-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_all.py
168 lines (139 loc) · 6.8 KB
/
train_all.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
import argparse
import os
import os.path as osp
import sys
sys.path.insert(0, os.path.dirname(os.path.realpath(__file__)) + '/../..')
import logging
import torch
import torch.distributed as dist
import torch.multiprocessing as mp
import torch.nn.functional as F
from config_all import config
from utils import common, dataloader, solver, model_opr
from dataset import get_dataset
from exps.rfdn.RFDNMGA import RFDNAll
from validate import validate_2
def init_dist(local_rank):
if mp.get_start_method(allow_none=True) != 'spawn':
mp.set_start_method('spawn', force=True)
torch.cuda.set_device(local_rank)
dist.init_process_group(backend="nccl", init_method='env://')
dist.barrier()
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--local_rank', type=int, default=0)
args = parser.parse_args()
# initialization
rank = 0
num_gpu = 1
distributed = False
if 'WORLD_SIZE' in os.environ:
num_gpu = int(os.environ['WORLD_SIZE'])
distributed = num_gpu > 1
if distributed:
rank = args.local_rank
init_dist(rank)
common.init_random_seed(config.DATASET.SEED + rank)
# set up dirs and log
exp_dir, cur_dir = osp.split(osp.split(osp.realpath(__file__))[0])
root_dir = osp.split(exp_dir)[0]
model_name='Model6_all'
log_dir = osp.join(root_dir, 'logs', cur_dir)
model_dir = osp.join(log_dir, 'models', model_name)
solver_dir = osp.join(log_dir, 'solvers', model_name)
if rank <= 0:
common.mkdir(log_dir)
ln_log_dir = osp.join(exp_dir, cur_dir, 'log')
if not osp.exists(ln_log_dir):
os.system('ln -s %s log' % log_dir)
common.mkdir(model_dir)
common.mkdir(solver_dir)
save_dir = osp.join(log_dir, 'saved_imgs')
common.mkdir(save_dir)
common.setup_logger('base', model_name, log_dir, 'train', level=logging.INFO, screen=True, to_file=True)
logger = logging.getLogger('base')
# dataset
train_dataset = get_dataset(config.DATASET)
train_loader = dataloader.train_loader(train_dataset, config, rank=rank, seed=config.DATASET.SEED,
is_dist=distributed)
if rank <= 0:
val_dataset = get_dataset(config.VAL)
val_loader = dataloader.val_loader(val_dataset, config, rank, 1)
data_len = val_dataset.data_len
# model
model = RFDNAll(in_nc=config.MODEL.IN_NC, nf=config.MODEL.NF,
num_modules=config.MODEL.NUM_MODULES, out_nc=config.MODEL.OUT_NC,
padding=config.MODEL.PADDING, size=config.MODEL.SIZE, upscale=config.MODEL.UPSCALE)
if rank <= 0:
print(model)
if config.CONTINUE_ITER:
model_path = osp.join(model_dir, '%d.pth' % config.CONTINUE_ITER)
if rank <= 0:
logger.info('[Continue] Iter: %d' % config.CONTINUE_ITER)
model_opr.load_model(model, model_path, strict=True, cpu=True)
else:
if config.INIT_COR_MODEL:
if rank <= 0:
logger.info('[Initialize] Model: %s' % config.INIT_COR_MODEL)
model_opr.load_corresponding_model(model, torch.load(config.INIT_COR_MODEL))
# solvers
optimizer = solver.make_mask_optimizer(config, model) # lr without X num_gpu
lr_scheduler = solver.CosineAnnealingLR_warmup_three(config, optimizer, config.SOLVER.BASE_LR, config.SOLVER.MASK_LR,
config.SOLVER.REFINE_LR)
device = torch.device(config.MODEL.DEVICE)
model.to(device)
if distributed:
model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[torch.cuda.current_device()],find_unused_parameters=True)
iteration = 0
if config.CONTINUE_ITER:
solver_path = osp.join(solver_dir, '%d.solver' % config.CONTINUE_ITER)
iteration = model_opr.load_solver(optimizer, lr_scheduler, solver_path)
max_iter = max_psnr = max_ssim = 0
for lr_img, hr_img in train_loader:
model.train()
iteration = iteration + 1
optimizer.zero_grad()
lr_img = lr_img.to(device)
hr_img = hr_img.to(device)
refine_img, coarse_img = model(lr_img, is_return_coarse=True)
loss_dict = compute_loss(refine_img, coarse_img, hr_img)
total_loss = sum(loss for loss in loss_dict.values())
total_loss.backward()
optimizer.step()
lr_scheduler.step()
if rank <= 0:
if iteration % config.LOG_PERIOD == 0 or iteration == config.SOLVER.MAX_ITER:
log_str = 'Iter: %d, base_LR: %.3e, mask_LR: %.3e , refine_LR: %.3e ' % (
iteration, optimizer.param_groups[0]['lr'], optimizer.param_groups[1]['lr']
, optimizer.param_groups[2]['lr'])
for key in loss_dict:
log_str += key + ': %.4f, ' % float(loss_dict[key])
logger.info(log_str)
if iteration % config.SAVE_PERIOD == 0 or iteration == config.SOLVER.MAX_ITER:
logger.info('[Saving] Iter: %d' % iteration)
model_path = osp.join(model_dir, '%d.pth' % iteration)
solver_path = osp.join(solver_dir, '%d.solver' % iteration)
model_opr.save_model(model, model_path)
model_opr.save_solver(optimizer, lr_scheduler, iteration, solver_path)
if iteration % config.VAL.PERIOD == 0 or iteration == config.SOLVER.MAX_ITER:
logger.info('[Validating] Iter: %d' % iteration)
model.eval()
with torch.no_grad():
psnr, ssim, psnr_c, ssim_c = validate_2(model, val_loader, config, device, iteration,
save_path=save_dir)
if psnr > max_psnr:
max_psnr, max_ssim, max_iter, m_psnr_c, m_ssim_c = psnr, ssim, iteration, psnr_c, ssim_c
logger.info('[Val Result] Iter: %d, PSNR: %.4f, SSIM: %.4f, PSNR: %.4f, SSIM: %.4f,' % (
iteration, psnr, ssim, psnr_c, ssim_c))
logger.info('[Best Result] Iter: %d, PSNR: %.4f, SSIM: %.4f, PSNR: %.4f, SSIM: %.4f,' % (
max_iter, max_psnr, max_ssim, m_psnr_c, m_ssim_c))
if iteration >= config.SOLVER.MAX_ITER:
break
if rank <= 0:
logger.info('Finish training process!')
logger.info('[Best Result] Iter: %d, PSNR: %.4f, SSIM: %.4f, PSNR: %.4f, SSIM: %.4f,' % (
max_iter, max_psnr, max_ssim, m_psnr_c, m_ssim_c))
def compute_loss(refine_img, coarse_img, hr_img):
return dict(l1_loss_1=10.0*F.l1_loss(refine_img, hr_img),l1_loss_2=F.l1_loss(coarse_img, hr_img))
if __name__ == '__main__':
main()