-
Notifications
You must be signed in to change notification settings - Fork 12
/
Copy pathsolver.py
2399 lines (2062 loc) · 125 KB
/
solver.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
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import os
import time
import datetime
from torch.autograd import grad
from torch.autograd import Variable
from torchvision.utils import save_image
from torchvision import transforms
from removalmodels.models import Generator
from removalmodels.models import GeneratorDiff, GeneratorDiffWithInp, GeneratorDiffAndMask, GeneratorDiffAndMask_V2, GeneratorBoxReconst, GeneratorOnlyMask, GeneratorMaskAndFeat, GeneratorMaskAndFeat_ImNetBackbone, GeneratorMaskAndFeat_ImNetBackbone_V2
from removalmodels.models import Discriminator_SN, Discriminator, DiscriminatorBBOX, DiscriminatorGlobalLocal, DiscriminatorGlobalLocal_SN, DiscriminatorGAP, DiscriminatorGAP_ImageNet
from removalmodels.models import BoxFeatEncoder, BoxFeatGenerator, VGGLoss
from PIL import Image
import gc
import random
from utils.utils import show
from collections import OrderedDict
###############################################################################
# Functions
###############################################################################
def weights_init(m):
classname = m.__class__.__name__
if classname.find('Conv') != -1:
m.weight.data.normal_(0.0, 0.02)
elif classname.find('BatchNorm2d') != -1:
m.weight.data.normal_(1.0, 0.02)
m.bias.data.fill_(0)
def save_checkpoint(state, fappend ='dummy', outdir = 'cv'):
filename = os.path.join(outdir,'checkpoint_stargan_'+fappend)
torch.save(state, filename)
class Solver(object):
def __init__(self, celebA_loader, rafd_loader, config, mode='train', pretrainedcv=None, mask_loader = None):
# Data loader
self.celebA_loader = celebA_loader
self.rafd_loader = rafd_loader
self.arch = vars(config)
self.differential_generator = config.differential_generator
self.mask_loader = mask_loader
self.use_maskprior_gan = getattr(config, 'use_maskprior_gan', 0)
self.use_gtmask_inp = getattr(config, 'use_gtmask_inp', 0)
self.boxprior = getattr(config, 'boxprior', 0)
self.maskprior_matchclass = getattr(config, 'maskprior_matchclass', 0)
self.alternate_mask_train= getattr(config, 'alternate_mask_train', 0)
self.train_g_every = getattr(config, 'train_g_every', 2)
self.mode = mode
# Model hyper-parameters
self.c_dim = config.c_dim
self.image_size = config.image_size
self.box_size = config.box_size
self.boxfeat_dim = config.boxfeat_dim
self.g_conv_dim = config.g_conv_dim
self.g_smooth_layers = config.g_smooth_layers
self.g_downsamp_layers= getattr(config, 'g_downsamp_layers', 2)
self.g_dil_start = config.g_dil_start
self.g_repeat_num = config.g_repeat_num
self.g_binary_mask= config.g_binary_mask
self.g_upsample_type= getattr(config, 'g_upsample_type', 't_conv')
self.g_nupsampFilters= getattr(config, 'g_nupsampFilters', 1)
self.g_pad_type = getattr(config, 'g_pad_type', 'zeros')
self.g_fixed_size= getattr(config, 'g_fixed_size', 0)
self.gen_fullimage = getattr(config, 'gen_fullimage', 0)
self.m_upsample_type = getattr(config, 'm_upsample_type', 'bilinear')
self.d_use_spectralnorm= getattr(config, 'd_use_spectralnorm', 0)
self.d_max_filters = getattr(config, 'd_max_filters', None)
self.d_max_filters_cls = getattr(config, 'd_max_filters_cls', None)
self.d_global_pool = getattr(config, 'd_global_pool', 'mean')
self.d_kernel_size = getattr(config, 'd_kernel_size', 4)
self.d_conv_dim = config.d_conv_dim
self.d_repeat_num = config.d_repeat_num
self.e_repeat_num = getattr(config, 'e_repeat_num',self.d_repeat_num-1)
self.d_init_stride= config.d_init_stride
self.d_train_repeat = config.d_train_repeat
self.d_local_supervision = getattr(config, 'd_local_supervision', 0)
self.d_patch_size = getattr(config, 'd_patch_size', 2)
self.e_norm_type = config.e_norm_type
self.e_ksize = config.e_ksize
self.lowres_mask = getattr(config, 'lowres_mask',0)
self.mask_size = getattr(config, 'mask_size',32)
self.mask_additional_cond = getattr(config, 'mask_additional_cond','image')
self.perclass_mask = getattr(config, 'perclass_mask',0)
self.mask_noinplabel = getattr(config, 'mask_noinplabel',0)
self.mask_normalize_byclass = getattr(config, 'mask_normalize_byclass',0)
self.use_seperate_classifier = config.use_seperate_classifier
self.use_gap_classifier = getattr(config, 'use_gap_classifier',0)
self.use_imagenet_pretrained = getattr(config, 'use_imagenet_pretrained',None)
self.use_imagenet_pretrained_mask = getattr(config, 'use_imagenet_pretrained_mask',None)
self.cond_inp_pnet = getattr(config, 'cond_inp_pnet',0)
self.cond_parallel_track = getattr(config, 'cond_parallel_track',0)
self.use_imnetmask_v2= getattr(config, 'use_imnetmask_v2',0)
self.use_bnorm= getattr(config, 'use_bnorm',0)
self.use_bnorm_mask = getattr(config, 'use_bnorm_mask',0) # 0 is no norm, 1 is batch norm, 2 is instance norm
self.adv_classifier = config.adv_classifier
self.adv_loss_type = getattr(config, 'adv_loss_type', 'wgan')
self.m_adv_loss_type = getattr(config, 'm_adv_loss_type', 'wgan')
self.train_boxreconst = config.train_boxreconst
self.full_image_encoder = getattr(config,'full_image_encoder', 0)
self.only_reconst_loss = getattr(config,'only_reconst_loss',0)
self.use_box_label = getattr(config,'use_box_label',0)
self.g_fine_tune= getattr(config,'g_fine_tune',0)
self.no_inpainter = getattr(config,'no_inpainter',0)
self.fixed_m = getattr(config,'fixed_m',0)
self.only_gt_mask = getattr(config,'only_gt_mask',0)
self.train_only_g= getattr(config,'train_only_g',0)
if self.train_only_g:
self.only_gt_mask = 1
self.dilateMask = getattr(config,'dilateMask',0)
self.train_g_wo_m = getattr(config, 'train_g_wo_m',0)
self.fixed_classifier = getattr(config,'fixed_classifier',0)
self.train_only_d = getattr(config,'train_only_d',0)
self.train_robust_d= getattr(config,'train_robust_d',0)
self.compositional_loss= getattr(config,'compositional_loss',0)
self.use_random_boxes= getattr(config,'use_random_boxes',0)
self.use_past_masks = getattr(config,'use_past_masks',0)
self.use_tv_inp = getattr(config,'use_tv_inp',0)
self.only_random_boxes_discr = getattr(config,'only_random_boxes_discr',0)
self.discrim_masked_image = getattr(config,'discrim_masked_image',0)
# Hyper-parameteres
self.use_topk_patch = config.use_topk_patch
self.lambda_cls = config.lambda_cls
self.onlytargetclsloss = getattr(config,'onlytargetclsloss',0)
self.useclsweights = getattr(config,'useclsweights',0)
self.clsweight_scale_decay= getattr(config,'clsweight_scale_decay',0)
self.clsweight_scale_init = getattr(config,'clsweight_scale_init',0)
self.lambda_rec = config.lambda_rec
self.lambda_feat_match = getattr(config, 'lambda_feat_match', 0)
self.lambda_vggloss = getattr(config, 'lambda_vggloss', 0)
self.use_style_loss = getattr(config, 'use_style_loss', 0)
self.lambda_smoothloss= getattr(config, 'lambda_smoothloss', 0)
self.lambda_tvloss= getattr(config, 'lambda_tvloss', 0.)
self.lambda_maskL1loss= getattr(config, 'lambda_maskL1loss', 0)
self.grad_weighted_l1loss= getattr(config, 'grad_weighted_l1loss', 0)
self.lambda_maskfake_loss= getattr(config, 'lambda_maskfake_loss', 1.)
self.vggloss_nw = getattr(config, 'vggloss_nw', 'vgg')
self.onlymaskgen= getattr(config, 'onlymaskgen', 0)
self.e_addnoise = getattr(config, 'e_addnoise', 0)
self.e_masked_image= getattr(config, 'e_masked_image', 0)
self.e_use_residual= getattr(config, 'e_use_residual', 0)
self.e_bias = getattr(config, 'e_bias', False)
self.load_encoder= getattr(config, 'load_encoder', 1)
self.load_generator= getattr(config, 'load_generator', 1)
self.load_discriminator = getattr(config, 'load_discriminator', 1)
self.adv_rec = config.adv_rec
self.lambda_gp = config.lambda_gp
self.g_lr = config.g_lr
self.e_lr = 1.0* getattr(config, 'e_lr', self.g_lr)
self.d_lr = config.d_lr
self.beta1 = config.beta1
self.beta2 = config.beta2
# Training settings
self.dataset = config.dataset
self.nc = 1 if self.dataset == 'mnist' else 3
self.num_epochs = config.num_epochs
self.num_epochs_decay = config.num_epochs_decay
self.decay_every = getattr(config, 'decay_every', 5)
self.num_iters = config.num_iters
self.num_iters_decay = config.num_iters_decay
self.batch_size = config.batch_size
self.use_tensorboard = config.use_tensorboard
self.pretrained_model = config.pretrained_model
# Test settings
self.test_model = config.test_model
# Path
self.log_path = config.log_path
self.sample_path = config.sample_path
self.model_save_path = config.model_save_path
self.result_path = config.result_path
# Step size
self.log_step = config.log_step
self.sample_step = config.sample_step
self.model_save_step = config.model_save_step
self.selected_attrs = config.selected_attrs
# Build tensorboard if use
self.build_model(pretrainedcv)
if self.use_tensorboard:
self.build_tensorboard()
if config.onlypretrained_discr and (self.mode=='train' or self.mode=='eval'):
self.load_pretrained_discr(config.onlypretrained_discr)
if getattr(config, 'onlypretrained_encoder',None):
self.load_pretrained_encoder(config.onlypretrained_encoder)
if getattr(config, 'onlypretrained_generator',None):
cvG = torch.load(config.onlypretrained_generator)
self.load_pretrained_generator(cvG)
def build_model(self, pretrainedcv):
# Define a generator and a discriminator
#-------------------------------
# Initialize the image generator
#-------------------------------
if self.train_boxreconst:
#------------------------------------------------------------------------------------------------
# This is the box generation mode
#------------------------------------------------------------------------------------------------
GenClass = GeneratorBoxReconst
self.G = GenClass(self.g_conv_dim, self.boxfeat_dim, self.g_repeat_num, self.g_downsamp_layers,
self.g_dil_start, self.g_upsample_type, self.g_pad_type, nc=self.nc,
n_upsamp_filt=self.g_nupsampFilters, gen_full_image=self.gen_fullimage)
else:
#--------------------------------------------------------------------------
# This is the full image (could be differential and masked) generation mode
#--------------------------------------------------------------------------
GenClass = GeneratorOnlyMask if self.onlymaskgen==1 else Generator if not self.differential_generator else GeneratorDiff if self.differential_generator==1 else GeneratorDiffWithInp if self.differential_generator==2 else GeneratorDiffAndMask if self.differential_generator==3 else GeneratorDiffAndMask_V2
self.G = GenClass(self.g_conv_dim, self.c_dim, self.g_repeat_num, self.g_smooth_layers, self.g_binary_mask)
#-------------------------------
# Initialize the real/fake discriminator and classifier
#-------------------------------
if (self.mode == 'train' or self.mode=='eval'):
if (self.train_boxreconst==1 or self.train_boxreconst==2):
DisClass = DiscriminatorGlobalLocal_SN if self.d_use_spectralnorm else DiscriminatorGlobalLocal
self.D = DisClass(self.image_size, self.box_size, self.d_conv_dim, self.c_dim, self.d_repeat_num,
self.d_repeat_num-1, nc=self.nc, d_kernel_size=self.d_kernel_size)
else:
DisClass = Discriminator_SN if self.d_use_spectralnorm else Discriminator
self.D = DisClass(self.image_size, self.d_conv_dim, self.c_dim, self.d_repeat_num,
self.d_init_stride, max_filters = self.d_max_filters, nc = self.nc*(1+self.discrim_masked_image),
d_kernel_size=self.d_kernel_size, patch_size = self.d_patch_size, use_tv_inp=self.use_tv_inp)
if self.use_seperate_classifier and not self.train_only_g:
if self.use_imagenet_pretrained is not None:
self.D_cls = DiscriminatorGAP_ImageNet(self.image_size, self.c_dim, net_type = self.use_imagenet_pretrained, max_filters = self.d_max_filters_cls, global_pool = self.d_global_pool)
else:
DisClass = DiscriminatorGAP if self.use_gap_classifier else DisClass
self.D_cls = DisClass(self.image_size, self.d_conv_dim, self.c_dim, 4 if self.use_gap_classifier else self.d_repeat_num ,
self.d_init_stride, max_filters = self.d_max_filters_cls, use_bnorm=self.use_bnorm)
else:
self.D_cls = None
else:
self.D = None
self.D_cls = None
if not self.train_only_g:
if ((self.train_boxreconst == 1) and not self.full_image_encoder) and (self.boxfeat_dim > 0):
#-----------------------------------------------------------------------------
# Initialize the box image encoder to pre-train the generator as an autoencoder
#-----------------------------------------------------------------------------
self.E = BoxFeatEncoder(self.box_size, self.e_ksize, self.d_conv_dim, self.boxfeat_dim,
self.e_repeat_num, (self.use_box_label*self.c_dim),
self.e_norm_type, nc=self.nc)
elif ((self.train_boxreconst == 2) or self.full_image_encoder) and (self.boxfeat_dim > 0):
#-----------------------------------------------------------------------------
# Initialize the feature generator network to do editing (adding/removing)
#-----------------------------------------------------------------------------
self.E = BoxFeatGenerator(self.image_size, self.e_ksize, self.d_conv_dim, self.boxfeat_dim,
c_dim=(self.use_box_label*self.c_dim)+(self.e_masked_image>1)*3,
use_residual=self.e_use_residual, nc=self.nc)
elif (self.train_boxreconst == 3):
#-----------------------------------------------------------------------------
# Initialize the mask and feature generator network to do editing (adding/removing)
#-----------------------------------------------------------------------------
#def __init__(self, conv_dim=64, c_dim=5, repeat_num=5, g_smooth_layers=0, binary_mask=0, out_feat_dim=256, up_sampling_type='bilinear', n_upsamp_filt=2):
maskgen = GeneratorMaskAndFeat if self.use_imagenet_pretrained_mask is None else GeneratorMaskAndFeat_ImNetBackbone_V2 if self.use_imnetmask_v2 else GeneratorMaskAndFeat_ImNetBackbone
self.E = maskgen(self.g_conv_dim, self.c_dim, self.e_repeat_num, 0,binary_mask = self.g_binary_mask,
out_feat_dim = self.boxfeat_dim, mask_size = self.mask_size if self.image_size == 128 else self.mask_size//2,
additional_cond = self.mask_additional_cond, per_classMask = self.perclass_mask,
noInpLabel= self.mask_noinplabel, mask_normalize = self.mask_normalize_byclass, nc = 4 if self.use_gtmask_inp else 3,
use_bias = self.e_bias, use_bnorm = self.use_bnorm_mask, cond_inp_pnet=self.cond_inp_pnet,
cond_parallel_track = self.cond_parallel_track)
else:
self.E = None
else:
self.E = None
if self.lambda_vggloss:
self.vggLoss = VGGLoss(network=self.vggloss_nw, imagenet_norm = False, use_style_loss = self.use_style_loss)
if self.use_maskprior_gan:
classify_type = 0 if self.c_dim ==1 else 2
self.mask_D = Discriminator(self.mask_size, self.d_conv_dim, self.c_dim, int(np.log2(self.mask_size) - 1),
self.d_init_stride, classify_branch = classify_type, max_filters = 256 if self.c_dim == 1 else 512,
nc=1)
if self.g_upsample_type!= 'deform':
self.G.apply(weights_init)
if self.mode == 'train' and not self.d_use_spectralnorm:
self.D.apply(weights_init)
if (self.D_cls is not None) and self.mode == 'train' and (not self.d_use_spectralnorm) and self.use_seperate_classifier and (self.use_imagenet_pretrained is None):
self.D_cls.apply(weights_init)
if (self.E is not None) and (self.use_imagenet_pretrained_mask is None):
self.E.apply(weights_init)
#if self.mask_normalize_byclass:
# self.E.final_Layer_mask.bias.data[-1] = 3.
if self.use_maskprior_gan:
self.mask_D.apply(weights_init)
# Optimizers
if self.mode == 'train':
if self.g_upsample_type== 'deform':
#split the parameters into groups
g_params = [{'params':[p[1] for p in self.G.named_parameters() if 'coordfilter' not in p[0]]}]
g_params.append({'params':[p[1] for p in self.G.named_parameters() if 'coordfilter' in p[0]], 'lr':self.g_lr*0.1})
else:
g_params = self.G.parameters()
self.g_optimizer = torch.optim.Adam(g_params, self.g_lr, [self.beta1, self.beta2])#, amsgrad=True)
self.d_optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad, self.D.parameters()), self.d_lr, [self.beta1, self.beta2])#, amsgrad=True)
if (self.D_cls is not None) and self.use_seperate_classifier:
self.dcls_optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad, self.D_cls.parameters()), self.d_lr, [self.beta1, self.beta2], weight_decay=0.01)#, amsgrad=True)
if self.E is not None:
self.e_optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad, self.E.parameters()), self.e_lr, [self.beta1, self.beta2])#, amsgrad=True)
if self.use_maskprior_gan:
self.mask_d_optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad, self.mask_D.parameters()), self.d_lr, [self.beta1, self.beta2])#, amsgrad=True)
# Print networks
self.print_network(self.G, 'G')
self.print_network(self.D, 'D')
if (self.D_cls is not None) and self.use_seperate_classifier:
self.print_network(self.D_cls, 'D_cls')
if self.E is not None:
self.print_network(self.E, 'E')
if self.use_maskprior_gan:
self.print_network(self.mask_D, 'mask_D')
# Before the training begins, save the requires_grad setting for all the params
for p in self.G.parameters():
p.requires_grad_orig = bool(p.requires_grad and ((self.train_boxreconst<2) or self.g_fine_tune))#
for p in self.D.parameters():
p.requires_grad_orig = p.requires_grad#
if (self.D_cls is not None) and self.use_seperate_classifier:
for p in self.D_cls.parameters():
p.requires_grad_orig = p.requires_grad#
if self.E is not None:
for p in self.E.parameters():
p.requires_grad_orig = p.requires_grad#
if self.use_maskprior_gan:
for p in self.mask_D.parameters():
p.requires_grad_orig = p.requires_grad#
# Start with trained model
if self.pretrained_model:
self.load_pretrained_model(pretrainedcv)
if torch.cuda.is_available():
if torch.cuda.device_count() > 1:
print 'Using multiple GPUs'
self.G = nn.DataParallel(self.G)
#self.vggLoss = nn.DataParallel(self.vggLoss)
self.G.cuda()
if (self.mode == 'train' or self.mode=='eval'):
self.D.cuda()
if (self.D_cls is not None) and self.use_seperate_classifier and (self.mode == 'train' or self.mode=='eval'):
self.D_cls.cuda()
if self.E is not None:
self.E.cuda()
if self.use_maskprior_gan and (self.mode == 'train' or self.mode=='eval'):
self.mask_D.cuda()
def getImageSizeMask(self, mask):
return F.upsample(mask, scale_factor=int(self.image_size/mask.size(-1)), mode=self.m_upsample_type) if mask.size(-1)!= self.image_size else mask
def print_network(self, model, name):
num_params = 0
for p in model.parameters():
if p.requires_grad == True:
num_params += p.numel()
print(name)
print(model)
print("The number of parameters: {}".format(num_params))
def load_pretrained_model(self, cv=None):
if cv is None:
cv = torch.load(self.pretrained_model)
if 'generator_state_dict' in cv and self.load_generator:
self.load_pretrained_generator(cv)
else:
print 'Generator not found'
if (self.mode == 'train' or self.mode=='eval') and self.load_discriminator: #and not self.d_use_spectralnorm:
if 'discriminator_state_dict' in cv:
self.D.load_state_dict(cv['discriminator_state_dict'])
else:
print 'Discriminator not found'
if (self.E is not None) and self.load_encoder:
if 'encoder_state_dict' in cv:
self.E.load_state_dict(cv['encoder_state_dict'])
else:
print 'Encoder not found'
if self.use_maskprior_gan and (self.mode == 'train'):
if 'mask_discriminator_state_dict' in cv:
self.mask_D.load_state_dict(cv['mask_discriminator_state_dict'])
else:
print 'Mask discriminator not found'
if (self.D_cls is not None) and self.use_seperate_classifier and self.load_discriminator:
if 'discriminator_cls_state_dict' in cv:
self.D_cls.load_state_dict(cv['discriminator_cls_state_dict'])
else:
print 'Object classifier not found'
print('loaded trained models (step: {})..!'.format(self.pretrained_model))
def load_pretrained_encoder(self, fname):
cv = torch.load(fname)
if (self.E is not None):
self.E.load_state_dict(cv['encoder_state_dict'])
def load_pretrained_generator(self, cv):
new_state_dict = OrderedDict()
for k, v in cv['generator_state_dict'].items():
# remove `module.`
name = k[7:] if k.startswith('module.') else k
new_state_dict[name] = v
cv['generator_state_dict'] = new_state_dict
if (self.G is not None):
self.G.load_state_dict(cv['generator_state_dict'])
def load_pretrained_discr(self, fname):
cv = torch.load(fname)
if 'discriminator_state_dict' in cv:
self.D.load_state_dict(cv['discriminator_state_dict'])
if (self.D_cls is not None) and self.use_seperate_classifier and 'discriminator_cls_state_dict' in cv:
self.D_cls.load_state_dict(cv['discriminator_cls_state_dict'])
if self.use_maskprior_gan and 'mask_discriminator_state_dict' in cv and (self.mode == 'train'):
self.mask_D.load_state_dict(cv['mask_discriminator_state_dict'])
print('loaded pre-trained discr (step: {})..!'.format(fname))
def build_tensorboard(self):
from logger import Logger
self.logger = Logger(self.log_path)
def update_lr(self, g_lr, d_lr, e_lr=None):
for param_group in self.g_optimizer.param_groups:
param_group['lr'] = g_lr
for param_group in self.d_optimizer.param_groups:
param_group['lr'] = d_lr
if self.use_seperate_classifier:
for param_group in self.dcls_optimizer.param_groups:
param_group['lr'] = d_lr
if (self.E is not None) and (e_lr is not None):
for param_group in self.e_optimizer.param_groups:
param_group['lr'] = e_lr
if self.use_maskprior_gan:
for param_group in self.mask_d_optimizer.param_groups:
param_group['lr'] = d_lr
self.g_lr = g_lr
self.d_lr = d_lr
self.e_lr = e_lr
def reset_grad(self):
self.g_optimizer.zero_grad()
self.d_optimizer.zero_grad()
if self.use_seperate_classifier:
self.dcls_optimizer.zero_grad()
if self.E is not None:
self.e_optimizer.zero_grad()
if self.use_maskprior_gan:
self.mask_d_optimizer.zero_grad()
def to_var(self, x, volatile=False):
return Variable(x.cuda() if torch.cuda.is_available() else x , volatile=volatile)
def denorm(self, x):
out = (x + 1) / 2
return out.clamp_(0, 1)
def threshold(self, x):
return (x>0.5).float()
def compute_accuracy(self, x, y, dataset):
x = F.sigmoid(x)
predicted = self.threshold(x)
correct = (predicted == y).float()
accuracy = torch.mean(correct, dim=0) * 100.0
return accuracy
def one_hot(self, labels, dim):
"""Convert label indices to one-hot vector"""
batch_size = labels.size(0)
out = torch.zeros(batch_size, dim)
out[np.arange(batch_size), labels.long()] = 1
return out
def make_coco_labels(self, real_c):
"""Generate domain labels for CelebA for debugging/testing.
if dataset == 'CelebA':
return single and multiple attribute changes
elif dataset == 'Both':
return single attribute changes
"""
y = np.eye(real_c.size(1))
fixed_c_list = []
# single object addition and removal
for i in range(2*self.c_dim):
fixed_c = real_c.clone()
for c in fixed_c:
if i%2:
c[i//2] = 0.
else:
c[i//2] = 1.
fixed_c_list.append(self.to_var(fixed_c, volatile=True))
# multi-attribute transfer (H+G, H+A, G+A, H+G+A)
#if self.dataset == 'CelebA':
# for i in range(4):
# fixed_c = real_c.clone()
# for c in fixed_c:
# if i in [0, 1, 3]: # Hair color to brown
# c[:3] = y[2]
# if i in [0, 2, 3]: # Gender
# c[3] = 0 if c[3] == 1 else 1
# if i in [1, 2, 3]: # Aged
# c[4] = 0 if c[4] == 1 else 1
# fixed_c_list.append(self.to_var(fixed_c, volatile=True))
return fixed_c_list
def make_celeb_labels(self, real_c):
"""Generate domain labels for CelebA for debugging/testing.
if dataset == 'CelebA':
return single and multiple attribute changes
elif dataset == 'Both':
return single attribute changes
"""
y = [torch.FloatTensor([1, 0, 0]), # black hair
torch.FloatTensor([0, 1, 0]), # blond hair
torch.FloatTensor([0, 0, 1])] # brown hair
fixed_c_list = []
# single attribute transfer
for i in range(self.c_dim):
fixed_c = real_c.clone()
for c in fixed_c:
if i < 3:
c[:3] = y[i]
else:
c[i] = 0 if c[i] == 1 else 1 # opposite value
fixed_c_list.append(self.to_var(fixed_c, volatile=True))
# multi-attribute transfer (H+G, H+A, G+A, H+G+A)
if self.dataset == 'CelebA':
for i in range(4):
fixed_c = real_c.clone()
for c in fixed_c:
if i in [0, 1, 3]: # Hair color to brown
c[:3] = y[2]
if i in [0, 2, 3]: # Gender
c[3] = 0 if c[3] == 1 else 1
if i in [1, 2, 3]: # Aged
c[4] = 0 if c[4] == 1 else 1
fixed_c_list.append(self.to_var(fixed_c, volatile=True))
return fixed_c_list
def compute_gradient_penalty(self, real_x, fake_x, discrim_type= 'regular', label=None, bbox=None):
rsz = real_x.shape[0:1] + (1,)*(real_x.ndimension()-1)
alpha = torch.rand(*rsz).cuda().expand_as(real_x)
interpolated = Variable(alpha * real_x.data + (1 - alpha) * fake_x.data, requires_grad=True)
if discrim_type == 'regular':
if bbox is None:
out, out_cls = self.D(interpolated)
else:
out, out_cls, _ = self.D(interpolated, self.getRandBoxWith(interpolated, bbox))
elif discrim_type == 'seperate':
out, out_cls = self.D_cls(interpolated)
elif discrim_type == 'mask':
out, _ = self.mask_D(interpolated , label = label)
## RS HACK
#interpout = torch.cat([out.view(out.size(0), -1), out_cls],dim=-1)
if not self.onlymaskgen:
interpout = out
else:
interpout = out_cls
grad = torch.autograd.grad(outputs=interpout,
inputs=interpolated,
grad_outputs=torch.ones(interpout.size()).cuda(),
retain_graph=True,
create_graph=True,
only_inputs=True)[0]
grad = grad.view(grad.size(0), -1)
grad_l2norm = torch.sqrt(torch.sum(grad ** 2, dim=1))
#import ipdb; ipdb.set_trace()
d_loss_gp = torch.mean((grad_l2norm - 1)**2)
return d_loss_gp
def compute_real_fake_loss_local(self, scores, loss_type, label, loss_for='discr', only_pos = False):
if loss_for == 'discr':
scores = scores.view(scores.size(0),-1)
label = (label.data.view(label.size(0),-1)>0.01).float()
score_p = scores[(1.-label).byte()]
if only_pos:
score_p = scores
score_n = []
n_n = 0
else:
score_n = scores[label.byte()]
n_n = score_n.numel()
n_p = score_p.numel()
d_loss_p = 0.
d_loss_n = 0.
if loss_type == 'lsgan':
# The Loss for least-square gan
if len(score_p):
d_loss_p = torch.pow(score_p - 1., 2).sum()
if len(score_n):
d_loss_n = torch.pow(score_n + 1 , 2).sum()
elif loss_type == 'wgan':
# The Loss for Wgan
if len(score_p):
d_loss_p += -(torch.sum(score_p))
if len(score_n):
sign = 1. if not only_pos else -1.
d_loss_n += sign*torch.sum(score_n)
else:
if len(score_p):
d_loss += (F.binary_cross_entropy_with_logits(score_p,torch.ones_like(score_p).detach()))
if len(score_n):
targ = torch.zeros_like(score_n) if not only_pos else torch.ones_like(score_n)
d_loss += (F.binary_cross_entropy_with_logits(score_n,targ.detach()))
return d_loss_p, d_loss_n, n_p, n_n
def compute_real_fake_loss(self, scores, loss_type, datasrc = 'real', loss_for='discr'):
if loss_for == 'discr':
if datasrc == 'real':
if loss_type == 'lsgan':
# The Loss for least-square gan
d_loss = torch.pow(scores - 1., 2).mean()
elif loss_type == 'hinge':
# Hinge loss used in the spectral GAN paper
d_loss = - torch.mean(torch.clamp(scores-1.,max=0.))
elif loss_type == 'wgan':
# The Loss for Wgan
d_loss = - torch.mean(scores)
else:
scores = scores.view(scores.size(0),-1).mean(dim=1)
d_loss = F.binary_cross_entropy_with_logits(scores, torch.ones_like(scores).detach())
else:
if loss_type == 'lsgan':
# The Loss for least-square gan
d_loss = torch.pow((scores),2).mean()
elif loss_type == 'hinge':
# Hinge loss used in the spectral GAN paper
d_loss = -torch.mean(torch.clamp(-scores-1.,max=0.))
elif loss_type == 'wgan':
# The Loss for Wgan
d_loss = torch.mean(scores)
else:
scores = scores.view(scores.size(0),-1).mean(dim=1)
d_loss = F.binary_cross_entropy_with_logits(scores, torch.zeros_like(scores).detach())
return d_loss
else:
if loss_type == 'lsgan':
# The Loss for least-square gan
g_loss = torch.pow(scores - 1., 2).mean()
elif (loss_type == 'wgan') or (loss_type == 'hinge') :
g_loss = - torch.mean(scores)
else:
scores = scores.view(scores.size(0),-1).mean(dim=1)
g_loss = F.binary_cross_entropy_with_logits(scores, torch.ones_like(scores).detach())
return g_loss
def forward_generator(self, real_x, boxImg=None, imagelabel = None, mask=None, boxlabel=None, bbox=None, get_feat=None, onlyMasks = False, mask_threshold = 0.3, withGTMask=False, dilate=None, n_iter=0):
if self.train_boxreconst == 3:
fake_x,_,mask = self.forward_fulleditor(real_x, imagelabel, onlyMasks=onlyMasks, mask_threshold = mask_threshold, gtMask=mask, withGTMask=withGTMask, dilate=dilate, n_iter = n_iter)
elif self.train_boxreconst > 0:
if get_feat:
fake_x, feat = self.forward_boxreconst(real_x, boxImg, mask, boxlabel, bbox,get_feat=get_feat)
else:
fake_x = self.forward_boxreconst(real_x, boxImg, mask, boxlabel, bbox)
elif self.train_boxreconst == 0:
fake_x, mask = self.G(real_x, boxlabel, out_diff = 1)
return fake_x, mask
def classify(self, x ):
if self.use_seperate_classifier:
return self.D_cls(x)
else:
return self.D(x)
def train(self):
# Set dataloader
if self.dataset in ['CelebA', 'coco', 'mnist', 'celebbox', 'pascal']:
self.data_loader = self.celebA_loader
else:
self.data_loader = self.rafd_loader
# The number of iterations per epoch
iters_per_epoch = len(self.data_loader)
fixed_x = []
real_c = []
for i, (images, labels) in enumerate(self.data_loader):
fixed_x.append(images)
real_c.append(labels)
if i == 3:
break
# Fixed inputs and target domain labels for debugging
fixed_x = torch.cat(fixed_x, dim=0)
fixed_x = self.to_var(fixed_x, volatile=True)
real_c = torch.cat(real_c, dim=0)
if self.dataset == 'CelebA':
fixed_c_list = self.make_celeb_labels(real_c)
elif self.dataset == 'coco':
fixed_c_list = self.make_coco_labels(real_c)
elif self.dataset == 'RaFD':
fixed_c_list = []
for i in range(self.c_dim):
fixed_c = self.one_hot(torch.ones(fixed_x.size(0)) * i, self.c_dim)
fixed_c_list.append(self.to_var(fixed_c, volatile=True))
# lr cache for decaying
g_lr = self.g_lr
d_lr = self.d_lr
e_lr = self.e_lr
# Start with trained model if exists
if self.pretrained_model:
start = int(self.pretrained_model.split('_')[-2])
else:
start = 0
if self.lambda_smoothloss:
smoothWeight = -1.0/9.*torch.ones((1,1,3,3))
smoothWeight[0,0,1,1] = 1.0
smoothWeight = Variable(smoothWeight,requires_grad=False).cuda()
# Start training
loss = {}
accLog = {}
avgexp= 0.95
start_time = time.time()
cocoAndCelebset= set(['CelebA', 'coco', 'pascal'])
for e in range(start, self.num_epochs):
for i, (real_x, real_label) in enumerate(self.data_loader):
# Generat fake labels randomly (target domain labels)
#plt.imshow(((real_x.numpy()[6,[0,1,2],:,:].transpose(1,2,0)+1.0)*255./2.0).astype(np.uint8)); plt.show()
cbsz = real_label.size(0)
if self.dataset == 'coco':
rand_idx = np.random.randint(0,real_label.size(1),cbsz)
aidx= np.arange(cbsz)
fake_label = real_label.clone()
# XXX: FIX THIS!!!!!
fake_label[aidx,rand_idx] = 1. - real_label[aidx,rand_idx] if not self.arch['only_remove_train'] else 0.
else:
rand_idx = torch.randperm(real_label.size(0))
fake_label = real_label[rand_idx]
for p in self.G.parameters():
p.requires_grad = False#
real_c = real_label.clone()
fake_c = fake_label.clone() #if (np.random.rand()> 0.5) or (self.dataset=='CelebA') else real_label.clone()
# Convert tensor to variable
real_x = self.to_var(real_x)
real_c = self.to_var(real_c) # input for the generator
fake_c = self.to_var(fake_c)
real_label = self.to_var(real_label) # this is same as real_c if dataset == 'CelebA'
fake_label = self.to_var(fake_label)
# ================== Train D ================== #
# Compute loss with real images
out_src, out_cls = self.D(real_x)
if self.use_seperate_classifier:
out_cls = self.D_cls(real_x)
d_loss_real = self.compute_real_fake_loss(out_src, self.adv_loss_type, datasrc = 'real')
if self.dataset in cocoAndCelebset:
d_loss_cls = F.binary_cross_entropy_with_logits(
out_cls.view(cbsz, -1), real_label, size_average=False) / cbsz
else:
d_loss_cls = F.cross_entropy(out_cls, real_label)
# Compute classification accuracy of the discriminator
if (i+1) % self.log_step == 0:
accuracies = self.compute_accuracy(out_cls.view(cbsz,-1), real_label, self.dataset)
accLog['acc'] = accLog['acc'] * 0.8 + (1-0.8)* accuracies.data.cpu().numpy() if 'acc' in accLog else accuracies.data.cpu().numpy()
log = ["{}: {:.2f}".format(self.selected_attrs[cati],acc) for cati,acc in enumerate(accLog['acc'])]
print(log)
# Compute loss with fake images
fake_x = self.G(real_x, fake_c)
fake_x = Variable(fake_x.data)
out_src, out_cls_fake = self.D(fake_x)
if self.use_seperate_classifier and self.adv_classifier:
out_cls_fake = self.D_cls(fake_x)
d_loss_fake = self.compute_real_fake_loss(out_src, self.adv_loss_type, datasrc = 'fake')
if self.adv_classifier:
d_loss_cls = d_loss_cls + (F.binary_cross_entropy_with_logits(
out_cls_fake.view(cbsz,-1), real_label, size_average=False) / cbsz)
# Backward + Optimize
if not self.use_seperate_classifier:
d_loss = (1.-self.onlymaskgen) * (d_loss_real + d_loss_fake) + self.lambda_cls * d_loss_cls
else:
d_loss = d_loss_real + d_loss_fake
self.reset_grad()
d_loss.backward()
self.d_optimizer.step()
if self.use_seperate_classifier:
d_loss_cls.backward()
self.dcls_optimizer.step()
# Compute gradient penalty
#if not self.onlymaskgen:
if not (self.d_use_spectralnorm):
d_loss_gp = self.compute_gradient_penalty(real_x, fake_x)
# Backward + Optimize
d_loss = self.lambda_gp * d_loss_gp
self.reset_grad()
d_loss.backward()
self.d_optimizer.step()
loss['D/loss_gp'] = loss['D/loss_gp'] * avgexp + (1-avgexp ) * d_loss_gp.data[0] if 'D/loss_gp' in loss else d_loss_gp.data[0]
# Logging
loss['D/loss_real'] = loss['D/loss_real'] * avgexp + (1-avgexp ) * d_loss_real.data[0] if 'D/loss_real' in loss else d_loss_real.data[0]
loss['D/loss_fake'] = loss['D/loss_fake'] * avgexp + (1-avgexp ) * d_loss_fake.data[0] if 'D/loss_fake' in loss else d_loss_fake.data[0]
loss['Wass_dist'] = loss['Wass_dist'] * avgexp + (1-avgexp ) * (-d_loss_real.data[0]-d_loss_fake.data[0]) if 'Wass_dist' in loss else (-d_loss_real.data[0]-d_loss_fake.data[0])
loss['D/loss_cls'] = loss['D/loss_cls'] * avgexp + (1-avgexp ) * d_loss_cls.data[0] if 'D/loss_cls' in loss else d_loss_cls.data[0]
#del grad, interpolated, out, out_cls, grad_l2norm, d_loss_gp, alpha
#gc.collect()
# ================== Train G ================== #
if (i+1) % self.d_train_repeat == 0:
for p in self.G.parameters():
p.requires_grad = p.requires_grad_orig#
for p in self.D.parameters():
p.requires_grad = False#
if self.use_seperate_classifier:
for p in self.D_cls.parameters():
p.requires_grad = False#
# Original-to-target and Target-to-original domain
#if ('G/loss_rec' in loss) and (loss['G/loss_rec'] < 0.22):
# import ipdb; ipdb.set_trace()
if self.dataset == 'coco':
fake_c = fake_label.clone()
fake_x, mask = self.G(real_x, fake_c, out_diff=True)
## Should this be detached!?
if self.lambda_rec:
rec_x = self.G(fake_x, real_c)
if self.adv_rec:
out_src_rec, out_cls_rec = self.D(rec_x)
g_loss_vgg =0.
if self.lambda_vggloss:
g_loss_vgg = self.vggLoss(fake_x, real_x)
g_loss_vgg += (mask[1].mean())
loss['G/l_vgg'] = loss['G/l_vgg'] * avgexp + (1-avgexp ) * (self.lambda_vggloss * g_loss_vgg.data[0]) if 'G/l_vgg' in loss else (self.lambda_vggloss*g_loss_vgg.data[0])
g_smooth_loss = 0.
if self.lambda_smoothloss:
g_smooth_loss = F.conv2d(F.pad(mask[1],(1,1,1,1),mode='replicate'),smoothWeight).mean()
loss['G/l_sm'] = loss['G/l_sm'] * avgexp + (1-avgexp ) * (self.lambda_smoothloss * g_smooth_loss.data[0]) if 'G/l_sm' in loss else (self.lambda_smoothloss*g_smooth_loss.data[0])
# Compute losses
out_src, out_cls = self.D(fake_x)
if self.use_seperate_classifier:
out_cls = self.D_cls(fake_x)
if not self.onlymaskgen:
g_loss_fake = self.compute_real_fake_loss(out_src, self.adv_loss_type, loss_for = 'generator')
#if self.use_topk_patch == 0:
# g_loss_fake = - torch.mean(out_src)
#else:
# g_loss_fake = - out_src.view(cbsz,-1).topk(self.use_topk_patch,dim=1,largest=False)[0].mean()
loss['G/loss_fake'] = loss['G/loss_fake'] * avgexp + (1-avgexp ) * g_loss_fake.data[0] if 'G/loss_fake' in loss else g_loss_fake.data[0]
else:
g_loss_fake = 0.
if self.lambda_rec:
g_loss_rec = torch.mean(torch.abs(real_x - rec_x))
if self.adv_rec:
g_loss_rec_fake = - torch.mean(out_src_rec) if self.use_topk_patch == 0 else - out_src_rec.view(cbsz,-1).topk(self.use_topk_patch,dim=1,largest=False)[0].mean()
g_loss_rec_cls = F.binary_cross_entropy_with_logits(
out_cls_rec.view(cbsz,-1), real_label, size_average=False) / cbsz
g_loss_rec += g_loss_rec_fake + g_loss_rec_cls
else:
g_loss_rec = 0.
if self.dataset in cocoAndCelebset:
g_loss_cls = F.binary_cross_entropy_with_logits(
out_cls.view(cbsz,-1), fake_label, size_average=False) / cbsz
else:
g_loss_cls = F.cross_entropy(out_cls, fake_label)
# Backward + Optimize
g_loss = g_loss_fake + self.lambda_rec * g_loss_rec + self.lambda_cls * g_loss_cls + self.lambda_vggloss* g_loss_vgg + self.lambda_smoothloss * g_smooth_loss
# Add L1 loss to keep the difference image from going beyond boundaries
if self.differential_generator:
g_loss_abs = (Variable((torch.abs(fake_x.data)>1.0).float(),requires_grad=False)*(torch.abs(fake_x))).mean()
g_loss += g_loss_abs
self.reset_grad()
g_loss.backward()
self.g_optimizer.step()
# Logging
if self.lambda_rec:
loss['G/loss_rec'] = loss['G/loss_rec'] * avgexp + (1-avgexp ) * g_loss_rec.data[0] if 'G/loss_rec' in loss else g_loss_rec.data[0]
loss['G/loss_cls'] = loss['G/loss_cls'] * avgexp + (1-avgexp ) * g_loss_cls.data[0] if 'G/loss_cls' in loss else g_loss_cls.data[0]
for p in self.D.parameters():
p.requires_grad = p.requires_grad_orig#
if self.use_seperate_classifier:
for p in self.D_cls.parameters():
p.requires_grad = True#
# Print out log info
if (i+1) % self.log_step == 0:
elapsed = time.time() - start_time
elapsed = str(datetime.timedelta(seconds=elapsed)).split('.')[0]
log = "Elapsed [{}], Epoch [{}/{}], Iter [{}/{}]".format(
elapsed, e+1, self.num_epochs, i+1, iters_per_epoch)
for tag, value in loss.items():
log += ", {}: {:.2f}".format(tag, value)
print(log)
if self.use_tensorboard:
for tag, value in loss.items():
self.logger.scalar_summary(tag, value, e * iters_per_epoch + i + 1)
# Translate fixed images for debugging
if (i+1) % self.sample_step == 0:
fake_image_list = [fixed_x]
for fixed_c in fixed_c_list:
fake_image_list.append(self.G(fixed_x, fixed_c))
fake_images = torch.cat(fake_image_list, dim=3)
save_image(self.denorm(fake_images.data),
os.path.join(self.sample_path, '{}_{}_fake.png'.format(e+1, i+1)),nrow=1, padding=0)
print('Translated images and saved into {}..!'.format(self.sample_path))
# Save model checkpoints
if (i+1) % self.model_save_step == 0:
checkpointData = {
'iter': e*iters_per_epoch+i,
'arch': self.arch,
'generator_state_dict':self.G.state_dict(),
'discriminator_state_dict': self.D.state_dict(),
'enc_dec_optimizer' : self.g_optimizer.state_dict(),
'discriminator_optimizer' : self.d_optimizer.state_dict(),
}
if self.use_seperate_classifier:
checkpointData['discriminator_cls_state_dict'] = self.D_cls.state_dict()
checkpointData['discriminator_cls_optimizer'] = self.dcls_optimizer.state_dict()
save_checkpoint(checkpointData, fappend = self.arch['fappend']+'_{}_{}.pth.tar'.format(e+1,i+1),
outdir = self.model_save_path)
#torch.save(self.G.state_dict(),
# os.path.join(self.model_save_path, '{}_{}_G.pth'.format(e+1, i+1)))
#torch.save(self.D.state_dict(),
# os.path.join(self.model_save_path, '{}_{}_D.pth'.format(e+1, i+1)))