-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathShmGANwithSSpecSeg.py
1341 lines (1093 loc) · 78.1 KB
/
ShmGANwithSSpecSeg.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
"""
# -----------------------------------------------------------
SHMGAN - Removal of Specular Highlights by a GAN Network
Uses Packages:
Python 3.8
CUDA 11.8
cuDnn 8.0
Tensorflow 2.8 + Keras 2.8
Tensorflow addons 0.17.1
(C) 2023 Atif Anwer, INSA Rouen, France
Email: atif.anwer@insa-rouen.fr
Folder Structure expected for the network to run:
├── PolarizedSource
├── ED // Estimated Diffuse folder pre-populated (for now)
├── I0 // Polar RGB images
├── I135
├── I45
└── I90
# -----------------------------------------------------------
"""
import gc
import os
import time
# from PIL import Image
from timeit import default_timer
import tensorflow as tf
from keras.optimizers import adam_v2
from matplotlib import pyplot as plt
from tensorflow.python.keras import (
regularizers,
)
from tensorflow.python.keras.initializers import RandomNormal
from tensorflow.python.keras.layers import (
AveragePooling2D,
Concatenate,
Conv2D,
Conv2DTranspose,
Dense,
Dropout,
Flatten,
GaussianNoise,
Input,
MaxPooling2D,
)
from tensorflow.python.keras.models import Model
from tensorflow_addons.layers import InstanceNormalization
from datasetLoader import datasetLoad
# Import SpecSeg network
from SpecSeg import SpecSeg
from utils import (
plot_single_image,
printProgressBar,
rescale_01,
)
print( "------------------------------------",
"\nTensoorflow version:", tf.__version__,
"\nKeras Version", tf.keras.__version__ ,
"\n------------------------------------",)
# Removes error when running Tensorflow on GPU
# os.environ['TF_XLA_FLAGS'] = '--tf_xla_enable_xla_devices'
# Reduces Tensorflow messages other than errors or important messages
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
os.environ['TF_ENABLE_AUTO_MIXED_PRECISION'] = '1'
# To debug @tf.functions in vscode (otherwise breakpoints dont work)
# tf.config.experimental_run_functions_eagerly(True)
tf.config.run_functions_eagerly(True)
tf.data.experimental.enable_debug_mode()
# ------------------------------------------------
# =============== THE MAIN CLASS ===============
# ------------------------------------------------
class ShmGANwithSSpecSeg( object ):
# ------------------------------------------------
#
# ██╗███╗ ██╗██╗████████╗
# ██║████╗ ██║██║╚══██╔══╝
# ██║██╔██╗ ██║██║ ██║
# ██║██║╚██╗██║██║ ██║
# ██║██║ ╚████║██║ ██║
# ╚═╝╚═╝ ╚═══╝╚═╝ ╚═╝
# ------------------------------------------------
def __init__( self, args ):
# only create Comet experiment if training, not testing
# Create an experiment with your api key
# if args.mode == 'train':
# self.comet_experiment = Experiment(
# api_key = "insert_your_api_key_here",
# project_name = "insert_your_project_name_here",
# workspace = "insert_your_workspace_here",
# auto_param_logging = True,
# auto_metric_logging = True,
# log_env_details = True,
# log_code = True, # code logging
# log_graph = True,
# log_env_gpu = True,
# log_env_host = True,
# log_env_cpu = True,
# auto_histogram_tensorboard_logging = True,
# auto_histogram_weight_logging = True,
# auto_histogram_gradient_logging = True,
# auto_histogram_activation_logging = True,
# # auto_histogram_epoch_rate=1,
# )
# self.comet_experiment.add_tag("insert_your_tags_here")
# Model configuration.
self.c_dim = args.c_dim
self.image_size = args.image_size # the size of the image after resizing
# Training configuration.
self.batch_size = args.batch_size # batch size for training
self.num_epochs = args.num_epochs
self.num_iteration_decay = args.num_iteration_decay
self.g_lr = args.g_lr
self.d_lr = args.d_lr
self.n_critic = args.n_critic
self.beta1 = args.beta1
self.beta2 = args.beta2
self.d_repeat_num = args.d_repeat_num
# Test configurations.
# self.test_iters = args.test_iters
# Miscellaneous.
self.mode = args.mode
# Directories.
self.data_dir = args.data_dir # the root folder containing polarimetric sub-folders
self.model_save_dir = args.model_save_dir
self.checkpoint_save_dir = args.checkpoint_save_dir
self.result_dir = args.result_dir
self.log_dir = args.log_dir
# self.lambda_recons = args.lambda_recons
# self.lambda_class = args.lambda_class
# Step size.
self.log_step = args.log_step
self.checkpoint_save_step = args.checkpoint_save_step
# Misc parameters
self.filter_size = args.filter_size
self.seed = 25
self.randomness = 0.50
self.dropout_amnt = 0.2 # ( 0.2 used in CollaGAN)
self.TARGET_LABELS = 0.90 # - Label smoothing by not using hard 1.0 value
# To use LSGAN
self.use_lsgan = True
self._graph=tf.Graph()
# using different optimizers for G and D
# To use decayed learning rate, replace the LR with this
decayed_lr = tf.keras.optimizers.schedules.ExponentialDecay( initial_learning_rate = self.g_lr, \
decay_steps = 10000, decay_rate = 00.95, \
staircase = False )
self.optimizer_G = adam_v2.Adam( learning_rate = decayed_lr, beta_1 = self.beta1, beta_2 = self.beta2 )
self.optimizer_D = adam_v2.Adam( learning_rate = decayed_lr, beta_1 = self.beta1, beta_2 = self.beta2 )
self.optimizer_SpecSeg = adam_v2.Adam( learning_rate = decayed_lr, beta_1 = self.beta1, beta_2 = self.beta2 )
# Use this only if Mixed precision = mixed_float16 is used. Otherwise dont use LossScaleOptimizer
# https://www.tensorflow.org/api_docs/python/tf/keras/mixed_precision/LossScaleOptimizer
# mixed_precision.set_global_policy('mixed_float16')
# self.optimizer_G = mixed_precision.LossScaleOptimizer( self.optimizer_G )
# self.optimizer_D = mixed_precision.LossScaleOptimizer( self.optimizer_D )
"""
Keep or delete old checkpoints
"""
# self.delete_old_checkpoints = args.delete_old_checkpoints
self.delete_old_checkpoints = False
# Only traing Discriminator at first. Train Generator after n epochs
self.train_G_after = 0
self.c_dim = 5
self.g_conv_dim = 64
self.g_repeat_num = 6
# for plotting histograms of gradients
self.gradmapD = {}
self.gradmapG = {}
self.init = RandomNormal(mean=0.0, stddev=0.02, seed=42) # suggested by DCGAN
# Initialize for use in training
self.random_flip = 0.0
# initializing the specular candidate matrix
self.specular_candidate = tf.zeros( [1, self.image_size, self.image_size, 1], dtype=tf.float32 )
# Trainable weights (?)
self.gamma = tf.Variable(initial_value=0, shape=(), trainable=True, dtype=tf.float32, name="Gamma")
# Zhao et al (2018) Loss function alpha as defined in paper
self.alpha = 0.84
# ------------------------------------------------
#
# ██████ ███████ ███ ██ ███████ ██████ █████ ████████ ██████ ██████
# ██ ██ ████ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██
# ██ ███ █████ ██ ██ ██ █████ ██████ ███████ ██ ██ ██ ██████
# ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██
# ██████ ███████ ██ ████ ███████ ██ ██ ██ ██ ██ ██████ ██ ██
#
# Generator has two inputs and one output
# INPUT: Concatenated multiple 5x Y-channel images (with concatenated label channels)
# OUTPUT: Single Image (Single Y channel only - concatenate with CbCr after generation) (generated_image)
# ------------------------------------------------
@tf.function(experimental_follow_type_hints=True, jit_compile=True)
def build_generator( self ):
# inp_images = INPUT2 >> Concatenated 5x Y Channel images
inp_images = Input( shape = (self.image_size, self.image_size, 10) )
# UNET Architecture inspired from CollaGAN
# inp o/p
# └── d1 -------------------------------- u4 ──┘
# └── d2 ----------------------- u3 ──┘
# └── d3 -------------- u2 ──┘
# └── d4 ----- u1 ──┘
# └─ im ─┘
# DOWNSAMPLE 5 times
N = self.filter_size
x = inp_images
for i in range( 2 ):
x = Conv2D( filters = N, kernel_size = 3, strides = 1, padding = "same", kernel_initializer=self.init, activation=tf.nn.leaky_relu, kernel_regularizer=regularizers.l2(0.001) )( x )
x = InstanceNormalization( axis = -1, epsilon=0.000001, center=True, beta_initializer=self.init )( x )
# x = LeakyReLU()( x )
down1 = x
self.attn_1, pooled = self.attention_layer( spec=self.specular_candidate, filter_size=N, pool=False )
x = AveragePooling2D( pool_size = (2, 2), strides = None, padding = 'same' )( x )
N *= 2
# ----d1
for i in range( 2 ):
x = Conv2D( filters = N, kernel_size = 3, strides = 1, padding = "same", kernel_initializer=self.init, activation=tf.nn.leaky_relu, kernel_regularizer=regularizers.l2(0.001) )( x )
x = InstanceNormalization( axis = -1, epsilon=0.000001, center=True, beta_initializer=self.init )( x )
down2 = x
self.attn_2, pooled = self.attention_layer( spec=pooled, filter_size=N, pool=True )
x = AveragePooling2D( pool_size = (2, 2), strides = None, padding = 'same' )( x )
N *= 2
# ----d2
for i in range( 2 ):
x = Conv2D( filters = N, kernel_size = 3, strides = 1, padding = "same", kernel_initializer=self.init, activation=tf.nn.leaky_relu, kernel_regularizer=regularizers.l2(0.001) )( x )
x = InstanceNormalization( axis = -1, epsilon=0.000001, center=True, beta_initializer=self.init )( x )
down3 = x
self.attn_3, pooled = self.attention_layer( spec=pooled, filter_size=N, pool=True )
x = AveragePooling2D( pool_size = (2, 2), strides = None, padding = 'same' )( x )
N *= 2
# ----d3
for i in range( 2 ):
x = Conv2D( filters = N, kernel_size = 3, strides = 1, padding = "same", kernel_initializer=self.init, activation=tf.nn.leaky_relu, kernel_regularizer=regularizers.l2(0.001) )( x )
x = InstanceNormalization( axis = -1, epsilon=0.000001, center=True, beta_initializer=self.init )( x )
down4 = x
self.attn_4, pooled = self.attention_layer( spec=pooled, filter_size=N, pool=True )
x = AveragePooling2D( pool_size = (2, 2), strides = None, padding = "same" )( x )
# ----d4
# Adding 1x1 conv layers
for i in range( 2 ):
x = Conv2D( filters = N, kernel_size = 1, strides = 1, padding = "same", kernel_initializer=self.init, activation=tf.nn.leaky_relu, kernel_regularizer=regularizers.l2(0.001) )( x )
x = InstanceNormalization( axis = -1, epsilon=0.000001, center=True, beta_initializer=self.init )( x )
# ----- 1x1 layers
# N *= 2
# Multiply the skip connection with the attention layer generated from the Mask.
# Note that attention layer is between [0,1] after sigmoid
# Note that the first call does not Maxpool the mask to retain dimension consistency
down4 = down4 + self.attn_4
down3 = down3 + self.attn_3
down2 = down2 + self.attn_2
down1 = down1 + self.attn_1
# # UPSAMPLE
# ----u1
# N /= 2
x = Conv2DTranspose( filters = N, kernel_size = 3, strides = 2, padding = "same", kernel_initializer=self.init, activation=tf.nn.leaky_relu )( x )
x = Concatenate()( [x, down4] )
for i in range( 2 ):
x = Conv2D( filters = N, kernel_size = 3, strides = 1, padding = "same", kernel_initializer=self.init, activation=tf.nn.leaky_relu, kernel_regularizer=regularizers.l2(0.001) )( x )
x = InstanceNormalization( axis = -1, epsilon=0.000001, center=True, beta_initializer=self.init )( x )
# ----u2
N /= 2
x = Conv2DTranspose( filters = N, kernel_size = 3, strides = 2, padding = "same", kernel_initializer=self.init, activation=tf.nn.leaky_relu )( x )
x = Concatenate()( [x, down3] )
for i in range( 2 ):
x = Conv2D( filters = N, kernel_size = 3, strides = 1, padding = "same", kernel_initializer=self.init, activation=tf.nn.leaky_relu, kernel_regularizer=regularizers.l2(0.001) )( x )
x = InstanceNormalization( axis = -1, epsilon=0.000001, center=True, beta_initializer=self.init )( x )
# ----u3
N /= 2
x = Conv2DTranspose( filters = N, kernel_size = 3, strides = 2, padding = "same", kernel_initializer=self.init, activation=tf.nn.leaky_relu )( x )
x = Concatenate()( [x, down2] )
for i in range( 2 ):
x = Conv2D( filters = N, kernel_size = 3, strides = 1, padding = "same", kernel_initializer=self.init, activation=tf.nn.leaky_relu, kernel_regularizer=regularizers.l2(0.001) )( x )
x = InstanceNormalization( axis = -1, epsilon=0.000001, center=True, beta_initializer=self.init )( x )
# ----u4
N /= 2
x = Conv2DTranspose( filters = N, kernel_size = 3, strides = 2, padding = "same", kernel_initializer=self.init, activation=tf.nn.leaky_relu )( x )
x = Concatenate()( [x, down1] )
for i in range( 2 ):
x = Conv2D( filters = N, kernel_size = 3, strides = 1, padding = "same", kernel_initializer=self.init, activation=tf.nn.leaky_relu, kernel_regularizer=regularizers.l2(0.001) )( x )
x = InstanceNormalization( axis = -1, epsilon=0.000001, center=True, beta_initializer=self.init )( x )
# Output is a single Y channel image
genOutput = Conv2D( filters = 1, kernel_size = 1, strides = 1, padding = "same", kernel_initializer=self.init, activation=tf.nn.leaky_relu)( x )
return Model( inp_images, genOutput, name = 'SHM_Generator' )
# ------------------------------------------------
#
# ██████ ██ ███████ ██████ ██████ ██ ███ ███ ██ ███ ██ █████ ████████ ██████ ██████
# ██ ██ ██ ██ ██ ██ ██ ██ ████ ████ ██ ████ ██ ██ ██ ██ ██ ██ ██ ██
# ██ ██ ██ ███████ ██ ██████ ██ ██ ████ ██ ██ ██ ██ ██ ███████ ██ ██ ██ ██████
# ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██
# ██████ ██ ███████ ██████ ██ ██ ██ ██ ██ ██ ██ ████ ██ ██ ██ ██████ ██ ██
#
# Discriminator has One inputs and two outputs
# INPUT: Image to discriminate (inp_img)
# OUTOUT1: Real or Fake (out_RealFake)
# OUTPUT2: Category of the image (out_cat_label)
# ------------------------------------------------
@tf.function(experimental_follow_type_hints=True, jit_compile=True)
def build_discriminator( self ):
"""Discriminator network with PatchGAN - taken from StarGAN implementation."""
# NOTE: Input image to discriminator is RGB after concatenating with Y with CbCr
# so 3 channels are only required without any labels concatenated
inp_img = Input( shape = (self.image_size, self.image_size, 3) )
x = inp_img
N = self.filter_size # default = 64... Should always be image_size/2 (?)
x = GaussianNoise( 0.1 )(x)
x = self.Conv_LReLU_IN ( x, N ) #64
x = self.Conv_LReLU_IN ( x, N * 2) #128
x = self.Conv_LReLU_IN ( x, N * 4) #256
x = self.Conv_LReLU_IN ( x, N * 8) #512
attn_disc, _ = self.attention_layer( spec=self.specular_candidate, filter_size=N * 8, pool=True, poolsize=(16,16) )
x = x + attn_disc
x = self.Conv_LReLU_IN ( x, N * 16) #1024
x = Dropout( rate=self.dropout_amnt )( x )
out_RealFake = Conv2D( filters = 1, kernel_size = 3, strides = 1, padding = 'same', \
use_bias = False, \
activation = tf.nn.leaky_relu, \
dtype = tf.float32, \
kernel_initializer = self.init )( x )
FC_Layer = Flatten()(x)
out_cat_label = Dense( units = self.c_dim, \
use_bias = False, \
kernel_initializer = self.init, \
dtype = tf.float32 )(FC_Layer) # (None, 5)
# The model is one input and two output (Real/Fake and its class)
D = Model( inp_img, [out_RealFake, out_cat_label], name = 'SHM_Discriminator' )
return D
# ----------------------------------------------
# Define the CNL block as in CollaGAN
# Conv2d + InstNorm + LeakyRELU called 2 times
@tf.function(experimental_follow_type_hints=True, jit_compile=True)
def Conv_LReLU_IN( self , x, N ):
x = Conv2D( filters = N, kernel_size = 3, strides = 2, padding = "same", kernel_initializer=self.init, use_bias=False, kernel_regularizer=regularizers.l2(0.001), activation=tf.nn.leaky_relu, dtype=tf.float32 )( x )
x = InstanceNormalization( axis = -1, epsilon=0.000001, center=True, beta_initializer=self.init )( x )
return x
# ----------------------------------------------
#
# █████ ████████ ████████ ███████ ███ ██ ████████ ██ ██████ ███ ██
# ██ ██ ██ ██ ██ ████ ██ ██ ██ ██ ██ ████ ██
# ███████ ██ ██ █████ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██
# ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██
# ██ ██ ██ ██ ███████ ██ ████ ██ ██ ██████ ██ ████
#
# ----------------------------------------------
# Generating the attention layer from the specular candidate from Shen function
# Note that the filter size is provided on call
# NOTE: Sigmoid activation included in the layer.
@tf.function(experimental_follow_type_hints=True, jit_compile=True)
def attention_layer( self, spec, filter_size, poolsize=(2,2) ,pool=True ):
if pool is True:
pooled = MaxPooling2D( pool_size = poolsize, strides = None, padding = "same" )( spec )
else:
pooled = self.specular_candidate
spec = Conv2D( filters = filter_size, kernel_size = 3, strides = 1, padding = "same", kernel_initializer=self.init, activation=tf.nn.leaky_relu, kernel_regularizer=regularizers.l2(0.001) )( pooled )
spec = Conv2D( filters = filter_size, kernel_size = 3, strides = 1, padding = "same", kernel_initializer=self.init, activation=tf.nn.leaky_relu, kernel_regularizer=regularizers.l2(0.001) )( spec )
return spec, pooled
# Sauce: https://github.com/taki0112/Self-Attention-GAN-Tensorflow
# Sauce: https://www-cxybb-com.translate.goog/article/qq_35586657/98875077?_x_tr_sl=zh-CN&_x_tr_tl=en&_x_tr_hl=en&_x_tr_pto=nui,sc
# Attention Mecanism from SAGAN paper. Calculates attention
@tf.function(experimental_follow_type_hints=True, jit_compile=True)
def sagan_attention( self, input_tensor, filter_size, poolsize=(2,2), pool=True ):
if pool is True:
input_tensor = MaxPooling2D( pool_size = poolsize, strides = None, padding = "same" )( input_tensor )
f = Conv2D( filters = filter_size // 8, kernel_size = 1, padding = "same", kernel_initializer=self.init )( input_tensor )
g = Conv2D( filters = filter_size // 8, kernel_size = 1, padding = "same", kernel_initializer=self.init )( input_tensor )
h = Conv2D( filters = filter_size, kernel_size = 1, padding = "same", kernel_initializer=self.init )( input_tensor )
input_shape = f.shape.as_list()
input_shape[0]
height = input_shape[1]
width = input_shape[2]
channels = filter_size // 8
f = tf.reshape(f, shape = [ height*width, channels] )
g = tf.reshape(g, shape = [ height*width, channels] )
h = tf.reshape(h, shape = [ height*width, filter_size] )
energy = tf.matmul( g, f, transpose_b=True)
# Note: Original SAGAN uses Softmax
attention = tf.nn.softmax( energy )
out = tf.matmul( attention, h )
gamma = self.gamma
# shape = input_tensor.get_shape().as_list()
# shape = shape[1:]
out = tf.reshape( out, shape= [ height , width, filter_size] )
output = gamma * out + input_tensor
return output, attention
"""
# ------------------------------------------------
#
# ████████ ██████ █████ ██ ███ ██ ███████ ████████ ███████ ██████
# ██ ██ ██ ██ ██ ██ ████ ██ ██ ██ ██ ██ ██
# ██ ██████ ███████ ██ ██ ██ ██ █████ ███████ ██ █████ ██████
# ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██
# ██ ██ ██ ██ ██ ██ ██ ████ ███████ ██ ███████ ██
#
# BUILD THE MODEL
# ------------------------------------------------
"""
# jit_compile=True
@tf.function(experimental_follow_type_hints=True, jit_compile=True)
def train_step( self, orig0, orig45, orig90, orig135, origED ):
# create zeros and ones for labels
tmp_zeros = tf.zeros( [1, self.image_size, self.image_size, 1], dtype=tf.float32 )
trg_ones = tf.ones( [1, self.image_size, self.image_size, 1], dtype=tf.float32 )
self.target_label_0deg = tf.Variable( [self.TARGET_LABELS,0,0,0,0], dtype=tf.float32 )
self.target_label_45deg = tf.Variable( [0,self.TARGET_LABELS,0,0,0], dtype=tf.float32 )
self.target_label_90deg = tf.Variable( [0,0,self.TARGET_LABELS,0,0], dtype=tf.float32 )
self.target_label_135deg = tf.Variable( [0,0,0,self.TARGET_LABELS,0], dtype=tf.float32 )
self.target_label_ED = tf.Variable( [0,0,0,0,self.TARGET_LABELS], dtype=tf.float32 )
ds1_yuv = self.custom_per_image_standardization( tf.image.rgb_to_yuv( orig0[:, :, :, :] ) )
ds2_yuv = self.custom_per_image_standardization( tf.image.rgb_to_yuv( orig45[:, :, :, :] ) )
ds3_yuv = self.custom_per_image_standardization( tf.image.rgb_to_yuv( orig90[:, :, :, :] ) )
ds4_yuv = self.custom_per_image_standardization( tf.image.rgb_to_yuv( orig135[:, :, :, :] ) )
ds5_yuv = self.custom_per_image_standardization( tf.image.rgb_to_yuv( origED[:, :, :, :] ) )
I0_Ych = ds1_yuv[:, :, :, 0, tf.newaxis]
I45_Ych = ds2_yuv[:, :, :, 0, tf.newaxis]
I90_Ych = ds3_yuv[:, :, :, 0, tf.newaxis]
I135_Ych = ds4_yuv[:, :, :, 0, tf.newaxis]
IED_Ych = ds5_yuv[:, :, :, 0, tf.newaxis]
# predict outside gradient tape
self.specular_candidate = (self.SpecSeg.predict(I90_Ych, verbose=0))
# using separate tapes for both D and G (sauce: pix2pix tutorial from TF site)
with tf.GradientTape( persistent=True ) as tape:
# NOTE: Generator is not trainiable in the first step, but D is trainable
self.G.trainable = False
self.D.trainable = True
"""--------------------G(1)-------------------"""
# STEP1: Generate Fake images from the Generator
# averaging the CB and CR channels to get the estimated CbCr to use later
averageCbCr = ( ds1_yuv[:, :, :, 1:] + ds2_yuv[:, :, :, 1:] + ds3_yuv[:, :, :, 1:] + ds4_yuv[:, :, :, 1:] + ds5_yuv[:, :, :, 1:] ) / 5.0
averageCbCr = tf.cast(averageCbCr, tf.float32)
# generating random numbers
RNG1 = tf.random.uniform( [] ) < self.randomness
RNG2 = tf.random.uniform( [] ) < self.randomness
RNG3 = tf.random.uniform( [] ) < self.randomness
RNG4 = tf.random.uniform( [] ) < self.randomness
RNG5 = tf.random.uniform( [] ) < self.randomness
# Generate inputs (Zero or orig image y ch) randomly for the training process
# Extracting the Y channels in a [1,rows,cols,5] stack
rand_inp1 = tf.cond( RNG1, lambda: tmp_zeros, lambda: I0_Ych )
rand_inp2 = tf.cond( RNG2, lambda: tmp_zeros, lambda: I45_Ych )
rand_inp3 = tf.cond( RNG3, lambda: tmp_zeros, lambda: I90_Ych )
rand_inp4 = tf.cond( RNG4, lambda: tmp_zeros, lambda: I135_Ych )
rand_inp5 = tf.cond( RNG5, lambda: tmp_zeros, lambda: IED_Ych )
# plt.imshow(tf.squeeze( ds1[:, :, :, 0, tf.newaxis] ), cmap=cm.gray) - Works
rand_input_Ych = tf.concat( [rand_inp1, rand_inp2, rand_inp3, rand_inp4, rand_inp5], axis = 3 )
# Generate a random number between 1 and 5 (maxval=6 required not 5 )
tf.random.uniform(shape=(), minval=1, maxval=6, dtype=tf.int32)
# making the G1 generate ED everytime ...
self.gen_input = tf.concat( [rand_input_Ych, tmp_zeros, tmp_zeros, tmp_zeros, tmp_zeros, trg_ones], axis = 3 )
self.target_img = origED
self.Target_angle_label = self.target_label_ED
# debug_plot( self.gen_input )
# Generate only 1 Y channel image, based on random target class.
self.gen_Y = self.G ( self.gen_input, training=False )
"""-------------------D(1)--------------------"""
# STEP2: Discriminate the Fake generated image and the Class of the fake image
# concatenate the generated Y channel with the average CbCr to get the complete Generated YCbCr image
self.gen_YCbCr = tf.concat( [self.gen_Y, averageCbCr], axis = 3 )
# plot_single_image ( self.gen_YCbCr )
# Normalize using std. ev
avg_stddev_arr = tf.reduce_mean(self.stddev_arr)
tf.reduce_mean(self.mean_arr)
self.gen_rgb_output = tf.image.yuv_to_rgb( ( tf.multiply(self.gen_YCbCr, avg_stddev_arr) ) * 255.0 )
# convert to RGB for the discriminator
self.gen_rgb = tf.image.yuv_to_rgb( self.gen_YCbCr )
# plot_single_image ( self.gen_Y )
# plot_single_image ( self.gen_rgb )
# DALEK: Discriminate!!
self.RealFake_gen_D1 , self.label_gen_D1 = self.D( self.gen_rgb, training=True )
"""-------------------D(2)--------------------"""
# STEP3: Discriminate the Target images and the Class of the target image
self.RealFake_target_D2 , self.label_target_D2 = self.D( self.target_img, training=True )
"""--------------------G(2)-------------------"""
# STEP5: Generate Fake images from the Generator
# NOTE: Generator is Now trainiable in the second step, but D is not
self.G.trainable = True
self.D.trainable = False
# Using previous Random values, re-substitute the generated images for cyclic consistency
# note that the dsx[] are the original batch of images passed
rand_inpA = tf.cond( RNG1, lambda: self.gen_Y, lambda: I0_Ych )
rand_inpB = tf.cond( RNG2, lambda: self.gen_Y, lambda: I45_Ych )
rand_inpC = tf.cond( RNG3, lambda: self.gen_Y, lambda: I90_Ych )
rand_inpD = tf.cond( RNG4, lambda: self.gen_Y, lambda: I135_Ych )
rand_inpE = tf.cond( RNG5, lambda: self.gen_Y, lambda: IED_Ych )
# append the randomly generated inputs to each channel. The zeros define the channel to be generated
cyc_Y1 = tf.concat( [tmp_zeros, rand_inpB, rand_inpC, rand_inpD, rand_inpE], axis = 3 )
cyc_Y2 = tf.concat( [rand_inpA, tmp_zeros, rand_inpC, rand_inpD, rand_inpE], axis = 3 )
cyc_Y3 = tf.concat( [rand_inpA, rand_inpB, tmp_zeros, rand_inpD, rand_inpE], axis = 3 )
cyc_Y4 = tf.concat( [rand_inpA, rand_inpB, rand_inpC, tmp_zeros, rand_inpE], axis = 3 )
cyc_Y5 = tf.concat( [rand_inpA, rand_inpB, rand_inpC, rand_inpD, tmp_zeros], axis = 3 )
# Append one-hot tensors as labels
cyclic_input1 = tf.concat( [cyc_Y1, trg_ones, tmp_zeros, tmp_zeros, tmp_zeros, tmp_zeros], axis = 3 )
cyclic_input2 = tf.concat( [cyc_Y2, tmp_zeros, trg_ones, tmp_zeros, tmp_zeros, tmp_zeros], axis = 3 )
cyclic_input3 = tf.concat( [cyc_Y3, tmp_zeros, tmp_zeros, trg_ones, tmp_zeros, tmp_zeros], axis = 3 )
cyclic_input4 = tf.concat( [cyc_Y4, tmp_zeros, tmp_zeros, tmp_zeros, trg_ones, tmp_zeros], axis = 3 )
cyclic_input5 = tf.concat( [cyc_Y5, tmp_zeros, tmp_zeros, tmp_zeros, tmp_zeros, trg_ones], axis = 3 )
# debug_plot( cyclic_input1 )
# debug_plot( cyclic_input2 )
# debug_plot( cyclic_input3 )
# debug_plot( cyclic_input4 )
# debug_plot( cyclic_input5 )
# Generate 5 Y channel CYCLIC images, to reconstruct the original images.
cyc_0_Y = self.G( cyclic_input1, training=True )
cyc_45_Y = self.G( cyclic_input2, training=True )
cyc_90_Y = self.G( cyclic_input3, training=True )
cyc_135_Y = self.G( cyclic_input4, training=True )
cyc_ED_Y = self.G( cyclic_input5, training=True )
"""--------------------D(3)-------------------"""
# STEP3: ?
# concatenate the generated CYCLIC Y channel with the average CbCr to get the complete Generated YCbCr image
cyc_gen0_yuv = tf.concat( [ cyc_0_Y, averageCbCr], axis = 3 )
cyc_gen45_yuv = tf.concat( [ cyc_45_Y, averageCbCr], axis = 3 )
cyc_gen90_yuv = tf.concat( [ cyc_90_Y, averageCbCr], axis = 3 )
cyc_gen135_yuv = tf.concat( [ cyc_135_Y, averageCbCr], axis = 3 )
cyc_genED_yuv = tf.concat( [ cyc_ED_Y, averageCbCr], axis = 3 )
# convert to RGB for the discriminator
self.cyc_gen0_rgb = tf.image.yuv_to_rgb( cyc_gen0_yuv )
self.cyc_gen45_rgb = tf.image.yuv_to_rgb( cyc_gen45_yuv )
self.cyc_gen90_rgb = tf.image.yuv_to_rgb( cyc_gen90_yuv )
self.cyc_gen135_rgb = tf.image.yuv_to_rgb( cyc_gen135_yuv )
self.cyc_genED_rgb = tf.image.yuv_to_rgb( cyc_genED_yuv )
# Train discriminator to differentiate the generated CYCLIC images
RealFake_cyc_gen0_D3 , label_cyc_gen0_D3 = self.D( self.cyc_gen0_rgb, training=False )
RealFake_cyc_gen45_D3 , label_cyc_gen45_D3 = self.D( self.cyc_gen45_rgb, training=False )
RealFake_cyc_gen90_D3 , label_cyc_gen90_D3 = self.D( self.cyc_gen90_rgb, training=False )
RealFake_cyc_gen135_D3, label_cyc_gen135_D3 = self.D( self.cyc_gen135_rgb, training=False )
RealFake_cyc_genED_D3 , label_cyc_genED_D3 = self.D( self.cyc_genED_rgb, training=False )
"""-------------------D(4)--------------------"""
# STEP1: Train Discriminator with real images only
# Discriminator outs for each of the images
# Discriminator outputs whether image was real/fake AND the type of the image
# REAL or FAKE , POLAR Ө = discriminator( real_img )
RealFake_orig0_D4 , self.label_orig0_D4 = self.D( orig0, training=False )
RealFake_orig45_D4 , self.label_orig45_D4 = self.D( orig45, training=False )
RealFake_orig90_D4 , self.label_orig90_D4 = self.D( orig90, training=False )
RealFake_orig135_D4 , self.label_orig135_D4 = self.D( orig135, training=False )
RealFake_origED_D4 , self.label_origED_D4 = self.D( origED, training=False )
# ==================================================================
#
# ██ ██████ ███████ ███████ ███████ ███████
# ██ ██ ██ ██ ██ ██ ██
# ██ ██ ██ ███████ ███████ █████ ███████
# ██ ██ ██ ██ ██ ██ ██
# ███████ ██████ ███████ ███████ ███████ ███████
# ==================================================================
"""----------------Losses--------------------"""
"""# STEP4 A: Generator GAN loss"""
# Total 4 losses (CollaGAN) or 4 (SHMGAN?)
# 1. Generator - 2x Genreator Loss (Real/Fake Image)
# -
# 2. Discriminator - 2x Classification Loss (Classified Labels)
#
# The loss is to compare the generated image from its actual targeted image.
# There will be 5 cyclic losses, one for each image since all 5 images are being generated
# real -> generated -> cyclic
# L1_loss = tf.keras.losses.MeanSquaredError()
# GENERATOR LOSS
# aka G_gan_loss_cyc
D3_RealFake_cyc1 = tf.math.reduce_mean( tf.math.squared_difference( RealFake_cyc_gen0_D3, self.TARGET_LABELS ) )
D3_RealFake_cyc2 = tf.math.reduce_mean( tf.math.squared_difference( RealFake_cyc_gen45_D3, self.TARGET_LABELS ) )
D3_RealFake_cyc3 = tf.math.reduce_mean( tf.math.squared_difference( RealFake_cyc_gen90_D3, self.TARGET_LABELS ) )
D3_RealFake_cyc4 = tf.math.reduce_mean( tf.math.squared_difference( RealFake_cyc_gen135_D3, self.TARGET_LABELS ) )
D3_RealFake_cyc5 = tf.math.reduce_mean( tf.math.squared_difference( RealFake_cyc_genED_D3, self.TARGET_LABELS ) )
self.D3_RealFake_cyc = D3_RealFake_cyc1 + D3_RealFake_cyc2 + D3_RealFake_cyc3 + D3_RealFake_cyc4 + D3_RealFake_cyc5
# Calculating loss for the generated image
self.D1_RealFake_loss = tf.math.reduce_mean( tf.math.squared_difference( self.RealFake_gen_D1, self.TARGET_LABELS ) )
self.G_gan_loss = ( self.D3_RealFake_cyc + self.D1_RealFake_loss ) / 6.0
# -------------------------------------------------
# oneHot lables will be of shape (1, 5)
oneHot_lbl1 = tf.reshape(tf.one_hot(tf.cast(0,tf.uint8),5),[1,5])
oneHot_lbl2 = tf.reshape(tf.one_hot(tf.cast(1,tf.uint8),5),[1,5])
oneHot_lbl3 = tf.reshape(tf.one_hot(tf.cast(2,tf.uint8),5),[1,5])
oneHot_lbl4 = tf.reshape(tf.one_hot(tf.cast(3,tf.uint8),5),[1,5])
oneHot_lbl5 = tf.reshape(tf.one_hot(tf.cast(4,tf.uint8),5),[1,5])
oneHot_lbl_target = tf.reshape(self.Target_angle_label,[1,5])
# NOTE: Warning: This op EXPECTS unscaled logits, since it performs a softmax on logits internally for efficiency.
# Do not call this op with the output of softmax, as it will produce incorrect results.
# https://www.tensorflow.org/api_docs/python/tf/nn/softmax_cross_entropy_with_logits
# Binary crossentropy = Sigmoid crossentropy
# Categorical crossentropy = Softmax crossentropy (for multi-class classification)
D3_classification_loss1 = tf.nn.softmax_cross_entropy_with_logits( labels = oneHot_lbl1, logits = label_cyc_gen0_D3 )
D3_classification_loss2 = tf.nn.softmax_cross_entropy_with_logits( labels = oneHot_lbl2, logits = label_cyc_gen45_D3 )
D3_classification_loss3 = tf.nn.softmax_cross_entropy_with_logits( labels = oneHot_lbl3, logits = label_cyc_gen90_D3 )
D3_classification_loss4 = tf.nn.softmax_cross_entropy_with_logits( labels = oneHot_lbl4, logits = label_cyc_gen135_D3 )
D3_classification_loss5 = tf.nn.softmax_cross_entropy_with_logits( labels = oneHot_lbl5, logits = label_cyc_genED_D3 )
self.D3_classification_loss = D3_classification_loss1 + D3_classification_loss2 + D3_classification_loss3 + D3_classification_loss4 + D3_classification_loss5
self.D1_classification_loss = tf.nn.softmax_cross_entropy_with_logits( labels = oneHot_lbl_target, logits = self.label_gen_D1, name="error_orig" )
self.G_clsf_loss = ( self.D3_classification_loss + self.D1_classification_loss ) / 6.0
# -------------------------------------------------
# DISCRIMINATOR LOSS:
D4_classification_loss1 = tf.nn.softmax_cross_entropy_with_logits( labels = oneHot_lbl1, logits = self.label_orig0_D4 )
D4_classification_loss2 = tf.nn.softmax_cross_entropy_with_logits( labels = oneHot_lbl2, logits = self.label_orig45_D4 )
D4_classification_loss3 = tf.nn.softmax_cross_entropy_with_logits( labels = oneHot_lbl3, logits = self.label_orig90_D4 )
D4_classification_loss4 = tf.nn.softmax_cross_entropy_with_logits( labels = oneHot_lbl4, logits = self.label_orig135_D4 )
D4_classification_loss5 = tf.nn.softmax_cross_entropy_with_logits( labels = oneHot_lbl5, logits = self.label_origED_D4 )
self.D4_classification_loss = D4_classification_loss1 + D4_classification_loss2 + D4_classification_loss3 + D4_classification_loss4 + D4_classification_loss5
# The Least Squares Generative Adversarial Network, or LSGAN for short, is an extension to the GAN architecture that addresses the problem of vanishing gradients and loss saturation.
# Using LSGAN for D2 and D4 outputs:
# D_loss = 0.5 * tf.reduce_mean((D_real-1)^2) + tf.reduce_mean(D_fake^2)
# G_loss = 0.5 * tf.reduce_mean((D_fake -1)^2)
# D2 + D1
self.D2_RealFake_target = tf.math.reduce_mean( tf.math.squared_difference( self.RealFake_target_D2 , self.TARGET_LABELS )) + tf.math.reduce_mean( tf.math.square( self.RealFake_gen_D1 ) )
# D4 + D3
D4_1 = tf.math.reduce_mean( tf.math.squared_difference( RealFake_orig0_D4 , self.TARGET_LABELS )) + tf.math.reduce_mean( tf.math.square( RealFake_cyc_gen0_D3 ) )
D4_2 = tf.math.reduce_mean( tf.math.squared_difference( RealFake_orig45_D4 , self.TARGET_LABELS )) + tf.math.reduce_mean( tf.math.square( RealFake_cyc_gen45_D3 ) )
D4_3 = tf.math.reduce_mean( tf.math.squared_difference( RealFake_orig90_D4 , self.TARGET_LABELS )) + tf.math.reduce_mean( tf.math.square( RealFake_cyc_gen90_D3 ) )
D4_4 = tf.math.reduce_mean( tf.math.squared_difference( RealFake_orig135_D4 , self.TARGET_LABELS )) + tf.math.reduce_mean( tf.math.square( RealFake_cyc_gen135_D3 ) )
D4_5 = tf.math.reduce_mean( tf.math.squared_difference( RealFake_origED_D4 , self.TARGET_LABELS )) + tf.math.reduce_mean( tf.math.square( RealFake_cyc_genED_D3 ) )
self.D4_RealFake_cyc = D4_1 + D4_2 + D4_3 + D4_4 + D4_5 + self.D2_RealFake_target
(self.D4_RealFake_cyc)/6.0 + (self.D4_classification_loss)/5.0
# -------------------- L1 Loss --------------------
# # Adding L1 loss for generator to force image generation to look like the original images
# L1_loss = tf.keras.losses.MeanSquaredError()
# L1_loss_G1 = L1_loss(self.gen_rgb, self.target_img )
# L1_loss_G2_1 = L1_loss(self.cyc_gen0_rgb, orig0 )
# L1_loss_G2_2 = L1_loss(self.cyc_gen45_rgb, orig45 )
# L1_loss_G2_3 = L1_loss(self.cyc_gen90_rgb, orig90 )
# L1_loss_G2_4 = L1_loss(self.cyc_gen135_rgb, orig135 )
# L1_loss_G2_5 = L1_loss(self.cyc_genED_rgb, origED )
# The cycle consistency loss is defined as the sum of the L1 distances between the real images from each domain and their generated counterparts.
# Source: https://github.com/AlamiMejjati/Unsupervised-Attention-guided-Image-to-Image-Translation/blob/master/losses.py
L1_loss_G1 = tf.reduce_mean( tf.abs( self.gen_rgb - self.target_img ))
L1_loss_G2_1 = tf.reduce_mean( tf.abs( self.cyc_gen0_rgb - orig0 ))
L1_loss_G2_2 = tf.reduce_mean( tf.abs( self.cyc_gen45_rgb - orig45 ))
L1_loss_G2_3 = tf.reduce_mean( tf.abs( self.cyc_gen90_rgb - orig90 ))
L1_loss_G2_4 = tf.reduce_mean( tf.abs( self.cyc_gen135_rgb - orig135 ))
L1_loss_G2_ED = tf.reduce_mean( tf.abs( self.cyc_genED_rgb - origED ))
self.L1_loss_Gen = ( L1_loss_G2_1 + L1_loss_G2_2 + L1_loss_G2_3 + L1_loss_G2_4 + L1_loss_G1) / 5.0 + ( L1_loss_G2_ED ) * 10.0
# -------------------- MS-SSIM --------------------
## SSIM loss for generator
# NOTE: SSIM is for greyscale only so using Y channel only
# Using maxval= 5 from the collaGAN MIR paper
# SSIM of YUV images (NOT RGB)
ssim1 = tf.image.ssim ( rescale_01(cyc_gen0_yuv) , rescale_01( ds1_yuv ) , 5 )
ssim2 = tf.image.ssim ( rescale_01(cyc_gen45_yuv) , rescale_01( ds2_yuv ) , 5 )
ssim3 = tf.image.ssim ( rescale_01(cyc_gen90_yuv) , rescale_01( ds3_yuv ) , 5 )
ssim4 = tf.image.ssim ( rescale_01(cyc_gen135_yuv) , rescale_01( ds4_yuv ) , 5 )
ssim5 = tf.image.ssim ( rescale_01(cyc_genED_yuv) , rescale_01( ds5_yuv ) , 5 )
# # RESCALING TO [0,1] BEFORE CALCULATING SSIM
# power_factors=[(0.0448, 0.2856, 0.3001)] # from TF documentation
# # ssim1 = tf.image.ssim_multiscale ( img1 = rescale_01( cyc_gen0_yuv) , img2 = rescale_01(ds1_yuv), max_val=1, power_factors = power_factors )
# # ssim2 = tf.image.ssim_multiscale ( img1 = rescale_01( cyc_gen45_yuv) , img2 = rescale_01(ds2_yuv), max_val=1, power_factors = power_factors )
# # ssim3 = tf.image.ssim_multiscale ( img1 = rescale_01( cyc_gen90_yuv) , img2 = rescale_01(ds3_yuv), max_val=1, power_factors = power_factors )
# # ssim4 = tf.image.ssim_multiscale ( img1 = rescale_01( cyc_gen135_yuv) , img2 = rescale_01(ds4_yuv), max_val=1, power_factors = power_factors )
# # ssim5 = tf.image.ssim_multiscale ( img1 = rescale_01( cyc_genED_yuv) , img2 = rescale_01(ds5_yuv), max_val=1, power_factors = power_factors )
# # self.ssim_cyc_loss = (ssim1 + ssim2 + ssim3 + ssim4 + ssim5) / 5.0
cyc_ssim_loss1 = tf.cond( RNG1 , lambda:0.0, lambda: -tf.math.log( ( 1.0 + ssim1) /2.0 ))
cyc_ssim_loss2 = tf.cond( RNG2 , lambda:0.0, lambda: -tf.math.log( ( 1.0 + ssim2) /2.0 ))
cyc_ssim_loss3 = tf.cond( RNG3 , lambda:0.0, lambda: -tf.math.log( ( 1.0 + ssim3) /2.0 ))
cyc_ssim_loss4 = tf.cond( RNG4 , lambda:0.0, lambda: -tf.math.log( ( 1.0 + ssim4) /2.0 ))
cyc_ssim_loss5 = tf.cond( RNG5 , lambda:0.0, lambda: -tf.math.log( ( 1.0 + ssim5) /2.0 ))
self.ssim_cyc_loss = (cyc_ssim_loss1 + cyc_ssim_loss2 + cyc_ssim_loss3 + cyc_ssim_loss4 + cyc_ssim_loss5 * 10) / 5.0
# #
# # ssim_loss_orig = -tf.math.log( (1.0+ssim_recon)/2.0)
# # ssim_loss = (ssim_loss_orig + ssim_cyc_loss)/6.
# # SSIM+L1 loss based on Zhao et al (2018) https://arxiv.org/pdf/1511.08861.pdf
# # Calculated as : self.alpha * loss_MSSSIM + (1-self.alpha) * loss_L1
# self.MS_SSIML1_loss = self.alpha * self.ssim_cyc_loss + ( 1-self.alpha ) * self.L1_loss_Gen
# -------------------- SPECULAR --------------------
# Specular Loss - To foce ED generation
Spec_loss1 = tf.reduce_mean( tf.math.square ( (cyc_gen0_yuv * self.specular_candidate) - (ds1_yuv * self.specular_candidate ) ) )
Spec_loss2 = tf.reduce_mean( tf.math.square ( (cyc_gen45_yuv * self.specular_candidate) - (ds2_yuv * self.specular_candidate ) ) )
Spec_loss3 = tf.reduce_mean( tf.math.square ( (cyc_gen90_yuv * self.specular_candidate) - (ds3_yuv * self.specular_candidate ) ) )
Spec_loss4 = tf.reduce_mean( tf.math.square ( (cyc_gen135_yuv * self.specular_candidate) - (ds4_yuv * self.specular_candidate ) ) )
Spec_lossED = tf.reduce_mean( tf.math.square ( (cyc_genED_yuv * self.specular_candidate) - (ds5_yuv * self.specular_candidate ) ) )
## L2 Norm inpainting loss
## Loss = L2_norm(1-Mask x (Igen - Igt)) + L2_norm(Mask x (Igen-Igt))
# Spec_loss1 = tf.norm( ( 1-self.specular_candidate ) * (cyc_gen0 - ds1_yuv) , ord=2 ) + tf.norm( self.specular_candidate * (cyc_gen0 - ds1_yuv) , ord=2 )
# Spec_loss2 = tf.norm( ( 1-self.specular_candidate ) * (cyc_gen45 - ds2_yuv) , ord=2 ) + tf.norm( self.specular_candidate * (cyc_gen45 - ds2_yuv) , ord=2 )
# Spec_loss3 = tf.norm( ( 1-self.specular_candidate ) * (cyc_gen90 - ds3_yuv) , ord=2 ) + tf.norm( self.specular_candidate * (cyc_gen90 - ds3_yuv) , ord=2 )
# Spec_loss4 = tf.norm( ( 1-self.specular_candidate ) * (cyc_gen135 - ds4_yuv) , ord=2 ) + tf.norm( self.specular_candidate * (cyc_gen135 - ds4_yuv) , ord=2 )
# Spec_loss5 = tf.norm( ( 1-self.specular_candidate ) * (cyc_genED - ds5_yuv) , ord=2 ) + tf.norm( self.specular_candidate * (cyc_genED - ds5_yuv) , ord=2 )
self.Spec_loss = ( Spec_loss1 + Spec_loss2 + Spec_loss3 + Spec_loss4 ) / 5.0 + ( Spec_lossED ) * 5.0
# -------------------- StyleTx --------------------
# Gram Matrix = [HxW, ch] * [HxW, ch]'
# calculated after actiavtion of the layer
# What if we take STYLE = ED and CONTENT = I0 ?
# CONTENT LOSS: (Content == I0 == RGB input)
self.content_loss = tf.reduce_mean( tf.math.square( cyc_genED_yuv - ds1_yuv ) )
# STYLE LOSS: (Style == ED)
factor = tf.cast(1/tf.math.square(2 * 9 * self.image_size * self.image_size), tf.float32 )
gram_STYLE = self.gram_matrix(cyc_genED_yuv)
gram_CONTENT = self.gram_matrix(ds5_yuv)
self.style_loss = tf.math.multiply(factor, tf.reduce_mean( tf.math.square( gram_STYLE - gram_CONTENT )))
# TOTAL LOSS:
style_weight = 100
content_weight = 1
self.total_NST_loss = style_weight * self.style_loss + content_weight * self.content_loss
# -------------------- TOTAL LOSSES --------------------
self.total_Generator_loss = ( self.D1_RealFake_loss + self.D3_RealFake_cyc ) / 6.0 + \
( self.L1_loss_Gen ) * 10.0 + \
( self.ssim_cyc_loss ) * 10.0 + \
( self.total_NST_loss ) * 10.0
# ( self.Spec_loss ) * 5.0 + \
# ( self.MS_SSIML1_loss ) + \
self.total_Discriminator_loss = ( self.D1_classification_loss + self.D3_classification_loss ) / 6.0 + \
( self.D2_RealFake_target + self.D4_RealFake_cyc ) / 6.0 + \
( self.D4_classification_loss ) * 0.5 + \
( self.total_NST_loss ) * 10
# ( self.Spec_loss ) * 5.0 + \
self.total_Classification_loss = ( self.D4_classification_loss + self.total_NST_loss) * 10
"""
# ------------------------------------------------
# GRADIENT-TAPE FOR LEARNING PARAMETERS
# Calculate the gradients for generator and discriminator
# NOTE: Gradient tape are calculated after the tf.Gradientape() function ends
# ------------------------------------------------
"""
# Unfreeze the weights for applying gradients - Important!
self.G.trainable = True
self.D.trainable = True
# https://www.tensorflow.org/api_docs/python/tf/keras/mixed_precision/LossScaleOptimizer
scaled_gradient_Discriminator = tape.gradient( [ self.total_Discriminator_loss, self.total_Classification_loss ], self.D.trainable_variables)
scaled_gradient_Discriminator = [(tf.clip_by_value(grad, clip_value_max=1.0, clip_value_min=-1.0)) for grad in scaled_gradient_Discriminator]
# unscaled_gradient_Discriminator = self.optimizer_D.get_unscaled_gradients(scaled_gradient_Discriminator)
self.optimizer_D.apply_gradients( zip( scaled_gradient_Discriminator, self.D.trainable_variables ), experimental_aggregate_gradients=True )
self.gradmapD = scaled_gradient_Discriminator
if self.epoch >= self.train_G_after:
# self.G.trainable = True
# self.D.trainable = False
scaled_gradient_Generator = tape.gradient( [ self.total_Generator_loss], self.G.trainable_variables )
scaled_gradient_Generator = [(tf.clip_by_value(grad, clip_value_max=1.0, clip_value_min=-1.0)) for grad in scaled_gradient_Generator]
# unscaled_gradient_Generator = self.optimizer_G.get_unscaled_gradients(scaled_gradient_Generator)
self.optimizer_G.apply_gradients( zip( scaled_gradient_Generator, self.G.trainable_variables ), experimental_aggregate_gradients=True )
self.gradmapG = scaled_gradient_Generator
return
"""
# ------------------------------------------------
#
# ███ ███ █████ ██ ███ ██
# ████ ████ ██ ██ ██ ████ ██
# ██ ████ ██ ███████ ██ ██ ██ ██
# ██ ██ ██ ██ ██ ██ ██ ██ ██
# ██ ██ ██ ██ ██ ██ ████
# ------------------------------------------------
"""
# jit_compile=True
@tf.function(experimental_follow_type_hints=True, jit_compile=True)
def train( self, args ):
# Set to train on CPU
# os.environ["CUDA_VISIBLE_DEVICES"]="-1"
# trainTime = time.time()
start_time = default_timer()
file_writer = tf.summary.create_file_writer( self.log_dir )
tf.Graph.finalize( self._graph )
# ------------------------------------------------
# Initialize and load the zipped dataset
# NOTE: Images returned will be resized, Normalized (TBD) and randomly flipped (TBD)
self.length_dataset, Dataset = datasetLoad( self )
# NOTE: Batch size is 1 for zipped dataset, so that 1 image from each polar angle is picked
# batched_dataset = Dataset.batch( 4 )
# ------------------------------------------------
self.G = self.build_generator( )
self.D = self.build_discriminator( )
# Print Model summary to console and file
self.G.summary()
self.D.summary()
with open('Generator_summary.txt', 'w') as f:
self.G.summary(print_fn=lambda x: f.write(x + '\n'))
with open('Discriminator_summary.txt', 'w') as f:
self.D.summary(print_fn=lambda x: f.write(x + '\n'))
# plot the model
# plot_model( self.D, to_file = 'discriminator_plot.png', show_shapes = True, show_layer_names = True )
# plot_model( self.G, to_file = 'generator_plot.png', show_shapes = True, show_layer_names = True )
# self.G.save('G2.h5')
# ------------------------------------------------
# Importing the SpecSeg model. The model is initialized with M,N,ch with 3 channels
self.SpecSeg = SpecSeg(self.image_size, self.image_size, 1)
self.SpecSeg = tf.keras.models.load_model('specsegv3_chkpt.h5',compile=False)
self.SpecSeg.summary()
with open('SpecSeg_summary.txt', 'w') as f:
self.SpecSeg.summary(print_fn=lambda x: f.write(x + '\n'))
# ------------------------------------------------
# Initiialize the checkpoint manager
checkpoint_dir = self.checkpoint_save_dir
ckpt = tf.train.Checkpoint( generator = self.G,
discriminator = self.D,
optimizer_D = self.optimizer_D,
optimizer_G = self.optimizer_G )
ckpt_manager = tf.train.CheckpointManager(ckpt, checkpoint_dir, max_to_keep=3)
# if old checkpoints exist and flag is true; del them
if self.delete_old_checkpoints is True and ckpt_manager.latest_checkpoint:
os.system("echo CLEANUP: Removing previous checkpoints")
os.system("rm /home/atif/Documents/checkpoints/*")
elif self.delete_old_checkpoints is False:
ckpt.restore(ckpt_manager.latest_checkpoint)
print ('Latest checkpoint restored!!')
# ------------------------------------------------
# Make the iterator from the zipped datasets
iterator = iter(Dataset)
batches_per_epoch = int( self.length_dataset/self.batch_size )
self.batch_step = 0
plt.close("all")
gc.collect()
# self.comet_experiment.set_model_graph(self.D)
# self.comet_experiment.set_model_graph(self.G)
# initialize confustion matrix
# confusion_matrix = self.comet_experiment.create_confusion_matrix(labels=["0°", "45°", "90°", "135°", "ED"])
for epoch in range(self.num_epochs):
self.epoch = epoch
print(f"\nStart of Training Epoch {self.epoch}")
# since batch will be 1 (to get all polar images in sequence), the number of batches
# will have to be equal to the length of the dataset. Also note that the randomness and shuffle
# are not in the dataset while loading, but introduced by randomizing input channels
# and randomizing lables while training (tf.cond ....)
for batch in range( batches_per_epoch-1 ):
self.batch_step +=1
# Randomly flip all images in the batch
self.random_flip = tf.random.uniform( [], dtype=tf.float16 ) >= 0.5
# Randomly generate target labels for more robustness instead of a hard value of 1
self.TARGET_LABELS = tf.random.uniform( [], minval=0.8, maxval=1.2, dtype=tf.float32 )
# Get the next set of images
element = iterator.get_next()
orig0 = element[0]
orig45 = element[1]
orig90 = element[2]
orig135 = element[3]
origED = element[4]
self.train_step( orig0, orig45, orig90, orig135, origED )
with file_writer.as_default():