
if resume_state: # resume training
model.resume_training(resume_state) # handle optimizers and schedulers
logger.info(f"Resuming training from epoch: {resume_state['epoch']}, iter: {resume_state['iter']}.")
start_epoch = resume_state['epoch']
current_iter = resume_state['iter']
else:
start_epoch = 0
current_iter = 0
# create message logger (formatted outputs)
msg_logger = MessageLogger(opt, current_iter, tb_logger)
# dataloader prefetcher
prefetch_mode = opt['datasets']['train'].get('prefetch_mode')
if prefetch_mode is None or prefetch_mode == 'cpu':
prefetcher = CPUPrefetcher(train_loader) #提高训练效率
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_timer, iter_timer = AvgTimer(), AvgTimer()
start_time = time.time()
for epoch in range(start_epoch, total_epochs + 1):
train_sampler.set_epoch(epoch)
prefetcher.reset()
train_data = prefetcher.next()
while train_data is not None:
data_timer.record()
current_iter += 1
if current_iter > total_iters:
break
# update learning rate warmup
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)
iter_timer.record()
if current_iter == 1:
# reset start time in msg_logger for more accurate eta_time
# not work in resume mode
msg_logger.reset_start_time()
# log 打印训练状态
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_timer.get_avg_time(), 'data_time': data_timer.get_avg_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):
for val_loader in val_loaders:
model.validation(val_loader, current_iter, tb_logger, opt['val']['save_val_img'])
data_timer.start()
iter_timer.start()
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:
for val_loader in val_loaders:
model.validation(val_loader, current_iter, tb_logger, opt['val']['save_img'])
if tb_logger:
tb_logger.close()
if __name__ == '__main__':
root_path = osp.abspath(osp.join(__file__, osp.pardir))
train_pipeline(root_path)