/
baseline.py
568 lines (431 loc) · 23.3 KB
/
baseline.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
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
import os, math, random, pdb, time, timeit
from tqdm import tqdm
import torch, torchfile
import torchvision.utils as utils
from torch.utils.data import random_split
from torch.utils.data.sampler import SubsetRandomSampler
from utils import *
from data.dataset_util import *
from baseline_models import Baseline_net, gaussian_blur, MultiScaleImageDiscriminator
from style_indicator import New_DA_Net_v1 as DA_Net
from torch.distributions.beta import Beta
from contextual_utils import contextual_loss, contextual_loss_v2
from AdaIN import Adaptive_IN
from evaluation import *
def size_arrange(x):
x_w, x_h = x.size(2), x.size(3)
if (x_w%2) != 0:
x_w = (x_w//2)*2
if (x_h%2) != 0:
x_h = (x_h//2)*2
if ( x_h > 1024 ) or (x_w > 1024) :
old_x_w = x_w
x_w = x_w//2
x_h = int(x_h*x_w/old_x_w)
return F.interpolate(x, size=(x_w, x_h))
def get_HH_LL(x):
pooled = torch.nn.functional.avg_pool2d(x, 2)
up_pooled = torch.nn.functional.interpolate(pooled, scale_factor=2, mode='nearest')
HH = x - up_pooled
LL = up_pooled
return HH, LL
def get_domainess(cur_iter, total_iter, batch):
alpha = np.exp((cur_iter - (0.5 * total_iter)) / (0.25 * total_iter))
distribution = Beta(alpha, 1)
#distribution = Beta(0.5, 0.5)
return distribution.sample((batch, 1))
class Baseline(object):
def __init__(self, args):
super(Baseline, self).__init__()
self.imsize = args.imsize #(512,1024)
self.batch_size = args.batch_size
self.cencrop = args.cencrop
self.cropsize = args.cropsize
self.num_workers = args.num_workers
self.content_dir = args.content_dir
self.style_dir = args.style_dir
self.lr = args.lr
self.train_result_dir = args.train_result_dir
self.DA_comment = args.DA_comment
self.ST_comment = args.ST_comment
self.max_iter = args.max_iter
self.check_iter = args.check_iter
self.args = args
self.is_da_train = args.is_da_train
self.is_st_train = args.is_st_train
#######################################
#### Model Load ####
#######################################
pretrained_vgg = torchfile.load('./baseline_checkpoints/vgg_normalised_conv4_1.t7')
self.network = Baseline_net(pretrained_vgg=pretrained_vgg)
self.network.cuda()
self.DA_Net = DA_Net(self.imsize)
self.DA_Net.cuda()
self.MSD_img = MultiScaleImageDiscriminator(nc=3, ndf=64)
self.MSD_img.cuda()
#######################################
#### Loss function, Optimizer ####
#######################################
for param in self.network.encoder.parameters():
param.requires_grad = False
betas=(0.5, 0.999)
self.dec_optim = torch.optim.Adam(
filter(lambda p: p.requires_grad, self.network.decoder.parameters()),
lr = self.lr,
betas=betas
)
self.Di_optim = torch.optim.Adam(self.MSD_img.parameters(), lr=self.lr, betas=betas, weight_decay=0.00001)
self.enc_optim = torch.optim.Adam(self.DA_Net.parameters(), lr=self.lr, betas=betas)
self.MSE_loss = torch.nn.MSELoss().cuda()
self.bce_loss = torch.nn.BCEWithLogitsLoss().cuda()
self.tv_weight = 1#1e-6
self.result_img_dir = os.path.join(self.train_result_dir, self.ST_comment, 'imgs')
self.result_log_dir = os.path.join(self.train_result_dir, self.DA_comment, 'log')
self.result_st_dir = os.path.join(self.train_result_dir, self.ST_comment, 'log')
os.makedirs(self.result_img_dir, exist_ok=True)
os.makedirs(self.result_log_dir, exist_ok=True)
os.makedirs(self.result_st_dir, exist_ok=True)
def train(self):
# ########################
# wandb.init(project="style_transfer")
# if self.is_da_train:
# wandb.run.name = self.DA_comment
# else:
# wandb.run.name = self.ST_comment
# wandb.config.update(self.args)
# ########################
#######################################
#### Data Loader (Photo, Artistic) ####
#######################################
self.data_set = MSCOCO(self.content_dir, self.imsize, self.cropsize, self.cencrop)
self.art_data_set = WiKiART(self.style_dir, self.imsize, self.cropsize, self.cencrop)
self.data_loader = torch.utils.data.DataLoader(self.data_set, batch_size=self.batch_size, shuffle=True, drop_last=True, num_workers=self.num_workers)
self.art_data_loader = torch.utils.data.DataLoader(self.art_data_set, batch_size=self.batch_size, shuffle=True, drop_last=True, num_workers=self.num_workers)
#######################################
#### Train Phase ####
#######################################
self.network.encoder.train(False)
self.network.decoder.train(True)
self.DA_Net.train(True)
content_weight = 1.0
style_weight = 0.0
d_lambda = 1
c_lambda = 5
_domainess = np.linspace(0, 1, self.max_iter)
best_distance=0.0
print(self.DA_comment," model train start!")
self.content_data_loader_iter = iter(self.data_loader)
self.art_data_loader_iter = iter(self.art_data_loader)
for iteration in range(self.max_iter):
try:
content = next(self.content_data_loader_iter).cuda()
art_ref = next(self.art_data_loader_iter).cuda()#.to(self.device)
except:
self.content_data_loader_iter = iter(self.data_loader)
self.art_data_loader_iter = iter(self.art_data_loader)
content = next(self.content_data_loader_iter).cuda()
art_ref = next(self.art_data_loader_iter).cuda()#.to(self.device)
if self.is_da_train:
self.adain = Adaptive_IN()
with torch.no_grad():
domainess = get_domainess(iteration, (self.max_iter), 1).cuda()
mix_img = (domainess)*(art_ref) + (1-domainess)*(content)
#mix_img = self.adain.interpolate(content, art_ref, domainess)
content_bce_loss = 0
style_bce_loss = 0
cls_loss = 0
domain_loss = 0
dist = []
c_dist = []
s_dist = []
#######################
####origin train code##
#######################
for level in [1,2,3]:
cont_feat = self.network.encoder.get_features(content, level)
cont_feat = self.DA_Net(cont_feat, level)
art_feat = self.network.encoder.get_features(art_ref, level)
art_feat = self.DA_Net(art_feat, level)
mix_feat = self.network.encoder.get_features(mix_img, level)
mix_feat = self.DA_Net(mix_feat, level)
content_bce_loss = self.bce_loss(cont_feat, torch.zeros_like(cont_feat).cuda())
style_bce_loss = self.bce_loss(art_feat, torch.ones_like(art_feat).cuda())
cls_loss += (content_bce_loss + style_bce_loss)
domain_loss += (1-domainess)*(torch.mean(torch.abs((cont_feat) - (mix_feat) ))) + \
(domainess)*(torch.mean(torch.abs((art_feat) - (mix_feat))))
############################
### StyleIndicator Acc ###
############################
with torch.no_grad():
dist.append(float(torch.mean(torch.abs(torch.sigmoid(cont_feat) - torch.sigmoid(art_feat)))))
c_dist.append(float(torch.mean(torch.abs(torch.sigmoid(cont_feat) - torch.sigmoid(mix_feat)))))
s_dist.append(float(torch.mean(torch.abs(torch.sigmoid(art_feat) - torch.sigmoid(mix_feat)))))
total_loss = d_lambda * domain_loss + c_lambda * cls_loss
self.enc_optim.zero_grad()
total_loss.backward()
self.enc_optim.step()
################################
#### Checkpoints ####
################################
if (iteration) % self.check_iter == 0:
print("%s: Iteration: [%d/%d]\tC_loss: %2.4f\tD_loss: %2.4f \tDomainess: %2.4f"%(time.ctime(), iteration, self.max_iter, (cls_loss).item(), domain_loss.item(), domainess))
print("level:%d dist:%2.4f c_dist:%2.4f s_dist:%2.4f"%(0, dist[0], c_dist[0], s_dist[0]))
print("level:%d dist:%2.4f c_dist:%2.4f s_dist:%2.4f"%(1, dist[1], c_dist[1], s_dist[1]))
print("level:%d dist:%2.4f c_dist:%2.4f s_dist:%2.4f"%(2, dist[2], c_dist[2], s_dist[2]))
print("AVG : dist:%2.4f c_dist:%2.4f s_dist:%2.4f"%(np.mean(dist), np.mean(c_dist), np.mean(s_dist)))
# wandb.log({
# #"L/content_bce_loss" : content_bce_loss.item(),
# #"L/style_bce_loss" : style_bce_loss.item(),
# "L/cls_loss" : cls_loss.item(),
# "L/domain_loss" : domain_loss.item(),
# "D/Average_distance" : np.mean(dist)
# })
del content_bce_loss, style_bce_loss, domain_loss, domainess
if ( np.mean(dist) >= 0.95 and np.mean(dist)>best_distance ) or (np.mean(dist) == 1.0):
best_distance = np.mean(dist)
torch.save({'iteration': iteration,
'state_dict': self.DA_Net.state_dict(),},
os.path.join(self.result_log_dir, 'model_'+str(iteration)+'.pth'))
if self.is_st_train:
################################
####Load pretrained_DA_model ###
#### Get Alpha value ####
################################
self.DA_Net.load_state_dict(torch.load(os.path.join(self.result_log_dir, 'style_indicator.pth'))['state_dict'])
###############################
####Step 2 : Reconstruction ###
###############################
empty_segment = np.asarray([])
cont_alphas = self.get_alphas(content)
content_recon = self.network(content, content, empty_segment, empty_segment, is_recon=True, alphas=cont_alphas, type='photo')
art_alphas = self.get_alphas(art_ref)
art_recon = self.network(art_ref, art_ref, empty_segment, empty_segment, is_recon=True, alphas=art_alphas, type='photo')
#########################
####Adversarial parts####
#########################
origin_gan_output = self.MSD_img(content)
recon_gan_output = self.MSD_img(content_recon.detach())
D_content_loss = self.bce_loss(origin_gan_output, ones_like(origin_gan_output)) + self.bce_loss(recon_gan_output, zeros_like(recon_gan_output))
origin_ref_gan_output = self.MSD_img(art_ref)
recon_ref_gan_output = self.MSD_img(art_recon.detach())
D_ref_loss = self.bce_loss(origin_ref_gan_output, ones_like(origin_ref_gan_output)) + self.bce_loss(recon_ref_gan_output, zeros_like(recon_ref_gan_output))
D_loss = D_content_loss + D_ref_loss
self.Di_optim.zero_grad()
D_loss.backward()
self.Di_optim.step()
tv_loss = TVloss(content_recon, self.tv_weight)
feature_recon_loss = []
feature_recon_loss_art = []
cx_loss_photo = 0
cx_loss_art = 0
for level in [1,2,3,4]:
cont_feat = self.network.encoder.get_features(content, level)
cont_recon_feat = self.network.encoder.get_features(content_recon, level)
feature_recon_loss.append(self.MSE_loss(cont_feat, cont_recon_feat))
if level == 4:
cx_loss_photo += (contextual_loss_v2(cont_feat, cont_recon_feat))
del cont_feat, cont_recon_feat
torch.cuda.empty_cache()
for level in [1,2,3,4]:
art_feat = self.network.encoder.get_features(art_ref, level)
art_recon_feat = self.network.encoder.get_features(art_recon, level)
feature_recon_loss_art.append(self.MSE_loss(art_feat, art_recon_feat))
if level == 4:
cx_loss_art += (contextual_loss_v2(art_feat, art_recon_feat))
content_recon_gan_output = self.MSD_img(content_recon)
art_recon_gan_output = self.MSD_img(art_recon)
G_loss = self.bce_loss(content_recon_gan_output, ones_like(content_recon_gan_output)) + self.bce_loss(art_recon_gan_output, ones_like(art_recon_gan_output))
total_loss = torch.mean(torch.stack(feature_recon_loss))*0.1 + torch.mean(torch.stack(feature_recon_loss_art))*0.1 + 1*(cx_loss_photo+cx_loss_art) + tv_loss + 0.1*G_loss
if torch.isnan(total_loss):
continue
self.dec_optim.zero_grad()
total_loss.backward()
self.dec_optim.step()
################################
#### Checkpoints ####
################################
wandb.log({
"L/feature_recon_loss" : torch.mean(torch.stack(feature_recon_loss)).item(),
"L/feature_recon_loss_art" : torch.mean(torch.stack(feature_recon_loss_art)).item(),
"L/total_loss" : total_loss.item(),
"L/D_loss" : D_loss.item(),
"L/G_loss" : G_loss.item(),
"L/cx_loss_photo" : cx_loss_photo.item(),
"L/cx_loss_art" : cx_loss_art.item(),
})
if (iteration) % self.check_iter == 0:
print("%s: Iteration: [%d/%d]\tC_loss: %2.4f"%(time.ctime(), iteration, self.max_iter, total_loss.item()))
print("Alphas : ", cont_alphas.mean(dim=1).cpu().data)
#wandb.log({"Recon_results": wandb.Image(denorm(torch.cat([content, content_recon, art_ref, art_recon]), nrow=self.batch_size))})
#wandb.log({"Fixed_results": wandb.Image(denorm(self.fixed_test(self.result_img_dir, iteration), nrow=8))})
if (iteration) % 2500 == 0:
torch.save({'iteration': iteration,
'state_dict': self.network.state_dict(),},
os.path.join(self.result_st_dir, 'model_'+str(iteration)+'.pth'))
def transfer(self, args):
self.DA_Net.load_state_dict(torch.load(os.path.join(self.result_log_dir, 'style_indicator.pth'))['state_dict'])
self.network.load_state_dict(torch.load(os.path.join(self.result_st_dir, 'decoder.pth'))['state_dict'])
content_set = Transfer_TestDataset(args.test_content, (256,512), self.cropsize, self.cencrop, type='art', is_test=True)
art_reference_set = Transfer_TestDataset(args.test_a_reference, (256,512), self.cropsize, self.cencrop, type='art', is_test=True)
photo_reference_set = Transfer_TestDataset(args.test_p_reference, (256,512), self.cropsize, self.cencrop, type='art', is_test=True)
content_loader = torch.utils.data.DataLoader(content_set, batch_size=self.batch_size, shuffle=False, drop_last=True, num_workers=self.num_workers)
art_reference_loader = torch.utils.data.DataLoader(art_reference_set, batch_size=self.batch_size, shuffle=False, drop_last=True, num_workers=self.num_workers)
photo_reference_loader = torch.utils.data.DataLoader(photo_reference_set, batch_size=self.batch_size, shuffle=False, drop_last=True, num_workers=self.num_workers)
self.DA_Net.train(False)
self.DA_Net.eval()
self.network.train(False)
self.network.eval()
dir_path = os.path.join(self.result_img_dir, 'transfer', self.DA_comment+'_'+self.ST_comment)
if not os.path.exists(dir_path):
os.makedirs(dir_path)
N = art_reference_set.__len__()
content_iter = iter(content_loader)
art_iter = iter(art_reference_loader)
photo_iter = iter(photo_reference_loader)
for iteration in range(N//self.batch_size):
with torch.no_grad():
empty_segment = np.asarray([])
content = next(content_iter).cuda()
a_reference = next(art_iter).cuda()
p_reference = next(photo_iter).cuda()
art_alphas = self.get_alphas(a_reference)
photo_alphas = self.get_alphas(p_reference)
art_stylized_output = self.network(content, a_reference, empty_segment, empty_segment, is_recon=True, alphas=art_alphas, type='photo')
print(str(iteration), 'art : ' , art_alphas)
photo_stylized_output = self.network(content, p_reference, empty_segment, empty_segment, is_recon=True, alphas=photo_alphas, type='photo')
print(str(iteration), 'photo : ' , photo_alphas)
imsave(art_stylized_output, os.path.join(dir_path, 'single_art_stylized_'+str(iteration)+'.png'), nrow=self.batch_size )
imsave(photo_stylized_output, os.path.join(dir_path, 'single_photo_stylized_'+str(iteration)+'.png'), nrow=self.batch_size )
del content, a_reference, p_reference, art_stylized_output, photo_stylized_output
torch.cuda.empty_cache()
time.sleep(0.2)
def transfer_user_guided(self, args):
self.DA_Net.load_state_dict(torch.load(os.path.join(self.result_log_dir, 'style_indicator.pth'))['state_dict'])
self.network.load_state_dict(torch.load(os.path.join(self.result_st_dir, 'decoder.pth'))['state_dict'])
content_set = Transfer_TestDataset(args.test_content, (256,512), self.cropsize, self.cencrop, type='art', is_test=True)
art_reference_set = Transfer_TestDataset(args.test_a_reference, (256,512), self.cropsize, self.cencrop, type='art', is_test=True)
photo_reference_set = Transfer_TestDataset(args.test_p_reference, (256,512), self.cropsize, self.cencrop, type='art', is_test=True)
content_loader = torch.utils.data.DataLoader(content_set, batch_size=self.batch_size, shuffle=False, drop_last=True, num_workers=self.num_workers)
art_reference_loader = torch.utils.data.DataLoader(art_reference_set, batch_size=self.batch_size, shuffle=False, drop_last=True, num_workers=self.num_workers)
photo_reference_loader = torch.utils.data.DataLoader(photo_reference_set, batch_size=self.batch_size, shuffle=False, drop_last=True, num_workers=self.num_workers)
self.DA_Net.train(False)
self.DA_Net.eval()
self.network.train(False)
self.network.eval()
dir_path = os.path.join(self.result_img_dir, 'transfer_user_guided', self.DA_comment+'_'+self.ST_comment)
if not os.path.exists(dir_path):
os.makedirs(dir_path)
N = art_reference_set.__len__()
content_iter = iter(content_loader)
art_iter = iter(art_reference_loader)
photo_iter = iter(photo_reference_loader)
for iteration in range(N//self.batch_size):
with torch.no_grad():
empty_segment = np.asarray([])
content = next(content_iter).cuda()
a_reference = next(art_iter).cuda()
p_reference = next(photo_iter).cuda()
alphas = torch.ones(args.batch_size, 1).repeat(1, 3)*args.alpha
art_stylized_output = self.network(content, a_reference, empty_segment, empty_segment, is_recon=True, alphas=alphas, type='photo')
photo_stylized_output = self.network(content, p_reference, empty_segment, empty_segment, is_recon=True, alphas=alphas, type='photo')
imsave(art_stylized_output, os.path.join(dir_path, 'single_art_stylized_'+str(iteration)+'.png'), nrow=self.batch_size )
imsave(photo_stylized_output, os.path.join(dir_path, 'single_photo_stylized_'+str(iteration)+'.png'), nrow=self.batch_size )
del content, a_reference, p_reference, art_stylized_output, photo_stylized_output
torch.cuda.empty_cache()
time.sleep(0.2)
def transfer_seg(self, args):
from baseline_models_seg import Baseline_net as Baseline_net_seg
pretrained_vgg = torchfile.load('./baseline_checkpoints/vgg_normalised_conv4_1.t7')
self.network = Baseline_net_seg(pretrained_vgg=pretrained_vgg)
self.network.cuda()
self.DA_Net.load_state_dict(torch.load(os.path.join(self.result_log_dir, 'style_indicator.pth'))['state_dict'])
self.network.load_state_dict(torch.load(os.path.join(self.result_st_dir, 'decoder.pth'))['state_dict'])
self.DA_Net.train(False)
self.DA_Net.eval()
self.network.train(False)
self.network.eval()
dir_path = os.path.join(self.result_img_dir, 'transfer_seg2', self.DA_comment+'_'+self.ST_comment)
if not os.path.exists(dir_path):
os.makedirs(dir_path)
for fname in tqdm.tqdm(range(0,60)):
content_fname = 'single_content_'+str(fname)+'.png'
style_fname = 'single_p_reference_'+str(fname)+'.png'
content_segment_fname = 'in'+str(fname)+'.png'
style_segment_fname = 'tar'+str(fname)+'.png'
output_fname = str(fname)+'.png'
with torch.no_grad():
_content = os.path.join(args.test_content, content_fname)
_style = os.path.join(args.test_p_reference, style_fname)
_content_segment = os.path.join(args.test_content_segment, content_segment_fname) if args.test_content_segment else None
_style_segment = os.path.join(args.test_p_reference_segment, style_segment_fname) if args.test_p_reference_segment else None
_output = os.path.join(dir_path, output_fname)
content = open_image(_content, 512).cuda()
style = open_image(_style, 512).cuda()
content_segment = load_segment(_content_segment, 512)
style_segment = load_segment(_style_segment, 512)
alphas = self.get_alphas(style)
stylized_output = self.network(content, style, content_segment, style_segment, is_recon=True, alphas=alphas, type='photo')
print(str(fname), alphas)
imsave(stylized_output, os.path.join(dir_path, 'single_photo_stylized_'+str(fname)+'.png'), nrow=self.batch_size )
#save_image(stylized_output, os.path.join(dir_path, 'single_photo_stylized_'+str(fname)+'.png'))
del content, style, stylized_output
torch.cuda.empty_cache()
time.sleep(0.2)
def interpolate(self, args):
#################################
########## Ours load ############
#################################
self.DA_Net.load_state_dict(torch.load(os.path.join(self.result_log_dir, 'model_'+str(self.DA_Net_trained_epoch)+'.pth'))['state_dict'])
self.network.load_state_dict(torch.load(os.path.join(self.result_st_dir, 'model_'+str(self.decoder_trained_epoch)+'.pth'))['state_dict'])
content_set = Transfer_TestDataset(args.test_content, (1080,1920), self.cropsize, self.cencrop, is_test=True)
art_reference_set = Transfer_TestDataset(args.test_a_reference, (256,512), self.cropsize, self.cencrop, type='art', is_test=True)
photo_reference_set = Transfer_TestDataset(args.test_p_reference, (256,512), self.cropsize, self.cencrop, type='art', is_test=True)
content_loader = torch.utils.data.DataLoader(content_set, batch_size=self.batch_size, shuffle=False, drop_last=True, num_workers=self.num_workers)
art_reference_loader = torch.utils.data.DataLoader(art_reference_set, batch_size=self.batch_size, shuffle=False, drop_last=True, num_workers=self.num_workers)
photo_reference_loader = torch.utils.data.DataLoader(photo_reference_set, batch_size=self.batch_size, shuffle=False, drop_last=True, num_workers=self.num_workers)
self.DA_Net.train(False)
self.DA_Net.eval()
self.network.train(False)
self.network.encoder.eval()
content_loader_iter = iter(content_loader)
art_reference_loader_iter = iter(art_reference_loader)
photo_reference_loader_iter = iter(photo_reference_loader)
empty_segment = np.asarray([])
dir_path = os.path.join(self.result_img_dir, 'interpolate', , self.DA_comment+'_'+self.ST_comment)
N = art_reference_set.__len__()
content_iter = iter(content_loader)
art_iter = iter(art_reference_loader)
photo_iter = iter(photo_reference_loader)
for iteration in range(N//self.batch_size):
with torch.no_grad():
empty_segment = np.asarray([])
content = next(content_iter).cuda()
a_reference = next(art_iter).cuda()
p_reference = next(photo_iter).cuda()
_dir_path = os.path.join(dir_path, str(iteration))
if not os.path.exists(_dir_path):
os.makedirs(_dir_path)
photo_diff = []
art_diff = []
imsave(content, os.path.join(_dir_path, 'content.png'), nrow=self.batch_size)
imsave(p_reference, os.path.join(_dir_path, 'p_reference.png'), nrow=self.batch_size)
imsave(a_reference, os.path.join(_dir_path, 'a_reference.png'), nrow=self.batch_size)
for itr, v in enumerate(np.linspace(0, 1, 11)):
alphas = torch.ones(args.batch_size, 1).repeat(1, 3)*v
photo_results = self.network(content, p_reference, empty_segment, empty_segment, is_recon=True, alphas=alphas, type='photo')
art_results = self.network(content, a_reference, empty_segment, empty_segment, is_recon=True, alphas=alphas, type='photo')
imsave(photo_results, os.path.join(_dir_path, 'photo_interpolate_stylized_'+str(v)+'.png'), nrow=self.batch_size)
imsave(art_results, os.path.join(_dir_path, 'art_interpolate_stylized_'+str(v)+'.png'), nrow=self.batch_size)
imsave(torch.sigmoid(get_HH_LL(photo_results)[0]*30), os.path.join(_dir_path, 'photo_HH_'+str(v)+'.png'), nrow=self.batch_size)
imsave(torch.sigmoid(get_HH_LL(art_results)[0]*30), os.path.join(_dir_path, 'art_HH_'+str(v)+'.png'), nrow=self.batch_size)
del photo_results, art_results
torch.cuda.empty_cache()
def get_alphas(self, imgs):
self.DA_Net.eval()
with torch.no_grad():
alphas = []
for level in [1,2,3]: #2,3,4
feat = self.network.encoder.get_features(imgs, level)
alphas.append(torch.sigmoid(self.DA_Net(feat, level)))
alphas = torch.stack(alphas).unsqueeze(0).cuda()
return alphas