forked from HRLTY/TP-GAN
-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathTP_GAN-Mar6FS_0.12.py
1061 lines (921 loc) · 61.2 KB
/
TP_GAN-Mar6FS_0.12.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
from __future__ import division
from __future__ import print_function
import os,sys
import time
from glob import glob
import tensorflow as tf
import numpy as np
from six.moves import xrange
from ops import *
from utils import *
from net_input_everything_featparts import *
from time import localtime, strftime
import random
import pickle
import subprocess
relu = tf.nn.relu
import scipy.misc
import shutil
from utils import pp, visualize, to_json
import tensorflow as tf
import logging#pt
#These parameters should provide a good initialization, but if for specific refinement, you can adjust them during training.
ALPHA_ADVER = 2e1
BELTA_FEATURE = 4e1 #3800 inital loss
UPDATE_G = 2 #optimize D once and UPDATE_G times G
UPDATE_D = 1
PART_W = 3
IDEN_W = 1e1
TV_WEIGHT = 1e-3
COND_WEIGHT = 0.3
L1_1_W = 1#0.5
L1_2_W = 1
L1_3_W = 1.5
RANGE= 60
RANGE_LOW = 0
SYM_W = 3e-1
CLIP_D = 0.1
L1 = True
MODE = 'fs60' #'f' feature loss enabled. 'v' -verification enanbled. 'o' original, 'm' masked is mandatory and no need to specify
UPDATE_DV = 1 #optimize DV net
DF = True #local discriminator 4x4
LOAD_60_LABEL = False #otherwise load frontal label
WITHOUT_CODEMAP = True
USE_MASK = False
ENABLE_SELECTION = True
RANDOM_VERIFY = False
RANK_MUL = 6
if WITHOUT_CODEMAP:
CHANNEL = 3
else:
CHANNEL = 6
flags = tf.app.flags
flags.DEFINE_integer("epoch", 250, "Epoch to train [25]")
flags.DEFINE_float("learning_rate", 1e-4, "Learning rate of for adam [0.0002]")
flags.DEFINE_float("beta1", 0.9, "Momentum term of adam [0.5]")
flags.DEFINE_integer("train_size", 1156, "The size of train images [np.inf]")
flags.DEFINE_integer("batch_size", 4, "The size of batch images [64]")
flags.DEFINE_integer("image_size", 128, "The size of image to use (will be center cropped) [108]")
flags.DEFINE_integer("output_size", 128, "The size of the output images to produce [64]")
flags.DEFINE_integer("c_dim", 3, "Dimension of image color. [3]")
flags.DEFINE_string("dataset", "MultiPIE", "The name of dataset [celebA, mnist, lsun]")
flags.DEFINE_string("checkpoint_dir", "checkpoint60", "Directory name to save the checkpoints [checkpoint]")
flags.DEFINE_string("sample_dir", "samples", "Directory name to save the image samples [samples]")
flags.DEFINE_boolean("is_train", True, "True for training, False for testing [False]")
flags.DEFINE_boolean("is_crop", True, "True for training, False for testing [False]")
flags.DEFINE_boolean("visualize", True, "True for visualizing, False for nothing [False]")
FLAGS = flags.FLAGS
class DCGAN(object):
def __init__(self, sess, image_size=128, is_crop=True,
batch_size=10, sample_size = 100, output_size=128,
y_dim=None, z_dim=100, gf_dim=64, df_dim=64,
gfc_dim=1024, dfc_dim=1024, c_dim=3, dataset_name='MultiPIE',
checkpoint_dir=None, sample_dir=None):
"""
Args:
sess: TensorFlow session
batch_size: The size of batch. Should be specified before training.
output_size: (optional) The resolution in pixels of the images. [64]
y_dim: (optional) Dimension of dim for y. [None]
z_dim: (optional) Dimension of dim for Z. [100]
gf_dim: (optional) Dimension of gen filters in first conv layer. [64]
df_dim: (optional) Dimension of discrim filters in first conv layer. [64]
gfc_dim: (optional) Dimension of gen units for for fully connected layer. [1024]
dfc_dim: (optional) Dimension of discrim units for fully connected layer. [1024]
c_dim: (optional) Dimension of image color. For grayscale input, set to 1. [3]
"""
self.test_batch_size = batch_size
self.save_interval = 500
self.sample_interval = 150
self.sess = sess
self.is_grayscale = (c_dim == 1)
self.batch_size = FLAGS.batch_size
self.sample_run_num = 15
self.testing = False
self.testingphase = 'FS'
self.testimg = True
if self.testing:
#self.batch_size = 10
self.testingphase = '60'#'gt50'
self.sample_run_num = 99999999
self.test_batch_size = self.batch_size
self.image_size = image_size
self.sample_size = sample_size
self.output_size = output_size
self.gf_dim = gf_dim
self.df_dim = df_dim
self.gfc_dim = gfc_dim
self.dfc_dim = dfc_dim
self.z_dim = 100
self.c_dim = c_dim
# batch normalization : deals with poor initialization helps gradient flow
random.seed()
self.DeepFacePath = './DeepFace168.pickle'
self.dataset_name = dataset_name
self.checkpoint_dir = checkpoint_dir
self.loadDeepFace()#pt
self.build_model()#pt
def build_model(self):
#hold all four
#Note: not true, if WITHOUT_CODEMAP is true, then here is pure images without codemap and 3 channels
#mirror concatenate
mc = lambda left : tf.concat_v2([left, left[:,:,::-1,:]], 3)
self.images_with_code = tf.placeholder(tf.float32, [self.batch_size] + [self.output_size, self.output_size, CHANNEL], name='images_with_code')
self.sample_images = tf.placeholder(tf.float32, [self.test_batch_size] + [self.output_size, self.output_size, CHANNEL], name='sample_images')
if WITHOUT_CODEMAP:
self.images = self.images_with_code
self.sample_images_nocode = self.sample_images
else:
self.images = tf.split(3, 2, self.images_with_code)[0]
self.sample_images_nocode = tf.split(3, 2, self.sample_images)[0]
self.g_images = self.images #tf.reduce_mean(self.images, axis=3, keep_dims=True)
self.g_samples = self.sample_images_nocode #tf.reduce_mean(self.sample_images_nocode, axis=3, keep_dims=True)
self.g32_images_with_code = tf.image.resize_bilinear(self.images_with_code, [32, 32])
self.g64_images_with_code = tf.image.resize_bilinear(self.images_with_code, [64, 64])
self.g32_sampleimages_with_code = tf.image.resize_bilinear(self.sample_images, [32, 32])
self.g64_sampleimages_with_code = tf.image.resize_bilinear(self.sample_images, [64, 64])
self.labels = tf.placeholder(tf.float32, [self.batch_size] + [self.output_size, self.output_size, 3], name='label_images')
self.poselabels = tf.placeholder(tf.int32, [self.batch_size])
self.idenlabels = tf.placeholder(tf.int32, [self.batch_size])
self.landmarklabels = tf.placeholder(tf.float32, [self.batch_size, 5*2])
self.g_labels = self.labels #tf.reduce_mean(self.labels, 3, keep_dims=True)
self.g8_labels = tf.image.resize_bilinear(self.g_labels, [8, 8])
self.g16_labels = tf.image.resize_bilinear(self.g_labels, [16, 16])
self.g32_labels = tf.image.resize_bilinear(self.g_labels, [32, 32])
self.g64_labels = tf.image.resize_bilinear(self.g_labels, [64, 64])
self.eyel = tf.placeholder(tf.float32, [self.batch_size, EYE_H, EYE_W, 3])
self.eyer = tf.placeholder(tf.float32, [self.batch_size, EYE_H, EYE_W, 3])
self.nose = tf.placeholder(tf.float32, [self.batch_size, NOSE_H, NOSE_W, 3])
self.mouth = tf.placeholder(tf.float32, [self.batch_size, MOUTH_H, MOUTH_W, 3])
self.eyel_label = tf.placeholder(tf.float32, [self.batch_size, EYE_H, EYE_W, 3])
self.eyer_label = tf.placeholder(tf.float32, [self.batch_size, EYE_H, EYE_W, 3])
self.nose_label = tf.placeholder(tf.float32, [self.batch_size, NOSE_H, NOSE_W, 3])
self.mouth_label = tf.placeholder(tf.float32, [self.batch_size, MOUTH_H, MOUTH_W, 3])
self.eyel_sam = tf.placeholder(tf.float32, [self.batch_size, EYE_H, EYE_W, 3])
self.eyer_sam = tf.placeholder(tf.float32, [self.batch_size, EYE_H, EYE_W, 3])
self.nose_sam = tf.placeholder(tf.float32, [self.batch_size, NOSE_H, NOSE_W, 3])
self.mouth_sam = tf.placeholder(tf.float32, [self.batch_size, MOUTH_H, MOUTH_W, 3])
#feats contains: self.feat128, self.feat64, self.feat32, self.feat16, self.feat8, self.feat
self.G_eyel,self.c_eyel = self.partRotator(self.eyel, "PartRotator_eyel")
self.G_eyer,self.c_eyer = self.partRotator(tf.concat_v2([self.eyer, self.eyel], axis=3), "PartRotator_eyer")
self.G_nose,self.c_nose = self.partRotator(self.nose, "PartRotator_nose")
self.G_mouth,self.c_mouth = self.partRotator(self.mouth, "PartRotator_mouth")
self.G_eyel_sam, self.c_eyel_sam = self.partRotator(self.eyel_sam, "PartRotator_eyel", reuse=True)
self.G_eyer_sam, self.c_eyer_sam = self.partRotator(tf.concat_v2([self.eyer_sam, self.eyel_sam],axis=3), "PartRotator_eyer", reuse=True)
self.G_nose_sam, self.c_nose_sam = self.partRotator(self.nose_sam, "PartRotator_nose", reuse=True)
self.G_mouth_sam, self.c_mouth_sam = self.partRotator(self.mouth_sam, "PartRotator_mouth", reuse=True)
self.z = tf.random_normal([self.batch_size, self.z_dim], mean=0.0, stddev=0.02, seed=2017)
#tf.placeholder(tf.float32, [self.batch_size, self.z_dim], name='z')
self.feats = self.generator(mc(self.images_with_code), self.batch_size, name="encoder")
self.feats += (mc(self.images_with_code), mc(self.g64_images_with_code), mc(self.g32_images_with_code),
self.G_eyel, self.G_eyer, self.G_nose, self.G_mouth,
self.c_eyel, self.c_eyer, self.c_nose, self.c_mouth,)
self.check_sel128, self.check_sel64, self.check_sel32, self.check_sel16, self.check_sel8, self.G, self.G2, self.G3 = \
self.decoder(*self.feats, batch_size=self.batch_size)
self.poselogits, self.identitylogits, self.Glandmark = self.FeaturePredict(self.feats[5])
sample_feats = self.generator(mc(self.sample_images),self.test_batch_size, name="encoder", reuse=True)
self.sample512 = sample_feats[-1]
sample_feats += (mc(self.sample_images), mc(self.g64_sampleimages_with_code), mc(self.g32_sampleimages_with_code),
self.G_eyel_sam, self.G_eyer_sam, self.G_nose_sam, self.G_mouth_sam,
self.c_eyel_sam, self.c_eyer_sam, self.c_nose_sam, self.c_mouth_sam,)
self.sample_generator = self.decoder(*sample_feats, batch_size=self.test_batch_size, reuse=True)
if not DF:
self.D, self.D_logits = self.discriminator(self.g_labels)
self.D_, self.D_logits_ = self.discriminator(self.G, reuse=True)
else:
print("Using local discriminator!")
self.D, self.D_logits = self.discriminatorLocal(self.g_labels)
self.D_, self.D_logits_ = self.discriminatorLocal(self.G, reuse=True)
self.logfile = 'loss.txt'
if 'f' in MODE:
#self.verify_images_masked = tf.mul(self.verify_images, self.masks_binary)
#can not apply mask !!!
# self.Dv, self.Dv_logits = self.discriminatorVerify(self.labels, self.verify_images)
_,_,_,_, self.G_pool5, self.Gvector = self.FeatureExtractDeepFace(tf.reduce_mean(self.G, axis=3, keep_dims=True))
_,_,_,_, self.label_pool5, self.labelvector = self.FeatureExtractDeepFace(tf.reduce_mean(self.g_labels, axis=3, keep_dims=True), reuse=True)
_,_,_,_, _, self.samplevector = self.FeatureExtractDeepFace(tf.reduce_mean(self.sample_images_nocode, axis=3, keep_dims=True), reuse=True)
#self.Dv, self.Dv_logits = self.discriminatorClassify(self.Gvector)
#self.dv_loss = tf.reduce_sum(tf.nn.sparse_softmax_cross_entropy_with_logits(self.Dv_logits, self.verify_labels))
self.dv_loss = tf.reduce_mean(tf.abs(self.Gvector-self.labelvector))
self.dv_loss += tf.reduce_mean(tf.reduce_sum(tf.reduce_sum(tf.abs(self.G_pool5-self.label_pool5),1),1))
self.logfile = 'loss_verify.txt'
#self.dv_sum = histogram_summary("dv_", self.Dv)
# self.d__sum = histogram_summary("d_", self.D_)
# self.d_sum = histogram_summary("d", self.D)
# self.G_sum = image_summary("G", self.G)
#basic loss
# self.d_loss_real = tf.reduce_mean(self.D_logits)
# self.d_loss_fake = -tf.reduce_mean(self.D_logits_)
# self.g_loss_adver = -tf.reduce_mean(self.D_logits_)
self.d_loss_real = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(self.D_logits, tf.ones_like(self.D) * 0.9))
self.d_loss_fake = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(self.D_logits_, tf.zeros_like(self.D_)))
self.g_loss_adver = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(self.D_logits_, tf.ones_like(self.D_) * 0.9))
#self.mark_regression_loss = tf.reduce_mean(tf.square(tf.abs(self.landmarklabels-self.Glandmark)))
#self.poseloss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(self.poselogits, self.poselabels))
self.idenloss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(self.identitylogits, self.idenlabels))
self.eyel_loss = tf.reduce_mean(tf.reduce_sum(tf.reduce_sum(tf.abs(self.c_eyel - self.eyel_label), 1), 1))
self.eyer_loss = tf.reduce_mean(tf.reduce_sum(tf.reduce_sum(tf.abs(self.c_eyer - self.eyer_label), 1), 1))
self.nose_loss = tf.reduce_mean(tf.reduce_sum(tf.reduce_sum(tf.abs(self.c_nose - self.nose_label), 1), 1))
self.mouth_loss = tf.reduce_mean(tf.reduce_sum(tf.reduce_sum(tf.abs(self.c_mouth - self.mouth_label), 1), 1))
#rotation L1 / L2 loss in g_loss
# one8 = tf.ones([1,8,4,1],tf.float32)
# mask8 = tf.concat_v2([one8, one8], 2)
# mask16 = tf.image.resize_nearest_neighbor(mask8, size=[16, 16])
# mask32 = tf.image.resize_nearest_neighbor(mask8, size=[32, 32])
# mask64 = tf.image.resize_nearest_neighbor(mask8, size=[64, 64])
# mask128 = tf.image.resize_nearest_neighbor(mask8, size=[128, 128])
#use L2 for 128, L1 for others. mask emphasize left side.
errL1 = tf.abs(self.G - self.g_labels) #* mask128
errL1_2 = tf.abs(self.G2 - self.g64_labels) #* mask64
errL1_3 = tf.abs(self.G3 - self.g32_labels) #* mask32
#errcheck8 = tf.abs(self.check_sel8 - self.g8_labels) #* mask8
#errcheck16 = tf.abs(self.check_sel16 - self.g16_labels) #* mask16
errcheck32 = tf.abs(self.check_sel32 - self.g32_labels) #* mask32
errcheck64 = tf.abs(self.check_sel64 - self.g64_labels) #* mask64
errcheck128 = tf.abs(self.check_sel128 - self.g_labels) #* mask128
self.weightedErrL1 = tf.reduce_mean(tf.reduce_sum(tf.reduce_sum(errL1, 1), 1))
self.symErrL1 = tf.reduce_mean(tf.reduce_sum(tf.reduce_sum(symL1(#self.processor(self.G)
tf.nn.avg_pool(self.G, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
), 1), 1))
self.weightedErrL2 = tf.reduce_mean(tf.reduce_sum(tf.reduce_sum(errL1_2, 1), 1))
self.symErrL2 = tf.reduce_mean(tf.reduce_sum(tf.reduce_sum(symL1(self.processor(self.G2)), 1), 1))
self.weightedErrL3 = tf.reduce_mean(tf.reduce_sum(tf.reduce_sum(errL1_3, 1), 1))
self.symErrL3 = tf.reduce_mean(tf.reduce_sum(tf.reduce_sum(symL1(self.processor(self.G3, reuse=True)), 1), 1))
cond_L12 = tf.abs(tf.image.resize_bilinear(self.G, [64,64]) - tf.stop_gradient(self.G2))
#self.condErrL12 = tf.reduce_mean(tf.reduce_sum(tf.reduce_sum(cond_L12, 1), 1))
#cond_L23 = tf.abs(tf.image.resize_bilinear(self.G2, [32,32]) - tf.stop_gradient(self.G3))
#self.condErrL23 = tf.reduce_mean(tf.reduce_sum(tf.reduce_sum(cond_L23, 1), 1))
self.tv_loss = tf.reduce_mean(total_variation(self.G))
#self.weightedErr_check8 = tf.reduce_mean(tf.reduce_sum(tf.reduce_sum(errcheck8, 1), 1))
#self.weightedErr_check16 = tf.reduce_mean(tf.reduce_sum(tf.reduce_sum(errcheck16, 1), 1))
# self.weightedErr_check32 = tf.reduce_mean(tf.reduce_sum(tf.reduce_sum(errcheck32, 1), 1))
# self.weightedErr_check64 = tf.reduce_mean(tf.reduce_sum(tf.reduce_sum(errcheck64, 1), 1))
# self.weightedErr_check128 = tf.reduce_mean(tf.reduce_sum(tf.reduce_sum(errcheck128, 1), 1))
# mean = tf.reduce_mean(tf.reduce_mean(self.G, 1,keep_dims=True), 2, keep_dims=True)
# self.stddev = tf.reduce_mean(tf.squared_difference(self.G, mean))
self.d_loss = self.d_loss_real + self.d_loss_fake
self.g_loss = L1_1_W * (self.weightedErrL1 + SYM_W * self.symErrL1) + L1_2_W * (self.weightedErrL2 + SYM_W * self.symErrL2) \
+ L1_3_W * (self.weightedErrL3 + SYM_W * self.symErrL3)
self.g_loss += BELTA_FEATURE * self.dv_loss + ALPHA_ADVER * self.g_loss_adver + IDEN_W * self.idenloss + self.tv_loss * TV_WEIGHT
self.rot_loss = PART_W * (self.eyel_loss + self.eyer_loss + self.nose_loss + self.mouth_loss)
#self.sel_loss = self.weightedErr_check32 + self.weightedErr_check64 + self.weightedErr_check128
#self.g_loss += self.sel_loss
self.var_file = open('var_log.txt', mode='a')
t_vars = [var for var in tf.trainable_variables() if 'FeatureExtractDeepFace' not in var.name \
and 'processor' not in var.name]
print('done!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!')
def isTargetVar(name, tokens):
for token in tokens:
if token in name:
return True
return False
dec128toks = ['dec128', 'recon128', 'check_img128']
self.d_vars = [var for var in t_vars if 'discriminatorLocal' in var.name]
self.all_g_vars = [var for var in t_vars if 'discriminatorLocal' not in var.name]
self.rot_vars = [var for var in t_vars if 'Rotator' in var.name]
self.sel_vars = [var for var in t_vars if 'select' in var.name]
self.dec_vars = [var for var in t_vars if 'decoder' in var.name and 'select' not in var.name]
self.enc_vars = [var for var in t_vars if 'encoder' in var.name]
self.pre_vars = [var for var in t_vars if 'FeaturePredict' in var.name]
#
self.se_vars = list(self.enc_vars); self.se_vars.extend(self.sel_vars)
self.ed_vars = list(self.dec_vars); self.ed_vars.extend(self.enc_vars);
self.ed_vars.extend(self.pre_vars); self.ed_vars.extend(self.rot_vars);
self.ed_vars.extend(self.sel_vars)
#self.rd_vars = list(self.dec_vars); self.rd_vars.extend([var for var in self.d_vars if isTargetVar(var.name, dec128toks)])
#print("-----enc and dec ---->", map(lambda x:x.name, self.ed_vars), sep='\n', file=var_file)
#print("-----enc and sel ---->", map(lambda x:x.name, self.se_vars), sep='\n', file=var_file)
#print("-----discrim ---->", map(lambda x:x.name, self.d_vars),sep='\n', file=var_file)
self.saver = tf.train.Saver(t_vars, max_to_keep=2)
def train(self, config):
"""Train DCGAN"""
#data = glob(os.path.join("./data", config.dataset, "*.jpg"))
data = MultiPIE(LOAD_60_LABEL=LOAD_60_LABEL, GENERATE_MASK=USE_MASK, RANDOM_VERIFY=RANDOM_VERIFY, MIRROR_TO_ONE_SIDE = True, source = self.testingphase)
#np.random.shuffle(data)
config.sample_dir += '{:05d}'.format(random.randint(1,100000))
#print(self.__dict__,66666666666)
d_optim = tf.train.AdamOptimizer(config.learning_rate, beta1=config.beta1) \
.minimize(self.d_loss, var_list=self.d_vars)
#clip_D = [p.assign(tf.clip_by_value(p, -CLIP_D, CLIP_D)) for p in self.d_vars]
g_dec_optim = tf.train.AdamOptimizer(config.learning_rate, beta1=config.beta1) \
.minimize(self.g_loss, var_list=self.ed_vars)
#g_enc_optim = tf.train.AdamOptimizer(config.learning_rate * 0.001, beta1=config.beta1) \
# .minimize(self.g_loss, var_list=self.enc_vars)
# s_optim = tf.train.AdamOptimizer(config.learning_rate, beta1=config.beta1) \
# .minimize(self.sel_loss, var_list=self.se_vars)
#g_sel_dec_optim = tf.train.RMSPropOptimizer(config.learning_rate) \
# .minimize(self.g_loss + self.sel_loss + self.rot_loss, var_list=self.all_g_vars)
rot_optim = tf.train.AdamOptimizer(config.learning_rate, beta1=config.beta1) \
.minimize(self.rot_loss, var_list=self.rot_vars)
init_op = tf.global_variables_initializer()
self.sess.run(init_op)
counter = random.randint(1,30)
start_time = time.time()
if self.load(self.checkpoint_dir):
print(" [*] Load SUCCESS")
else:
print(" [!] Load failed...")
print(555555555555555555555,self.test_batch_size,self.sample_run_num)
#self.test_batch_size * self.sample_run_num分别是10,15
sample_images, filenames ,sample_eyel, sample_eyer, sample_nose, sample_mouth, \
sample_labels, sample_leyel, sample_leyer, sample_lnose, sample_lmouth, sample_iden = data.test_batch(self.test_batch_size * self.sample_run_num)
if not self.testing:
sample_imagesT, filenamesT ,sample_eyelT, sample_eyerT, sample_noseT, sample_mouthT, \
sample_labelsT, sample_leyelT, sample_leyerT, sample_lnoseT, sample_lmouthT, sample_idenT = data.test_batch(self.test_batch_size * self.sample_run_num * RANK_MUL, Pose=RANGE)
if WITHOUT_CODEMAP:
sample_images = sample_images[..., 0:3]
#append loss log to file
self.f = open(self.logfile, mode='a')
self.f.write('----'+strftime("%a, %d %b %Y %H:%M:%S +0000", localtime())+' BEGINS----MODE:'+MODE+'-----\n')
print("start training!")
for epoch in xrange(config.epoch):
#data = glob(os.path.join("./data", config.dataset, "*.jpg"))
batch_idxs = min(data.size, config.train_size) // self.batch_size
for idx in xrange(0, batch_idxs):
#load data from MultiPIE
batch_images_with_code, batch_labels, batch_masks, verify_images, verify_labels, \
batch_pose, batch_iden, batch_landmarks,\
batch_eyel, batch_eyer, batch_nose, batch_mouth,\
batch_eyel_label, batch_eyer_label, batch_nose_label, batch_mouth_label \
= data.next_image_and_label_mask_batch(self.batch_size, imageRange=RANGE, imageRangeLow=RANGE_LOW)
# batch_images = batch_images_with_code[:,:,:,0:3] #discard codes
if WITHOUT_CODEMAP:
batch_images_with_code = batch_images_with_code[..., 0:3]
# needs self.G(needing images with code) and real images
for _ in range(UPDATE_D):
# Update D network
_ = self.sess.run([d_optim,],
feed_dict={ self.images_with_code: batch_images_with_code,
self.labels : batch_labels,
self.eyel : batch_eyel, self.eyer : batch_eyer, self.nose: batch_nose, self.mouth: batch_mouth,
})
for _ in range(UPDATE_G):
# Update G network
# Run g_optim twice to make sure that d_loss does not go to zero (different from paper)
_ = self.sess.run([rot_optim, g_dec_optim,],
# _ = self.sess.run([g_sel_dec_optim],
feed_dict={self.images_with_code: batch_images_with_code,
self.labels : batch_labels,
self.eyel : batch_eyel, self.eyer : batch_eyer, self.nose: batch_nose, self.mouth: batch_mouth,
self.poselabels : batch_pose, self.idenlabels: batch_iden, self.landmarklabels: batch_landmarks,
self.eyel_label : batch_eyel_label, self.eyer_label : batch_eyer_label,
self.nose_label: batch_nose_label, self.mouth_label : batch_mouth_label
})
counter += 1
print('.',end='');sys.stdout.flush()
if(counter % 5 == 0):
self.evaluate(epoch, idx, batch_idxs, start_time, 'train',
batch_images_with_code, batch_eyel, batch_eyer, batch_nose, batch_mouth,
batch_labels, batch_eyel_label, batch_eyer_label, batch_nose_label, batch_mouth_label, batch_iden);
if np.mod(counter, self.sample_interval) == self.sample_interval-1:
for i in range(self.sample_run_num):
print(i, end=' ')
currentBatchSamples = sample_images[i*self.test_batch_size:(i+1)*self.test_batch_size,...]
currentBatchEyel = sample_eyel[i*self.test_batch_size:(i+1)*self.test_batch_size,...]
currentBatchEyer = sample_eyer[i*self.test_batch_size:(i+1)*self.test_batch_size,...]
currentBatchNose = sample_nose[i*self.test_batch_size:(i+1)*self.test_batch_size,...]
currentBatchMouth = sample_mouth[i*self.test_batch_size:(i+1)*self.test_batch_size,...]
samples = self.sess.run(
self.sample_generator,
feed_dict={ self.sample_images: currentBatchSamples,
self.eyel_sam : currentBatchEyel,
self.eyer_sam : currentBatchEyer,
self.nose_sam : currentBatchNose,
self.mouth_sam : currentBatchMouth
})
savedtest = save_images(currentBatchSamples if WITHOUT_CODEMAP else currentBatchSamples[...,0:3], [100, 100],
'./{}/{:02d}_{:04d}/train{}_'.format(config.sample_dir, epoch, idx, i), suffix='')
savedoutput = save_images(samples[5], [100, 100],
'./{}/{:02d}_{:04d}/train{}_'.format(config.sample_dir, epoch, idx, i),suffix='_128')
savedoutput = save_images(samples[6], [100, 100],
'./{}/{:02d}_{:04d}/train{}_'.format(config.sample_dir, epoch, idx, i),suffix='_64')
savedoutput = save_images(samples[7], [100, 100],
'./{}/{:02d}_{:04d}/train{}_'.format(config.sample_dir, epoch, idx, i),suffix='_32')
print("[{} completed{} and saved {}.]".format(config.sample_dir, savedtest*self.sample_run_num, savedoutput*self.sample_run_num))
#testing accuracy
savedir = 'tem_test'
if not os.path.exists(savedir):
os.mkdir(savedir)
else:
#subprocess.call(['rm' '-f' 'tem_test/*'])
shutil.rmtree(savedir, ignore_errors=True)
os.mkdir(savedir)
print("cleaned tem_test!")
listfid = open('probef.txt','w')
for i in range(self.sample_run_num * RANK_MUL):
currentBatchSamples = sample_imagesT[i*self.test_batch_size:(i+1)*self.test_batch_size,...]
currentBatchEyel = sample_eyelT[i*self.test_batch_size:(i+1)*self.test_batch_size,...]
currentBatchEyer = sample_eyerT[i*self.test_batch_size:(i+1)*self.test_batch_size,...]
currentBatchNose = sample_noseT[i*self.test_batch_size:(i+1)*self.test_batch_size,...]
currentBatchMouth = sample_mouthT[i*self.test_batch_size:(i+1)*self.test_batch_size,...]
currentBatchLSamples = sample_labelsT[i*self.test_batch_size:(i+1)*self.test_batch_size,...]
currentBatchLEyel = sample_leyelT[i*self.test_batch_size:(i+1)*self.test_batch_size,...]
currentBatchLEyer = sample_leyerT[i*self.test_batch_size:(i+1)*self.test_batch_size,...]
currentBatchLNose = sample_lnoseT[i*self.test_batch_size:(i+1)*self.test_batch_size,...]
currentBatchLMouth = sample_lmouthT[i*self.test_batch_size:(i+1)*self.test_batch_size,...]
currentBatchIden = sample_idenT[i*self.test_batch_size:(i+1)*self.test_batch_size,...]
if i % (4 * RANK_MUL) == 0:
self.evaluate(epoch, idx, batch_idxs, start_time, 'test',
currentBatchSamples, currentBatchEyel, currentBatchEyer, currentBatchNose, currentBatchMouth,
currentBatchLSamples, currentBatchLEyel, currentBatchLEyer, currentBatchLNose, currentBatchLMouth, currentBatchIden);
samples = self.sess.run(
self.sample_generator,
feed_dict={ self.sample_images: currentBatchSamples,
self.eyel_sam : currentBatchEyel,
self.eyer_sam : currentBatchEyer,
self.nose_sam : currentBatchNose,
self.mouth_sam : currentBatchMouth
})
samplevectors = self.samplevector.eval({self.sample_images_nocode : samples[5]})
for k in range(samplevectors.shape[0]):
filename = filenamesT[i*self.test_batch_size + k]
savefilename = filename.replace('.png','.feat')
label = filename[0:3]
listfid.write(savedir + '/' + filename +' '+ label + '\n')
result = samplevectors[k,:]
f = open(savedir +'/'+ savefilename,'wb')
f.write(result)
f.close()
listfid.close()
print("[{} saved {} feats. calling comparision..]".format(savedir, self.test_batch_size * self.sample_run_num))
output, err = subprocess.Popen('./evaluation_rank.sh', stdout=subprocess.PIPE, shell=True).communicate()
tobePrint = '-------!' + ''.join([rank for rank in output.splitlines() if rank.startswith('Rank-1')]) + '!-------'
print(err, tobePrint)
self.var_file.write(tobePrint)
self.var_file.flush()
if np.mod(counter, self.save_interval) == self.save_interval-1:
self.save(config.checkpoint_dir, counter)
else:
print('test samples reading complete')
batchnum = sample_images.shape[0] // self.test_batch_size #current test batch size
savedtest = 0
savedoutput = 0
sample_dir = 'testall'
for i in range(batchnum):
print('generating test result batch{}'.format(i))
ind = (i*self.test_batch_size, (i+1)*self.test_batch_size)
if self.testimg:#Save images
samples = self.sess.run(
self.sample_generator,
feed_dict={ self.sample_images: sample_images[ind[0]:ind[1],:,:,:],
self.eyel_sam : sample_eyel[ind[0]:ind[1],...],
self.eyer_sam : sample_eyer[ind[0]:ind[1],...],
self.nose_sam : sample_nose[ind[0]:ind[1],...],
self.mouth_sam : sample_mouth[ind[0]:ind[1],...]}
)
colorgt = sample_images[ind[0]:ind[1],:,:,0:3]
savedtest += save_images(colorgt, [128, 128],
'./{}/'.format(sample_dir),isOutput=False, filelist=filenames[ind[0]:ind[1]])
savedoutput += save_images(samples[5], [128, 128],
'./{}/'.format(sample_dir),isOutput=True, filelist=filenames[ind[0]:ind[1]])
print("[{} completed{} and saved {}.]".format(sample_dir, savedtest, savedoutput))
else:#save features
savedir = 'testall_f'#'gt50_f'
samples = self.sess.run(
self.sample512,
feed_dict={ self.sample_images: sample_images[ind[0]:ind[1],:,:,:],})
listfid = open('probef.txt','a')
for j in range(self.test_batch_size):
filename = filenames[ind[0]+j]
savefilename = filename.replace('.png','.feat')
label = filename[0:3]
listfid.write(savedir + '/' + filename +' '+ label + '\n')
result = samples[j,0:448]
if not os.path.exists(savedir):
os.mkdir(savedir)
f = open(savedir +'/'+ savefilename,'wb')
f.write(result)
f.close()
print("saved %d files!" % (self.test_batch_size * (i+1)))
listfid.close()
def processor(self, images, reuse=False):
#accept 3 channel images, output orginal 3 channels and 3 x 4 gradient map-> 15 channels
with tf.variable_scope("processor") as scope:
if reuse:
scope.reuse_variables()
input_dim = images.get_shape()[-1]
gradientKernel = gradientweight()
output_dim = gradientKernel.shape[-1]
print("processor:", output_dim)
k_hw = gradientKernel.shape[0]
init = tf.constant_initializer(value=gradientKernel, dtype=tf.float32)
w = tf.get_variable('w', [k_hw, k_hw, input_dim, output_dim],
initializer=init)
conv = tf.nn.conv2d(images, w, strides=[1, 1, 1, 1], padding='SAME')
#conv = conv * 2
return tf.concat_v2([images, conv], 3)
def FeaturePredict(self, featvec, reuse=False):
with tf.variable_scope("FeaturePredict") as scope:
if reuse:
scope.reuse_variables()
identitylogits = linear(Dropout(featvec, keep_prob=0.3, is_training= not self.testing), output_size=340, scope='idenLinear', bias_start=0.1, with_w=True)[0]
return None, identitylogits, None
def discriminatorLocal(self, images, reuse=False):
with tf.variable_scope("discriminatorLocal") as scope:
if reuse:
scope.reuse_variables()
h0 = lrelu(conv2d(images, self.df_dim, name='d_h0_conv'))
#64
h1 = lrelu(batch_norm(conv2d(h0, self.df_dim*2, name='d_h1_conv'), name='d_bn1'))
#32
h2 = lrelu(batch_norm(conv2d(h1, self.df_dim*4, name='d_h2_conv'), name='d_bn2'))
#16
h3 = lrelu(batch_norm(conv2d(h2, self.df_dim*8, name='d_h3_conv'), name='d_bn3'))
# #8x8
h3r1 = resblock(h3, name = "d_h3_conv_res1")
h4 = lrelu(batch_norm(conv2d(h3r1, self.df_dim*8, name='d_h4_conv'), name='d_bn4'))
h4r1 = resblock(h4, name = "d_h4_conv_res1")
h5 = conv2d(h4r1, 1, k_h=1, k_w=1, d_h=1, d_w=1, name='d_h5_conv')
h6 = tf.reshape(h5, [self.batch_size, -1])
#fusing 512 feature map to one layer prediction.
return h6, h6 #tf.nn.sigmoid(h6), h6
def decoder(self, feat128, feat64, feat32, feat16, feat8, featvec,
g128_images_with_code, g64_images_with_code, g32_images_with_code,
eyel, eyer, nose, mouth, c_eyel, c_eyer, c_nose, c_mouth, batch_size = 10, name="decoder", reuse = False):
sel_feat_capacity = self.gf_dim
with tf.variable_scope(name) as scope:
if reuse:
scope.reuse_variables()
initial_all = tf.concat_v2([featvec, self.z], 1)
initial_8 = relu(tf.reshape(linear(initial_all, output_size=8*8*self.gf_dim,scope='initial8', bias_start=0.1, with_w=True)[0],
[batch_size, 8, 8, self.gf_dim]))
initial_32 = relu(deconv2d(initial_8, [batch_size, 32, 32, self.gf_dim // 2], d_h=4, d_w=4, name="initial32"))
initial_64 = relu(deconv2d(initial_32, [batch_size, 64, 64, self.gf_dim // 4], name="initial64"))
initial_128 = relu(deconv2d(initial_64, [batch_size, 128, 128, self.gf_dim // 8], name="initial128"))
before_select8 = resblock(tf.concat_v2([initial_8, feat8], 3), k_h=2, k_w=2, name = "select8_res_1")
#selection T module
reconstruct8 = resblock(resblock(before_select8, k_h=2, k_w=2, name="dec8_res1"), k_h=2, k_w=2, name="dec8_res2")
#selection F module
reconstruct16_deconv = relu(batch_norm(deconv2d(reconstruct8, [batch_size, 16, 16, self.gf_dim*8], name="g_deconv16"), name="g_bnd1"))
before_select16 = resblock(feat16, name = "select16_res_1")
reconstruct16 = resblock(resblock(tf.concat_v2([reconstruct16_deconv, before_select16], 3), name="dec16_res1"), name="dec16_res2")
reconstruct32_deconv = relu(batch_norm(deconv2d(reconstruct16, [batch_size, 32, 32, self.gf_dim*4], name="g_deconv32"), name="g_bnd2"))
before_select32 = resblock(tf.concat_v2([feat32, g32_images_with_code, initial_32], 3), name = "select32_res_1")
reconstruct32 = resblock(resblock(tf.concat_v2([reconstruct32_deconv, before_select32], 3), name="dec32_res1"), name="dec32_res2")
img32 = tf.nn.tanh(conv2d(reconstruct32, 3, d_h=1, d_w=1, name="check_img32"))
reconstruct64_deconv = relu(batch_norm(deconv2d(reconstruct32, [batch_size, 64, 64, self.gf_dim*2], name="g_deconv64"), name="g_bnd3"))
before_select64 = resblock(tf.concat_v2([feat64, g64_images_with_code, initial_64], 3), k_h=5, k_w=5, name = "select64_res_1")
reconstruct64 = resblock(resblock(tf.concat_v2([reconstruct64_deconv, before_select64,
tf.image.resize_bilinear(img32, [64,64])], 3), name="dec64_res1"), name="dec64_res2")
img64 = tf.nn.tanh(conv2d(reconstruct64, 3, d_h=1, d_w=1, name="check_img64"))
reconstruct128_deconv = relu(batch_norm(deconv2d(reconstruct64, [batch_size, 128, 128, self.gf_dim], name="g_deconv128"), name="g_bnd4"))
before_select128 = resblock(tf.concat_v2([feat128, initial_128, g128_images_with_code],3), k_h = 7, k_w = 7, name = "select128_res_1")
reconstruct128 = resblock(tf.concat_v2([reconstruct128_deconv, before_select128,
self.partCombiner(eyel, eyer, nose, mouth),
self.partCombiner(c_eyel, c_eyer, c_nose, c_mouth),
tf.image.resize_bilinear(img64, [128,128])], 3), k_h=5, k_w=5, name="dec128_res1")
reconstruct128_1 = lrelu(batch_norm(conv2d(reconstruct128, self.gf_dim, k_h=5, k_w=5, d_h=1, d_w=1, name="recon128_conv"), name="recon128_bnc"))
reconstruct128_1_r = resblock(reconstruct128_1, name="dec128_res2")
reconstruct128_2 = lrelu(batch_norm(conv2d(reconstruct128_1_r, self.gf_dim/2, d_h=1, d_w=1, name="recon128_conv2"),name="recon128_bnc2"))
img128 = tf.nn.tanh(conv2d(reconstruct128_2, 3, d_h=1, d_w=1, name="check_img128"))
return img128, img64, img32, img32, img32, img128, img64, img32
def generator(self, images, batch_size, name = "generator", reuse = False):
with tf.variable_scope(name) as scope:
if reuse:
scope.reuse_variables()
# imgs: input: IMAGE_SIZE x IMAGE_SIZE x CHANNEL
# return labels: IMAGE_SIZE x IMAGE_SIZE x 3
# U-Net structure, slightly different from the original on the location of relu/lrelu
#128x128
c0 = lrelu(conv2d(images, self.gf_dim, k_h=7, k_w=7, d_h=1, d_w=1, name="g_conv0"))
c0r = resblock(c0, k_h=7, k_w=7, name="g_conv0_res")
c1 = lrelu(batch_norm(conv2d(c0r, self.gf_dim, k_h=5, k_w=5, name="g_conv1"),name="g_bnc1"))
#64x64
c1r = resblock(c1, k_h=5, k_w=5, name="g_conv1_res")
c2 = lrelu(batch_norm(conv2d(c1r, self.gf_dim*2, name='g_conv2'),name="g_bnc2"))
#32x32
c2r = resblock(c2, name="g_conv2_res")
c3 = lrelu(batch_norm(conv2d(c2r, self.gf_dim*4, name='g_conv3'),name="g_bnc3"))
#16x16
c3r = resblock(c3, name="g_conv3_res")
c4 = lrelu(batch_norm(conv2d(c3r, self.gf_dim*8, name='g_conv4'),name="g_bnc4"))
#8x8
c4r = resblock(c4, name="g_conv4_res")
# c5 = lrelu(batch_norm(conv2d(c4r, self.gf_dim*8, name='g_conv5'),name="g_bnc5"))
# #4x4
# #2x2
# c6r = resblock(c6,k_h=2, k_w=2, name="g_conv6_res")
c4r2 = resblock(c4r, name="g_conv4_res2")
c4r3 = resblock(c4r2, name="g_conv4_res3")
c4r4 = resblock(c4r3, name="g_conv4_res4")
c4r4_l = tf.reshape(c4r4,[batch_size, -1])
c7_l = linear(c4r4_l, output_size=512,scope='feature', bias_start=0.1, with_w=True)[0]
c7_l_m = tf.maximum(c7_l[:, 0:256], c7_l[:, 256:])
return c0r, c1r, c2r, c3r, c4r4, c7_l_m
def partRotator(self, images, name, batch_size=10, reuse=False):
#HW 40x40, 32x40, 32x48
with tf.variable_scope(name) as scope:
if reuse:
scope.reuse_variables()
c0 = lrelu(conv2d(images, self.gf_dim, d_h=1, d_w=1, name="p_conv0"))
c0r = resblock(c0, name="p_conv0_res")
c1 = lrelu(batch_norm(conv2d(c0r, self.gf_dim*2, name="p_conv1"),name="p_bnc1"))
#down1
c1r = resblock(c1, name="p_conv1_res")
c2 = lrelu(batch_norm(conv2d(c1r, self.gf_dim*4, name='p_conv2'),name="p_bnc2"))
#down2
c2r = resblock(c2, name="p_conv2_res")
c3 = lrelu(batch_norm(conv2d(c2r, self.gf_dim*8, name='p_conv3'),name="p_bnc3"))
#down3 5x5, 4x5, 4x6
c3r = resblock(c3, name="p_conv3_res")
c3r2 = resblock(c3r, name="p_conv3_res2")
shape = c3r2.get_shape().as_list()
d1 = lrelu(batch_norm(deconv2d(c3r2, [shape[0], shape[1] * 2, shape[2] * 2, self.gf_dim*4], name="p_deconv1"), name="p_bnd1"))
#up1
after_select_d1 = lrelu(batch_norm(conv2d(tf.concat_v2([d1, c2r], axis=3), self.gf_dim*4, d_h=1, d_w=1, name="p_deconv1_s"),name="p_bnd1_s"))
d1_r = resblock(after_select_d1, name="p_deconv1_res")
d2 = lrelu(batch_norm(deconv2d(d1_r, [shape[0], shape[1] * 4, shape[2] * 4, self.gf_dim*2], name="p_deconv2"), name="p_bnd2"))
#up2
after_select_d2 = lrelu(batch_norm(conv2d(tf.concat_v2([d2, c1r], axis=3), self.gf_dim*2, d_h=1, d_w=1, name="p_deconv2_s"),name="p_bnd2_s"))
d2_r = resblock(after_select_d2, name="p_deconv2_res")
d3 = lrelu(batch_norm(deconv2d(d2_r, [shape[0], shape[1] * 8, shape[2] * 8, self.gf_dim], name="p_deconv3"), name="p_bnd3"))
#up3
after_select_d3 = lrelu(batch_norm(conv2d(tf.concat_v2([d3, c0r], axis=3), self.gf_dim, d_h=1, d_w=1, name="p_deconv3_s"),name="p_bnd3_s"))
d3_r = resblock(after_select_d3, name="p_deconv3_res")
check_part = tf.nn.tanh(conv2d(d3_r, 3, d_h=1, d_w=1, name="p_check"))
return d3_r, check_part
def partCombiner(self, eyel, eyer, nose, mouth):
'''
x y
43.5823 41.0000
86.4177 41.0000
64.1165 64.7510
47.5863 88.8635
82.5904 89.1124
this is the mean locaiton of 5 landmarks
'''
eyel_p = tf.pad(eyel, [[0,0], [int(41 - EYE_H / 2 - 1), int(IMAGE_SIZE - (41+EYE_H/2 - 1))], [int(44 - EYE_W / 2 - 1), int(IMAGE_SIZE - (44+EYE_W/2-1))], [0,0]])
eyer_p = tf.pad(eyer, [[0,0], [int(41 - EYE_H / 2 - 1), int(IMAGE_SIZE - (41+EYE_H/2 - 1))], [int(86 - EYE_W / 2 - 1), int(IMAGE_SIZE - (86+EYE_W/2-1))], [0,0]])
nose_p = tf.pad(nose, [[0,0], [int(65 - NOSE_H / 2 - 1), int(IMAGE_SIZE - (65+NOSE_H/2 - 1))], [int(64 - NOSE_W / 2 - 1), int(IMAGE_SIZE - (64+NOSE_W/2-1))], [0,0]])
month_p = tf.pad(mouth, [[0,0], [int(89 - MOUTH_H / 2 - 1), int(IMAGE_SIZE - (89+MOUTH_H/2 - 1))], [int(65 - MOUTH_W / 2 - 1), int(IMAGE_SIZE - (65+MOUTH_W/2-1))], [0,0]])
eyes = tf.maximum(eyel_p, eyer_p)
eye_nose = tf.maximum(eyes, nose_p)
return tf.maximum(eye_nose, month_p)
def evaluate(self,epoch, idx, batch_idxs, start_time, mode,
batch_images_with_code, batch_eyel, batch_eyer, batch_nose, batch_mouth,
batch_labels, batch_eyel_label, batch_eyer_label, batch_nose_label, batch_mouth_label, batch_iden):
errD = self.d_loss.eval({self.images_with_code: batch_images_with_code, self.labels: batch_labels,
self.eyel : batch_eyel, self.eyer : batch_eyer, self.nose: batch_nose, self.mouth: batch_mouth,})
errG_L = self.weightedErrL1.eval({self.images_with_code: batch_images_with_code, self.labels : batch_labels,
self.eyel : batch_eyel, self.eyer : batch_eyer, self.nose: batch_nose, self.mouth: batch_mouth,})
errG_L2 = self.weightedErrL2.eval({self.images_with_code: batch_images_with_code, self.labels : batch_labels,
self.eyel : batch_eyel, self.eyer : batch_eyer, self.nose: batch_nose, self.mouth: batch_mouth,})
errG_L3 = self.weightedErrL3.eval({self.images_with_code: batch_images_with_code, self.labels : batch_labels,
self.eyel : batch_eyel, self.eyer : batch_eyer, self.nose: batch_nose, self.mouth: batch_mouth,})
errG_adver = self.g_loss_adver.eval({self.images_with_code: batch_images_with_code,
self.eyel : batch_eyel, self.eyer : batch_eyer, self.nose: batch_nose, self.mouth: batch_mouth,})
errtv = self.tv_loss.eval({self.images_with_code: batch_images_with_code, self.labels : batch_labels,
self.eyel : batch_eyel, self.eyer : batch_eyer, self.nose: batch_nose, self.mouth: batch_mouth,})
errG_sym = self.symErrL1.eval({self.images_with_code: batch_images_with_code,
self.eyel : batch_eyel, self.eyer : batch_eyer, self.nose: batch_nose, self.mouth: batch_mouth,})
errG2_sym = self.symErrL2.eval({self.images_with_code: batch_images_with_code,
self.eyel : batch_eyel, self.eyer : batch_eyer, self.nose: batch_nose, self.mouth: batch_mouth,})
errG3_sym = self.symErrL3.eval({self.images_with_code: batch_images_with_code,
self.eyel : batch_eyel, self.eyer : batch_eyer, self.nose: batch_nose, self.mouth: batch_mouth,})
errcheck32 = 0#self.weightedErr_check32.eval({self.images_with_code: batch_images_with_code, self.labels : batch_labels, })
errcheck64 = 0#self.weightedErr_check64.eval({self.images_with_code: batch_images_with_code, self.labels : batch_labels, })
errcheck128 = 0#self.weightedErr_check128.eval({self.images_with_code: batch_images_with_code, self.labels : batch_labels, })
erreyel = self.eyel_loss.eval({self.eyel: batch_eyel, self.eyel_label: batch_eyel_label })
erreyer = self.eyer_loss.eval({self.eyel: batch_eyel, self.eyer: batch_eyer, self.eyer_label: batch_eyer_label})
errnose = self.nose_loss.eval({self.nose: batch_nose, self.nose_label: batch_nose_label})
errmouth = self.mouth_loss.eval({self.mouth : batch_mouth, self.mouth_label: batch_mouth_label})
erriden = self.idenloss.eval({self.images_with_code: batch_images_with_code, self.idenlabels : batch_iden,})
if 'f' in MODE:
errDv = self.dv_loss.eval({self.images_with_code: batch_images_with_code, self.labels : batch_labels,
self.eyel : batch_eyel, self.eyer : batch_eyer, self.nose: batch_nose, self.mouth: batch_mouth,})
else:
errDv = 0
err_total_G = L1_1_W * (errG_L + errG_sym * SYM_W) + L1_2_W * (errG_L2 + errG2_sym * SYM_W) + L1_3_W * (errG_L3 + errG3_sym * SYM_W) \
+ ALPHA_ADVER * errG_adver + errDv * BELTA_FEATURE + IDEN_W*erriden #+ COND_WEIGHT*(errCondL12 + errCondL23) + errtv * TV_WEIGHT
errfeat_total = errcheck32 + errcheck64+errcheck128 + PART_W * (erreyel + erreyer + errnose + errmouth)
tobePrint = "%s Epo[%2d][%4d/%4d][t%4.2f] d_l:%.4f" % (MODE + 'T' if mode == 'test' else '', epoch, idx, batch_idxs, time.time() - start_time, errD)
tobePrint += " fstol:%.0f cg32:%.0f cg64:%.0f cg128:%.0f el:%.0f er:%.0f no:%.0f mo:%.0f\n" % \
(errfeat_total, errcheck32, errcheck64,errcheck128, erreyel, erreyer, errnose, errmouth)
tobePrint += "g_l:%.0f gL1:%.0f(sym:%.0f) gL2:%.0f(sym:%.0f) gL3:%.0f(sym:%.0f) gadv:%.4f dv_l:%.2f iden:%.4f, tv:%.0f " \
% (err_total_G, errG_L, errG_sym, errG_L2, errG2_sym, errG_L3, errG3_sym, errG_adver, errDv, erriden, errtv)
if 'f' in MODE:
tobePrint += 'L:{} G:{}'.format(ALPHA_ADVER, BELTA_FEATURE)
self.f.write(tobePrint+'\n')
self.f.flush()
print(tobePrint)
#DEEPFACE MODEL BEGINS---
def loadDeepFace(self, DeepFacePath= None):
if DeepFacePath is None:
path = sys.modules[self.__class__.__module__].__file__
# print path
path = os.path.abspath(os.path.join(path, os.pardir))
# print path
path = os.path.join(path, "DeepFace168.pickle")
DeepFacePath = path
logging.info("Load npy file from '%s'.", DeepFacePath)
if not os.path.isfile(DeepFacePath):
logging.error(("File '%s' not found. "), DeepFacePath)
sys.exit(1)
#with open(DeepFacePath,'r') as file:
# self.data_dict = pickle.load(file)
with open(DeepFacePath,'rb') as file: # Binary read
u = pickle._Unpickler(file)
u.encoding = 'latin1'
p = u.load()
self.data_dict = p
print("Deep Face pickle data file loaded")
def FeatureExtractDeepFace(self, images, name = "FeatureExtractDeepFace", reuse=False):
#Preprocessing: from color to gray(reduce_mean)
with tf.variable_scope(name) as scope:
if reuse:
scope.reuse_variables()
conv1 = self._conv_layer(images, name='conv1')
#print(3, type(3))
slice1_1, slice1_2 = tf.split(3, 2, conv1)
eltwise1 = tf.maximum(slice1_1, slice1_2)
pool1 = tf.nn.max_pool(eltwise1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1],
padding='SAME')
conv2_1 = self._conv_layer(pool1, name='conv2_1')
slice2_1_1, slice2_1_2 = tf.split(3, 2, conv2_1)
eltwise2_1 = tf.maximum(slice2_1_1, slice2_1_2)
conv2_2 = self._conv_layer(eltwise2_1, name='conv2_2')
slice2_2_1, slice2_2_2 = tf.split(3, 2, conv2_2)
eltwise2_2 = tf.maximum(slice2_2_1, slice2_2_2)
res2_1 = pool1 + eltwise2_2
conv2a = self._conv_layer(res2_1, name='conv2a')
slice2a_1, slice2a_2 = tf.split(3, 2, conv2a)
eltwise2a = tf.maximum(slice2a_1, slice2a_2)
conv2 = self._conv_layer(eltwise2a, name='conv2')
slice2_1, slice2_2 = tf.split(3, 2, conv2)
eltwise2 = tf.maximum(slice2_1, slice2_2)
pool2 = tf.nn.max_pool(eltwise2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1],
padding='SAME')
conv3_1 = self._conv_layer(pool2, name='conv3_1')
slice3_1_1, slice3_1_2 = tf.split(3, 2, conv3_1)
eltwise3_1 = tf.maximum(slice3_1_1, slice3_1_2)
conv3_2 = self._conv_layer(eltwise3_1, name='conv3_2')
slice3_2_1, slice3_2_2 = tf.split(3, 2, conv3_2)
eltwise3_2 = tf.maximum(slice3_2_1, slice3_2_2)
res3_1 = pool2 + eltwise3_2
conv3_3 = self._conv_layer(res3_1, name='conv3_3')
slice3_3_1, slice3_3_2 = tf.split(3, 2, conv3_3)
eltwise3_3 = tf.maximum(slice3_3_1, slice3_3_2)
conv3_4 = self._conv_layer(eltwise3_3, name='conv3_4')
slice3_4_1, slice3_4_2 = tf.split(3, 2, conv3_4)
eltwise3_4 = tf.maximum(slice3_4_1, slice3_4_2)
res3_2 = res3_1 + eltwise3_4
conv3a = self._conv_layer(res3_2, name='conv3a')
slice3a_1, slice3a_2 = tf.split(3, 2, conv3a)
eltwise3a = tf.maximum(slice3a_1, slice3a_2)
conv3 = self._conv_layer(eltwise3a, name='conv3')
slice3_1, slice3_2 = tf.split(3, 2, conv3)
eltwise3 = tf.maximum(slice3_1, slice3_2)
pool3 = tf.nn.max_pool(eltwise3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1],
padding='SAME')
conv4_1 = self._conv_layer(pool3, name='conv4_1')
slice4_1_1, slice4_1_2 = tf.split(3, 2, conv4_1)
eltwise4_1 = tf.maximum(slice4_1_1, slice4_1_2)
conv4_2 = self._conv_layer(eltwise4_1, name='conv4_2')
slice4_2_1, slice4_2_2 = tf.split(3, 2, conv4_2)
eltwise4_2 = tf.maximum(slice4_2_1, slice4_2_2)
res4_1 = pool3 + eltwise4_2
conv4_3 = self._conv_layer(res4_1, name='conv4_3')
slice4_3_1, slice4_3_2 = tf.split(3, 2, conv4_3)
eltwise4_3 = tf.maximum(slice4_3_1, slice4_3_2)
conv4_4 = self._conv_layer(eltwise4_3, name='conv4_4')
slice4_4_1, slice4_4_2 = tf.split(3, 2, conv4_4)
eltwise4_4 = tf.maximum(slice4_4_1, slice4_4_2)
res4_2 = res4_1 + eltwise4_4
conv4_5 = self._conv_layer(res4_2, name='conv4_5')
slice4_5_1, slice4_5_2 = tf.split(3, 2, conv4_5)
eltwise4_5 = tf.maximum(slice4_5_1, slice4_5_2)
conv4_6 = self._conv_layer(eltwise4_5, name='conv4_6')
slice4_6_1, slice4_6_2 = tf.split(3, 2, conv4_6)
eltwise4_6 = tf.maximum(slice4_6_1, slice4_6_2)
res4_3 = res4_2 + eltwise4_6
conv4a = self._conv_layer(res4_3, name='conv4a')
slice4a_1, slice4a_2 = tf.split(3, 2, conv4a)
eltwise4a = tf.maximum(slice4a_1, slice4a_2)
conv4 = self._conv_layer(eltwise4a, name='conv4')
slice4_1, slice4_2 = tf.split(3, 2, conv4)
eltwise4 = tf.maximum(slice4_1, slice4_2)
conv5_1 = self._conv_layer(eltwise4, name='conv5_1')
slice5_1_1, slice5_1_2 = tf.split(3, 2, conv5_1)
eltwise5_1 = tf.maximum(slice5_1_1, slice5_1_2)
conv5_2 = self._conv_layer(eltwise5_1, name='conv5_2')
slice5_2_1, slice5_2_2 = tf.split(3, 2, conv5_2)
eltwise5_2 = tf.maximum(slice5_2_1, slice5_2_2)
res5_1 = eltwise4 + eltwise5_2
conv5_3 = self._conv_layer(res5_1, name='conv5_3')
slice5_3_1, slice5_3_2 = tf.split(3, 2, conv5_3)
eltwise5_3 = tf.maximum(slice5_3_1, slice5_3_2)
conv5_4 = self._conv_layer(eltwise5_3, name='conv5_4')
slice5_4_1, slice5_4_2 = tf.split(3, 2, conv5_4)
eltwise5_4 = tf.maximum(slice5_4_1, slice5_4_2)
res5_2 = res5_1 + eltwise5_4
conv5_5 = self._conv_layer(res5_2, name='conv5_5')
slice5_5_1, slice5_5_2 = tf.split(3, 2, conv5_5)
eltwise5_5 = tf.maximum(slice5_5_1, slice5_5_2)
conv5_6 = self._conv_layer(eltwise5_5, name='conv5_6')
slice5_6_1, slice5_6_2 = tf.split(3, 2, conv5_6)
eltwise5_6 = tf.maximum(slice5_6_1, slice5_6_2)
res5_3 = res5_2 + eltwise5_6
conv5_7 = self._conv_layer(res5_3, name='conv5_7')
slice5_7_1, slice5_7_2 = tf.split(3, 2, conv5_7)
eltwise5_7 = tf.maximum(slice5_7_1, slice5_7_2)
conv5_8 = self._conv_layer(eltwise5_7, name='conv5_8')
slice5_8_1, slice5_8_2 = tf.split(3, 2, conv5_8)
eltwise5_8 = tf.maximum(slice5_8_1, slice5_8_2)
res5_4 = res5_3 + eltwise5_8
conv5a = self._conv_layer(res5_4, name='conv5a')
slice5a_1, slice5a_2 = tf.split(3, 2, conv5a)
eltwise5a = tf.maximum(slice5a_1, slice5a_2)
conv5 = self._conv_layer(eltwise5a, name='conv5')
slice5_1, slice5_2 = tf.split(3, 2, conv5)
eltwise5 = tf.maximum(slice5_1, slice5_2)
pool4 = tf.nn.max_pool(eltwise5, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1],
padding='SAME')
pool4_transposed = tf.transpose(pool4, perm=[0, 3, 1, 2])
# pool4_reshaped = tf.reshape(pool4_transposed, shape=[tf.shape(pool4)[0],-1])
fc1 = self._fc_layer(pool4_transposed, name="fc1")
slice_fc1_1, slice_fc1_2 = tf.split(1, 2, fc1)
eltwise_fc1 = tf.maximum(slice_fc1_1, slice_fc1_2)
return eltwise1, eltwise2, eltwise3, eltwise5, pool4, eltwise_fc1
#DEEPFACE NET ENDS---
#DEEPFACE OPS BEGINS---
def _conv_layer(self, input_, output_dim=96,
k_h=3, k_w=3, d_h=1, d_w=1, stddev=0.02,
name="conv2d"):
#Note: currently kernel size and input output channel num are decided by loaded filter weights.
#only strides are decided by calling param.
with tf.variable_scope(name) as scope:
filt = self.get_conv_filter(name)
conv = tf.nn.conv2d(input_, filt, strides=[1, d_h, d_w, 1], padding='SAME')
conv_biases = self.get_bias(name)
return tf.nn.bias_add(conv, conv_biases)
return conv
def _fc_layer(self, bottom, name="fc1", num_classes=None):
with tf.variable_scope(name) as scope:
#shape = bottom.get_shape().as_list()
if name == 'fc1':
filt = self.get_fc_weight(name)
bias = self.get_bias(name)
reshaped_bottom = tf.reshape(bottom,[tf.shape(bottom)[0],-1])
return tf.matmul(reshaped_bottom, filt) + bias
def get_conv_filter(self, name):
init = tf.constant_initializer(value=self.data_dict[name][0],
dtype=tf.float32)
shape = self.data_dict[name][0].shape
var = tf.get_variable(name="filter", initializer=init, shape=shape)
return var
def get_bias(self, name, num_classes=None):
bias_wights = self.data_dict[name][1]
shape = self.data_dict[name][1].shape
init = tf.constant_initializer(value=bias_wights,
dtype=tf.float32)
var = tf.get_variable(name="biases", initializer=init, shape=shape)