forked from shreshthsaini/SR-DDPM
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathsr.py
199 lines (174 loc) · 8.69 KB
/
sr.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
import torch
import data as Data
import model as Model
import argparse
import logging
import core.logger as Logger
import core.metrics as Metrics
from tensorboardX import SummaryWriter
import os
import numpy as np
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('-c', '--config', type=str, default='config/sr_sr3_16_128_AnimeF.json',
help='JSON file for configuration')
parser.add_argument('-p', '--phase', type=str, choices=['train', 'val'],
help='train OR generation)', default='train')
parser.add_argument('-gpu', '--gpu_ids', type=str, default=None)
parser.add_argument('-debug', '-d', action='store_true')
parser.add_argument('-log_eval', action='store_true')
# parse configs
args = parser.parse_args()
opt = Logger.parse(args)
# Convert to NoneDict, which return None for missing key.
opt = Logger.dict_to_nonedict(opt)
# logging
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = True
Logger.setup_logger(None, opt['path']['log'],
'train', level=logging.INFO, screen=True)
Logger.setup_logger('val', opt['path']['log'], 'val', level=logging.INFO)
logger = logging.getLogger('base')
logger.info(Logger.dict2str(opt))
tb_logger = SummaryWriter(log_dir=opt['path']['tb_logger'])
# dataset
for phase, dataset_opt in opt['datasets'].items():
if phase == 'train' and args.phase != 'val':
train_set = Data.create_dataset(dataset_opt, phase)
train_loader = Data.create_dataloader(
train_set, dataset_opt, phase)
elif phase == 'val':
val_set = Data.create_dataset(dataset_opt, phase)
val_loader = Data.create_dataloader(
val_set, dataset_opt, phase)
logger.info('Initial Dataset Finished')
# model
diffusion = Model.create_model(opt)
logger.info('Created Diffusion model')
# Train
current_step = diffusion.begin_step
current_epoch = diffusion.begin_epoch
n_iter = opt['train']['n_iter']
if opt['path']['resume_state']:
logger.info('Resuming training from epoch: {}, iter: {}.'.format(
current_epoch, current_step))
diffusion.set_new_noise_schedule(
#opt['model']['beta_schedule']['train'], schedule_phase="train")
opt['model']['beta_schedule'][opt['phase']], schedule_phase=opt['phase'])
if opt['phase'] == 'train':
while current_step < n_iter:
current_epoch += 1
for _, train_data in enumerate(train_loader):
# due to ICC profile issue in iamgenet; if encourter image icc issue. we skip them
current_step += 1
if current_step > n_iter:
break
diffusion.feed_data(train_data)
diffusion.optimize_parameters()
# log
if current_step % opt['train']['print_freq'] == 0:
logs = diffusion.get_current_log()
message = '<epoch:{:3d}, iter:{:8,d}> '.format(
current_epoch, current_step)
for k, v in logs.items():
message += '{:s}: {:.4e} '.format(k, v)
tb_logger.add_scalar(k, v, current_step)
logger.info(message)
# validation
if current_step % opt['train']['val_freq'] == 0:
avg_psnr = 0.0
idx = 0
result_path = '{}/{}'.format(opt['path']
['results'], current_epoch)
os.makedirs(result_path, exist_ok=True)
diffusion.set_new_noise_schedule(
opt['model']['beta_schedule']['val'], schedule_phase='val')
for _, val_data in enumerate(val_loader):
idx += 1
diffusion.feed_data(val_data)
diffusion.test(continous=False)
visuals = diffusion.get_current_visuals()
sr_img = Metrics.tensor2img(visuals['SR']) # uint8
hr_img = Metrics.tensor2img(visuals['HR']) # uint8
lr_img = Metrics.tensor2img(visuals['LR']) # uint8
fake_img = Metrics.tensor2img(visuals['INF']) # uint8
# generation
Metrics.save_img(
hr_img, '{}/{}_{}_hr.png'.format(result_path, current_step, idx))
Metrics.save_img(
sr_img, '{}/{}_{}_sr.png'.format(result_path, current_step, idx))
Metrics.save_img(
lr_img, '{}/{}_{}_lr.png'.format(result_path, current_step, idx))
Metrics.save_img(
fake_img, '{}/{}_{}_inf.png'.format(result_path, current_step, idx))
tb_logger.add_image(
'Iter_{}'.format(current_step),
np.transpose(np.concatenate(
(fake_img, sr_img, hr_img), axis=1), [2, 0, 1]),
idx)
avg_psnr += Metrics.calculate_psnr(
sr_img, hr_img)
avg_psnr = avg_psnr / idx
diffusion.set_new_noise_schedule(
opt['model']['beta_schedule']['train'], schedule_phase='train')
# log
logger.info('# Validation # PSNR: {:.4e}'.format(avg_psnr))
logger_val = logging.getLogger('val') # validation logger
logger_val.info('<epoch:{:3d}, iter:{:8,d}> psnr: {:.4e}'.format(
current_epoch, current_step, avg_psnr))
# tensorboard logger
tb_logger.add_scalar('psnr', avg_psnr, current_step)
if current_step % opt['train']['save_checkpoint_freq'] == 0:
logger.info('Saving models and training states.')
diffusion.save_network(current_epoch, current_step)
# save model
logger.info('End of training.')
else:
logger.info('Begin Model Evaluation.')
avg_psnr = 0.0
avg_ssim = 0.0
idx = 0
result_path = '{}'.format(opt['path']['results'])
os.makedirs(result_path, exist_ok=True)
for _, val_data in enumerate(val_loader):
idx += 1
diffusion.feed_data(val_data)
diffusion.test(continous=True)
visuals = diffusion.get_current_visuals()
hr_img = Metrics.tensor2img(visuals['HR']) # uint8
lr_img = Metrics.tensor2img(visuals['LR']) # uint8
fake_img = Metrics.tensor2img(visuals['INF']) # uint8
sr_img_mode = 'grid'
if sr_img_mode == 'single':
# single img series
sr_img = visuals['SR'] # uint8
sample_num = sr_img.shape[0]
for iter in range(0, sample_num):
Metrics.save_img(
Metrics.tensor2img(sr_img[iter]), '{}/{}_{}_sr_{}.png'.format(result_path, current_step, idx, iter))
else:
# grid img
sr_img = Metrics.tensor2img(visuals['SR']) # uint8
Metrics.save_img(
sr_img, '{}/{}_{}_sr_process.png'.format(result_path, current_step, idx))
Metrics.save_img(
Metrics.tensor2img(visuals['SR'][-1]), '{}/{}_{}_sr.png'.format(result_path, current_step, idx))
Metrics.save_img(
hr_img, '{}/{}_{}_hr.png'.format(result_path, current_step, idx))
Metrics.save_img(
lr_img, '{}/{}_{}_lr.png'.format(result_path, current_step, idx))
Metrics.save_img(
fake_img, '{}/{}_{}_inf.png'.format(result_path, current_step, idx))
# generation
eval_psnr = Metrics.calculate_psnr(Metrics.tensor2img(visuals['SR'][-1]), hr_img)
eval_ssim = Metrics.calculate_ssim(Metrics.tensor2img(visuals['SR'][-1]), hr_img)
avg_psnr += eval_psnr
avg_ssim += eval_ssim
avg_psnr = avg_psnr / idx
avg_ssim = avg_ssim / idx
# logging the train and validation information
logger.info('# Validation # PSNR: {:.4e}'.format(avg_psnr))
logger.info('# Validation # SSIM: {:.4e}'.format(avg_ssim))
logger_val = logging.getLogger('val')
logger_val.info('<epoch:{:3d}, iter:{:8,d}> psnr: {:.4e}, ssim：{:.4e}'.format(
current_epoch, current_step, avg_psnr, avg_ssim))