-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
executable file
·194 lines (173 loc) · 9.04 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
import argparse
import collections
import copy
import glob
import os
import os.path as osp
import time
from collections import ChainMap
import numpy as np
import torch.utils.data
from tensorboardX import SummaryWriter
from torch import optim
from torchvision import transforms
from tqdm import tqdm
from Vision import Model, file, build_dataset, Augmenter, Normalizer, collate_fn, Resizer, build_optimizer, \
collect_env, get_root_logger, ComputeLoss, build_generator, get_by_key, model_info, \
map_compute, initialize_weights, tensorboard_event_vis
parser = argparse.ArgumentParser()
parser.add_argument('--config', '-c', type=str, default='Vision/configs/retinanet_fpn_r50.yaml')
parser.add_argument('--resume_from', type=str, default=None)
parser.add_argument('--pretrained', type=str, default=None)
parser.add_argument('--exclude_layer', type=list, default=['fc.weight', 'fc.bias'],
help='layers weights to del,default is repvgg.')
parser.add_argument('--epochs', type=int, default=100, help='number of total epochs')
parser.add_argument('--best_map', type=int, default=0.3)
parser.add_argument('--num_workers', type=int, default=8)
parser.add_argument('--n_gpu', type=str, default='0', help='GPU device ids')
parser.add_argument('--log_iter', type=int, default=20)
parser.add_argument('--eval_epoch', type=int, default=1)
parser.add_argument('--freeze', type=int or list, default=0, help='Number of layers to freeze(list(int) or int). '
'backbone=6, all=15')
parser.add_argument('--deploy', type=bool, default=False, help='switch the model mode(False for train,True to fuse '
'model).')
parser.add_argument('--use_dbb', type=bool, default=False, help='whether to use diverse branch block.')
parser.add_argument('--use_cot', type=bool, default=False, help='whether to use CoTLayer')
opt = parser.parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = opt.n_gpu
if opt.resume_from is not None:
config = file.load(glob.glob(f'{opt.resume_from}/*.json')[0])
weights = torch.load(glob.glob(f'{opt.resume_from}/epoch_23_0.705.pth')[0])
else:
config = file.load(opt.config)
weights = None
data_cfg = config.get('dataset')
train_cfg = ChainMap(data_cfg.get('base'), data_cfg.get('train'))
eval_cfg = ChainMap(data_cfg.get('base'), data_cfg.get('eval'))
# logger
timestamp = time.strftime('%Y%m%d_%H%M%S', time.localtime())
logpath = f'runs/{config.get("cfg")}/{train_cfg.get("cfg")}/{timestamp}'.lower()
if not osp.exists(logpath):
os.makedirs(logpath)
file.save(config, osp.join(logpath, f'{timestamp}.json'))
writer = SummaryWriter(logdir=logpath, filename_suffix=f'_{timestamp}')
log_file = osp.join(logpath, f'{timestamp}.log')
logger = get_root_logger(name=f'{train_cfg["cfg"]}'.lower(), log_file=log_file, log_level='INFO',
format_='color_format')
env_info_dict = collect_env()
env_info = '\n'.join([f'{k}: {v}' for k, v in env_info_dict.items()])
dash_line = '-' * 60 + '\n'
logger.info('Environment info:\n' + dash_line + env_info + '\n' + dash_line)
# model
model = Model(config['model'], logger=logger, deploy=opt.deploy, use_dbb=opt.use_dbb, use_cot=opt.use_cot)
initialize_weights(model)
if weights is not None:
model.load_pretrained(weights)
elif opt.pretrained is not None:
model.load_pretrained(torch.load(opt.pretrained), exclude=opt.exclude_layer)
if torch.cuda.is_available():
model = torch.nn.DataParallel(model).cuda()
else:
model = torch.nn.DataParallel(model)
device = next(model.parameters()).device
# Freeze module
freeze = [f'model.{x}.' for x in range(opt.freeze)] if isinstance(opt.freeze,
int) else opt.freeze if opt.freeze else None # layers to freeze
logger.info(f'Frozen module: {freeze}' + '\n' + dash_line)
if freeze is not None:
for k, v in model.named_parameters():
v.requires_grad = True # train all layers
if any(x in k for x in freeze):
v.requires_grad = False
model_info(model, verbose=True, logger=logger)
# dataset
train_dataset = build_dataset(cfg=train_cfg, transform=transforms.Compose([
Resizer(img_sizes=train_cfg.get('img_sizes')),
Normalizer()]))
train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=train_cfg.get('batch_size'), collate_fn=collate_fn,
shuffle=True, pin_memory=True)
# sampler = WeightedRandomSampler: 根据各样本权重选择数据, RandomSampler：shuffle=True时调用 # https://zhuanlan.zhihu.com/p/117270644
eval_dataset = build_dataset(cfg=eval_cfg, transform=transforms.Compose([Resizer(img_sizes=eval_cfg.get('img_sizes')),
Normalizer()]))
eval_loader = torch.utils.data.DataLoader(eval_dataset, batch_size=eval_cfg.get('batch_size'), collate_fn=collate_fn)
# optimizer
optimizer = build_optimizer(config.get('optimizer'), params=model.parameters())
# scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, patience=2, verbose=True)
# scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=5, eta_min=0.001)
# scheduler = optim.lr_scheduler.CosineAnnealingWarmRestarts(optimizer, T_0=5, T_mult=2)
scheduler = optim.lr_scheduler.MultiStepLR(optimizer, milestones=[30, 60], gamma=0.1)
compute_loss = ComputeLoss(config['loss'], config['generator'])
generater = build_generator(config['generator'])
names = get_by_key(config, 'CLASSES_NAME')
best_map = opt.best_map
for epoch in range(opt.epochs):
model.training = True
model.train()
# epoch_loss = []
if epoch == 2:
for k, v in model.named_parameters():
v.requires_grad = True # train all layers
for batch_step, data in enumerate(train_loader):
global_steps = len(train_loader) * epoch + batch_step
images, annots = data['img'].to(device), data['annot'].to(device)
anchors = generater.grid_anchors(images)
optimizer.zero_grad()
# for pram in optimizer.param_groups:
# lr = pram['lr']
lr = scheduler.get_last_lr()[0]
preds = model.forward(images, visualize=False)
losses_dict = compute_loss(preds, annots, anchors)
loss = losses_dict['total_loss']
if bool(loss == 0):
continue
for k, v in losses_dict.items():
writer.add_scalar(k, v, global_step=global_steps)
writer.add_scalar('lr', lr, global_step=epoch)
# flood = (loss - b).abs() + b
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 1)
optimizer.step()
# epoch_loss.append(float(loss))
logg_info = f'Epoch: %3s | Iteration: %8s/%4s/%-4s' % (epoch + 1, global_steps, batch_step, len(train_loader))
for k, v in losses_dict.items():
logg_info += f' | %-6s: %1.5f' % (k, v)
if batch_step % opt.log_iter == 0:
logger.info(logg_info + f' | lr: %1.5f' % lr)
del loss
# eval loss
# with torch.no_grad():
# for batch_step, data in tqdm(enumerate(eval_loader), 'eval...'):
# _steps = len(eval_loader) * epoch + batch_step
# images, annots = data['img'].to(device), data['annot'].to(device)
# anchors = generater.grid_anchors(images)
# preds = model.forward(images, visualize=False)
# losses_dict = compute_loss(preds, annots, anchors)
# for k, v in losses_dict.items():
# writer.add_scalar(f'eval_{k}', v, global_step=_steps)
# torch.save(model.state_dict(), f'{logpath}/epoch_{epoch+1}.pth')
# model_fused = copy.deepcopy(model)
# for module in model_fused.modules():
# if hasattr(module, 'switch_to_deploy'):
# module.switch_to_deploy()
# torch.save(model_fused.state_dict(), f'{logpath}/fused_last_epoch.pth')
scheduler.step()
config['epoch'] = epoch + 1
config['steps'] = len(train_loader) * (epoch + 1)
config['optimizer']['lr'] = lr # 保存当前学习率
file.save(config, osp.join(logpath, f'{timestamp}.json'))
if (epoch + 1) % opt.eval_epoch == 0:
torch.cuda.empty_cache()
model.training = False
model.eval()
map50 = map_compute(model, eval_loader, logger, iou_thresh=0.5, conf_thresh=0.01, nc=len(names[1:]), plots=True,
save_dir=logpath, loss_computer=compute_loss, writer=writer, epoch=epoch,
generater=generater, class_name=names, multi_label=False, agnostic=False, retur=True)
if map50 > best_map:
best_map = map50
# torch.save(model.state_dict(), '{}/epoch_{}_{:.3f}.pth'.format(logpath, epoch + 1, map50))
torch.save(model.state_dict(), '{}/best.pth'.format(logpath))
torch.save(model.state_dict(), '{}/last.pth'.format(logpath))
writer.export_scalars_to_json(osp.join(logpath, f'{timestamp}_scalars.json'))
model.eval()
# tensorboard_event_vis()(glob.glob(f'{logpath}/events.out*')[0], 'total_loss', f'{logpath}/total_loss.txt',
# f'{logpath}/total_loss', 'upper right')