-
Notifications
You must be signed in to change notification settings - Fork 123
/
model_non_linear_3DMM_proxy.py
3733 lines (2739 loc) · 215 KB
/
model_non_linear_3DMM_proxy.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
import os
import time
import csv
import random
from random import randint
from math import floor
from glob import glob
import tensorflow as tf
import tensorflow.contrib.slim as slim
import numpy as np
from six.moves import xrange
from progress.bar import Bar
from rendering_ops import *
from ops import *
from utils import *
#SUBJECT_NUM_MTPIE = 200
#SUBJECT_NUM_MTPIE_FULL = 347
#SUBJECT_NUM_CASIA = 10575
TRI_NUM = 105840
VERTEX_NUM = 53215
TRI_NUM_REMESH6k = 12281
VERTEX_NUM_REMESH6k = 6248
#TRI_NUM_REDUCE = 12281
VERTEX_NUM_REDUCE = 39111
WARPING = True
CONST_PIXELS_NUM = 20
HISTRORY_SZ = 1000
class DCGAN(object):
def __init__(self, sess, config, devices=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.sess = sess
self.c_dim = config.c_dim
print('------------------------------')
print(config.gpu.split(','))
self.gpu_num = len(config.gpu.split(','))
print(self.gpu_num)
self.is_using_landmark = config.is_using_landmark
self.is_using_symetry = config.is_using_symetry
self.is_using_res_symetry = config.is_using_res_symetry
self.is_using_recon = config.is_using_recon
self.is_using_frecon = config.is_using_frecon
self.is_using_graddiff = config.is_using_graddiff
self.is_gt_m = config.is_gt_m
self.is_using_linear = False #config.is_using_linear
self.is_batchwise_white_shading = config.is_batchwise_white_shading
self.is_const_albedo = config.is_const_albedo
self.is_const_local_albedo = config.is_const_local_albedo
self.is_smoothness = config.is_smoothness
self.is_using_GAN = config.is_using_GAN
self.is_using_L2_GAN = config.is_using_L2_GAN
self.is_using_patchGAN = config.is_using_patchGAN
self.is_random_gan_labels = config.is_random_gan_labels
self.is_combine = config.is_combine
self.is_2d_normalize = True
self.is_partbase_albedo = config.is_partbase_albedo
self.is_2d_shape = True #config.is_2d_shape
self.batch_size = config.batch_size
self.image_size = config.image_size
self.sample_size = config.sample_size
self.output_size = config.output_size
self.texture_size = [192, 224]
self.z_dim = config.z_dim
self.gf_dim = config.gf_dim
self.df_dim = config.df_dim
self.gfc_dim = config.gfc_dim
self.dfc_dim = config.dfc_dim
self.shape_loss = config.shape_loss if hasattr(config, 'shape_loss') else "l2"
self.tex_loss = 'l1' #config.tex_loss if hasattr(config, 'tex_loss') else "l1"
self.mDim = 8
self.poseDim = 7
self.shapeDim = 199
self.expDim = 29
self.texDim = 40
self.ilDim = 9 * 3
self.is_reduce = config.is_reduce
if self.is_reduce:
self.vertexNum = VERTEX_NUM_REDUCE
else:
self.vertexNum = VERTEX_NUM
self.landmark_num = 68
# batch normalization : deals with poor initialization helps gradient flow
self.bns = {}
self.d_bn0_0 = batch_norm(name='d_k_bn0_0')
self.d_bn0_1 = batch_norm(name='d_k_bn0_1')
self.d_bn0_2 = batch_norm(name='d_k_bn0_2')
self.d_bn1_0 = batch_norm(name='d_k_bn1_0')
self.d_bn1_1 = batch_norm(name='d_k_bn1_1')
self.d_bn1_2 = batch_norm(name='d_k_bn1_2')
self.d_bn1_3 = batch_norm(name='d_k_bn1_3')
self.d_bn2_0 = batch_norm(name='d_k_bn2_0')
self.d_bn2_1 = batch_norm(name='d_k_bn2_1')
self.d_bn2_2 = batch_norm(name='d_k_bn2_2')
self.d_bn3_0 = batch_norm(name='d_k_bn3_0')
self.d_bn3_1 = batch_norm(name='d_k_bn3_1')
self.d_bn3_2 = batch_norm(name='d_k_bn3_2')
self.d_bn3_3 = batch_norm(name='d_k_bn3_3')
self.d_bn4_0 = batch_norm(name='d_k_bn4_0')
self.d_bn4_1 = batch_norm(name='d_k_bn4_1')
self.d_bn4_2 = batch_norm(name='d_k_bn4_2')
self.d_bn4_1_l = batch_norm(name='d_k_bn4_1_l')
self.d_bn4_2_l = batch_norm(name='d_k_bn4_2_l')
self.d_bn4_1_r = batch_norm(name='d_k_bn4_1_r')
self.d_bn4_2_r = batch_norm(name='d_k_bn4_2_r')
self.d_bn4_1_p = batch_norm(name='d_k_bn4_1_p')
self.d_bn4_2_p = batch_norm(name='d_k_bn4_2_P')
self.d_bn4_1_a = batch_norm(name='d_k_bn4_1_a')
self.d_bn4_2_a = batch_norm(name='d_k_bn4_2_a')
self.d_bn5 = batch_norm(name='d_k_bn5')
#self.g2_bn0_0 = batch_norm(name='g_h2_bn0_0')
#self.g2_bn0_1 = batch_norm(name='g_h2_bn0_1')
#self.g2_bn0_2 = batch_norm(name='g_h2_bn0_2')
self.dataset_name = config.dataset
self.checkpoint_dir = config.checkpoint_dir
self.samples_dir = config.samples_dir
#model_dir = "%s_%s_%s_%s_%s_%s_%s" % (self.dataset_name, self.batch_size, self.output_size, self.gf_dim, self.gfc_dim, self.df_dim, self.dfc_dim)
if not os.path.exists(self.samples_dir+"/"+self.model_dir):
os.makedirs(self.samples_dir+"/"+self.model_dir)
if not os.path.exists(self.checkpoint_dir+"/"+self.model_dir):
os.makedirs(self.checkpoint_dir+"/"+self.model_dir)
self.devices = devices
self.setupParaStat()
self.setupValData()
self.build_model()
def build_model(self):
def filename2image(input_filenames, offset_height = None, offset_width = None, target_height=None, target_width=None):
batch_size = len(input_filenames)
if offset_height != None:
offset_height = tf.split(offset_height, batch_size)
offset_width = tf.split(offset_width, batch_size)
images = []
for i in range(batch_size):
file_contents = tf.read_file(input_filenames[i])
image = tf.image.decode_png(file_contents, channels=3)
if offset_height != None:
image = tf.image.crop_to_bounding_box(image, tf.reshape(offset_height[i], []), tf.reshape(offset_width[i], []), target_height, target_width)
images.append(image)
return tf.cast(tf.stack(images), tf.float32)
#self.subject_num = SUBJECT_NUM_CASIA
#self.id_labels_CASIA = tf.placeholder(tf.float32, [self.batch_size, self.subject_num], name='id_labels_CASIA')
self.m_300W_labels = tf.placeholder(tf.float32, [self.batch_size, self.mDim], name='m_300W_labels')
self.shape_300W_labels = tf.placeholder(tf.float32, [self.batch_size, self.vertexNum * 3], name='shape_300W_labels')
self.texture_300W_labels = tf.placeholder(tf.float32, [self.batch_size, self.texture_size[0], self.texture_size[1], self.c_dim], name='tex_300W_labels')
#self.exp_300W_labels = tf.placeholder(tf.float32, [self.batch_size, self.expDim], name='exp_300W_labels')
#self.il_300W_labels = tf.placeholder(tf.float32, [self.batch_size, self.ilDim], name='lighting_300W_labels')
self.input_offset_height = tf.placeholder(tf.int32, [self.batch_size], name='input_offset_height')
self.input_offset_width = tf.placeholder(tf.int32, [self.batch_size], name='input_offset_width')
self.input_images_fn_300W = [tf.placeholder(dtype=tf.string) for _ in range(self.batch_size)]
self.input_masks_fn_300W = [tf.placeholder(dtype=tf.string) for _ in range(self.batch_size)]
self.texture_labels_fn_300W = [tf.placeholder(dtype=tf.string) for _ in range(self.batch_size)]
self.texture_masks_fn_300W = [tf.placeholder(dtype=tf.string) for _ in range(self.batch_size)]
#self.input_images_300W = filename2image(self.input_images_fn_300W, offset_height = self.input_offset_height, offset_width = self.input_offset_width, target_height=self.output_size, target_width=self.output_size)
#self.input_images_300W = self.input_images_300W /127.5 - 1
#self.input_masks_300W = filename2image(self.input_masks_fn_300W, offset_height = self.input_offset_height, offset_width = self.input_offset_width, target_height=self.output_size, target_width=self.output_size)
#self.input_masks_300W = self.input_masks_300W / 255.0
#self.texture_300W_labels = filename2image(self.texture_labels_fn_300W)
#self.texture_300W_labels = self.texture_300W_labels / 127.5 - 1
#self.texture_mask_300W_labels = filename2image(self.texture_masks_fn_300W)
#self.texture_mask_300W_labels = self.texture_mask_300W_labels / 255.0
# For const alb loss
self.albedo_indexes_x1 = tf.placeholder(tf.int32, [self.batch_size, CONST_PIXELS_NUM, 1], name='idexes_x1')
self.albedo_indexes_y1 = tf.placeholder(tf.int32, [self.batch_size, CONST_PIXELS_NUM, 1], name='idexes_y1')
self.albedo_indexes_x2 = tf.placeholder(tf.int32, [self.batch_size, CONST_PIXELS_NUM, 1], name='idexes_x2')
self.albedo_indexes_y2 = tf.placeholder(tf.int32, [self.batch_size, CONST_PIXELS_NUM, 1], name='idexes_y2')
self.const_alb_mask = load_const_alb_mask()
def rendering(texture, m ,shape, background=None, extra_mask=None):
image, mask = warp_texture(texture, m, shape, output_size=self.output_size, is_reduce = self.is_reduce)
mask = tf.expand_dims(mask, -1)
if extra_mask != None:
mask = tf.multiply(extra_mask, mask)
if background != None:
image = tf.multiply(image, mask) + tf.multiply(background, 1 - mask)
return image, mask
def rendering_wflow(texture, flow_u, flow_v, flow_mask, background=None, extra_mask=None):
image = bilinear_sampler(texture, flow_v, flow_u)
mask = tf.expand_dims(flow_mask, -1)
if extra_mask != None:
mask = tf.multiply(extra_mask, mask)
if background != None:
image = tf.multiply(image, mask) + tf.multiply(background, 1 - mask)
return image, mask
def alb2tex(albedo, shade):
return 2.0*tf.multiply( (albedo + 1.0)/2.0, shade) - 1
def model_and_loss(input_images_fn_300W, input_masks_fn_300W, texture_labels_fn_300W, texture_masks_fn_300W, input_offset_height, input_offset_width, m_300W_labels, shape_300W_labels, albedo_indexes_x1, albedo_indexes_y1, albedo_indexes_x2, albedo_indexes_y2):
batch_size = self.batch_size / self.gpu_num
input_images_300W_ = filename2image(input_images_fn_300W, offset_height = input_offset_height, offset_width = input_offset_width, target_height=self.output_size, target_width=self.output_size)
input_images_300W = input_images_300W_ /127.5 - 1
input_masks_300W = filename2image(input_masks_fn_300W, offset_height = input_offset_height, offset_width = input_offset_width, target_height=self.output_size, target_width=self.output_size)
input_masks_300W = input_masks_300W / 255.0
texture_300W_labels = filename2image(texture_labels_fn_300W)
texture_300W_labels = texture_300W_labels / 127.5 - 1
texture_mask_300W_labels = filename2image(texture_masks_fn_300W)
texture_mask_300W_labels = texture_mask_300W_labels / 255.0
## ------------------------- Network ---------------------------
shape_fx_300W, alb_fx_300W, m_300W, il_300W = self.generator_encoder( input_images_300W, is_reuse=False)
if False: #v2
shape_base_300W, shape_2d_base_300W, shape_res_300W, shape_2d_res_300W = self.generator_decoder_shape(shape_fx_300W, is_reuse=False, is_training=True)
shape_2d_300W = shape_2d_base_300W + shape_2d_res_300W
shape_300W = shape_base_300W + shape_res_300W
albedo_base_300W, albedo_res_300W = self.generator_decoder_albedo(alb_fx_300W, is_reuse=False, is_training=True)
albedo_300W = albedo_base_300W + albedo_res_300W
else: #v3
shape_base_300W, shape_2d_base_300W, shape_300W, shape_2d_300W = self.generator_decoder_shape(shape_fx_300W, is_reuse=False, is_training=True)
shape_2d_res_300W = shape_2d_300W - shape_2d_base_300W
shape_res_300W = shape_300W - shape_base_300W
albedo_base_300W, albedo_300W = self.generator_decoder_albedo(alb_fx_300W, is_reuse=False, is_training=True)
albedo_res_300W = albedo_300W - albedo_base_300W
m_300W_full = self.m2full(m_300W)
m_300W_labels_full = self.m2full(m_300W_labels)
shape_base_300W_full = self.shape2full(shape_base_300W)
shape_300W_full = self.shape2full(shape_300W)
shape_300W_labels_full = self.shape2full(shape_300W_labels)
if self.is_gt_m:
m_4syn = m_300W_labels_full
else:
m_4syn = m_300W_full
# Shade
shade_base_300W, rotated_normal_2d = generate_shade(il_300W, m_4syn, shape_base_300W_full, self.texture_size, is_reduce= self.is_reduce, is_with_normal=True)
shade_300W = generate_shade(il_300W, m_4syn, shape_300W_full, self.texture_size, is_reduce= self.is_reduce)
# Texture
texture_base_300W = alb2tex(albedo_base_300W, shade_base_300W)
texture_mix_s0a1_300W = alb2tex(albedo_300W, shade_base_300W)
texture_mix_s1a0_300W = alb2tex(albedo_base_300W, shade_300W)
texture_300W = alb2tex(albedo_300W, shade_300W)
# Rendering
G_flow_base_u, G_flow_base_v, G_flow_base_300W_mask = warping_flow(m_4syn, shape_base_300W_full, output_size=self.output_size, is_reduce = self.is_reduce)
G_flow_u, G_flow_v, G_flow_300W_mask = warping_flow(m_4syn, shape_300W_full, output_size=self.output_size, is_reduce = self.is_reduce)
G_images_base_300W, G_images_base_300W_mask = rendering_wflow(texture_base_300W, G_flow_base_u, G_flow_base_v, G_flow_base_300W_mask, background=input_images_300W, extra_mask=input_masks_300W)
G_images_mix_s0a1_300W, _ = rendering_wflow(texture_mix_s0a1_300W, G_flow_base_u, G_flow_base_v, G_flow_base_300W_mask, background=input_images_300W, extra_mask=input_masks_300W)
G_images_mix_s1a0_300W, _ = rendering_wflow(texture_mix_s1a0_300W, G_flow_u, G_flow_v, G_flow_300W_mask, background=input_images_300W, extra_mask=input_masks_300W)
G_images_300W, G_images_300W_mask = rendering_wflow(texture_300W, G_flow_u, G_flow_v, G_flow_300W_mask, background=input_images_300W, extra_mask=input_masks_300W)
# Landmarks
landmark_u_300W, landmark_v_300W = compute_landmarks(m_4syn, shape_300W_full, output_size=self.output_size, is_reduce = self.is_reduce)
landmark_u_base_300W, landmark_v_base_300W = compute_landmarks(m_300W_full, shape_base_300W_full, output_size=self.output_size, is_reduce = self.is_reduce)
landmark_u_base2_300W, landmark_v_base2_300W = compute_landmarks(m_300W_full, shape_300W_full, output_size=self.output_size, is_reduce = self.is_reduce)
landmark_u_300W_labels, landmark_v_300W_labels = compute_landmarks(m_300W_labels_full, shape_300W_labels_full, output_size=self.output_size, is_reduce = self.is_reduce)
## ------------------------- Losses ---------------------------
g_loss = tf.zeros(1)
G_loss_shape = 10*norm_loss(shape_base_300W, shape_300W_labels, loss_type = self.shape_loss)
G_loss_m = 5*norm_loss(m_300W, m_300W_labels, loss_type = 'l2')
g_loss += G_loss_m + G_loss_shape
texture_vis_mask = tf.cast(tf.not_equal(texture_300W_labels, tf.ones_like(texture_300W_labels)*(-1)), tf.float32)
texture_vis_mask = tf.multiply(texture_vis_mask, texture_mask_300W_labels)
texture_ratio = tf.reduce_sum(texture_vis_mask) / (batch_size* self.texture_size[0] * self.texture_size[1] * self.c_dim)
uv_mask_tf = tf.expand_dims(tf.expand_dims(tf.constant( self.uv_mask, dtype = tf.float32 ), 0), -1)
if self.is_batchwise_white_shading:
mean_shade = tf.reduce_mean( tf.multiply(shade_base_300W, uv_mask_tf) , axis=[0,1,2]) * 16384 / 10379
G_loss_white_shading = 10*norm_loss(mean_shade, 0.99*tf.ones([1, 3], dtype=tf.float32), loss_type = "l2")
else:
G_loss_white_shading = tf.zeros(1)
g_loss += G_loss_white_shading
G_loss_texture = (norm_loss(texture_base_300W, texture_300W_labels, mask = texture_vis_mask, loss_type = self.tex_loss) +
#+norm_loss(texture_300W, texture_300W_labels, mask = texture_vis_mask, loss_type = self.tex_loss) +
norm_loss(texture_mix_s0a1_300W, texture_300W_labels, mask = texture_vis_mask, loss_type = self.tex_loss) +
norm_loss(texture_mix_s1a0_300W, texture_300W_labels, mask = texture_vis_mask, loss_type = self.tex_loss)) / 2 / texture_ratio
G_loss_recon_base = 10*norm_loss(G_images_base_300W, input_images_300W, loss_type = self.tex_loss ) / (tf.reduce_sum(G_images_base_300W_mask)/ (batch_size* self.output_size * self.output_size))
G_loss_recon_mix_s0a1 = 10*norm_loss(G_images_mix_s0a1_300W, input_images_300W, loss_type = self.tex_loss ) / (tf.reduce_sum(G_images_base_300W_mask)/ (batch_size* self.output_size * self.output_size))
G_loss_recon_mix_s1a0 = 10*norm_loss(G_images_mix_s1a0_300W, input_images_300W, loss_type = self.tex_loss ) / (tf.reduce_sum(G_images_300W_mask) / (batch_size* self.output_size * self.output_size))
G_loss_recon = 10*norm_loss(G_images_300W, input_images_300W, loss_type = self.tex_loss ) / (tf.reduce_sum(G_images_300W_mask) / (batch_size* self.output_size * self.output_size))
if self.is_smoothness:
G_loss_smoothness = 5e5*norm_loss( (shape_2d_base_300W[:, :-2, 1:-1, :] + shape_2d_base_300W[:, 2:, 1:-1, :] + shape_2d_base_300W[:, 1:-1, :-2, :] + shape_2d_base_300W[:, 1:-1, 2:, :])/4.0,
shape_2d_base_300W[:, 1:-1, 1:-1, :], loss_type = 'l2') + \
norm_loss( (shape_2d_300W[:, :-2, 1:-1, :] + shape_2d_300W[:, 2:, 1:-1, :] + shape_2d_300W[:, 1:-1, :-2, :] + shape_2d_300W[:, 1:-1, 2:, :])/4.0,
shape_2d_300W[:, 1:-1, 1:-1, :], loss_type = 'l2')
else:
G_loss_smoothness = tf.zeros(1)
g_loss += G_loss_smoothness
G_landmark_loss = (tf.reduce_mean(tf.nn.l2_loss(landmark_u_base_300W - landmark_u_300W_labels )) + tf.reduce_mean(tf.nn.l2_loss(landmark_v_base_300W - landmark_v_300W_labels ))) / self.landmark_num / batch_size / 50 + \
(tf.reduce_mean(tf.nn.l2_loss(landmark_u_base2_300W - landmark_u_300W_labels )) + tf.reduce_mean(tf.nn.l2_loss(landmark_v_base2_300W - landmark_v_300W_labels ))) / self.landmark_num / batch_size / 50 + \
(tf.reduce_mean(tf.nn.l2_loss(landmark_u_300W - landmark_u_300W_labels )) + tf.reduce_mean(tf.nn.l2_loss(landmark_v_300W - landmark_v_300W_labels ))) / self.landmark_num / batch_size / 50
#G_landmark_loss = ((tf.reduce_mean(tf.abs(landmark_u_base_300W - landmark_u_300W_labels )) + tf.reduce_mean(tf.abs(landmark_v_base_300W - landmark_v_300W_labels ))) / self.landmark_num / batch_size + \
# (tf.reduce_mean(tf.abs(landmark_u_base2_300W - landmark_u_300W_labels )) + tf.reduce_mean(tf.abs(landmark_v_base2_300W - landmark_v_300W_labels ))) / self.landmark_num / batch_size + \
# (tf.reduce_mean(tf.abs(landmark_u_300W - landmark_u_300W_labels )) + tf.reduce_mean(tf.abs(landmark_v_300W - landmark_v_300W_labels ))) / self.landmark_num / batch_size )* 1e2
if self.is_using_symetry:
albedo_base_300W_flip = tf.map_fn(lambda img: tf.image.flip_left_right(img), albedo_base_300W)
G_loss_symetry = 10*norm_loss(tf.maximum(tf.abs(albedo_base_300W-albedo_base_300W_flip), 0.01), 0, loss_type = self.tex_loss)
else:
G_loss_symetry = tf.zeros(1)
g_loss += G_loss_symetry
if self.is_using_res_symetry:
_, _, shape_res_300W_z = tf.split(shape_2d_res_300W, num_or_size_splits=3, axis=3)
_, _, rotated_normal_z = tf.split(rotated_normal_2d, num_or_size_splits=3, axis=3)
shape_res_300W_z_flip = tf.map_fn(lambda img: tf.image.flip_left_right(img), shape_res_300W_z)
G_loss_res_symetry = 100*norm_loss(shape_res_300W_z, tf.stop_gradient(shape_res_300W_z_flip), mask= tf.cast( tf.less(rotated_normal_z, 0), tf.float32), loss_type = 'l1')
else:
G_loss_res_symetry = tf.zeros(1)
g_loss += G_loss_res_symetry
G_loss_res = norm_loss(shape_res_300W, 0, loss_type = 'l1') + norm_loss(albedo_res_300W, 0, loss_type = 'l1')
g_loss += G_loss_res
if self.is_const_albedo:
albedo_1 = get_pixel_value(albedo_base_300W, albedo_indexes_x1, albedo_indexes_y1)
albedo_2 = get_pixel_value(albedo_base_300W, albedo_indexes_x2, albedo_indexes_y2)
G_loss_albedo_const = 10*norm_loss( tf.maximum(tf.abs(albedo_1- albedo_2), 0.01), 0, loss_type = self.tex_loss)
else:
G_loss_albedo_const = tf.zeros(1)
g_loss += G_loss_albedo_const
if self.is_const_local_albedo:
local_albedo_alpha = 0.9
texture_300W_labels_chromaticity = (texture_300W_labels + 1.0)/2.0
texture_300W_labels_chromaticity = tf.divide(texture_300W_labels_chromaticity, tf.reduce_sum(texture_300W_labels_chromaticity, axis=[-1], keep_dims=True) + 1e-6)
w_u = tf.stop_gradient(tf.exp(-15*tf.norm( texture_300W_labels_chromaticity[:, :-1, :, :] - texture_300W_labels_chromaticity[:, 1:, :, :], ord='euclidean', axis=-1, keep_dims=True)) * texture_vis_mask[:, :-1, :, :] )
G_loss_local_albedo_const_u = tf.reduce_mean(norm_loss( albedo_base_300W[:, :-1, :, :], albedo_base_300W[:, 1:, :, :], loss_type = 'l2,1', reduce_mean=False, p=0.8) * w_u) / tf.reduce_sum(w_u+1e-6)
w_v = tf.stop_gradient(tf.exp(-15*tf.norm( texture_300W_labels_chromaticity[:, :, :-1, :] - texture_300W_labels_chromaticity[:, :, 1:, :], ord='euclidean', axis=-1, keep_dims=True)) * texture_vis_mask[:, :, :-1, :] )
G_loss_local_albedo_const_v = tf.reduce_mean(norm_loss( albedo_base_300W[:, :, :-1, :], albedo_base_300W[:, :, 1:, :], loss_type = 'l2,1', reduce_mean=False, p=0.8) * w_v) / tf.reduce_sum(w_v+1e-6)
G_loss_local_albedo_const = (G_loss_local_albedo_const_u + G_loss_local_albedo_const_v)*10
else:
G_loss_local_albedo_const = tf.zeros(1)
g_loss += G_loss_local_albedo_const
if self.is_using_recon:
g_loss += (G_loss_recon_base + 2*G_loss_recon_mix_s0a1 + 2*G_loss_recon_mix_s1a0)/2 #G_loss_recon +
else:
g_loss += G_loss_texture*4
g_loss += G_loss_texture
G_loss_frecon = tf.zeros(1)
if self.is_using_frecon:
from vgg_face import vgg_face
#G_images_300W_vgg = (tf.multiply(G_images_300W, G_images_300W_mask)+ 1.0)*127.5
G_images_mix_s1a0_300W_vgg = (G_images_mix_s1a0_300W + 1.0)*127.5
G_images_mix_s0a1_300W_vgg = (G_images_mix_s0a1_300W + 1.0)*127.5
input_images_300W_vgg = (input_images_300W + 1.0)*127.5
input_features, _, _ = vgg_face('vgg-face.mat', input_images_300W_vgg)
recon_mix_s1a0_features, _, _ = vgg_face('vgg-face.mat', G_images_mix_s1a0_300W_vgg)
recon_mix_s0a1_features, _, _ = vgg_face('vgg-face.mat', G_images_mix_s0a1_300W_vgg)
layer_names = ['conv1_1', 'conv2_1', 'conv3_1', 'conv4_1', 'conv5_1'] #['conv1_2', 'conv2_2', 'conv3_3', 'conv4_3', 'conv5_3']
layer_weights = [50, 500, 2000, 3000, 250]
for i in range(len(layer_names)):
layer_name = layer_names[i]
layer_weight = layer_weights[i]
G_loss_frecon += norm_loss(input_features[layer_name], recon_mix_s1a0_features[layer_name], loss_type='l2') / (layer_weight*layer_weight * len(layer_names))*2
G_loss_frecon += norm_loss(input_features[layer_name], recon_mix_s0a1_features[layer_name], loss_type='l2') / (layer_weight*layer_weight * len(layer_names))*2
g_loss = g_loss + G_loss_frecon
G_loss_graddiff = tf.zeros(1)
if self.is_using_graddiff:
G_images_300W_grad_u = G_images_300W[:, :-1, :, :] - G_images_300W[:, 1:, :, :]
input_images_300W_grad_u = input_images_300W[:, :-1, :, :] - input_images_300W[:, 1:, :, :]
G_images_300W_grad_v = G_images_300W[:, :, :-1, :] - G_images_300W[:, :, 1:, :]
input_images_300W_grad_v = input_images_300W[:, :, :-1, :] - input_images_300W[:, :, 1:, :]
G_loss_graddiff = 10* ( norm_loss(G_images_300W_grad_u, input_images_300W_grad_u, loss_type = 'l1' ) +
norm_loss(G_images_300W_grad_u, input_images_300W_grad_u, loss_type = 'l1' ) ) #/ (batch_size* self.output_size * self.output_size)
g_loss = g_loss + G_loss_graddiff
G_loss_shade_mag = norm_loss(shade_base_300W, 1, mask= tf.cast(tf.greater(texture_base_300W , 1), dtype=tf.float32) , loss_type = 'l1')
g_loss += G_loss_shade_mag
'''
# GAN loss
if self.is_using_GAN:
if not self.is_using_L2_GAN:
self.d_loss_real = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=self.images_D_logit, labels= 0.99*tf.ones_like(self.images_D_logit)))
self.d_loss_fake = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=self.G_images_D_logit, labels= tf.zeros_like(self.G_images_D_logit)))
self.d_loss_fake_his = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=self.G_images_D_logit, labels= tf.zeros_like(self.G_images_D_logit)))
self.d_loss = self.d_loss_real + self.d_loss_fake + self.d_loss_fake_his
self.G_loss_adversarial = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=self.G_images_D_logit, labels= tf.ones_like(self.G_images_D_logit)))
else:
if self.is_random_gan_labels:
self.d_loss_real = norm_loss(self.images_D_logit, tf.ones_like(self.images_D_logit) + tf.random_uniform(shape = self.images_D_logit.get_shape(), minval=-0.1, maxval=0.1, dtype=tf.float32), loss_type = 'l2')
#self.d_loss_fake = norm_loss(self.G_images_D_logit, tf.random_uniform(shape = self.G_images_D_logit.get_shape(), minval=-0.2, maxval=0.2, dtype=tf.float32), loss_type = 'l2')
#self.d_loss_fake_his = norm_loss(self.G_images_his_D_logit, tf.random_uniform(shape = self.G_images_his_D_logit.get_shape(), minval=-0.2, maxval=0.2, dtype=tf.float32), loss_type = 'l2')
else:
self.d_loss_real = norm_loss(self.images_D_logit, tf.ones_like(self.images_D_logit), loss_type = 'l2')
self.d_loss_fake = norm_loss(self.G_images_D_logit, tf.zeros_like(self.G_images_D_logit), loss_type = 'l2')
self.d_loss_fake_his = norm_loss(self.G_images_his_D_logit, tf.zeros_like(self.G_images_his_D_logit), loss_type = 'l2')
self.d_loss = (self.d_loss_real*2 + self.d_loss_fake + self.d_loss_fake_his)/4
self.G_loss_adversarial = norm_loss(self.G_images_D_logit, tf.ones_like(self.G_images_D_logit), loss_type = 'l2')
self.g_loss = self.g_loss + self.G_loss_adversarial
'''
if self.is_using_landmark:
g_loss_wlandmark = G_landmark_loss + g_loss
else:
g_loss_wlandmark = g_loss
return g_loss, g_loss_wlandmark, G_loss_m, G_loss_shape, G_loss_texture, G_loss_recon, G_loss_recon_base, G_loss_recon_mix_s1a0, G_loss_recon_mix_s0a1, G_loss_frecon, G_loss_graddiff, G_landmark_loss, \
G_loss_symetry, G_loss_res_symetry, G_loss_white_shading, G_loss_albedo_const, G_loss_smoothness, G_loss_local_albedo_const, G_loss_res, G_loss_shade_mag, \
G_images_300W, G_images_base_300W, G_images_mix_s1a0_300W, G_images_mix_s0a1_300W, texture_300W, texture_base_300W, texture_mix_s1a0_300W, texture_mix_s0a1_300W, albedo_300W, albedo_base_300W, shade_300W, shade_base_300W, texture_300W_labels, input_images_300W
g_loss, g_loss_wlandmark, G_loss_m, G_loss_shape, G_loss_texture, G_loss_recon, G_loss_recon_base, G_loss_recon_mix_s1a0, G_loss_recon_mix_s0a1, G_loss_frecon, G_loss_graddiff, G_landmark_loss, \
G_loss_symetry, G_loss_res_symetry, G_loss_white_shading, G_loss_albedo_const, G_loss_smoothness, G_loss_local_albedo_const, G_loss_res, G_loss_shade_mag, \
G_images_300W, G_images_base_300W, G_images_mix_s1a0_300W, G_images_mix_s0a1_300W, texture_300W, texture_base_300W, texture_mix_s1a0_300W, texture_mix_s0a1_300W, albedo_300W, albedo_base_300W, shade_300W, shade_base_300W, texture_300W_labels, input_images_300W \
= make_parallel(model_and_loss, self.gpu_num,
input_images_fn_300W= self.input_images_fn_300W, input_masks_fn_300W=self.input_masks_fn_300W,
texture_labels_fn_300W=self.texture_labels_fn_300W, texture_masks_fn_300W=self.texture_masks_fn_300W,
input_offset_height=self.input_offset_height, input_offset_width=self.input_offset_width,
m_300W_labels = self.m_300W_labels, shape_300W_labels=self.shape_300W_labels,
albedo_indexes_x1= self.albedo_indexes_x1, albedo_indexes_y1 = self.albedo_indexes_y1,
albedo_indexes_x2=self.albedo_indexes_x2, albedo_indexes_y2 = self.albedo_indexes_y2)
#with tf.variable_scope(tf.get_variable_scope(), reuse=True):
# var_name = 'w_u'
self.w_u = tf.zeros([]) #get_variable(var_name)
t_vars = tf.trainable_variables()
weights_vars = [var for var in t_vars if ('conv/w' in var.name or 'Matrix' in var.name or 'kernel' in var.name) ]
print('weights_vars')
for var in weights_vars:
print(var.name)
self.regu_loss = tf.add_n([ tf.nn.l2_loss(var) for var in weights_vars ]) * 1e-6
self.g_loss = tf.reduce_mean(g_loss) #+ self.regu_loss
self.g_loss_wlandmark = tf.reduce_mean(g_loss_wlandmark)
self.G_loss_m = tf.reduce_mean(G_loss_m)
self.G_loss_shape = tf.reduce_mean(G_loss_shape)
self.G_loss_texture = tf.reduce_mean(G_loss_texture)
self.G_loss_recon = tf.reduce_mean(G_loss_recon)
self.G_loss_recon_base = tf.reduce_mean(G_loss_recon_base)
self.G_loss_recon_mix_s1a0 = tf.reduce_mean(G_loss_recon_mix_s1a0)
self.G_loss_recon_mix_s0a1 = tf.reduce_mean(G_loss_recon_mix_s0a1)
self.G_loss_frecon = tf.reduce_mean(G_loss_frecon)
self.G_loss_graddiff = tf.reduce_mean(G_loss_graddiff)
self.G_landmark_loss = tf.reduce_mean(G_landmark_loss)
self.G_loss_symetry = tf.reduce_mean(G_loss_symetry)
self.G_loss_res_symetry = tf.reduce_mean(G_loss_res_symetry)
self.G_loss_white_shading = tf.reduce_mean(G_loss_white_shading)
self.G_loss_albedo_const = tf.reduce_mean(G_loss_albedo_const)
self.G_loss_res = tf.reduce_mean(G_loss_res)
self.G_loss_local_albedo_const = tf.reduce_mean(G_loss_local_albedo_const)
self.G_loss_smoothness = tf.reduce_mean(G_loss_smoothness)
self.G_loss_shade_mag = tf.reduce_mean(G_loss_shade_mag)
self.G_images_300W = tf.clip_by_value(tf.concat(G_images_300W, axis=0), -1, 1)
self.G_images_mix_s1a0_300W = tf.clip_by_value(tf.concat(G_images_mix_s1a0_300W, axis=0), -1, 1)
self.G_images_mix_s0a1_300W = tf.clip_by_value(tf.concat(G_images_mix_s0a1_300W, axis=0), -1, 1)
self.G_images_base_300W = tf.clip_by_value(tf.concat(G_images_base_300W, axis=0), -1, 1)
self.texture_300W = tf.clip_by_value(tf.concat(texture_300W, axis=0), -1, 1)
self.texture_base_300W = tf.clip_by_value(tf.concat(texture_base_300W, axis=0), -1, 1)
self.texture_mix_s1a0_300W = tf.clip_by_value(tf.concat(texture_mix_s1a0_300W, axis=0), -1, 1)
self.texture_mix_s0a1_300W = tf.clip_by_value(tf.concat(texture_mix_s0a1_300W, axis=0), -1, 1)
self.albedo_300W = tf.concat(albedo_300W, axis=0)
self.albedo_base_300W = tf.concat(albedo_base_300W, axis=0)
self.shade_300W = tf.concat(shade_300W, axis=0)
self.shade_base_300W = tf.concat(shade_base_300W, axis=0)
self.texture_300W_labels = tf.concat(texture_300W_labels, axis=0)
self.input_images_300W = tf.concat(input_images_300W, axis=0)
'''
if self.is_using_GAN:
self.input_good_images_300W = tf.placeholder(tf.float32, [self.batch_size, self.output_size, self.output_size, self.c_dim], name='input_good_images_300W')
self.G_images_300W_his = tf.placeholder(tf.float32, [self.batch_size, self.output_size, self.output_size, self.c_dim], name='history_images_300W')
if self.is_combine:
self.all_D_logits = self.discriminator( tf.concat( [self.input_good_images_300W, self.G_images_300W, self.G_images_300W_his], axis = 0 ), is_reuse = False )
self.images_D_logit, self.G_images_D_logit, self.G_images_his_D_logit = tf.split(self.all_D_logits, num_or_size_splits = 3)
else:
self.images_D_logit = self.discriminator(self.input_good_images_300W, is_reuse = False)
self.G_images_D_logit = self.discriminator(self.G_images_300W, is_reuse = True)
self.G_images_his_D_logit = self.discriminator(self.G_images_300W_his, is_reuse = True)
'''
# Sampler
#self.sample_images= tf.placeholder(tf.float32, [self.sample_size] + [self.output_size, self.output_size, self.c_dim], name='sample_images')
#self.sample_m_ph = tf.placeholder(tf.float32, [self.sample_size, self.mDim], name='mm')
#self.sample_shape_ph = tf.placeholder(tf.float32, [self.sample_size, self.vertexNum * 3], name='ssss')
#self.sample_texture_ph = tf.placeholder(tf.float32, [self.sample_size, self.texture_size, self.texture_size, self.c_dim], name='tttt')
#self.landmark_u, self.landmark_v = compute_landmarks(self.sample_m_ph, self.sample_shape_ph, output_size=self.output_size, is_reduce = self.is_reduce)
#self.s_shape, self.s_shade, self.s_shaded_texture, self.s_texture, self.s_m, self.s_img, self.s_overlay_img = self.sampler(self.sample_images)
self.d_vars = [var for var in t_vars if 'd_' in var.name]
self.g_vars = [var for var in t_vars if 'g_' in var.name]
self.g_vars_v6 = [var for var in self.g_vars if 'v6' in var.name]
self.g_en_vars = [var for var in t_vars if 'g_k' in var.name]
self.g_tex_de_vars = [var for var in t_vars if 'g_h' in var.name]
self.g_shape_de_vars = [var for var in t_vars if 'g_s' in var.name]
def setupParaStat(self):
if self.is_reduce:
self.tri = load_3DMM_tri_reduce()
self.vertex_tri = load_3DMM_vertex_tri_reduce()
self.vt2pixel_u, self.vt2pixel_v = load_FaceAlignment_vt2pixel_reduce()
self.uv_tri, self.uv_mask = load_FaceAlignment_tri_2d_reduce(with_mask = True)
else:
self.tri = load_3DMM_tri()
self.vertex_tri = load_3DMM_vertex_tri()
self.vt2pixel_u, self.vt2pixel_v = load_FaceAlignment_vt2pixel()
self.uv_tri, self.uv_mask = load_FaceAlignment_tri_2d(with_mask = True)
# Basis
mu_shape, w_shape = load_FaceAlignment_basic('shape', is_reduce = self.is_reduce)
mu_exp, w_exp = load_FaceAlignment_basic('exp', is_reduce = self.is_reduce)
self.mean_shape = mu_shape + mu_exp
if self.is_2d_normalize:
#self.mean_shape = np.tile(np.array([0, 0, 6e4]), VERTEX_NUM)
self.std_shape = np.tile(np.array([1e4, 1e4, 1e4]), self.vertexNum)
else:
#self.mean_shape = np.load('mean_shape.npy')
self.std_shape = np.load('std_shape.npy')
self.mean_shape_tf = tf.constant(self.mean_shape, tf.float32)
self.std_shape_tf = tf.constant(self.std_shape, tf.float32)
self.mean_m = np.load('mean_m.npy')
self.std_m = np.load('std_m.npy')
self.mean_m_tf = tf.constant(self.mean_m, tf.float32)
self.std_m_tf = tf.constant(self.std_m, tf.float32)
self.w_shape = w_shape
self.w_exp = w_exp
#mu_tex, w_tex = load_FaceAlignment_2dbasic('tex')
##mu_tex = mu_tex.reshape(-1, 128, 128, 3)
#self.mu_tex = mu_tex#/127.5-1
#self.w_tex = w_tex#/127.5-1
#self.mu_tex_const = tf.constant(mu_tex.reshape(-1, 128, 128, 3)/127.5-1 )
def m2full(self, m):
return m * self.std_m_tf + self.mean_m_tf
def shape2full(self, shape):
return shape * self.std_shape_tf + self.mean_shape_tf
def setupTrainingData(self):
# Training data - 300W
dataset = ['AFW', 'AFW_Flip', 'HELEN', 'HELEN_Flip', 'IBUG', 'IBUG_Flip', 'LFPW', 'LFPW_Flip']
dataset_num = len(dataset)
images = [0] * dataset_num
pid = [0] * dataset_num
m = [0] * dataset_num
pose = [0] * dataset_num
shape = [0] * dataset_num
exp = [0] * dataset_num
tex_para = [0] * dataset_num
tex = [0] * dataset_num
il = [0] * dataset_num
alb = [0] * dataset_num
mask = [0] * dataset_num
for i in range(dataset_num):
#images_AFLW2000, pid_AFLW2000, m_AFLW2000, pose_AFLW2000, shape_AFLW2000, exp_AFLW2000, tex_para_AFLW2000, _, = load_FaceAlignment_dataset_recrop_sz224('AFW', with_sh = False)
images[i], pid[i], m[i], pose[i], shape[i], exp[i], tex_para[i], _, = load_FaceAlignment_dataset_recrop_sz224(dataset[i], False) #_, tex[i], il[i], alb[i], mask[i]
self.images_300W = np.concatenate(images, axis=0)
images = None
#self.textures_300W = np.concatenate(tex, axis=0)
#tex = None
#self.albedos_300W = np.concatenate(alb, axis=0)
#alb = None
#self.masks_300W = np.concatenate(mask, axis=0)
#mask = None
all_m = np.concatenate(m, axis=0)
#self.mean_m = np.mean(all_m, axis=0)
#self.std_m = np.std(all_m, axis=0)
#np.save('mean_m.npy', self.mean_m)
#np.save('std_m.npy', self.std_m)
#print("Save mean m--------------------------------")
all_shape_para = np.concatenate(shape, axis=0)
all_exp_para = np.concatenate(exp, axis=0)
self.all_tex_para = np.concatenate(tex_para, axis=0)
self.pids_300W = np.concatenate(pid, axis=0)
#self.all_il = np.concatenate(il, axis=0)
self.all_m = np.divide(np.subtract(all_m, self.mean_m), self.std_m)
self.mean_shape_para = np.mean(all_shape_para, axis=0)
self.std_shape_para = np.std(all_shape_para, axis=0)
self.all_shape_para = all_shape_para #np.divide(np.subtract(all_shape_para, self.mean_shape_para), self.std_shape_para)
self.mean_exp_para = np.mean(all_exp_para, axis=0)
self.std_exp_para = np.std(all_exp_para, axis=0)
self.all_exp_para = all_exp_para #np.divide(np.subtract(all_exp_para, self.mean_exp_para), self.std_exp_para)
'''
# Basis
if self.is_remesh:
mu_shape, w_shape = load_FaceAlignment_basic_remesh6k('shape')
mu_exp, w_exp = load_FaceAlignment_basic_remesh6k('exp')
else:
mu_shape, w_shape = load_FaceAlignment_basic('shape')
mu_exp, w_exp = load_FaceAlignment_basic('exp')
var_shape_para = np.var(all_shape_para, axis=0)
w_shape_square = np.square(w_shape)
var_exp_para = np.var(all_exp_para, axis=0)
w_exp_square = np.square(w_exp)
var_shape = np.matmul(w_shape_square, var_shape_para) + np.matmul(w_exp_square, var_exp_para)
self.std_shape =np.sqrt(var_shape)
self.mean_shape = mu_shape+mu_exp
if self.is_remesh:
np.save('std_shape_remesh6k.npy', self.std_shape)
np.save('mean_shape_remesh6k.npy', self.mean_shape)
'''
def setupValData(self):
return
# Samples data - AFLW200
self.images_AFLW2000, pid_AFLW2000, m_AFLW2000, pose_AFLW2000, shape_AFLW2000, exp_AFLW2000, _, _, _, self.tex_AFLW2000 = load_FaceAlignment_dataset_recrop('AFLW2000')
self.AFLW2000_m = np.divide(np.subtract(m_AFLW2000, self.mean_m), self.std_m)
self.AFLW2000_shape_para = shape_AFLW2000
self.AFLW2000_exp_para = exp_AFLW2000
def train(self, config):
np.random.seed(0)
tf.set_random_seed(0)
# Training data
self.setupTrainingData()
valid_300W_idx = range(self.images_300W.shape[0])
print("Valid images %d/%d" % ( len(valid_300W_idx), self.images_300W.shape[0] ))
np.random.shuffle(valid_300W_idx)
# Pairwise preparing
num_300W = len(self.pids_300W)
self.is_using_pairwise = False
if self.is_using_pairwise:
id_dict = {}
for i in range(num_300W):
pid = int(self.pids_300W[i])
if (pid not in id_dict):
id_dict[pid] = []
id_dict[pid].append(i)
# Optimizer
#elf.setupLossFunctions()
print('Optim start')
history = np.zeros(1)
if self.is_using_GAN:
real_indices = load_FaceAlignment_datasets_good()
real_id_dict = {}
for i in range(num_300W):
if real_indices[i]:
pid = int(self.pids_300W[i])
if (pid not in real_id_dict):
real_id_dict[pid] = []
real_id_dict[pid].append(i)
real_indices = [int(i) for i in range(real_indices.shape[0]) if ( real_indices[i])]
# History
if os.path.exists(self.history_file):
history = np.load(self.history_file)
else:
history = np.zeros([HISTRORY_SZ, self.output_size, self.output_size, self.c_dim], dtype=np.float32)
d_optim = tf.train.AdamOptimizer(config.learning_rate, beta1=config.beta1).minimize(self.d_loss, var_list=self.d_vars, colocate_gradients_with_ops=True)
g_optim = tf.train.AdamOptimizer(config.learning_rate, beta1=config.beta1).minimize(self.g_loss, var_list=self.g_vars, colocate_gradients_with_ops=True)
g_en_optim = tf.train.AdamOptimizer(config.learning_rate, beta1=config.beta1).minimize(self.g_loss_wlandmark, var_list=self.g_en_vars, colocate_gradients_with_ops=True)
tf.global_variables_initializer().run()
#self.g_global_vars = [var for var in tf.global_variables() if not ('v6' in var.name) ] # and 'w_v' not in var.name]
#self.saver = tf.train.Saver(self.g_global_vars, keep_checkpoint_every_n_hours=1, max_to_keep = 0)
self.saver = tf.train.Saver(tf.trainable_variables(), keep_checkpoint_every_n_hours=1, max_to_keep = 10) #tf.trainable_variables(),
#self.saver = tf.train.Saver([var for var in tf.global_variables() if 'res' not in var.name], keep_checkpoint_every_n_hours=1, max_to_keep = 10)
"""Train DCGAN"""
could_load, checkpoint_counter = self.load(self.checkpoint_dir)
if could_load:
epoch0 = checkpoint_counter + 1
print(" [*] Load SUCCESS")
else:
epoch0 = 1
print(" [!] Load failed...")
self.saver = tf.train.Saver(keep_checkpoint_every_n_hours=1, max_to_keep = 10)
start_time = time.time()
for epoch in xrange(epoch0, config.epoch):
batch_idxs = min(len(valid_300W_idx), config.train_size) // config.batch_size
print('--------------------------------')
for idx in xrange(0, batch_idxs):
batch_idx = valid_300W_idx[idx*config.batch_size:(idx+1)*config.batch_size] #valid_300W_idx[idx*config.batch_size:(idx+1)*config.batch_size]
if self.is_using_pairwise:
for i in range( int(config.batch_size/2)) :
pid = int(self.pids_300W[ batch_idx[i] ])
idx_list = id_dict.get(pid)
batch_idx[int(config.batch_size/2) + i] = idx_list[random.randint(0, len(idx_list)-1)]
batch_300W_images_fn = [self.images_300W[batch_idx[i]] for i in range(config.batch_size)]
tx = np.random.random_integers(0, 32, size=config.batch_size)
ty = np.random.random_integers(0, 32, size=config.batch_size)
delta_m = np.zeros([config.batch_size, 8]);
delta_m[:,6] = np.divide(ty, self.std_m[6]);
delta_m[:,7] = np.divide(32 - tx, self.std_m[7]);
batch_m = self.all_m[batch_idx,:] - delta_m;
#batch_il = self.all_il[batch_idx,:]
batch_shape_para = self.all_shape_para[batch_idx,:]
batch_exp_para = self.all_exp_para[batch_idx,:]
batch_tex_para = self.all_tex_para[batch_idx,:]
batch_shape = np.divide( np.matmul(batch_shape_para, np.transpose(self.w_shape)) + np.matmul(batch_exp_para, np.transpose(self.w_exp)), self.std_shape)
ffeed_dict={ self.m_300W_labels: batch_m, self.shape_300W_labels: batch_shape, self.input_offset_height: tx, self.input_offset_width: ty}
for i in range(self.batch_size):
ffeed_dict[self.input_images_fn_300W[i]] = DATA_DIR + 'image/'+ batch_300W_images_fn[i]
ffeed_dict[self.input_masks_fn_300W[i]] = DATA_DIR + 'mask_img/'+ batch_300W_images_fn[i]
ffeed_dict[self.texture_labels_fn_300W[i]] = DATA_DIR + 'texture/'+ image2texture_fn(batch_300W_images_fn[i])
ffeed_dict[self.texture_masks_fn_300W[i]] = DATA_DIR + 'mask/'+ image2texture_fn(batch_300W_images_fn[i])
if self.is_const_albedo:
indexes1 = np.random.randint(low=0, high=self.const_alb_mask.shape[0], size=[self.batch_size* CONST_PIXELS_NUM])
indexes2 = np.random.randint(low=0, high=self.const_alb_mask.shape[0], size=[self.batch_size* CONST_PIXELS_NUM])
ffeed_dict[self.albedo_indexes_x1] = np.reshape(self.const_alb_mask[indexes1, 1], [self.batch_size, CONST_PIXELS_NUM, 1])
ffeed_dict[self.albedo_indexes_y1] = np.reshape(self.const_alb_mask[indexes1, 0], [self.batch_size, CONST_PIXELS_NUM, 1])
ffeed_dict[self.albedo_indexes_x2] = np.reshape(self.const_alb_mask[indexes2, 1], [self.batch_size, CONST_PIXELS_NUM, 1])
ffeed_dict[self.albedo_indexes_y2] = np.reshape(self.const_alb_mask[indexes2, 0], [self.batch_size, CONST_PIXELS_NUM, 1])
'''
self.sess.run(g_en_optim, feed_dict=ffeed_dict)
_, g_loss, G_loss_m, G_loss_shape, G_loss_texture, G_loss_recon, G_loss_recon_base, G_loss_frecon, G_loss_graddiff, G_landmark_loss, \
G_loss_symetry, G_loss_white_shading, G_loss_albedo_const, G_loss_local_albedo_const, G_loss_smoothness, G_loss_regu, G_loss_res, w_u, \
albedo_300W, albedo_base_300W, texture_300W, shade_300W, shade_base_300W, G_images_300W, G_images_base_300W, batch_300W_textures, batch_300W_images = \
self.sess.run([g_optim, self.g_loss_wlandmark, self.G_loss_m, self.G_loss_shape, self.G_loss_texture, self.G_loss_recon, self.G_loss_recon_base, self.G_loss_frecon, self.G_loss_graddiff, self.G_landmark_loss, \
self.G_loss_symetry, self.G_loss_white_shading, self.G_loss_albedo_const, self.G_loss_local_albedo_const, self.G_loss_smoothness, self.regu_loss, self.G_loss_res, self.w_u, \
self.albedo_300W, self.albedo_base_300W, self.texture_300W, self.shade_300W, self.shade_base_300W, self.G_images_300W, self.G_images_base_300W, self.texture_300W_labels, self.input_images_300W], feed_dict=ffeed_dict)
print("Epoch [%2d][%4d/%4d] time: %4.0fs, G_loss:%2.4f (m:%.4f, shape:%.4f, tex:%.4f, recon:%.4f, base_recon:%.4f , frecon:%.4f, graddiff:%.4f, land:%.4f, sym:%.4f, wshade:%.4f, alb_const:%.4f, alb_l_const:%.4f, smooth:%.4f, res:%.4f, regu:%.4f)"\
% (epoch, idx, batch_idxs, time.time() - start_time, g_loss, G_loss_m, G_loss_shape, G_loss_texture, G_loss_recon, G_loss_recon_base, G_loss_frecon, G_loss_graddiff, G_landmark_loss, G_loss_symetry, G_loss_white_shading, G_loss_albedo_const, G_loss_local_albedo_const, G_loss_smoothness, G_loss_res, G_loss_regu))
if np.mod(idx+1, 2) == 0:
save_images(G_images_300W, [-1, 8], './{:s}/{:s}/pred_image_{:02d}_{:06d}.png'.format(self.samples_dir, self.model_dir, epoch, idx))
save_images(G_images_base_300W, [-1, 8], './{:s}/{:s}/pred_image_{:02d}_{:06d}_base.png'.format(self.samples_dir, self.model_dir, epoch, idx))
#save_images(batch_300W_textures, [-1, 8], './{:s}/{:s}/texture_{:02d}_{:06d}.png'.format(self.samples_dir, self.model_dir, epoch, idx))
save_images(batch_300W_images, [-1, 8], './{:s}/{:s}/pred_image_{:02d}_{:06d}_in.png'.format(self.samples_dir, self.model_dir, epoch, idx))
save_images(texture_300W, [-1, 8], './{:s}/{:s}/pred_shaded_texture_{:02d}_{:06d}.png'.format(self.samples_dir, self.model_dir, epoch, idx))
#save_images(texture_300W, [8, 8], './{:s}/{:s}/pred_texture_{:02d}_{:06d}.png'.format(self.samples_dir, self.model_dir, epoch, idx))
save_images(albedo_300W, [-1, 8], './{:s}/{:s}/pred_albedo_{:02d}_{:06d}.png'.format(self.samples_dir, self.model_dir, epoch, idx))
save_images(albedo_base_300W, [-1, 8], './{:s}/{:s}/pred_albedo_{:02d}_{:06d}_base.png'.format(self.samples_dir, self.model_dir, epoch, idx))
save_images(shade_300W/2, [-1, 8], './{:s}/{:s}/pred_shade_{:02d}_{:06d}.png'.format(self.samples_dir, self.model_dir, epoch, idx), inverse = False)
save_images(shade_base_300W/2, [-1, 8], './{:s}/{:s}/pred_shade_{:02d}_{:06d}_base.png'.format(self.samples_dir, self.model_dir, epoch, idx), inverse = False)
if np.mod(idx+1, 100) == 0:
self.save(config.checkpoint_dir, epoch)
'''
if not self.is_using_GAN:
#Not using GAN
if np.mod(idx, 25) == 0:
_, g_loss, G_loss_m, G_loss_shape, G_loss_texture, G_loss_recon, G_loss_recon_base, G_loss_recon_mix_s1a0, G_loss_recon_mix_s0a1, G_loss_frecon, G_loss_graddiff, G_landmark_loss, \
G_loss_symetry, G_loss_res_symetry, G_loss_white_shading, G_loss_albedo_const, G_loss_local_albedo_const, G_loss_smoothness, G_loss_regu, G_loss_res, G_loss_shade_mag, w_u = \
self.sess.run([g_en_optim, self.g_loss_wlandmark, self.G_loss_m, self.G_loss_shape, self.G_loss_texture, self.G_loss_recon, self.G_loss_recon_base, self.G_loss_recon_mix_s1a0, self.G_loss_recon_mix_s0a1, self.G_loss_frecon, self.G_loss_graddiff, self.G_landmark_loss, \
self.G_loss_symetry, self.G_loss_res_symetry, self.G_loss_white_shading, self.G_loss_albedo_const, self.G_loss_local_albedo_const, self.G_loss_smoothness, self.regu_loss, self.G_loss_res, self.G_loss_shade_mag, self.w_u],
feed_dict=ffeed_dict)
print("Epoch [%2d][%4d/%4d] time: %4.0fs, G_loss:%2.4f (m:%.4f, shape:%.4f, tex:%.4f, recon:%.4f, base_r:%.4f, mix_s1a0_r:%.4f, mix_s0a1_r:%.4f, frecon:%.4f, graddiff:%.4f, land:%.4f, sym:%.4f, res_sym:%.4f, wshade:%.4f, magshade:%.4f, alb_const:%.4f, alb_l_const:%.4f, smooth:%.4f, res:%.4f, regu:%.4f)"\
% (epoch, idx, batch_idxs, time.time() - start_time, g_loss, G_loss_m, G_loss_shape, G_loss_texture, G_loss_recon, G_loss_recon_base, G_loss_recon_mix_s1a0, G_loss_recon_mix_s0a1, G_loss_frecon, G_loss_graddiff, G_landmark_loss, G_loss_symetry, G_loss_res_symetry, G_loss_white_shading, G_loss_shade_mag, G_loss_albedo_const, G_loss_local_albedo_const, G_loss_smoothness, G_loss_res, G_loss_regu))
self.sess.run([g_en_optim], feed_dict=ffeed_dict)
self.sess.run([g_optim], feed_dict=ffeed_dict)
else:
# Good images for GAN
real_batch_idx = np.zeros(config.batch_size)
for i in range( int(config.batch_size)) :
pid = int(self.pids_300W[ batch_idx[i] ])
idx_list = real_id_dict.get(pid)
real_batch_idx[i] = idx_list[random.randint(0, len(idx_list)-1)]