-
Notifications
You must be signed in to change notification settings - Fork 6
/
sample.py
163 lines (139 loc) · 6.52 KB
/
sample.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
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 core.wandb_logger import WandbLogger
from tensorboardX import SummaryWriter
import os
import numpy as np
import wandb
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('-c', '--config', type=str, default='config/sample_sr3_128.json',
help='JSON file for configuration')
parser.add_argument('-p', '--phase', type=str, choices=['train', 'val'],
help='Run either train(training) or val(generation)', default='train')
parser.add_argument('-gpu', '--gpu_ids', type=str, default=None)
parser.add_argument('-debug', '-d', action='store_true')
parser.add_argument('-enable_wandb', action='store_true')
parser.add_argument('-log_wandb_ckpt', 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'])
# Initialize WandbLogger
if opt['enable_wandb']:
wandb_logger = WandbLogger(opt)
val_step = 0
else:
wandb_logger = None
# 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)
logger.info('Initial Dataset Finished')
# model
diffusion = Model.create_model(opt)
logger.info('Initial Model Finished')
# Train
current_step = diffusion.begin_step
current_epoch = diffusion.begin_epoch
n_iter = opt['train']['n_iter']
sample_sum = opt['datasets']['val']['data_len']
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'][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):
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)
if wandb_logger:
wandb_logger.log_metrics(logs)
# validation
if current_step % opt['train']['val_freq'] == 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 idx in range(sample_sum):
diffusion.sample(continous=False)
visuals = diffusion.get_current_visuals(sample=True)
sample_img = Metrics.tensor2img(
visuals['SAM']) # uint8
# generation
Metrics.save_img(
sample_img, '{}/{}_{}_sr.png'.format(result_path, current_step, idx))
tb_logger.add_image(
'Iter_{}'.format(current_step),
np.transpose(sample_img, [2, 0, 1]),
idx)
if wandb_logger:
wandb_logger.log_image(f'validation_{idx}', sample_img)
diffusion.set_new_noise_schedule(
opt['model']['beta_schedule']['train'], schedule_phase='train')
if current_step % opt['train']['save_checkpoint_freq'] == 0:
logger.info('Saving models and training states.')
diffusion.save_network(current_epoch, current_step)
if wandb_logger and opt['log_wandb_ckpt']:
wandb_logger.log_checkpoint(current_epoch, current_step)
# save model
logger.info('End of training.')
else:
logger.info('Begin Model Evaluation.')
result_path = '{}'.format(opt['path']['results'])
os.makedirs(result_path, exist_ok=True)
sample_imgs = []
for idx in range(sample_sum):
idx += 1
diffusion.sample(continous=True)
visuals = diffusion.get_current_visuals(sample=True)
show_img_mode = 'grid'
if show_img_mode == 'single':
# single img series
sample_img = visuals['SAM'] # uint8
sample_num = sample_img.shape[0]
for iter in range(0, sample_num):
Metrics.save_img(
Metrics.tensor2img(sample_img[iter]), '{}/{}_{}_sample_{}.png'.format(result_path, current_step, idx, iter))
else:
# grid img
sample_img = Metrics.tensor2img(visuals['SAM']) # uint8
Metrics.save_img(
sample_img, '{}/{}_{}_sample_process.png'.format(result_path, current_step, idx))
Metrics.save_img(
Metrics.tensor2img(visuals['SAM'][-1]), '{}/{}_{}_sample.png'.format(result_path, current_step, idx))
sample_imgs.append(Metrics.tensor2img(visuals['SAM'][-1]))
if wandb_logger:
wandb_logger.log_images('eval_images', sample_imgs)