-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy pathtrainer.py
482 lines (383 loc) · 22.8 KB
/
trainer.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
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
import os
import time
import datetime
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.backends.cudnn as cudnn
from tensorboardX import SummaryWriter
from util import utils
from dataloader import dataset
from models.loss import ColorLoss, VGGLoss
import models.loss as L
from models.optimizer.adamw import AdamW
from models.TrainingModule import TrainingModule, LossManager, pack_network_output, pack_gt_data
from models.utils import create_generator, create_discriminator, create_generator_val
class Trainer(TrainingModule):
def __init__(self, opt, num_gpus, rank = None, world_size = None):
super(Trainer, self).__init__(opt = opt,
num_gpus = num_gpus,
rank = rank,
world_size = world_size)
self.Training_config = self.opt.Training_config
self.optim_config = self.opt.Optimizer
self.G = create_generator(opt.GNet)
self.D = create_discriminator(opt.DNet)
# to support multi gpu or distributed training
self.G = self.wrapper(self.G)
self.D = self.wrapper(self.D)
self.LM = LossManager(self.opt.Loss, num_gpus = num_gpus)
self._init_dataloader()
self._init_optim()
def _init_dataloader(self):
# Define the dataset
train_dataset = dataset.SLRGB2RGB_dataset(self.opt.Dataset, 'train')
val_dataset = dataset.SLRGB2RGB_dataset(self.opt.Dataset, 'val')
print('The overall number of training images:', len(train_dataset))
print('The overall number of validation images:', len(val_dataset))
# Define the dataloader
self.train_loader = torch.utils.data.DataLoader(train_dataset, batch_size = self.Training_config.train_batch_size, shuffle = True, num_workers = self.Training_config.num_workers, pin_memory = True)
self.val_loader = torch.utils.data.DataLoader(val_dataset, batch_size = self.Training_config.val_batch_size, shuffle = False, num_workers = self.Training_config.num_workers, pin_memory = True)
def _init_optim(self):
if self.optim_config.name == "Adam":
self.optim_G = torch.optim.Adam(self.G.parameters(), lr = self.optim_config.args.lr_g, betas = (self.optim_config.args.b1, self.optim_config.args.b2), weight_decay = self.optim_config.args.weight_decay)
self.optim_D = torch.optim.Adam(self.D.parameters(), lr = self.optim_config.args.lr_d, betas = (self.optim_config.args.b1, self.optim_config.args.b2), weight_decay = self.optim_config.args.weight_decay)
elif self.optim_config.name == "SGD":
self.optim_G = torch.optim.SGD(self.G.parameters(), lr = self.optim_config.args.lr_g)
self.optim_D = torch.optim.SGD(self.D.parameters(), lr = self.optim_config.args.lr_d)
def train(self):
# Count start time
iters_done = 0
# self._validate(1)
for epoch in range(self.Training_config.start_idx + 1, self.Training_config.epochs):
print('epoch ', epoch)
# Record learning rate
for param_group in self.optim_G.param_groups:
self.add_scalar('lr', param_group['lr'], epoch)
for i, data in enumerate(self.train_loader):
print(i, self.device)
short_img = data['short_img'].to(self.device)
long_img = data['long_img'].to(self.device)
RGBout_img = data['RGBout_img'].to(self.device)
gt_long_img = data['gt_long_img'].to(self.device)
# process patch
if len(short_img.shape) == 5:
_, _, C, H, W = short_img.shape
short_img = short_img.view(-1, C, H, W)
long_img = long_img.view(-1, C, H, W)
RGBout_img = RGBout_img.view(-1, C, H, W)
gt_long_img = gt_long_img.view(-1, C, H, W)
outs = self.G(short_img, long_img)
if isinstance(outs, list):
out = outs[0]
else:
out = outs
# ============ pack gt data ============
gt_dict = pack_gt_data([RGBout_img, gt_long_img])
# ============ pack output =============
outputs = pack_network_output(outs, self.opt.GNet.name)
# ============== train G ===============
if self.LM.train_GAN:
fake_scalar = self.D(short_img, long_img, out)
else:
fake_scalar = None
G_loss, G_loss_info = self.LM(outputs, gt_dict, fake_scalar)
if self.LM.distill_training:
distill_loss, distill_info, big_outs = self.LM.distill_loss(outs, short_img, long_img)
G_loss += distill_loss
G_loss_info.update(distill_info)
self.optim_G.zero_grad()
G_loss.backward()
self.optim_G.step()
# ============== train D ===============
if self.LM.train_GAN:
fake_scalar_d = self.D(short_img, long_img, out.detach())
true_scalar_d = self.D(short_img, long_img, RGBout_img)
D_loss, D_loss_info = self.LM.d_loss(fake_scalar_d, true_scalar_d)
self.optim_D.zero_grad()
D_loss.backward()
self.optim_D.step()
else:
D_loss, D_loss_info = 0., {}
# record loss
if iters_done % self.Training_config.show_loss_iter == 0:
self.add_scalars(main_tag = 'G_loss', tag_scalar_dict = G_loss_info, global_step = iters_done)
self.add_scalars(main_tag = 'D_loss', tag_scalar_dict = D_loss_info, global_step = iters_done)
if iters_done % self.Training_config.show_img_iter == 0:
vis_imgs = [short_img, long_img, RGBout_img, gt_long_img]
if isinstance(outs, list):
for j in range(len(outs)):
if outs[j] is not None and outs[j].size()[1] == 3:
vis_imgs.append(outs[j].clamp(0.0, 1.0))
else:
vis_imgs.append(outs)
self.visual_image('train_image', vis_imgs, iters_done)
# show gradient image
if isinstance(outs, list):
vis_imgs = []
for j in range(len(outs)):
if outs[j] is not None and outs[j].size()[1] == 1:
vis_imgs.append(outs[j].clamp(0.0, 1.0))
if len(vis_imgs) > 0:
self.visual_image('train_feat', vis_imgs, iters_done)
# show large model output
if self.LM.distill_training:
vis_imgs = []
vis_feats = []
for j in range(len(big_outs)):
if big_outs[j] is not None:
if big_outs[j].size()[1] == 3:
vis_imgs.append(big_outs[j].clamp(0.0, 1.0))
elif big_outs[j].size()[1] == 1:
vis_feats.append(big_outs[j].clamp(0.0, 1.0))
if len(vis_imgs) > 0:
self.visual_image('distill_image', vis_imgs, iters_done)
if len(vis_feats) > 0:
self.visual_image('distill_feat', vis_feats, iters_done)
self._save_G(self.opt, epoch, iters_done, len(self.train_loader), self.G)
if self.LM.train_GAN:
self._save_D(self.opt, epoch, iters_done, len(self.train_loader), self.D)
self._adjust_learning_rate(self.optim_config, (epoch + 1), iters_done, self.optim_G, self.opt.Optimizer.args.lr_g)
if self.LM.train_GAN:
self._adjust_learning_rate(self.optim_config, (epoch + 1), iters_done, self.optim_D, self.opt.Optimizer.args.lr_d)
iters_done += 1
if epoch % 1 == 0:
img_list = [short_img, long_img, out, RGBout_img]
name_list = ['inshort', 'inlong', 'pred', 'gt']
utils.save_sample_png(sample_folder = os.path.join(self.save_folder, 'sample'), sample_name = 'train_epoch%d' % (epoch + 1), img_list = img_list, name_list = name_list, pixel_max_cnt = 255)
# if epoch % 5 == 0:
self._validate(epoch)
def _validate(self, epoch):
self.G.eval()
val_PSNR = 0
val_SSIM = 0
val_day_PSNR, val_night_PSNR = 0, 0
val_day_SSIM, val_night_SSIM = 0, 0
num_of_val_image = 0
num_of_val_day, num_of_val_night = 0, 0
for j, dataset in enumerate(self.val_loader):
short_img = dataset['short_img'].to(self.device)
long_img = dataset['long_img'].to(self.device)
RGBout_img = dataset['RGBout_img'].to(self.device)
short_paths = dataset['in_short_path']
# process patch
if len(short_img.shape) == 5:
_, _, C, H, W = short_img.shape
short_img = short_img.view(-1, C, H, W)
long_img = long_img.view(-1, C, H, W)
RGBout_img = RGBout_img.view(-1, C, H, W)
out = self.G(short_img, long_img)
if isinstance(out, list):
out = out[0]
# deblur_out = out[1]
# deblur_out = deblur_out.clamp(0.0, 1.0)
else:
# deblur_out = None
pass
out = out.clamp(0.0, 1.0)
num_of_val_image += short_img.shape[0]
val_PSNR += utils.psnr(out, RGBout_img, 1) * short_img.shape[0]
val_SSIM += utils.ssim(out, RGBout_img) * short_img.shape[0]
# calculate psnr, ssim for day and night data
day_outs, day_gts = [], []
night_outs, night_gts = [], []
for idx in range(len(short_paths)):
if 'day' in short_paths[idx]:
day_outs.append(out[idx].unsqueeze(dim = 0))
day_gts.append(RGBout_img[idx].unsqueeze(dim = 0))
elif 'night' in short_paths[idx]:
night_outs.append(out[idx].unsqueeze(dim = 0))
night_gts.append(RGBout_img[idx].unsqueeze(dim = 0))
if len(day_outs) >= 1:
day_outs = torch.cat(day_outs, dim = 0)
day_gts = torch.cat(day_gts, dim = 0)
val_day_PSNR += utils.psnr(day_outs, day_gts, 1) * day_outs.shape[0]
val_day_SSIM += utils.ssim(day_outs, day_gts) * day_outs.shape[0]
num_of_val_day += day_outs.shape[0]
if len(night_outs) >= 1:
night_outs = torch.cat(night_outs, dim = 0)
night_gts = torch.cat(night_gts, dim = 0)
val_night_PSNR += utils.psnr(night_outs, night_gts, 1) * night_outs.shape[0]
val_night_SSIM += utils.ssim(night_outs, night_gts) * night_outs.shape[0]
num_of_val_night += night_outs.shape[0]
if j % 10 == 0:
print('val: %d | epoch: %d' % (j, epoch))
val_PSNR = val_PSNR / num_of_val_image
val_SSIM = val_SSIM / num_of_val_image
val_day_PSNR = val_day_PSNR / num_of_val_day
val_day_SSIM = val_day_SSIM / num_of_val_day
val_night_PSNR = val_night_PSNR / num_of_val_night
val_night_SSIM = val_night_SSIM / num_of_val_night
self.add_scalar('val_PSNR', val_PSNR, global_step = epoch)
self.add_scalar('val_SSIM', val_SSIM, global_step = epoch)
self.add_scalar('val_day_PSNR', val_day_PSNR, global_step = epoch)
self.add_scalar('val_day_SSIM', val_day_SSIM, global_step = epoch)
self.add_scalar('val_night_PSNR', val_night_PSNR, global_step = epoch)
self.add_scalar('val_night_SSIM', val_night_SSIM, global_step = epoch)
self.G.train()
print('val: epoch: %d, psnr: %.3f, ssim: %.3f' % (epoch, val_PSNR, val_SSIM))
class TwoPhaseTrainer(TrainingModule):
def __init__(self, opt, num_gpus, rank = None, world_size = None):
super(TwoPhaseTrainer, self).__init__(opt = opt,
num_gpus = num_gpus,
rank = rank,
world_size = world_size)
self.Training_config = self.opt.Training_config
self.optim_config = self.opt.Optimizer
if self.Training_config.phase == 'deblur':
self.G = create_generator(opt.DeblurNet)
elif self.Training_config.phase == 'denoise':
self.G = create_generator(opt.DenoiseNet)
if opt.DeblurNet.finetune_path:
self.deblurNet = create_generator_val(opt.DeblurNet, opt.DeblurNet.finetune_path)
else:
self.deblurNet = create_generator_val(opt.DeblurNet)
for param in self.deblurNet.parameters():
param.requires_grad = False
self.deblurNet = self.wrapper(self.deblurNet)
else:
raise ValueError('phase should be deblur or denoise, but found %s.' % opt.phase)
# to support multi gpu or distributed training
self.G = self.wrapper(self.G)
self.LM = LossManager(self.opt.Loss, num_gpus = num_gpus)
self._init_dataloader()
self._init_optim()
def _init_dataloader(self):
# Define the dataset
train_dataset = dataset.TP_dataset_v1(self.opt.Dataset, 'train')
val_dataset = dataset.TP_dataset_v1(self.opt.Dataset, 'val')
print('The overall number of training images:', len(train_dataset))
print('The overall number of validation images:', len(val_dataset))
# Define the dataloader
self.train_loader = torch.utils.data.DataLoader(train_dataset, batch_size = self.Training_config.train_batch_size, shuffle = True, num_workers = self.Training_config.num_workers, pin_memory = True)
self.val_loader = torch.utils.data.DataLoader(val_dataset, batch_size = self.Training_config.val_batch_size, shuffle = False, num_workers = self.Training_config.num_workers, pin_memory = True)
def _init_optim(self):
if self.optim_config.name == "Adam":
self.optim_G = torch.optim.Adam(self.G.parameters(), lr = self.optim_config.args.lr_g, betas = (self.optim_config.args.b1, self.optim_config.args.b2), weight_decay = self.optim_config.args.weight_decay)
# self.optim_D = torch.optim.Adam(self.D.parameters(), lr = self.optim_config.args.lr_d, betas = (self.optim_config.args.b1, self.optim_config.args.b2), weight_decay = self.optim_config.args.weight_decay)
elif self.optim_config.name == "SGD":
self.optim_G = torch.optim.SGD(self.G.parameters(), lr = self.optim_config.args.lr_g)
# self.optim_D = torch.optim.SGD(self.D.parameters(), lr = self.optim_config.args.lr_d)
elif self.optim_config.name == "Adamw":
self.optim_G = AdamW(self.G.parameters(), lr = self.optim_config.args.lr_g, betas = (self.optim_config.args.b1, self.optim_config.args.b2), weight_decay = self.optim_config.args.weight_decay)
def train(self):
# Count start time
iters_done = 0
for epoch in range(self.Training_config.start_idx + 1, self.Training_config.epochs):
print('epoch', epoch)
# self._validate(epoch)
for param_group in self.optim_G.param_groups:
self.add_scalar('lr', param_group['lr'], epoch)
for i, data in enumerate(self.train_loader):
print(i, self.device)
# ============= Get data ==============
short_img = data['short_img'].to(self.device)
long_img = data['long_img'].to(self.device)
RGBout_img = data['RGBout_img'].to(self.device)
gt_long_img = data['gt_long_img'].to(self.device)
if len(short_img.shape) == 5:
_, _, C, H, W = short_img.shape
short_img = short_img.view(-1, C, H, W)
long_img = long_img.view(-1, C, H, W)
RGBout_img = RGBout_img.view(-1, C, H, W)
gt_long_img = gt_long_img.view(-1, C, H, W)
down_short_img, down_long_img, down_out_img, down_gtlong_img = \
self.train_loader.dataset.downsample_tensors([short_img, long_img, RGBout_img, gt_long_img])
# ============== forward ===============
if self.Training_config.phase == 'deblur':
outs = self.G(down_short_img, down_long_img)
elif self.Training_config.phase == 'denoise':
deblur_out = self.deblurNet(down_short_img, down_long_img).detach()
deblur_out = F.upsample(deblur_out, size = (short_img.shape[2], short_img.shape[3]), mode = 'bilinear', align_corners = False)
short_img, long_img, RGBout_img, deblur_out = \
self.train_loader.dataset.crop_tensor_patch([short_img, long_img, RGBout_img, deblur_out])
deblur_out = self.train_loader.dataset.augment_tensor_patch(deblur_out, RGBout_img)
deblur_out = deblur_out.clamp(0.0, 1.0)
outs = self.G(short_img, long_img, deblur_out)
# ============ pack gt data ============
if self.Training_config.phase == 'deblur':
gt_dict = pack_gt_data([down_out_img, down_gtlong_img])
elif self.Training_config.phase == 'denoise':
gt_dict = pack_gt_data([RGBout_img, gt_long_img])
# ============ pack output =============
if self.Training_config.phase == 'deblur':
outputs = pack_network_output(outs, self.opt.DeblurNet.name)
elif self.Training_config.phase == 'denoise':
outputs = pack_network_output(outs, self.opt.DenoiseNet.name)
# ============== train G ===============
G_loss, G_loss_info = self.LM(outputs, gt_dict, None)
self.optim_G.zero_grad()
G_loss.backward()
self.optim_G.step()
# record loss
if iters_done % self.Training_config.show_loss_iter == 0:
self.add_scalars(main_tag = 'G_loss', tag_scalar_dict = G_loss_info, global_step = iters_done)
if iters_done % self.Training_config.show_img_iter == 0:
if self.Training_config.phase == 'deblur':
vis_imgs = [down_short_img, down_long_img, down_out_img]
elif self.Training_config.phase == 'denoise':
vis_imgs = [short_img, long_img, RGBout_img, deblur_out]
if isinstance(outs, list):
for j in range(len(outs)):
if outs[j] is not None and outs[j].size()[1] == 3:
vis_imgs.append(outs[j].clamp(0.0, 1.0))
else:
outs = outs.clamp(0.0, 1.0)
vis_imgs.append(outs)
self.visual_image('train_img', vis_imgs, iters_done)
self._save_G(self.opt, epoch, iters_done, len(self.train_loader), self.G)
if self.LM.train_GAN:
self._save_D(self.opt, epoch, iters_done, len(self.train_loader), self.D)
self._adjust_learning_rate(self.optim_config, (epoch + 1), iters_done, self.optim_G, self.opt.Optimizer.args.lr_g)
if self.LM.train_GAN:
self._adjust_learning_rate(self.optim_config, (epoch + 1), iters_done, self.optim_D, self.opt.Optimizer.args.lr_d)
iters_done += 1
self._validate(epoch)
def _validate(self, epoch):
self.G.eval()
val_PSNR, val_SSIM, num_of_val_image = 0, 0, 0
for j, data in enumerate(self.val_loader):
# ============= Get data ===============
short_img = data['short_img'].to(self.device)
long_img = data['long_img'].to(self.device)
RGBout_img = data['RGBout_img'].to(self.device)
gt_long_img = data['gt_long_img'].to(self.device)
if len(short_img.shape) == 5:
_, _, C, H, W = short_img.shape
short_img = short_img.view(-1, C, H, W)
long_img = long_img.view(-1, C, H, W)
RGBout_img = RGBout_img.view(-1, C, H, W)
gt_long_img = gt_long_img.view(-1, C, H, W)
down_short_img, down_long_img, down_out_img, down_gtlong_img = \
self.train_loader.dataset.downsample_tensors([short_img, long_img, RGBout_img, gt_long_img])
# ============== forward ===============
with torch.no_grad():
if self.Training_config.phase == 'deblur':
print(down_short_img.shape, down_long_img.shape)
outs = self.G(down_short_img, down_long_img)
elif self.Training_config.phase == 'denoise':
deblur_out = self.deblurNet(down_short_img, down_long_img).detach()
deblur_out = F.upsample(deblur_out, size = (H, W), mode = 'bilinear', align_corners = False)
deblur_out = deblur_out.clamp(0.0, 1.0)
short_img, long_img, RGBout_img, deblur_out = \
self.train_loader.dataset.crop_tensor_patch([short_img, long_img, RGBout_img, deblur_out])
outs = self.G(short_img, long_img, deblur_out)
if isinstance(outs, list):
out = outs[0]
else:
out = outs
out = out.clamp(0.0, 1.0)
num_of_val_image += out.shape[0]
if self.Training_config.phase == 'deblur':
val_PSNR += utils.psnr(out, down_out_img, 1) * out.shape[0]
val_SSIM += utils.ssim(out, down_out_img) * out.shape[0]
elif self.Training_config.phase == 'denoise':
val_PSNR += utils.psnr(out, RGBout_img, 1) * out.shape[0]
val_SSIM += utils.ssim(out, RGBout_img) * out.shape[0]
if j % 10 == 0:
print('val: %d | epoch: %d' % (j, epoch))
val_PSNR = val_PSNR / num_of_val_image
val_SSIM = val_SSIM / num_of_val_image
self.add_scalar('val_PSNR', val_PSNR, global_step = epoch)
self.add_scalar('val_SSIM', val_SSIM, global_step = epoch)
self.G.train()