-
Notifications
You must be signed in to change notification settings - Fork 4
/
main.py
400 lines (351 loc) · 18.6 KB
/
main.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
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
import os, time, torch, imageio, csv
import numpy as np
import torch.backends.cudnn as cudnn
import torch.nn as nn
import torch.optim as optim
from PIL import Image
from torch.autograd import Variable
from torch.utils.data import DataLoader
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
from torch.utils.tensorboard import SummaryWriter
import utils, shutil
import model as architecture
import data.common as common
from data.dataloader import dataloader as DatasetLoader
from option import args
from data import data
import src.degradation as degradation
from tqdm import tqdm
from src.cal_complexity import profile_origin
import warnings
# from src.SSIM import SSIM
warnings.filterwarnings("ignore")
from src.img_postprocessing import USMSharp
from pytorch_wavelets import DWTForward, DWTInverse
# from src.bicubic import bicubic
def main():
global opt, scaling
opt = utils.print_args(args)
opt.normalize_mean = torch.from_numpy(np.array([0.466, 0.448, 0.403])).float().view(1, 3, 1, 1).cuda()
opt.normalize_std = torch.from_numpy(np.array([0.242, 0.234, 0.246])).float().view(1, 3, 1, 1).cuda()
# if opt.n_colors == 3:
# if opt.data_train == 'DF2K':
# # DF2K data normalize
# normalize_mean = torch.from_numpy(np.array([0.466, 0.448, 0.403])).float().view(1, 3, 1, 1)
# normalize_std = torch.from_numpy(np.array([0.242, 0.234, 0.246])).float().view(1, 3, 1, 1)
# elif opt.data_train == 'ImageNet':
# # imagenet data normalize
# normalize_mean = torch.from_numpy(np.array([0.485, 0.456, 0.406])).float().view(1, 3, 1, 1)
# normalize_std = torch.from_numpy(np.array([0.229, 0.224, 0.225])).float().view(1, 3, 1, 1)
if opt.cuda and not torch.cuda.is_available():
raise Exception("No GPU found, please run without --cuda")
print("Random Seed: ", opt.seed)
torch.manual_seed(opt.seed)
if opt.cuda:
torch.cuda.manual_seed(opt.seed)
cudnn.benchmark = True
print('===> Building SR_Model')
print("===> Building model")
model = {
'SR': architecture.Generator(opt.n_colors, opt.net_type, opt.n_channels,
opt.n_blocks, opt.n_units, opt.n_layers, opt.growth_rate,
opt.act, opt.use_CoMo, opt.scale)
}
optimizer = {
'SR': None
}
scheduler = {
'SR': None
}
if opt.train.lower() == 'train':
# model['SR'] = utils.load_checkpoint(opt.resume_SR, model['SR'], opt.cuda, opt.n_GPUs)
if os.path.isfile(opt.resume_SR):
print('===> Loading Checkpoint from {}'.format(opt.resume_SR))
ckp = torch.load(opt.resume_SR)
CKP_BEST = torch.load(opt.resume_SR)
model['SR'].load_state_dict(ckp, strict=False)
else:
print('===> No Checkpoint in {}'.format(opt.resume_SR))
CKP_BEST = None
Avg_PSNR_BEST = 0
print('===> Calculating NumParams & FLOPs')
input = torch.FloatTensor(1, opt.n_colors, 480 // opt.scale, 360 // opt.scale)
macs, params = profile_origin(model['SR'], inputs=(input,), verbose=False)
print('-------------SR Model-------------')
print('\tParam = {:.3f}K\n\tFLOPs = {:.3f}G on {}'.format(params * 1e-3, macs * 1e-9, input[0].shape))
torch.cuda.empty_cache()
print("===> Setting GPU")
for item in model:
if opt.n_GPUs > 1 and opt.cuda:
model[item] = torch.nn.DataParallel(model[item]).cuda()
para = filter(lambda x: x.requires_grad, model[item].module.parameters())
else:
model[item] = model[item].cuda() if opt.cuda else model[item]
para = filter(lambda x: x.requires_grad, model[item].parameters())
optimizer[item] = opt.optimizer[item]([{'params': para, 'initial_lr': opt.lr[item]}], lr=opt.lr[item])
if opt.n_epochs > opt.lr_gamma_1[item]:
scheduler[item] = optim.lr_scheduler.StepLR(optimizer[item],
last_epoch=opt.start_epoch,
step_size=opt.lr_gamma_1[item],
gamma=opt.lr_gamma_2[item])
else:
scheduler[item] = optim.lr_scheduler.CosineAnnealingWarmRestarts(optimizer[item],
last_epoch=opt.start_epoch,
T_0=opt.lr_gamma_1[item],
T_mult=1,
eta_min=opt.lr_gamma_2[item])
model[item].train()
start_epoch = opt.start_epoch if opt.start_epoch >= 0 else 0
if os.path.exists(opt.model_path + '/' + 'runs'):
shutil.rmtree(opt.model_path + '/' + 'runs')
writer = SummaryWriter(opt.model_path + '/runs')
print('===> Testing')
for valid_set in opt.data_valid:
print('\t on {} datasets'.format(valid_set))
result_path = opt.model_path + '/Results/{}'.format(valid_set)
if not os.path.exists(result_path):
os.makedirs(result_path)
PSNR, SSIM, Time = validation(valid_set, result_path, model)
if valid_set == 'Urban100':
Avg_PSNR_BEST = PSNR
writer.add_scalar('PSNR/{}'.format(valid_set), PSNR, start_epoch)
writer.add_scalar('SSIM/{}'.format(valid_set), SSIM, start_epoch)
# # Avg_PSNR_BEST = 0
print('===> Building Training dataloader')
trainset = DatasetLoader(opt)
train_dataloader = DataLoader(trainset, batch_size=opt.batch_size, shuffle=True,
num_workers=opt.threads, pin_memory=False)
for epoch in range(start_epoch + 1, opt.n_epochs + 1):
# if (epoch - 1) % 1 == 0:
# print('===> Building Training dataloader')
# trainset = DatasetLoader(opt)
# train_dataloader = DataLoader(trainset, batch_size=opt.batch_size, shuffle=True,
# num_workers=opt.threads, pin_memory=False)
torch.cuda.empty_cache()
print('===> Training')
train(train_dataloader, optimizer, model, epoch, writer)
scheduler['SR'].step()
torch.cuda.empty_cache()
print('===> Testing')
PSNR_ALL = {}
SSIM_ALL = {}
print('\t on {} datasets'.format(opt.data_valid[0]))
result_path = opt.model_path + '/Results/{}'.format(opt.data_valid[0])
if not os.path.exists(result_path):
os.makedirs(result_path)
Avg_PSNR, SSIM, Time = validation(opt.data_valid[0], result_path, model)
PSNR_ALL[opt.data_valid[0]] = Avg_PSNR
SSIM_ALL[opt.data_valid[0]] = SSIM
torch.cuda.empty_cache()
if ((Avg_PSNR_BEST - Avg_PSNR) / Avg_PSNR_BEST > 0.005) or np.isnan(Avg_PSNR):
model['SR'].load_state_dict(CKP_BEST)
print('Poor Updating !!!')
elif Avg_PSNR_BEST < Avg_PSNR:
for valid_set in opt.data_valid[1:]:
print('\t on {} datasets'.format(valid_set))
result_path = opt.model_path + '/Results/{}'.format(valid_set)
if not os.path.exists(result_path):
os.makedirs(result_path)
PSNR, SSIM, Time = validation(valid_set, result_path, model)
PSNR_ALL[valid_set] = PSNR
SSIM_ALL[valid_set] = SSIM
torch.cuda.empty_cache()
print('Best Updating !!!')
Avg_PSNR_BEST = Avg_PSNR
model_path= opt.model_path + '/Checkpoints/model_epoch_best.pth'
torch.save(model['SR'].state_dict(), model_path)
print('Checkpoint saved to {}'.format(model_path))
CKP_BEST = torch.load(model_path)
for valid_set in opt.data_valid:
writer.add_scalar('PSNR/{}'.format(valid_set), PSNR_ALL[valid_set], epoch)
writer.add_scalar('SSIM/{}'.format(valid_set), SSIM_ALL[valid_set], epoch)
else:
for valid_set in opt.data_valid[1:]:
print('\t on {} datasets'.format(valid_set))
result_path = opt.model_path + '/Results/{}'.format(valid_set)
if not os.path.exists(result_path):
os.makedirs(result_path)
PSNR, SSIM, Time = validation(valid_set, result_path, model)
PSNR_ALL[valid_set] = PSNR
SSIM_ALL[valid_set] = SSIM
torch.cuda.empty_cache()
print('General Updating !!!')
model_path = opt.model_path + '/Checkpoints/model_epoch_last.pth'
torch.save(model['SR'].state_dict(), model_path)
print('Checkpoint saved to {}'.format(model_path))
# CKP_BEST = model['SR'].state_dict().copy()
CKP_BEST = torch.load(model_path)
for valid_set in opt.data_valid:
writer.add_scalar('PSNR/{}'.format(valid_set), PSNR_ALL[valid_set], epoch)
writer.add_scalar('SSIM/{}'.format(valid_set), SSIM_ALL[valid_set], epoch)
writer.close()
elif opt.train == 'Test':
model['SR'] = utils.load_checkpoint(opt.model_path + '.pth', model['SR'], opt.cuda, opt.n_GPUs)
print("===> Setting GPU")
for item in model:
if opt.n_GPUs > 1 and opt.cuda:
model[item] = torch.nn.DataParallel(model[item]).cuda()
else:
model[item] = model[item].cuda() if opt.cuda else model[item]
model[item].eval()
for i in range(len(opt.data_valid)):
valid_path = opt.dir_data + 'Test/' + opt.data_valid[i]
with open(opt.model_path + '/SR Results/' + opt.data_valid[i] + '_x{:d}.csv'.format(opt.scale), 'w', newline='') as f:
f_csv = csv.writer(f)
f_csv.writerow(['image_name', 'PSNR', 'SSIM', 'Time'])
validation(valid_path, model, scale=opt.scale, f_csv=f_csv)
torch.cuda.empty_cache()
else:
raise InterruptedError
def train(training_dataloader, optimizer, model, epoch, writer):
criterion_MAE = nn.L1Loss(reduction='mean').cuda()
criterion_LInf = architecture.LInf_Loss(reduction='mean').cuda()
# coef_RGB2Y = torch.tensor([65.481, 128.553, 24.966]).reshape(1, 3, 1, 1).cuda()
# criterion_PSNR_Y = architecture.PSNRLoss(toY=True).cuda()
# criterion_Char = architecture.L1_Charbonnier_loss().cuda()
# criterion_MAPE = architecture.MAPELoss(reduction='mean').cuda()
# criterion_SSIM = SSIM(size_average=False).cuda()
# xfm = DWTForward(J=1, wave='haar', mode='zero').cuda()
# ifm = DWTInverse(wave='haar', mode='zero').cuda()
for item in model:
model[item].train()
# Bic_Resize = bicubic().cuda()
scaler = torch.cuda.amp.GradScaler()
lambda_ext = 0.02
# sharpen = USMSharp(radius=21, sigma=2.0).cuda()
with tqdm(total=len(training_dataloader), ncols=140) as pbar:
for iteration, (LR_img, HR_img) in enumerate(training_dataloader):
if HR_img.shape[0] == opt.batch_size:
niter = (epoch - 1) * len(training_dataloader) + iteration
HR_img = Variable(HR_img, volatile=False)
LR_img = Variable(LR_img)
if opt.cuda:
HR_img = HR_img.cuda()
LR_img = LR_img.cuda()
HR_img = (HR_img - opt.normalize_mean) / opt.normalize_std
LR_img = (LR_img - opt.normalize_mean) / opt.normalize_std
optimizer['SR'].zero_grad()
with torch.cuda.amp.autocast():
SR_img = model['SR'](LR_img)
# loss = -criterion_SSIM(SR_img, HR_img)
loss_basic = criterion_MAE(SR_img, HR_img)
# loss = criterion_MAE(torch.log(SR_img + 0.5),
# Variable(torch.log(HR_img + 0.5), requires_grad=False))
# sharpen = USMSharp(radius=21, sigma=np.random.uniform(1.0, 3.0)).cuda()
# loss_extension = criterion_MAE(sharpen(SR_img) - SR_img,
# Variable(sharpen(HR_img) - SR_img, requires_grad=False))
# loss_DN = criterion_PSNR(img_denoised, Variable(img_clean, requires_grad=False))
# loss_extension = 3 * criterion_MAE(torch.sum(SR_img * coef_RGB2Y, dim=1) / 255.,
# torch.sum(HR_img * coef_RGB2Y, dim=1) / 255.)
# x_L, x_H = xfm(SR_img)
loss_extension = criterion_LInf(SR_img, HR_img)
# loss_extension = 0
scaler.scale(loss_basic + lambda_ext*loss_extension).backward()
if opt.use_CoMo:
for group in optimizer['SR'].param_groups:
a = np.zeros(len(group["params"]))
i = 0
param_group_copy = ['' for _ in range(len(group["params"]))]
for param in group["params"]:
if param.grad != None:
if param.grad.shape != param.shape:
param_group_copy[i] = param.grad.data
ref_batch_grad = param.grad.data.view(opt.batch_size,
param.data.shape[0], param.data.shape[1],
param.data.shape[2], param.data.shape[3])
param.grad.data = torch.mean(ref_batch_grad, dim=0)
a[i] = 1
i += 1
# scaler.unscale_(optimizer['SR'])
# torch.nn.utils.clip_grad_value_(model['SR'].parameters(), 0.1)
scaler.step(optimizer['SR'])
scaler.update()
for group in optimizer['SR'].param_groups:
for i in range(len(a)):
if a[i] == 1:
para = group["params"][i]
para.grad.data = param_group_copy[i]
else:
# scaler.unscale_(optimizer['SR'])
# torch.nn.utils.clip_grad_value_(model['SR'].parameters(), 0.1)
scaler.step(optimizer['SR'])
scaler.update()
time.sleep(0.01)
pbar.update(1)
pbar.set_postfix(Epoch=epoch,
LeRate='{:.6f}'.format(optimizer['SR'].param_groups[0]['lr']),
Loss='{:.3f}+{}*{:.3f}'.format(loss_basic, lambda_ext, loss_extension))
if (iteration + 1) % 50 == 0:
if 'loss_basic' in locals().keys():
writer.add_scalar('Loss/loss_basic', loss_basic, niter)
if 'loss_extension' in locals().keys():
writer.add_scalar('Loss/loss_extension', loss_extension, niter)
torch.cuda.empty_cache()
if np.isnan(loss_basic.data.cpu().numpy()):
break
def validation(valid_set, result_path, model):
for item in model:
model[item].eval()
count = 0
Avg_PSNR = 0
Avg_SSIM = 0
Avg_Time = 0
path_lr = opt.dir_data + '/Test/{}/LR_bicubic/X{}'.format(valid_set, opt.scale)
path_hr = opt.dir_data + '/Test/{}/HR'.format(valid_set)
file_lr = os.listdir(path_lr)
file_hr = os.listdir(path_hr)
file_lr.sort()
file_hr.sort()
length = file_lr.__len__()
if opt.cuda:
start = torch.cuda.Event(enable_timing=True)
end = torch.cuda.Event(enable_timing=True)
else:
Time = 0
torch.cuda.empty_cache()
with torch.no_grad():
with tqdm(total=length, ncols=140) as pbar:
for idx_img in range(length):
torch.cuda.empty_cache()
img_name, ext = os.path.splitext(file_lr[idx_img])
LR_img = imageio.imread(os.path.join(path_lr, file_lr[idx_img]))
HR_img = imageio.imread(os.path.join(path_hr, file_hr[idx_img]))
LR_img, HR_img = common.set_channel([LR_img, HR_img], opt.n_colors)
LR_img = common.np2Tensor(LR_img, opt.value_range)
HR_img = common.np2Tensor(HR_img, opt.value_range)
LR_img = Variable(LR_img).unsqueeze(0)
if opt.cuda:
LR_img = LR_img.cuda()
start.record()
LR_img = (LR_img - opt.normalize_mean) / opt.normalize_std
with torch.cuda.amp.autocast():
SR_img = model['SR'].forward(LR_img.half())
SR_img = SR_img * opt.normalize_std + opt.normalize_mean
end.record()
torch.cuda.synchronize()
Time = start.elapsed_time(end) * 1e-3
SR_img = SR_img.data[0].cpu()
PSNR = utils.calc_PSNR(SR_img, HR_img, opt.value_range, shave=opt.scale)
SSIM = utils.calc_SSIM(SR_img, HR_img, opt.value_range, shave=opt.scale)
Avg_PSNR += PSNR
Avg_SSIM += SSIM
Avg_Time += Time
count = count + 1
# if opt.n_colors > 1:
# SR_img = SR_img.mul(255).clamp(0, 255).round()
# SR_img = SR_img.numpy().astype(np.uint8)
# SR_img = SR_img.transpose((1, 2, 0))
# SR_img = Image.fromarray(SR_img)
# else:
# SR_img = SR_img[0, :, :].mul(opt.value_range).clamp(0, opt.value_range).round().numpy().astype(
# np.uint8)
# SR_img = Image.fromarray(SR_img).convert('L')
#
# SR_img.save(result_path + '/' + img_name + '.png')
time.sleep(0.01)
pbar.update(1)
pbar.set_postfix(Eval='|{:.3f}|{:.5f}|'.format(Avg_PSNR / count, Avg_SSIM / count),
Time='{:.3f}ms'.format(Avg_Time / count * 1000))
return Avg_PSNR / count, Avg_SSIM / count, Avg_Time / count
if __name__ == '__main__':
main()