-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrain.py
191 lines (168 loc) · 7.84 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
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Created by: jasonseu
# Created on: 2021-8-9
# Email: [email protected]
#
# Copyright © 2021 - CPSS Group
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
import os
import sys
import time
import random
import traceback
import torch
import torch.nn as nn
import torch.distributed as dist
import torch.multiprocessing as mp
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
from lib.utils import *
from lib.metrics import *
from lib.dataset import MLDataset
from models.factory import create_model
class Trainer(object):
def __init__(self, cfg, world_size, rank):
super(Trainer, self).__init__()
self.distributed = world_size > 1
batch_size = cfg.batch_size // world_size if self.distributed else cfg.batch_size
train_dataset = MLDataset(cfg, cfg.train_path, training=True)
val_dataset = MLDataset(cfg, cfg.test_path, training=False)
if self.distributed:
self.train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset)
val_sampler = torch.utils.data.distributed.DistributedSampler(val_dataset)
else:
self.train_sampler = val_sampler = None
self.train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=(self.train_sampler is None),
num_workers=4, sampler=self.train_sampler)
self.val_loader = DataLoader(val_dataset, batch_size=batch_size, shuffle=False, num_workers=4, sampler=val_sampler)
torch.cuda.set_device(rank)
self.model = create_model(cfg.model, cfg=cfg)
self.model.cuda(rank)
if cfg.pretrained_model is not None:
state_dict = torch.load(cfg.pretrained_model)
new_state_dict = {}
for k, v in state_dict.items():
k = k[7:]
if k.startswith('query_embed') or k.startswith('net') or k.startswith('fc'):
continue
new_state_dict[k] = v
self.model.load_state_dict(new_state_dict, strict=False)
print('loaded pretrained model from {}!'.format(cfg.pretrained_model))
self.ema_model = ModelEma(self.model, decay=cfg.ema_decay)
if self.distributed:
self.model = nn.SyncBatchNorm.convert_sync_batchnorm(self.model)
self.model = nn.parallel.DistributedDataParallel(self.model, device_ids=[rank], find_unused_parameters=True)
parameters = self.model.parameters()
self.optimizer = get_optimizer(parameters, cfg)
self.lr_scheduler = get_lr_scheduler(self.optimizer, cfg, steps_per_epoch=len(self.train_loader))
self.criterion = get_loss_fn(cfg)
self.voc_mAP = VOCmAP(cfg.num_classes, year='2012', ignore_path=cfg.ignore_path)
self.voc_ema_mAP = VOCmAP(cfg.num_classes, year='2012', ignore_path=cfg.ignore_path)
self.cfg = cfg
self.best_mAP = 0
self.global_step = 0
self.notdist_or_rank0 = (not self.distributed) or (self.distributed and rank == 0)
if self.notdist_or_rank0:
self.logger = get_logger(cfg.log_path, __name__)
self.logger.info(train_dataset.transform)
self.logger.info(val_dataset.transform)
self.writer = SummaryWriter(log_dir=cfg.exp_dir)
def run(self):
patience = 0
for epoch in range(self.cfg.max_epochs):
if self.distributed:
self.train_sampler.set_epoch(epoch)
self.train(epoch)
mAP = self.validation(epoch)
if self.best_mAP < mAP and self.notdist_or_rank0:
torch.save(self.ema_model.state_dict(), self.cfg.ckpt_ema_best_path)
self.best_mAP = mAP
patience = 0
else:
patience += 1
if self.cfg.estop and patience > 2:
break
if self.notdist_or_rank0:
self.logger.info('\ntraining over, best validation score: {} mAP'.format(self.best_mAP))
def train(self, epoch):
scaler = torch.cuda.amp.GradScaler(enabled=self.cfg.amp)
self.model.train()
for _, batch in enumerate(self.train_loader):
batch_begin = time.time()
imgs = batch['img'].cuda()
targets = batch['target'].cuda()
with torch.cuda.amp.autocast(enabled=self.cfg.amp):
ret = self.model(imgs, targets)
bce, ctr = ret['bce'], ret['ctr']
loss = bce + self.cfg.lamda * ctr
self.optimizer.zero_grad()
scaler.scale(loss).backward()
scaler.step(self.optimizer)
scaler.update()
dur = time.time() - batch_begin
self.lr_scheduler.step()
self.ema_model.update(self.model)
if self.global_step % (len(self.train_loader) // 6) == 0 and self.notdist_or_rank0:
lr = get_lr(self.optimizer)
self.writer.add_scalar('Loss/train', loss, self.global_step)
self.writer.add_scalar('bce/train', bce, self.global_step)
self.writer.add_scalar('ctr/train', ctr, self.global_step)
self.writer.add_scalar('lr', lr, self.global_step)
self.logger.info('TRAIN [epoch {}] loss: {:4f} bce: {:4f} ctr: {:4f} lr:{:.6f} time:{:.4f}'
.format(epoch, loss, bce, ctr, lr, dur))
self.global_step += 1
@torch.no_grad()
def validation(self, epoch):
self.model.eval()
self.ema_model.eval()
self.voc_mAP.reset()
self.voc_ema_mAP.reset()
for batch in self.val_loader:
imgs = batch['img'].cuda()
targets = batch['target'].cuda()
logits = self.model(imgs)['logits']
scores = torch.sigmoid(logits)
logits = self.ema_model(imgs)['logits']
ema_scores = torch.sigmoid(logits)
if self.distributed:
scores = concat_all_gather(scores)
ema_scores = concat_all_gather(ema_scores)
targets = concat_all_gather(targets)
targets = targets.cpu().numpy()
scores = scores.detach().cpu().numpy()
self.voc_mAP.update(scores, targets)
ema_scores = ema_scores.detach().cpu().numpy()
self.voc_ema_mAP.update(ema_scores, targets)
if self.distributed:
dist.barrier()
_, mAP = self.voc_mAP.compute()
_, ema_mAP = self.voc_ema_mAP.compute()
if self.notdist_or_rank0:
self.writer.add_scalar('mAP/val', mAP, self.global_step)
self.writer.add_scalar('ema_mAP/val', ema_mAP, self.global_step)
self.logger.info("VALID [epoch {}] mAP: {:.4f} ema_mAP: {:.4f} best mAP: {:.4f}"
.format(epoch, mAP, ema_mAP, max(ema_mAP, self.best_mAP)))
return ema_mAP
def main_worker(local_rank, ngpus_per_node, cfg, port=None):
world_size = ngpus_per_node # only single node is enough.
if ngpus_per_node > 1:
init_method = 'tcp://127.0.0.1:{}'.format(port)
dist.init_process_group(backend='nccl', init_method=init_method, world_size=world_size, rank=local_rank)
trainer = Trainer(cfg, world_size, local_rank)
trainer.run()
if __name__ == "__main__":
args = get_args()
cfg = prepare_env(args, sys.argv)
try:
ngpus_per_node = torch.cuda.device_count()
if ngpus_per_node > 1:
port = get_port()
setup_seed(cfg.seed)
mp.spawn(main_worker, args=(ngpus_per_node, cfg, port,), nprocs=ngpus_per_node)
else:
setup_seed(cfg.seed)
main_worker(0, ngpus_per_node, cfg)
except (Exception, KeyboardInterrupt):
print(traceback.format_exc())
if not os.path.exists(cfg.ckpt_ema_best_path):
clear_exp(cfg.exp_dir)