-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
218 lines (194 loc) · 12.2 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
import argparse
import logging
import os
import re
import time
import imgaug as ia
import pytorch_lightning as pl
import torch
import torch.utils.data
from pytorch_lightning.loggers import TensorBoardLogger
from datasets import get_dataset
from models import get_model
from strategies import get_strategy
from utils.exp_logger import ExpLogger
from utils.project_utils import load_config, process_config, save_config, set_logger
from utils.training_utils import ConcatDataset
parser = argparse.ArgumentParser()
parser.add_argument('-c', '--config', type=str, default='fundus/relay', help='Which config is loaded from configs')
parser.add_argument('-g', '--gpus', type=str, default=None, help='Gpu ids to use, like "1,2,3". If None, use cpu')
parser.add_argument('-n', '--note', type=str, default='default',
help='Note to identify this experiment, like "first_version"... Should not contain space')
parser.add_argument('-v', '--verbose', action='store_true', help='If set, log debug level, else info level')
args = parser.parse_args()
if args.gpus is not None:
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpus
exp_path = os.path.join('./exps', args.config, args.note)
def main():
# torch.autograd.set_detect_anomaly(True)
config_file = os.path.join('configs', args.config + '.yaml')
ori_config = load_config(config_file)
config = process_config(ori_config)
# backup config
save_config(os.path.join(exp_path, 'original_config_back.yaml'), ori_config)
save_config(os.path.join(exp_path, 'processed_config.yaml'), config)
# create logger to file and console
logging_path = os.path.join(exp_path, 'logging')
set_logger(logging_folder=logging_path, verbose=args.verbose)
logging_paths = {
'tensorboard': os.path.join(exp_path, 'tensorboard_summary'),
'checkpoint': os.path.join(exp_path, 'checkpoints'),
'exp_info_csv': os.path.join(exp_path, 'exp_results')
}
console_logger = logging.getLogger('PL_training')
tb_logger = TensorBoardLogger(logging_paths['tensorboard'], name='', version='', default_hp_metric=False)
exp_info_logger = ExpLogger(logging_paths['exp_info_csv'], name='', version='')
console_logger.info('Total number of tasks: %d' % len(config))
for task_conf in config:
console_logger.info('Begin to run task %s' % task_conf['dataset']['kwargs']['task_id'])
if task_conf['random_seed'] is not None:
seed = task_conf['random_seed'] + task_conf['dataset']['kwargs']['task_id']
pl.seed_everything(seed, workers=True)
ia.seed(seed)
# joint training and validation set
if task_conf['train'].get('joint_training', False):
train_datasets = []
val_datasets = []
for task_conf_ in config:
train_datasets.append(get_dataset(
task_conf_['dataset']['name'], train=True, **task_conf_['dataset']['kwargs']))
val_datasets.append(get_dataset(
task_conf_['dataset']['name'], train=False, **task_conf_['dataset']['kwargs']))
train_dataset = ConcatDataset(train_datasets, shuffle=True)
val_dataset = ConcatDataset(val_datasets, shuffle=False)
else:
train_dataset = get_dataset(
task_conf['dataset']['name'], train=True, **task_conf['dataset']['kwargs'])
val_dataset = get_dataset(
task_conf['dataset']['name'], train=False, **task_conf['dataset']['kwargs'])
test_datasets = []
for test_dataset_conf in task_conf['test_datasets']:
test_datasets.append(get_dataset(
test_dataset_conf['name'], train=False, **test_dataset_conf['kwargs']))
# prepare loaders
train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=task_conf['train']['batch_size'],
shuffle=True, num_workers=task_conf['num_workers'],
pin_memory=task_conf['train'].get('pin_memory', False),
prefetch_factor=task_conf['train'].get('prefetch', 2))
val_loader = torch.utils.data.DataLoader(val_dataset, batch_size=task_conf['train']['batch_size'],
shuffle=False, num_workers=task_conf['num_workers'],
pin_memory=task_conf['train'].get('pin_memory', False),
prefetch_factor=task_conf['train'].get('prefetch', 2))
test_loaders = [torch.utils.data.DataLoader(
test_dataset, batch_size=task_conf['train']['batch_size'], shuffle=False,
num_workers=task_conf['num_workers'], pin_memory=task_conf['train'].get('pin_memory', False))
for test_dataset in test_datasets]
num_task_epochs = task_conf['train'].get('num_epochs', None)
num_task_mimgs = task_conf['train'].get('num_mimgs', None)
assert (num_task_epochs is None) ^ (num_task_mimgs is None)
if num_task_mimgs is not None:
num_task_epochs = ((num_task_mimgs * 1000 * 1000) // len(train_loader.dataset)) + 1
additional_model_kwargs = {}
if 'solver_steps_ratio' not in task_conf['model']['kwargs']:
if task_conf['train'].get('solver_num_mimgs', None) is not None:
assert num_task_mimgs is not None
additional_model_kwargs['solver_steps_ratio'] = task_conf['train']['solver_num_mimgs'] / num_task_mimgs
elif task_conf['train'].get('solver_num_epochs', None) is not None:
assert num_task_epochs is not None
additional_model_kwargs['solver_steps_ratio'] = \
task_conf['train']['solver_num_epochs'] / num_task_epochs
model = get_model(task_conf['model']['name'], **dict(**task_conf['model']['kwargs'], **additional_model_kwargs))
if model.task != train_dataset.task:
console_logger.critical('Model task %s is different from dataset task %s'
% (model.task, train_dataset.task))
exit(-1)
cl_strategy = get_strategy(name=task_conf['strategy']['name'], model=model, task_conf=task_conf,
num_task_epochs=num_task_epochs, **task_conf['strategy']['kwargs'])
load_file = None
resume_file = None
if task_conf['checkpoint'].get('resume', None) is not None:
resume_file = os.path.join(logging_paths['checkpoint'], task_conf['checkpoint']['resume'])
elif task_conf['checkpoint'].get('load', None) is not None:
load_file = os.path.join(logging_paths['checkpoint'], task_conf['checkpoint']['load'])
save_dir = os.path.join(logging_paths['checkpoint'], task_conf['checkpoint']['save']['dir'])
if os.path.exists(save_dir):
if resume_file is None:
if 'last.ckpt' in os.listdir(save_dir):
console_logger.warning('Task %d seems to be finished because last.ckpt exists. Try next task'
% task_conf['dataset']['kwargs']['task_id'])
continue
ckpt_files = [f for f in os.listdir(save_dir) if f.endswith('.ckpt') and 'last' not in f
and 'epoch=' in f]
if len(ckpt_files) > 0:
ckpt_files = sorted(ckpt_files, key=lambda s: int(re.match(r'.*epoch=(\d+).*', s).group(1)))
resume_file = os.path.join(save_dir, ckpt_files[-1])
console_logger.warning('Model ckpt %s detected in save dir and will be used for resuming'
% resume_file)
current_task_max_epochs = num_task_epochs
if resume_file is not None:
state = torch.load(resume_file, map_location=torch.device('cpu'))
all_tasks_max_epochs_ = state['all_tasks_max_epochs']
current_task_max_epochs_ = state['current_task_max_epochs']
if current_task_max_epochs < current_task_max_epochs_:
console_logger.warning('num_epoch in config is smaller than that from resume file')
exit(0)
else:
all_tasks_max_epochs = all_tasks_max_epochs_ - current_task_max_epochs_ + current_task_max_epochs
del state
elif load_file is not None:
state = torch.load(load_file, map_location=torch.device('cpu'))
all_tasks_max_epochs_ = state['all_tasks_max_epochs']
all_tasks_actual_epochs_ = state['epoch']
last_task_max_epochs_ = state['current_task_max_epochs']
if all_tasks_actual_epochs_ != all_tasks_max_epochs_:
console_logger.warning('Training of the last task is not finished, '
'which is supposed to train %d epochs but only trained %d epochs.'
'We will skip and continue to train current task.'
% (last_task_max_epochs_, all_tasks_actual_epochs_
- (all_tasks_max_epochs_ - last_task_max_epochs_)))
all_tasks_max_epochs_ = all_tasks_actual_epochs_
all_tasks_max_epochs = all_tasks_max_epochs_ + current_task_max_epochs
del state
else:
all_tasks_max_epochs = current_task_max_epochs
checkpoint_saving_callback = pl.callbacks.ModelCheckpoint(
dirpath=os.path.join(logging_paths['checkpoint'], task_conf['checkpoint']['save']['dir']),
every_n_val_epochs=task_conf['train'].get('check_val_every_n_epoch', 1),
)
trainer = pl.Trainer(gpus=None if args.gpus is None else -1,
num_nodes=1,
accelerator=None if args.gpus is None else 'ddp',
benchmark=True,
max_epochs=all_tasks_max_epochs,
sync_batchnorm=task_conf['train'].get('sync_batchnorm', False),
gradient_clip_val=task_conf['train']['gradient_clip'],
accumulate_grad_batches=task_conf['train'].get('accumulate_grad_batches', 1),
precision=task_conf['train'].get('precision', 32),
callbacks=[checkpoint_saving_callback],
# plugins=pl.plugins.DDPPlugin(find_unused_parameters=False),
resume_from_checkpoint=resume_file,
check_val_every_n_epoch=task_conf['train'].get('check_val_every_n_epoch', 1),
logger=[tb_logger, exp_info_logger],
log_every_n_steps=task_conf['train']['log_every_n_steps'],
reload_dataloaders_every_epoch=task_conf['train'].get('reload_dataloaders_every_epoch',
False),
fast_dev_run=task_conf.get('fast_dev_run', False),
num_sanity_val_steps=0,
move_metrics_to_cpu=True,
)
if load_file is not None:
state = torch.load(load_file, map_location=torch.device('cpu'))
trainer.checkpoint_connector.restore_model_state(cl_strategy, state)
trainer.checkpoint_connector.restore_training_state(state, load_optimizer_states=False)
console_logger.info('Begin to train and val task %s' % task_conf['dataset']['kwargs']['task_id'])
console_logger.info('Current task num_epochs: %d. Total num_epochs: %d'
% (current_task_max_epochs, all_tasks_max_epochs))
trainer.fit(cl_strategy, train_dataloader=train_loader, val_dataloaders=val_loader)
console_logger.info('Begin to test %d viewed tasks after training of task %s'
% (len(task_conf['test_datasets']), task_conf['dataset']['kwargs']['task_id']))
trainer.test(cl_strategy, test_dataloaders=test_loaders, ckpt_path=None)
trainer.save_checkpoint(
os.path.join(logging_paths['checkpoint'], task_conf['checkpoint']['save']['dir'], 'last.ckpt'))
time.sleep(5)
if __name__ == '__main__':
main()