-
Notifications
You must be signed in to change notification settings - Fork 30
/
Copy pathtrain_fp16.py
266 lines (240 loc) · 10.3 KB
/
train_fp16.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
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
import sys
import os
curPath = os.path.abspath(os.path.dirname(__file__))
rootPath = os.path.split(curPath)[0]
sys.path.append(rootPath)
import argparse
import torch
import time
import os.path as osp
from torch.nn.parallel import DataParallel
import torch.nn as nn
import collections
from torch.autograd import Variable
import visdom
from UW.utils import Config
from UW.core.Models import build_network
from UW.core.Datasets import build_dataset, build_dataloader
from UW.core.Optimizer import build_optimizer, build_scheduler
from UW.utils import (mkdir_or_exist, get_root_logger,
save_epoch, save_latest, save_item, load_part,
resume, load)
from UW.core.Losses import build_loss
from UW.utils.Visualizer import Visualizer
from UW.utils.save_image import normimage, normPRED
import numpy as np
from apex import amp
from apex.parallel import convert_syncbn_model
from apex.parallel import DistributedDataParallel as DDP
from tensorboardX import SummaryWriter
from getpass import getuser
from socket import gethostname
def get_host_info():
return f'{getuser()}@{gethostname()}'
def parse_args():
parser = argparse.ArgumentParser(description='Train a detector')
parser.add_argument('--config',type=str, default='/home/dong/GitHub_Frame/UW/config/UWCNN.py',
help='train config file path')
parser.add_argument('--work_dir', help='the dir to save logs and models,')
group_gpus = parser.add_mutually_exclusive_group()
group_gpus.add_argument(
'--gpus',
type=int,
help='number of gpus to use '
'(only applicable to non-distributed training)')
group_gpus.add_argument(
'--gpu-ids',
type=int,
nargs='+',
help='ids of gpus to use '
'(only applicable to non-distributed training)')
args = parser.parse_args()
return args
if __name__ == '__main__':
args = parse_args()
cfg = Config.fromfile(args.config)
if args.work_dir is not None:
# update configs according to CLI args if args.work_dir is not None
cfg.work_dir = args.work_dir
elif cfg.get('work_dir', None) is None:
# use config filename as default work_dir if cfg.work_dir is None
cfg.work_dir = osp.join('./work_dirs',
osp.splitext(osp.basename(args.config))[0])
if args.gpu_ids is not None:
cfg.gpu_ids = args.gpu_ids
else:
cfg.gpu_ids = range(1) if args.gpus is None else range(args.gpus)
mata = dict()
# make dirs
mkdir_or_exist(osp.abspath(cfg.work_dir))
timestamp = time.strftime('%Y%m%d_%H%M%S', time.localtime())
cfg.log_file = osp.join(cfg.work_dir, f'{timestamp}.log')
# create text log
logger = get_root_logger(log_file=cfg.log_file, log_level=cfg.log_level)
dash_line = '-' * 60 + '\n'
logger.info(dash_line)
logger.info(f'Config:\n{cfg.pretty_text}')
# build model
model = build_network(cfg.model, train_cfg=cfg.train_cfg, test_cfg=cfg.test_cfg)
logger.info('-' * 20 + 'finish build model' + '-' * 20)
logger.info('Total Parameters: %d, Trainable Parameters: %s',
model.net_parameters['Total'],
str(model.net_parameters['Trainable']))
optimizer = build_optimizer(model, cfg.optimizer)
if cfg.resume_from:
start_epoch, ite_num = resume(cfg.resume_from, model, optimizer, logger, )
elif cfg.load_from:
load(cfg.load_from, model, logger)
# model = convert_syncbn_model(model)
# build dataset
datasets = build_dataset(cfg.data.train)
logger.info('-' * 20 + 'finish build dataset' + '-' * 20)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# put model on gpu
if torch.cuda.is_available():
# device = torch.device(cfg.gpu_ids)
model = model.to(device)
model.train()
model, optimizer = amp.initialize(model, optimizer, opt_level="O1")
# model = DDP(model, delay_allreduce=True)
# model = DDP(model, delay_allreduce=True)
# create data_loader
data_loader = build_dataloader(
datasets,
cfg.data.samples_per_gpu,
cfg.data.workers_per_gpu,
len(cfg.gpu_ids))
logger.info('-' * 20 + 'finish build dataloader' + '-' * 20)
# create optimizer
Scheduler = build_scheduler(cfg.lr_config)
logger.info('-' * 20 + 'finish build optimizer' + '-' * 20)
visualizer = Visualizer()
vis = visdom.Visdom()
criterion_ssim_loss = build_loss(cfg.loss_ssim)
criterion_l1_loss = build_loss(cfg.loss_l1)
ite_num = 0
start_epoch = 1 # start range at 1-1 = 0
running_loss = 0.0
running_tar_loss = 0.0
ite_num4val = 0
max_iters = cfg.total_epoch * len(data_loader)
print("---start training...")
scheduler = Scheduler(optimizer, cfg)
# before run
t = time.time()
log_dir = osp.join(cfg.work_dir, 'tf_logs')
write = SummaryWriter(log_dir)
logger.info('Start running, host: %s, work_dir: %s',
get_host_info(), cfg.work_dir)
logger.info('max: %d epochs, %d iters', cfg.total_epoch, max_iters)
amp_handle = amp.init(enabled=True)
for epoch in range(start_epoch-1, cfg.total_epoch):
# before epoch
logger.info('\nStart Epoch %d -------- ', epoch+1)
for i, data in enumerate(data_loader):
# before iter
data_time = time.time()-t
ite_num = ite_num + 1
ite_num4val = ite_num*cfg.data.samples_per_gpu
inputs, gt = data['image'], data['gt']
out_rgb = model(inputs)
optimizer.zero_grad()
# loss_up_1 = criterion_l1_loss(up_1_out, gt)
# loss_up_2 = criterion_l1_loss(up_2_out, gt)
# loss_up_3 = criterion_l1_loss(up_3_out, gt)
# loss_fft = (loss_up_1 + loss_up_2 + loss_up_3) / 6
loss_l1 = criterion_l1_loss(out_rgb, gt)
loss_ssim = criterion_ssim_loss(out_rgb, gt)
loss = loss_l1 + loss_ssim
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
# loss.backward()
optimizer.step()
write.add_scalar('loss_l1', loss_l1, ite_num)
write.add_scalar('loss_ssim', loss_ssim, ite_num)
logger.info('Epoch: [%d][%d/%d] lr: %f time: %.3f loss_l1: %f loss_ssim: %f loss: %f',
epoch+1, ite_num, max_iters, optimizer.param_groups[0]['lr'],
data_time, loss_l1, loss_ssim, loss)
losses = collections.OrderedDict()
losses['loss_l1'] = loss_l1.data.cpu()
losses['loss_ssim'] = loss_ssim.data.cpu()
losses['total_loss'] = loss.data.cpu()
visualizer.plot_current_losses(epoch + 1,
float(i) / len(data_loader),
losses)
# after iter
time_ = time.time() - t
t = time.time()
if ite_num4val % 5 == 0:
# pred_1 = inputs[0:1, 0:1, :, :]
# pred_1 = normPRED(pred_1)
# pred_2 = inputs[0:1, 1:2, :, :]
# pred_2 = normPRED(pred_2)
# pred_3 = inputs[0:1, 2:3, :, :]
# pred_3 = normPRED(pred_3)
# inputs_show = torch.cat([pred_1, pred_2, pred_3], dim=1)
# inputs_show = inputs_show[0].cpu().float().numpy() * 255
#
# gt_1 = gt[0:1, 0:1, :, :]
# gt_1 = normPRED(gt_1)
# gt_2 = gt[0:1, 1:2, :, :]
# gt_2 = normPRED(gt_2)
# gt_3 = gt[0:1, 2:3, :, :]
# gt_3 = normPRED(gt_3)
# gt_show = torch.cat([gt_1, gt_2, gt_3], dim=1)
# gt_show = gt_show[0].cpu().float().numpy() * 255
#
# pred_1 = out_rgb[0:1, 0:1, :, :]
# # pred_1 = normPRED(pred_1)
# pred_2 = out_rgb[0:1, 1:2, :, :]
# # pred_2 = normPRED(pred_2)
# pred_3 = out_rgb[0:1, 2:3, :, :]
# # pred_3 = normPRED(pred_3)
# outputs_show = torch.cat([pred_1, pred_2, pred_3], dim=1)
# outputs_show = Variable(outputs_show[0], requires_grad=False).cpu().float().numpy() * 255
#
# pred_1 = out_rgb[0:1, 0:1, :, :]
# pred_1 = normPRED(pred_1)
# pred_2 = out_rgb[0:1, 1:2, :, :]
# pred_2 = normPRED(pred_2)
# pred_3 = out_rgb[0:1, 2:3, :, :]
# pred_3 = normPRED(pred_3)
# outputs_show1 = torch.cat([pred_1, pred_2, pred_3], dim=1)
# outputs_show1 = Variable(outputs_show1[0], requires_grad=False).cpu().float().numpy() * 255
inputshow = normimage(inputs)
gtshow = normimage(gt)
outshow = normimage(out_rgb)
shows = []
# shows.append(inputs_show)
# shows.append(gt_show)
# shows.append(outputs_show)
# shows.append(outputs_show1)
shows.append(inputshow.transpose([2, 0, 1]))
shows.append(gtshow.transpose([2, 0, 1]))
shows.append(outshow.transpose([2, 0, 1]))
vis.images(shows, nrow=4, padding=3, win=1, opts=dict(title='Output images'))
ite_num4val = 0
if ite_num % 100 == 0:
save_latest(model, optimizer, cfg.work_dir, epoch, ite_num)
model.train()
if epoch % 20 == 0 or epoch == cfg.total_epoch - 1:
# print('-'*30, 'saving model')
save_epoch(model, optimizer, cfg.work_dir, epoch, ite_num)
model.train()
# after eppoch
# update learning rate
# print(optimizer.param_groups[0]['lr'])
# if cfg.lr_config.step[1] >= (epoch+1) >= cfg.lr_config.step[0]:
scheduler.step()
# after run
write.close()
# print(optimizer.param_groups[0]['lr'])
print()
save_epoch(model, optimizer, cfg.work_dir, epoch, ite_num)
logger.info('Finish Training')
# import matplotlib.pyplot as plt
# plt.plot(list(range(start_epoch-1, cfg.total_epoch)), lr_list)
# plt.xlabel("epoch")
# plt.ylabel("lr")
# plt.title("CosineAnnealingLR")
# plt.show()