-
Notifications
You must be signed in to change notification settings - Fork 20
/
train.py
212 lines (195 loc) · 9.15 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
import math
import argparse, yaml
import utils
import os
from tqdm import tqdm
import logging
import sys
import time
import importlib
import glob
parser = argparse.ArgumentParser(description='ELAN')
## yaml configuration files
parser.add_argument('--config', type=str, default=None, help = 'pre-config file for training')
parser.add_argument('--resume', type=str, default=None, help = 'resume training or not')
if __name__ == '__main__':
args = parser.parse_args()
if args.config:
opt = vars(args)
yaml_args = yaml.load(open(args.config), Loader=yaml.FullLoader)
opt.update(yaml_args)
## set visibel gpu
gpu_ids_str = str(args.gpu_ids).replace('[','').replace(']','')
os.environ['CUDA_DEVICE_ORDER'] = 'PCI_BUS_ID'
os.environ['CUDA_VISIBLE_DEVICES'] = '{}'.format(gpu_ids_str)
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.optim.lr_scheduler import MultiStepLR, StepLR
from datas.utils import create_datasets
## select active gpu devices
device = None
if args.gpu_ids is not None and torch.cuda.is_available():
print('use cuda & cudnn for acceleration!')
print('the gpu id is: {}'.format(args.gpu_ids))
device = torch.device('cuda')
torch.backends.cudnn.benchmark = True
else:
print('use cpu for training!')
device = torch.device('cpu')
torch.set_num_threads(args.threads)
## create dataset for training and validating
train_dataloader, valid_dataloaders = create_datasets(args)
## definitions of model
try:
model = utils.import_module('models.{}_network'.format(args.model, args.model)).create_model(args)
except Exception:
raise ValueError('not supported model type! or something')
model = nn.DataParallel(model).to(device)
## definition of loss and optimizer
loss_func = nn.L1Loss()
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr)
scheduler = MultiStepLR(optimizer, milestones=args.decays, gamma=args.gamma)
## load pretrain
if args.pretrain is not None:
print('load pretrained model: {}!'.format(args.pretrain))
ckpt = torch.load(args.pretrain)
model.load_state_dict(ckpt['model_state_dict'])
## resume training
start_epoch = 1
if args.resume is not None:
ckpt_files = glob.glob(os.path.join(args.resume, 'models', "*.pt"))
if len(ckpt_files) != 0:
ckpt_files = sorted(ckpt_files, key=lambda x: int(x.replace('.pt','').split('_')[-1]))
ckpt = torch.load(ckpt_files[-1])
prev_epoch = ckpt['epoch']
start_epoch = prev_epoch + 1
model.load_state_dict(ckpt['model_state_dict'])
optimizer.load_state_dict(ckpt['optimizer_state_dict'])
scheduler.load_state_dict(ckpt['scheduler_state_dict'])
stat_dict = ckpt['stat_dict']
## reset folder and param
experiment_path = args.resume
log_name = os.path.join(experiment_path, 'log.txt')
experiment_model_path = os.path.join(experiment_path, 'models')
print('select {}, resume training from epoch {}.'.format(ckpt_files[-1], start_epoch))
else:
## auto-generate the output logname
experiment_name = None
timestamp = utils.cur_timestamp_str()
if args.log_name is None:
experiment_name = '{}-{}-x{}-{}'.format(args.model, 'fp32', args.scale, timestamp)
else:
experiment_name = '{}-{}'.format(args.log_name, timestamp)
experiment_path = os.path.join(args.log_path, experiment_name)
log_name = os.path.join(experiment_path, 'log.txt')
stat_dict = utils.get_stat_dict()
## create folder for ckpt and stat
if not os.path.exists(experiment_path):
os.makedirs(experiment_path)
experiment_model_path = os.path.join(experiment_path, 'models')
if not os.path.exists(experiment_model_path):
os.makedirs(experiment_model_path)
## save training paramters
exp_params = vars(args)
exp_params_name = os.path.join(experiment_path, 'config.yml')
with open(exp_params_name, 'w') as exp_params_file:
yaml.dump(exp_params, exp_params_file, default_flow_style=False)
## print architecture of model
time.sleep(3) # sleep 3 seconds
sys.stdout = utils.ExperimentLogger(log_name, sys.stdout)
print(model)
sys.stdout.flush()
## start training
timer_start = time.time()
for epoch in range(start_epoch, args.epochs+1):
epoch_loss = 0.0
stat_dict['epochs'] = epoch
model = model.train()
opt_lr = scheduler.get_last_lr()
print('##==========={}-training, Epoch: {}, lr: {} =============##'.format('fp32', epoch, opt_lr))
for iter, batch in enumerate(train_dataloader):
optimizer.zero_grad()
lr, hr = batch
lr, hr = lr.to(device), hr.to(device)
sr = model(lr)
loss = loss_func(sr, hr)
loss.backward()
optimizer.step()
epoch_loss += float(loss)
if (iter + 1) % args.log_every == 0:
cur_steps = (iter+1)*args.batch_size
total_steps = len(train_dataloader.dataset)
fill_width = math.ceil(math.log10(total_steps))
cur_steps = str(cur_steps).zfill(fill_width)
epoch_width = math.ceil(math.log10(args.epochs))
cur_epoch = str(epoch).zfill(epoch_width)
avg_loss = epoch_loss / (iter + 1)
stat_dict['losses'].append(avg_loss)
timer_end = time.time()
duration = timer_end - timer_start
timer_start = timer_end
print('Epoch:{}, {}/{}, loss: {:.4f}, time: {:.3f}'.format(cur_epoch, cur_steps, total_steps, avg_loss, duration))
if epoch % args.test_every == 0:
torch.set_grad_enabled(False)
test_log = ''
model = model.eval()
for valid_dataloader in valid_dataloaders:
avg_psnr, avg_ssim = 0.0, 0.0
name = valid_dataloader['name']
loader = valid_dataloader['dataloader']
for lr, hr in tqdm(loader, ncols=80):
lr, hr = lr.to(device), hr.to(device)
sr = model(lr)
# quantize output to [0, 255]
hr = hr.clamp(0, 255)
sr = sr.clamp(0, 255)
# conver to ycbcr
if args.colors == 3:
hr_ycbcr = utils.rgb_to_ycbcr(hr)
sr_ycbcr = utils.rgb_to_ycbcr(sr)
hr = hr_ycbcr[:, 0:1, :, :]
sr = sr_ycbcr[:, 0:1, :, :]
# crop image for evaluation
hr = hr[:, :, args.scale:-args.scale, args.scale:-args.scale]
sr = sr[:, :, args.scale:-args.scale, args.scale:-args.scale]
# calculate psnr and ssim
psnr = utils.calc_psnr(sr, hr)
ssim = utils.calc_ssim(sr, hr)
avg_psnr += psnr
avg_ssim += ssim
avg_psnr = round(avg_psnr/len(loader) + 5e-3, 2)
avg_ssim = round(avg_ssim/len(loader) + 5e-5, 4)
stat_dict[name]['psnrs'].append(avg_psnr)
stat_dict[name]['ssims'].append(avg_ssim)
if stat_dict[name]['best_psnr']['value'] < avg_psnr:
stat_dict[name]['best_psnr']['value'] = avg_psnr
stat_dict[name]['best_psnr']['epoch'] = epoch
if stat_dict[name]['best_ssim']['value'] < avg_ssim:
stat_dict[name]['best_ssim']['value'] = avg_ssim
stat_dict[name]['best_ssim']['epoch'] = epoch
test_log += '[{}-X{}], PSNR/SSIM: {:.2f}/{:.4f} (Best: {:.2f}/{:.4f}, Epoch: {}/{})\n'.format(
name, args.scale, float(avg_psnr), float(avg_ssim),
stat_dict[name]['best_psnr']['value'], stat_dict[name]['best_ssim']['value'],
stat_dict[name]['best_psnr']['epoch'], stat_dict[name]['best_ssim']['epoch'])
# print log & flush out
print(test_log)
sys.stdout.flush()
# save model
saved_model_path = os.path.join(experiment_model_path, 'model_x{}_{}.pt'.format(args.scale, epoch))
# torch.save(model.state_dict(), saved_model_path)
torch.save({
'epoch': epoch,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'scheduler_state_dict': scheduler.state_dict(),
'stat_dict': stat_dict
}, saved_model_path)
torch.set_grad_enabled(True)
# save stat dict
## save training paramters
stat_dict_name = os.path.join(experiment_path, 'stat_dict.yml')
with open(stat_dict_name, 'w') as stat_dict_file:
yaml.dump(stat_dict, stat_dict_file, default_flow_style=False)
## update scheduler
scheduler.step()