-
Notifications
You must be signed in to change notification settings - Fork 73
/
train_distributed.py
213 lines (188 loc) · 8.77 KB
/
train_distributed.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
import os
import sys
import PIL
import time
import random
import logging
import datetime
import os.path as osp
import torch
import torch.backends
import torch.nn as nn
import torch.backends.cudnn
import torch.distributed as dist
from utils.loss import OhemCELoss
from utils.utils import prepare_seed
from utils.utils import time_for_file
from evaluate_distributed import MscEval
from dataloaders import make_data_loader
from utils.logger import Logger, setup_logger
from utils.optimizer_distributed import Optimizer
from config_utils.retrain_config import config_factory
from retrain_model.build_autodeeplab import Retrain_Autodeeplab
from config_utils.re_train_autodeeplab import obtain_retrain_autodeeplab_args
def main():
args = obtain_retrain_autodeeplab_args()
torch.cuda.set_device(args.local_rank)
cfg = config_factory['resnet_cityscapes']
if not os.path.exists(cfg.respth):
os.makedirs(cfg.respth)
dist.init_process_group(
backend='nccl',
init_method='tcp://127.0.0.1:{}'.format(cfg.port),
world_size=torch.cuda.device_count(),
rank=args.local_rank
)
setup_logger(cfg.respth)
logger = logging.getLogger()
rand_seed = random.randint(0, args.manualSeed)
prepare_seed(rand_seed)
if args.local_rank == 0:
log_string = 'seed-{}-time-{}'.format(rand_seed, time_for_file())
train_logger = Logger(args, log_string)
train_logger.log('Arguments : -------------------------------')
for name, value in args._get_kwargs():
train_logger.log('{:16} : {:}'.format(name, value))
train_logger.log("Python version : {}".format(sys.version.replace('\n', ' ')))
train_logger.log("Pillow version : {}".format(PIL.__version__))
train_logger.log("PyTorch version : {}".format(torch.__version__))
train_logger.log("cuDNN version : {}".format(torch.backends.cudnn.version()))
train_logger.log("random_seed : {}".format(rand_seed))
if args.checkname is None:
args.checkname = 'deeplab-' + str(args.backbone)
# dataset
kwargs = {'num_workers': args.workers, 'pin_memory': True, 'drop_last': True}
train_loader, args.num_classes, sampler = make_data_loader(args=args, **kwargs)
# model
model = Retrain_Autodeeplab(args)
model.train()
model.cuda()
model = nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank, ], output_device=args.local_rank,
find_unused_parameters=True).cuda()
n_min = cfg.ims_per_gpu * cfg.crop_size[0] * cfg.crop_size[1] // 16
criterion = OhemCELoss(thresh=cfg.ohem_thresh, n_min=n_min).cuda()
max_iteration = int(cfg.max_epoch * len(train_loader))
# max_iteration = int(1500000 * 4 // cfg.gpus)
it = 0
# optimizer
optimizer = Optimizer(model, cfg.lr_start, cfg.momentum, cfg.weight_decay, cfg.warmup_steps,
cfg.warmup_start_lr, max_iteration, cfg.lr_power)
if dist.get_rank() == 0:
print('======optimizer launch successfully , max_iteration {:}!======='.format(max_iteration))
# train loop
loss_avg = []
start_time = glob_start_time = time.time()
# for it in range(cfg.max_iter):
if args.resume is not None:
checkpoint = torch.load(args.resume, map_location='cpu')
if checkpoint['iter'] is not None:
args.train_mode = 'iter'
start_iter = checkpoint['iter']
n_epoch = checkpoint['epoch']
elif checkpoint['epoch'] is not None:
args.train_mode = 'epoch'
model.module.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'], checkpoint['iter'])
else:
if args.train_mode == 'iter':
start_iter = 0
n_epoch = 0
elif args.train_mode == 'epoch':
start_epoch = 0
if args.train_mode is 'iter':
diter = iter(train_loader)
for it in range(start_iter, cfg.max_iter):
try:
sample = next(diter)
except StopIteration:
n_epoch += 1
sampler.set_epoch(n_epoch)
diter = iter(train_loader)
sample = next(diter)
im, lb = sample['image'].cuda(), sample['label'].cuda()
lb = torch.squeeze(lb, 1)
optimizer.zero_grad()
logits = model(im)
loss = criterion(logits, lb)
loss.backward()
optimizer.step()
loss_avg.append(loss.item())
# print training log message
if it % cfg.msg_iter == 0 and not it == 0 and dist.get_rank() == 0:
loss_avg = sum(loss_avg) / len(loss_avg)
lr = optimizer.lr
ed = time.time()
t_intv, glob_t_intv = ed - start_time, ed - glob_start_time
eta = int((max_iteration - it) * (glob_t_intv / it))
eta = str(datetime.timedelta(seconds=eta))
msg = ', '.join(['iter: {it}/{max_iteration}', 'lr: {lr:4f}', 'loss: {loss:.4f}', 'eta: {eta}', 'time: {time:.4f}',
]).format(it=it, max_iteration=max_iteration, lr=lr, loss=loss_avg, time=t_intv, eta=eta)
# TODO : now the logger.info will error if iter > 350000, so use print haha
if max_iteration > 350000:
logger.info(msg)
else:
print(msg)
loss_avg = []
it += 1
if (cfg.msg_iter is not None) and (it % cfg.msg_iter == 0) and (it != 0):
if args.verbose:
logger.info('evaluating the model of iter:{}'.format(it))
model.eval()
evaluator = MscEval(cfg, args)
mIOU, loss = evaluator(model, criteria=criterion, multi_scale=False)
logger.info('mIOU is: {}, loss_eval is {}'.format(mIOU, loss))
model.cpu()
save_name = 'iter_{}_naive_model.pth'.format(it)
save_pth = osp.join(cfg.respth, save_name)
state = model.module.state_dict() if hasattr(model, 'module') else model.state_dict()
checkpoint = {'state_dict': state, 'epoch': n_epoch, 'iter': it, 'optimizer': optimizer.optim.state_dict()}
if dist.get_rank() == 0:
torch.save(state, save_pth)
logger.info('model of iter {} saved to: {}'.format(it, save_pth))
model.cuda()
model.train()
elif args.train_mode is 'epoch':
for epoch in range(start_epoch, cfg.max_epoch):
for i, sample in enumerate(train_loader):
im = sample['image'].cuda()
lb = sample['label'].cuda()
lb = torch.squeeze(lb, 1)
optimizer.zero_grad()
logits = model(im)
loss = criterion(logits, lb)
loss.backward()
optimizer.step()
loss_avg.append(loss.item())
# print training log message
if i % cfg.msg_iter == 0 and not (i == 0 and epoch == 0) and dist.get_rank() == 0:
loss_avg = sum(loss_avg) / len(loss_avg)
lr = optimizer.lr
ed = time.time()
t_intv, glob_t_intv = ed - start_time, ed - glob_start_time
eta = int((max_iteration - it) * (glob_t_intv / it))
eta = str(datetime.timedelta(seconds=eta))
msg = ', '.join(['iter: {it}/{max_iteration}', 'lr: {lr:4f}', 'loss: {loss:.4f}', 'eta: {eta}', 'time: {time:.4f}',
]).format(it=it, max_iteration=max_iteration, lr=lr, loss=loss_avg, time=t_intv, eta=eta)
logger.info(msg)
loss_avg = []
# save model and optimizer each epoch
if args.verbose:
logger.info('evaluating the model of iter:{}'.format(it))
model.eval()
evaluator = MscEval(cfg, args)
mIOU, loss = evaluator(model, criteria=criterion, multi_scale=False)
logger.info('mIOU is: {}, loss_eval is {}'.format(mIOU, loss))
model.cpu()
save_name = 'iter_{}_naive_model.pth'.format(it)
save_pth = osp.join(cfg.respth, save_name)
state = model.module.state_dict() if hasattr(model, 'module') else model.state_dict()
checkpoint = {'state_dict': state, 'epoch': n_epoch, 'iter': it, 'optimizer': optimizer.state_dict()}
if dist.get_rank() == 0:
torch.save(state, save_pth)
logger.info('model of iter {} saved to: {}'.format(it, save_pth))
model.cuda()
model.train()
else:
raise NotImplementedError
if __name__ == "__main__":
main()