-
Notifications
You must be signed in to change notification settings - Fork 1
/
SM-SGE.py
1304 lines (1181 loc) · 57.2 KB
/
SM-SGE.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
import time
import numpy as np
import tensorflow as tf
import os, sys
from models import MGRN_S
from utils import process_i as process
from tensorflow.python.layers.core import Dense
from sklearn.preprocessing import label_binarize
dataset = 'IAS'
split = 'B'
pretext = 'recon'
RN_dir = 'RN/' # save MLP models for person Re-ID
pre_dir = 'Pre-Trained/' # save self-supervised SM-SGE models
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
pre_epochs = 200 # epochs for self-supervised training
sample_num = 1 # number of sampling rounds (r)
nb_nodes = 20 # number of nodes in joint-scale graph
nhood = 1 # structural relation learning (nhood=1 for neighbor nodes)
fusion_lambda = 1 # collaboration fusion coefficient
ft_size = 3 # originial node feature dimension (D)
time_step = 6 # sequence length (f)
LSTM_embed = 256 # number of hidden units per layer in LSTM (D_{h})
num_layers = 2 # number of LSTM layers
# training params
batch_size = 256
nb_epochs = 100000
patience = 150 # patience for early stopping
lr = 0.005 # learning rate
hid_units = [8] # numbers of hidden units per each attention head in each layer
Ps = [8, 1] # additional entry for the output layer
residual = False
nonlinearity = tf.nn.elu
model = MGRN_S
tf.app.flags.DEFINE_string('dataset', 'BIWI', "Dataset: BIWI, IAS, KS20 or KGBD")
tf.app.flags.DEFINE_string('length', '6', "4, 6, 8 or 10")
tf.app.flags.DEFINE_string('split', '', "for IAS-Lab testing splits")
tf.app.flags.DEFINE_string('gpu', '0', "GPU number")
tf.app.flags.DEFINE_string('task', 'pre', "prediction") # we regard reconstruction as a prediction with 0 pred_margin
tf.app.flags.DEFINE_string('frozen', '1', "Frozen LSTM states") # frozen encoded graph states for person Re-ID
tf.app.flags.DEFINE_string('pre_epochs', '200', "epochs for pre-training")
tf.app.flags.DEFINE_string('pre_train', '1', "pre-train or not")
tf.app.flags.DEFINE_string('s_range', 'all', "1, f/2, f-1, all")
tf.app.flags.DEFINE_string('pred_margin', '0', "0, 1, 2")
tf.app.flags.DEFINE_string('view', '', "test different views on CASIA B")
tf.app.flags.DEFINE_string('reverse', '0', "use reverse sequences")
tf.app.flags.DEFINE_string('bi', '0', "bi-directional prediction")
tf.app.flags.DEFINE_string('save_flag', '1', "save model or not")
tf.app.flags.DEFINE_string('consecutive_pre', '0', "consecutive_pre")
tf.app.flags.DEFINE_string('single_level', '0', "single_level")
tf.app.flags.DEFINE_string('global_att', '0', "global_att")
tf.app.flags.DEFINE_string('last_pre', '0', "last_pre")
tf.app.flags.DEFINE_string('random_sample', '0', "random_sample")
tf.app.flags.DEFINE_string('ord_sample', '0', "ord_sample")
tf.app.flags.DEFINE_string('concate', '0', "concate")
tf.app.flags.DEFINE_string('struct_only', '0', "struct_only")
tf.app.flags.DEFINE_string('abla', '0', "abla")
tf.app.flags.DEFINE_string('P', '8', "P")
tf.app.flags.DEFINE_string('n_hood', '1', "n_hood")
tf.app.flags.DEFINE_string('probe_type', '', "probe.gallery")
tf.app.flags.DEFINE_string('no_MSR', '0', "using MSR or not")
tf.app.flags.DEFINE_string('patience', '100', "epochs for early stopping")
tf.app.flags.DEFINE_string('sample_num', '1', "sampling times")
tf.app.flags.DEFINE_string('fusion_lambda', '1', "collaboration fusion coefficient")
tf.app.flags.DEFINE_string('loss', 'l1', "use l1, l2 or MSE loss")
FLAGS = tf.app.flags.FLAGS
# check parameters
if FLAGS.dataset not in ['IAS', 'KGBD', 'KS20']:
raise Exception('Dataset must be IAS, KGBD, or KS20.')
if not FLAGS.gpu.isdigit() or int(FLAGS.gpu) < 0:
raise Exception('GPU number must be a positive integer.')
if FLAGS.length not in ['4', '6', '8', '10']:
raise Exception('Length number must be 4, 6, 8 or 10.')
if FLAGS.split not in ['', 'A', 'B']:
raise Exception('Datset split must be "A" (for IAS-A), "B" (for IAS-B), "" (for other datasets).')
if float(FLAGS.fusion_lambda) < 0 or float(FLAGS.fusion_lambda) > 1:
raise Exception('Collaboration Fusion coefficient must be not less than 0 or not larger than 1.')
if FLAGS.pre_train not in ['1', '0']:
raise Exception('Pre-train Flag must be 0 or 1.')
if FLAGS.save_flag not in ['1', '0']:
raise Exception('Save_flag must be 0 or 1.')
os.environ["CUDA_VISIBLE_DEVICES"] = FLAGS.gpu
dataset = FLAGS.dataset
# optimal paramters
if dataset == 'KS20':
batch_size = 64
LSTM_embed = 256
lr = 0.0005
elif dataset == 'IAS' and split == 'A':
batch_size = 128
LSTM_embed = 256
lr = 0.0025
elif dataset == 'IAS' and split == 'B':
batch_size = 64
LSTM_embed = 256
lr = 0.0025
elif dataset == 'CASIA_B':
batch_size = 128
LSTM_embed = 256
lr = 0.0005
# patience = 50
if int(FLAGS.pre_epochs) != 200:
pre_epochs = int(FLAGS.pre_epochs)
elif dataset == 'KGBD':
batch_size = 256
LSTM_embed = 128
lr = 0.0025
pre_epochs = 160
if dataset == 'IAS' or dataset == 'BIWI':
pre_epochs = 120
if dataset == 'CASIA_B':
pre_epochs = 160
if FLAGS.probe_type != '':
pre_epochs = 200
time_step = int(FLAGS.length)
k_values = list(range(1, time_step - int(FLAGS.pred_margin) + 1))
k_filter = -1
if FLAGS.s_range == '1':
k_filter = 1
elif FLAGS.s_range == 'f/2':
k_filter = int(time_step//2)
elif FLAGS.s_range == 'f-1':
k_filter = time_step-1
elif FLAGS.s_range == 'f':
k_filter = time_step
fusion_lambda = float(FLAGS.fusion_lambda)
split = FLAGS.split
pretext = FLAGS.task
levels = len(k_values)
frozen= FLAGS.frozen
nhood = int(FLAGS.n_hood)
pred_margin = FLAGS.pred_margin
s_range = FLAGS.s_range
save_flag = FLAGS.save_flag
patience = int(FLAGS.patience)
sample_num = int(FLAGS.sample_num)
consecutive_pre = False
single_level = False
global_att = False
MG_only = False
last_pre = False
ord_sample = False
concate = False
struct_only = False
abla = False
P = '8'
only_ske_embed = False
three_layer = False
embed_half = False
one_layer = False
CAGEs_embed = False
no_interp = False
no_multi_pred = False
change = '_formal'
if FLAGS.loss != 'l1':
change += '_loss_' + FLAGS.loss
if sample_num != 1:
change = '_sample_num_' + str(sample_num)
if FLAGS.probe_type != '':
change = '_CME'
if FLAGS.fusion_lambda != '1':
change = '_lambda_' + FLAGS.fusion_lambda
if FLAGS.consecutive_pre == '1':
consecutive_pre = True
if FLAGS.single_level == '1':
single_level = True
if FLAGS.global_att == '1':
global_att = True
if FLAGS.last_pre == '1':
last_pre = True
if FLAGS.ord_sample == '1':
ord_sample = True
if FLAGS.concate == '1':
concate = True
if FLAGS.struct_only == '1':
struct_only = True
if FLAGS.P != '8':
P = FLAGS.P
Ps = [int(P), 1]
try:
os.mkdir(RN_dir)
except:
pass
try:
os.mkdir(pre_dir)
except:
pass
if consecutive_pre:
pre_dir += '_consec_pre'
RN_dir += '_consec_pre'
if single_level:
fusion_lambda = 0
pre_dir += '_single'
RN_dir += '_single'
if global_att:
pre_dir += '_global_att'
RN_dir += '_global_att'
if MG_only:
pre_dir += '_MG_only'
RN_dir += '_MG_only'
if last_pre:
pre_dir += '_last_pre'
RN_dir += '_last_pre'
if ord_sample:
pre_dir += '_ord_sample'
RN_dir += '_ord_sample'
if concate:
pre_dir += '_concate'
RN_dir += '_concate'
if frozen == '1':
pre_dir += '_frozen'
RN_dir += '_frozen'
if struct_only:
pre_dir += '_struct_only'
RN_dir += '_struct_only'
if P != '8':
pre_dir += '_P_' + P
RN_dir += '_P_' + P
if pretext == 'none':
pre_epochs = 0
I_nodes = 39
if dataset == 'KS20':
nb_nodes = 25
I_nodes = 49
if dataset == 'CASIA_B':
nb_nodes = 14
I_nodes = 27
if FLAGS.view != '':
view_dir = '_view_' + FLAGS.view
else:
view_dir = ''
print('Dataset: ' + dataset)
print('----- Opt. hyperparams -----')
print('pre_train_epochs: ' + str(pre_epochs))
print('nhood: ' + str(nhood))
print('skeleton_nodes: ' + str(nb_nodes))
print('seqence_length: ' + str(time_step))
print('pretext: ' + str(pretext))
print('fusion_lambda: ' + str(fusion_lambda))
print('batch_size: ' + str(batch_size))
print('lr: ' + str(lr))
print('view: ' + FLAGS.view)
print('P: ' + FLAGS.P)
print('fusion_lambda: ' + FLAGS.fusion_lambda)
print('loss_type: ' + FLAGS.loss)
print('patience: ' + FLAGS.patience)
print('save_flag: ' + FLAGS.save_flag)
print('----- Archi. hyperparams -----')
print('structural relation matrix number: ' + str(Ps[0]))
print('LSTM_embed_num: ' + str(LSTM_embed))
print('LSTM_layer_num: ' + str(num_layers))
"""
Obtain training and testing data in hyper-joint-scale, joint-scale, part-scale, and body-scale.
Generate corresponding adjacent matrix and bias.
"""
if FLAGS.probe_type == '':
X_train_J, X_train_P, X_train_B, X_train_I, y_train, X_test_J, X_test_P, X_test_B, X_test_I, y_test, \
adj_J, biases_J, adj_P, biases_P, adj_B, biases_B, adj_I, biases_I, nb_classes = \
process.gen_train_data(dataset=dataset, split=split, time_step=time_step,
nb_nodes=nb_nodes, nhood=nhood, global_att=global_att, batch_size=batch_size, view=FLAGS.view, reverse=FLAGS.reverse)
else:
from utils import process_cme as process
X_train_J, X_train_P, X_train_B, X_train_I, y_train, X_test_J, X_test_P, X_test_B, X_test_I, y_test, \
adj_J, biases_J, adj_P, biases_P, adj_B, biases_B, adj_I, biases_I, nb_classes = \
process.gen_train_data(dataset=dataset, split=split, time_step=time_step,
nb_nodes=nb_nodes, nhood=nhood, global_att=global_att, batch_size=batch_size,
reverse=FLAGS.reverse, PG_type=FLAGS.probe_type.split('.')[0])
print('## [Probe].[Gallery]', FLAGS.probe_type)
if FLAGS.pre_train == '1':
with tf.Graph().as_default():
with tf.name_scope('Input'):
lbl_in = tf.placeholder(dtype=tf.int32, shape=(batch_size, nb_classes))
J_in = tf.placeholder(dtype=tf.float32, shape=(batch_size*time_step, nb_nodes, ft_size))
P_in = tf.placeholder(dtype=tf.float32, shape=(batch_size * time_step, 10, ft_size))
B_in = tf.placeholder(dtype=tf.float32, shape=(batch_size * time_step, 5, ft_size))
# Interpolation
I_in = tf.placeholder(dtype=tf.float32, shape=(batch_size * time_step, I_nodes, ft_size))
J_bias_in = tf.placeholder(dtype=tf.float32, shape=(1, nb_nodes, nb_nodes))
P_bias_in = tf.placeholder(dtype=tf.float32, shape=(1, 10, 10))
B_bias_in = tf.placeholder(dtype=tf.float32, shape=(1, 5, 5))
I_bias_in = tf.placeholder(dtype=tf.float32, shape=(1, I_nodes, I_nodes))
attn_drop = tf.placeholder(dtype=tf.float32, shape=())
ffd_drop = tf.placeholder(dtype=tf.float32, shape=())
is_train = tf.placeholder(dtype=tf.bool, shape=())
with tf.name_scope("Multi_Scale"), tf.variable_scope("Multi_Scale", reuse=tf.AUTO_REUSE):
def SRL(J_in, J_bias_in, nb_nodes):
W_h = tf.Variable(tf.random_normal([3, hid_units[-1]]))
b_h = tf.Variable(tf.zeros(shape=[hid_units[-1], ]))
J_h = tf.reshape(J_in, [-1, ft_size])
J_h = tf.matmul(J_h, W_h) + b_h
J_h = tf.reshape(J_h, [batch_size*time_step, nb_nodes, hid_units[-1]])
if not concate:
J_seq_ftr = model.inference(J_h, 0, nb_nodes, is_train,
attn_drop, ffd_drop,
bias_mat=J_bias_in,
hid_units=hid_units, n_heads=Ps,
residual=residual, activation=nonlinearity, r_pool=True)
else:
J_seq_ftr = model.inference(J_h, 0, nb_nodes, is_train,
attn_drop, ffd_drop,
bias_mat=J_bias_in,
hid_units=hid_units, n_heads=Ps,
residual=residual, activation=nonlinearity, r_pool=False)
return J_seq_ftr
def CRL(s1, s2, s1_num, s2_num, hid_in):
r_unorm = tf.matmul(s2, tf.transpose(s1, [0, 2, 1]))
att_w = tf.nn.softmax(r_unorm)
att_w = tf.expand_dims(att_w, axis=-1)
s1 = tf.reshape(s1, [s1.shape[0], 1, s1.shape[1], hid_in])
c_ftr = tf.reduce_sum(att_w * s1, axis=2)
c_ftr = tf.reshape(c_ftr, [-1, hid_in])
att_w = tf.reshape(att_w, [-1, s1_num * s2_num])
return r_unorm, c_ftr
def MGRN(J_in, P_in, B_in, I_in, J_bias_in, P_bias_in, B_bias_in, I_bias_in, hid_in, hid_out):
h_J_seq_ftr = SRL(J_in=J_in, J_bias_in=J_bias_in, nb_nodes=nb_nodes)
h_P_seq_ftr = SRL(J_in=P_in, J_bias_in=P_bias_in, nb_nodes=10)
h_B_seq_ftr = SRL(J_in=B_in, J_bias_in=B_bias_in, nb_nodes=5)
h_I_seq_ftr = SRL(J_in=I_in, J_bias_in=I_bias_in, nb_nodes=I_nodes)
h_J_seq_ftr = tf.reshape(h_J_seq_ftr, [-1, nb_nodes, hid_in])
h_P_seq_ftr = tf.reshape(h_P_seq_ftr, [-1, 10, hid_in])
h_B_seq_ftr = tf.reshape(h_B_seq_ftr, [-1, 5, hid_in])
h_I_seq_ftr = tf.reshape(h_I_seq_ftr, [-1, I_nodes, hid_in])
W_cs_12 = tf.Variable(tf.random_normal([hid_in, hid_out]))
W_cs_23 = tf.Variable(tf.random_normal([hid_in, hid_out]))
W_cs_13 = tf.Variable(tf.random_normal([hid_in, hid_out]))
W_cs_I = tf.Variable(tf.random_normal([hid_in, hid_out]))
W_cs_01 = tf.Variable(tf.random_normal([hid_in, hid_out]))
W_cs_02 = tf.Variable(tf.random_normal([hid_in, hid_out]))
W_cs_03 = tf.Variable(tf.random_normal([hid_in, hid_out]))
W_self_0 = tf.Variable(tf.random_normal([hid_in, hid_out]))
W_self_1 = tf.Variable(tf.random_normal([hid_in, hid_out]))
W_self_2 = tf.Variable(tf.random_normal([hid_in, hid_out]))
W_self_3 = tf.Variable(tf.random_normal([hid_in, hid_out]))
self_a_0, self_r_0 = CRL(h_I_seq_ftr, h_I_seq_ftr, I_nodes, I_nodes, hid_in)
self_a_1, self_r_1 = CRL(h_J_seq_ftr, h_J_seq_ftr, nb_nodes, nb_nodes, hid_in)
self_a_2, self_r_2 = CRL(h_P_seq_ftr, h_P_seq_ftr, 10, 10, hid_in)
self_a_3, self_r_3 = CRL(h_B_seq_ftr, h_B_seq_ftr, 5, 5, hid_in)
h_J_seq_ftr = tf.reshape(h_J_seq_ftr, [-1, hid_in])
h_P_seq_ftr = tf.reshape(h_P_seq_ftr, [-1, hid_in])
h_B_seq_ftr = tf.reshape(h_B_seq_ftr, [-1, hid_in])
h_I_seq_ftr = tf.reshape(h_I_seq_ftr, [-1, hid_in])
h_J_seq_ftr = tf.reshape(h_J_seq_ftr, [-1, nb_nodes, hid_in])
h_P_seq_ftr = tf.reshape(h_P_seq_ftr, [-1, 10, hid_in])
h_B_seq_ftr = tf.reshape(h_B_seq_ftr, [-1, 5, hid_in])
h_I_seq_ftr = tf.reshape(h_I_seq_ftr, [-1, I_nodes, hid_in])
a_12, r_12 = CRL(h_P_seq_ftr, h_J_seq_ftr, 10, nb_nodes, hid_in)
a_13, r_13 = CRL(h_B_seq_ftr, h_J_seq_ftr, 5, nb_nodes, hid_in)
a_01, r_01 = CRL(h_J_seq_ftr, h_I_seq_ftr, nb_nodes, I_nodes, hid_in)
a_02, r_02 = CRL(h_P_seq_ftr, h_I_seq_ftr, 10, I_nodes, hid_in)
a_03, r_03 = CRL(h_B_seq_ftr, h_I_seq_ftr, 5, I_nodes, hid_in)
a_23, r_23 = CRL(h_B_seq_ftr, h_P_seq_ftr, 5, 10, hid_in)
h_J_seq_ftr = tf.reshape(h_J_seq_ftr, [-1, hid_in])
h_P_seq_ftr = tf.reshape(h_P_seq_ftr, [-1, hid_in])
h_B_seq_ftr = tf.reshape(h_B_seq_ftr, [-1, hid_in])
h_I_seq_ftr = tf.reshape(h_I_seq_ftr, [-1, hid_in])
if not struct_only:
h_J_seq_ftr = h_J_seq_ftr + float(FLAGS.fusion_lambda) * (tf.matmul(self_r_1, W_self_1) + tf.matmul(r_12, W_cs_12) + tf.matmul(r_13, W_cs_13))
h_I_seq_ftr = h_I_seq_ftr + float(FLAGS.fusion_lambda) * (tf.matmul(self_r_0, W_self_0) + tf.matmul(r_01, W_cs_01) + tf.matmul(r_02, W_cs_02) + tf.matmul(r_03, W_cs_03))
h_P_seq_ftr = h_P_seq_ftr + float(FLAGS.fusion_lambda) * (tf.matmul(self_r_2, W_self_2) + tf.matmul(r_23, W_cs_23))
h_B_seq_ftr = h_B_seq_ftr + float(FLAGS.fusion_lambda) * (tf.matmul(self_r_3, W_self_3))
h_J_seq_ftr = tf.reshape(h_J_seq_ftr, [-1, nb_nodes, hid_out])
h_P_seq_ftr = tf.reshape(h_P_seq_ftr, [-1, 10, hid_out])
h_B_seq_ftr = tf.reshape(h_B_seq_ftr, [-1, 5, hid_out])
h_I_seq_ftr = tf.reshape(h_I_seq_ftr, [-1, I_nodes, hid_out])
return h_B_seq_ftr, h_P_seq_ftr, h_J_seq_ftr, h_I_seq_ftr
if not concate:
h_B_seq_ftr, h_P_seq_ftr, h_J_seq_ftr, h_I_seq_ftr = MGRN(J_in, P_in, B_in,
I_in, J_bias_in,
P_bias_in,
B_bias_in,
I_bias_in,
hid_units[-1],
hid_units[-1])
else:
h_B_seq_ftr, h_P_seq_ftr, h_J_seq_ftr = MSGAM(J_in, P_in, B_in, J_bias_in, P_bias_in, B_bias_in,
hid_units[-1] * Ps[0], hid_units[-1] * Ps[0])
h_J_seq_ftr = tf.reshape(h_J_seq_ftr, [-1, hid_units[-1]])
h_P_seq_ftr = tf.reshape(h_P_seq_ftr, [-1, hid_units[-1]])
h_B_seq_ftr = tf.reshape(h_B_seq_ftr, [-1, hid_units[-1]])
h_I_seq_ftr = tf.reshape(h_I_seq_ftr, [-1, hid_units[-1]])
ftr_in = J_in
P_ftr_in = P_in
B_ftr_in = B_in
I_ftr_in = I_in
J_seq_ftr = tf.reshape(h_J_seq_ftr, [batch_size, time_step, -1])
P_seq_ftr = tf.reshape(h_P_seq_ftr, [batch_size, time_step, -1])
B_seq_ftr = tf.reshape(h_B_seq_ftr, [batch_size, time_step, -1])
I_seq_ftr = tf.reshape(h_I_seq_ftr, [batch_size, time_step, -1])
J_T_m_ftr = []
J_Pred_tar = []
J_Test_ftr = []
J_skes_in = tf.reshape(ftr_in, [batch_size, time_step, -1])
# part-scale
P_T_m_ftr = []
P_Pred_tar = []
P_Test_ftr = []
P_skes_in = tf.reshape(P_ftr_in, [batch_size, time_step, -1])
# body-scale
B_T_m_ftr = []
B_Pred_tar = []
B_Test_ftr = []
B_skes_in = tf.reshape(B_ftr_in, [batch_size, time_step, -1])
# interpolation
I_T_m_ftr = []
I_Pred_tar = []
I_Test_ftr = []
I_skes_in = tf.reshape(I_ftr_in, [batch_size, time_step, -1])
print_flag = False
for i in range(sample_num):
for k in range(1, len(k_values)+2):
seq_ind = np.arange(batch_size).reshape(-1, 1)
seq_ind = np.tile(seq_ind, [1, k]).reshape(-1, 1)
if k <= time_step - int(pred_margin):
# in order
if ord_sample:
T_m = np.arange(k).reshape(-1, 1)
else:
T_m = np.random.choice(time_step - int(pred_margin), size=[k], replace=False).reshape(-1, 1)
T_m = np.sort(T_m, axis=0)
if pretext == 'pre':
Pred_t = T_m + int(pred_margin)
T_m = T_m.astype(dtype=np.int32)
Pred_t = Pred_t.astype(dtype=np.int32)
T_m = np.tile(T_m, [batch_size, 1]).reshape(-1, 1)
# T_m_tar.append(T_m)
Pred_t = np.tile(Pred_t, [batch_size, 1]).reshape(-1, 1)
T_m = np.hstack([seq_ind, T_m])
Pred_t = np.hstack([seq_ind, Pred_t])
J_sampled_seq_ftr = tf.gather_nd(J_seq_ftr, T_m)
# P, B
P_sampled_seq_ftr = tf.gather_nd(P_seq_ftr, T_m)
B_sampled_seq_ftr = tf.gather_nd(B_seq_ftr, T_m)
I_sampled_seq_ftr = tf.gather_nd(I_seq_ftr, T_m)
J_Pred_t_seq = tf.gather_nd(J_skes_in, Pred_t)
J_sampled_seq_ftr = tf.reshape(J_sampled_seq_ftr, [batch_size, k, -1])
J_Pred_t_seq = tf.reshape(J_Pred_t_seq, [batch_size, k, -1])
# P,B
P_Pred_t_seq = tf.gather_nd(P_skes_in, Pred_t)
P_sampled_seq_ftr = tf.reshape(P_sampled_seq_ftr, [batch_size, k, -1])
P_Pred_t_seq = tf.reshape(P_Pred_t_seq, [batch_size, k, -1])
B_Pred_t_seq = tf.gather_nd(B_skes_in, Pred_t)
B_sampled_seq_ftr = tf.reshape(B_sampled_seq_ftr, [batch_size, k, -1])
B_Pred_t_seq = tf.reshape(B_Pred_t_seq, [batch_size, k, -1])
I_Pred_t_seq = tf.gather_nd(I_skes_in, Pred_t)
I_sampled_seq_ftr = tf.reshape(I_sampled_seq_ftr, [batch_size, k, -1])
I_Pred_t_seq = tf.reshape(I_Pred_t_seq, [batch_size, k, -1])
J_T_m_ftr.append(J_sampled_seq_ftr)
J_Pred_tar.append(J_Pred_t_seq)
#
P_T_m_ftr.append(P_sampled_seq_ftr)
P_Pred_tar.append(P_Pred_t_seq)
#
B_T_m_ftr.append(B_sampled_seq_ftr)
B_Pred_tar.append(B_Pred_t_seq)
I_T_m_ftr.append(I_sampled_seq_ftr)
I_Pred_tar.append(I_Pred_t_seq)
if i == 0:
T_m_test = np.arange(k).reshape(-1, 1)
# print(T_m_test)
T_m_test = T_m_test.astype(dtype=np.int32)
T_m_test = np.tile(T_m_test, [batch_size, 1]).reshape(-1, 1)
T_m_test = np.hstack([seq_ind, T_m_test])
J_test_seq_ftr = tf.gather_nd(J_seq_ftr, T_m_test)
J_test_seq_ftr = tf.reshape(J_test_seq_ftr, [batch_size, k, -1])
J_Test_ftr.append(J_test_seq_ftr)
#
P_test_seq_ftr = tf.gather_nd(P_seq_ftr, T_m_test)
P_test_seq_ftr = tf.reshape(P_test_seq_ftr, [batch_size, k, -1])
P_Test_ftr.append(P_test_seq_ftr)
B_test_seq_ftr = tf.gather_nd(B_seq_ftr, T_m_test)
B_test_seq_ftr = tf.reshape(B_test_seq_ftr, [batch_size, k, -1])
B_Test_ftr.append(B_test_seq_ftr)
I_test_seq_ftr = tf.gather_nd(I_seq_ftr, T_m_test)
I_test_seq_ftr = tf.reshape(I_test_seq_ftr, [batch_size, k, -1])
I_Test_ftr.append(I_test_seq_ftr)
if FLAGS.bi == '1':
for i in range(sample_num):
for k in range(1, len(k_values) + 2):
seq_ind = np.arange(batch_size).reshape(-1, 1)
seq_ind = np.tile(seq_ind, [1, k]).reshape(-1, 1)
if k <= time_step - int(pred_margin):
# in order
if ord_sample:
T_m = np.arange(k).reshape(-1, 1)
else:
T_m = np.random.choice(time_step - int(pred_margin), size=[k], replace=False).reshape(-1, 1)
# print (T_m)
# print(T_m.shape)
# if not random_sample:
T_m = np.sort(T_m, axis=0)
# Reverse
T_m = np.sort(-T_m)
T_m += time_step
if pretext == 'pre':
Pred_t = T_m - int(pred_margin)
# no used
# elif pretext == 'recon':
# Pred_t = T_m
# elif pretext == 'rev':
# Pred_t = np.sort(-T_m)
# Pred_t = -Pred_t
print(T_m)
print(Pred_t)
T_m = T_m.astype(dtype=np.int32)
Pred_t = Pred_t.astype(dtype=np.int32)
T_m = np.tile(T_m, [batch_size]).reshape(-1, 1)
# T_m_tar.append(T_m)
Pred_t = np.tile(Pred_t, [batch_size]).reshape(-1, 1)
T_m = np.hstack([seq_ind, T_m])
Pred_t = np.hstack([seq_ind, Pred_t])
J_sampled_seq_ftr = tf.gather_nd(J_seq_ftr, T_m)
# P, B
P_sampled_seq_ftr = tf.gather_nd(P_seq_ftr, T_m)
B_sampled_seq_ftr = tf.gather_nd(B_seq_ftr, T_m)
I_sampled_seq_ftr = tf.gather_nd(I_seq_ftr, T_m)
#
J_Pred_t_seq = tf.gather_nd(J_skes_in, Pred_t)
J_sampled_seq_ftr = tf.reshape(J_sampled_seq_ftr, [batch_size, k, -1])
J_Pred_t_seq = tf.reshape(J_Pred_t_seq, [batch_size, k, -1])
# P,B
P_Pred_t_seq = tf.gather_nd(P_skes_in, Pred_t)
P_sampled_seq_ftr = tf.reshape(P_sampled_seq_ftr, [batch_size, k, -1])
P_Pred_t_seq = tf.reshape(P_Pred_t_seq, [batch_size, k, -1])
B_Pred_t_seq = tf.gather_nd(B_skes_in, Pred_t)
B_sampled_seq_ftr = tf.reshape(B_sampled_seq_ftr, [batch_size, k, -1])
B_Pred_t_seq = tf.reshape(B_Pred_t_seq, [batch_size, k, -1])
I_Pred_t_seq = tf.gather_nd(I_skes_in, Pred_t)
I_sampled_seq_ftr = tf.reshape(I_sampled_seq_ftr, [batch_size, k, -1])
I_Pred_t_seq = tf.reshape(I_Pred_t_seq, [batch_size, k, -1])
#
# sorted random frames
J_T_m_ftr.append(J_sampled_seq_ftr)
J_Pred_tar.append(J_Pred_t_seq)
#
P_T_m_ftr.append(P_sampled_seq_ftr)
P_Pred_tar.append(P_Pred_t_seq)
#
B_T_m_ftr.append(B_sampled_seq_ftr)
B_Pred_tar.append(B_Pred_t_seq)
I_T_m_ftr.append(I_sampled_seq_ftr)
I_Pred_tar.append(I_Pred_t_seq)
if FLAGS.bi == '1':
sample_num = sample_num * 2
with tf.name_scope("MSR"), tf.variable_scope("MSR", reuse=tf.AUTO_REUSE):
J_cell = tf.contrib.rnn.MultiRNNCell([tf.contrib.rnn.LSTMCell(LSTM_embed) for _ in range(num_layers)])
P_cell = tf.contrib.rnn.MultiRNNCell([tf.contrib.rnn.LSTMCell(LSTM_embed) for _ in range(num_layers)])
B_cell = tf.contrib.rnn.MultiRNNCell([tf.contrib.rnn.LSTMCell(LSTM_embed) for _ in range(num_layers)])
I_cell = tf.contrib.rnn.MultiRNNCell([tf.contrib.rnn.LSTMCell(LSTM_embed) for _ in range(num_layers)])
J_all_pred_loss = []
# P, B
P_all_pred_loss = []
B_all_pred_loss = []
I_all_pred_loss = []
#
a0_all_pred_loss = []
a1_all_pred_loss = []
a2_all_pred_loss = []
# all_mask_loss = []
J_en_outs = []
J_en_outs_whole = []
J_en_outs_test = []
#
P_en_outs = []
P_en_outs_whole = []
P_en_outs_test = []
B_en_outs = []
B_en_outs_whole = []
B_en_outs_test = []
I_en_outs = []
I_en_outs_whole = []
I_en_outs_test = []
#
ske_en_outs = []
# all_mask_acc = []
with tf.name_scope("J_pred"), tf.variable_scope("J_pred", reuse=tf.AUTO_REUSE):
J_W1_pred = tf.Variable(tf.random_normal([LSTM_embed, LSTM_embed]))
J_b1_pred = tf.Variable(tf.zeros(shape=[LSTM_embed, ]))
J_W2_pred = tf.Variable(tf.random_normal([LSTM_embed, nb_nodes * 3]))
J_b2_pred = tf.Variable(tf.zeros(shape=[nb_nodes * 3, ]))
P_W1_pred = tf.Variable(tf.random_normal([LSTM_embed, LSTM_embed]))
P_b1_pred = tf.Variable(tf.zeros(shape=[LSTM_embed, ]))
P_W2_pred = tf.Variable(tf.random_normal([LSTM_embed, 10 * 3]))
P_b2_pred = tf.Variable(tf.zeros(shape=[10 * 3, ]))
B_W1_pred = tf.Variable(tf.random_normal([LSTM_embed, LSTM_embed]))
B_b1_pred = tf.Variable(tf.zeros(shape=[LSTM_embed, ]))
B_W2_pred = tf.Variable(tf.random_normal([LSTM_embed, 5 * 3]))
B_b2_pred = tf.Variable(tf.zeros(shape=[5 * 3, ]))
for i in range((levels - int(pred_margin)) * sample_num):
J_sampled_seq_ftr = J_T_m_ftr[i]
J_encoder_output, J_encoder_state = tf.nn.dynamic_rnn(J_cell, J_sampled_seq_ftr, dtype=tf.float32)
J_encoder_output = tf.reshape(J_encoder_output, [-1, LSTM_embed])
# Skeleton-level Prediction
J_pred_embedding_J = tf.nn.relu(tf.matmul(J_encoder_output, J_W1_pred) + J_b1_pred)
#
J_pred_skeleton = tf.matmul(J_pred_embedding_J, J_W2_pred) + J_b2_pred
J_pred_skeleton = tf.reshape(J_pred_skeleton, [batch_size, k_values[i % levels], nb_nodes * 3])
# P, B
P_pred_embedding_J = tf.nn.relu(tf.matmul(J_encoder_output, P_W1_pred) + P_b1_pred)
P_pred_skeleton = tf.matmul(P_pred_embedding_J, P_W2_pred) + P_b2_pred
P_pred_skeleton = tf.reshape(P_pred_skeleton, [batch_size, k_values[i % levels], 10 * 3])
B_pred_embedding_J = tf.nn.relu(tf.matmul(J_encoder_output, B_W1_pred) + B_b1_pred)
B_pred_skeleton = tf.matmul(B_pred_embedding_J, B_W2_pred) + B_b2_pred
B_pred_skeleton = tf.reshape(B_pred_skeleton, [batch_size, k_values[i % levels], 5 * 3])
if FLAGS.loss == 'l1':
J_pred_loss = tf.reduce_mean(tf.losses.absolute_difference(J_pred_skeleton, J_Pred_tar[i]))
# P, B
# tf.nn.l1_loss
P_pred_loss = tf.reduce_mean(tf.losses.absolute_difference(P_pred_skeleton, P_Pred_tar[i]))
B_pred_loss = tf.reduce_mean(tf.losses.absolute_difference(B_pred_skeleton, B_Pred_tar[i]))
elif FLAGS.loss == 'MSE':
J_pred_loss = tf.reduce_mean(tf.losses.mean_squared_error(J_pred_skeleton, J_Pred_tar[i]))
# P, B
# tf.nn.l1_loss
P_pred_loss = tf.reduce_mean(tf.losses.mean_squared_error(P_pred_skeleton, P_Pred_tar[i]))
B_pred_loss = tf.reduce_mean(tf.losses.mean_squared_error(B_pred_skeleton, B_Pred_tar[i]))
elif FLAGS.loss == 'l2':
J_pred_loss = tf.reduce_mean(tf.nn.l2_loss(J_pred_skeleton - J_Pred_tar[i]))
# P, B
# tf.nn.l1_loss
P_pred_loss = tf.reduce_mean(tf.nn.l2_loss(P_pred_skeleton - P_Pred_tar[i]))
B_pred_loss = tf.reduce_mean(tf.nn.l2_loss(B_pred_skeleton - B_Pred_tar[i]))
if k_filter == -1 or (k_filter !=-1 and i == 0):
if no_multi_pred:
J_all_pred_loss.append(J_pred_loss)
else:
J_all_pred_loss.append(J_pred_loss+P_pred_loss+B_pred_loss)
J_encoder_output = tf.reshape(J_encoder_output, [batch_size, k_values[i % levels], -1])
# Average
J_en_outs.append(J_encoder_output[:, -1, :])
# en_outs.append(encoder_output)
# if i == sample_num * levels - 1:
# J_en_outs_whole = J_encoder_output
for i in range(levels+1):
J_test_seq_ftr = J_Test_ftr[i]
J_encoder_output, J_encoder_state = tf.nn.dynamic_rnn(J_cell, J_test_seq_ftr, dtype=tf.float32)
J_encoder_output = tf.reshape(J_encoder_output, [-1, LSTM_embed])
# Skeleton-level Prediction
J_encoder_output = tf.reshape(J_encoder_output, [batch_size, i+1, -1])
# en_outs_test.append(encoder_output)
J_en_outs_test.append(J_encoder_output[:, -1, :])
with tf.name_scope("P_pred"), tf.variable_scope("P_pred", reuse=tf.AUTO_REUSE):
P_W1_pred = tf.Variable(tf.random_normal([LSTM_embed, LSTM_embed]))
P_b1_pred = tf.Variable(tf.zeros(shape=[LSTM_embed, ]))
P_W2_pred = tf.Variable(tf.random_normal([LSTM_embed, 10 * 3]))
P_b2_pred = tf.Variable(tf.zeros(shape=[10 * 3, ]))
B_W1_pred = tf.Variable(tf.random_normal([LSTM_embed, LSTM_embed]))
B_b1_pred = tf.Variable(tf.zeros(shape=[LSTM_embed, ]))
B_W2_pred = tf.Variable(tf.random_normal([LSTM_embed, 5 * 3]))
B_b2_pred = tf.Variable(tf.zeros(shape=[5 * 3, ]))
for i in range((levels - int(pred_margin)) * sample_num):
# if k_filter !=-1 and i != 0:
# continue
P_sampled_seq_ftr = P_T_m_ftr[i]
P_encoder_output, P_encoder_state = tf.nn.dynamic_rnn(P_cell, P_sampled_seq_ftr, dtype=tf.float32)
P_encoder_output = tf.reshape(P_encoder_output, [-1, LSTM_embed])
# Skeleton-level Prediction
P_pred_embedding_P = tf.nn.relu(tf.matmul(P_encoder_output, P_W1_pred) + P_b1_pred)
P_pred_skeleton = tf.matmul(P_pred_embedding_P, P_W2_pred) + P_b2_pred
P_pred_skeleton = tf.reshape(P_pred_skeleton, [batch_size, k_values[i % levels], 10 * 3])
B_pred_embedding_P = tf.nn.relu(tf.matmul(P_encoder_output, B_W1_pred) + B_b1_pred)
B_pred_skeleton = tf.matmul(B_pred_embedding_P, B_W2_pred) + B_b2_pred
B_pred_skeleton = tf.reshape(B_pred_skeleton, [batch_size, k_values[i % levels], 5 * 3])
if k_filter == -1 or (k_filter !=-1 and i == 0):
if FLAGS.loss == 'l1':
P_pred_loss = tf.reduce_mean(tf.losses.absolute_difference(P_pred_skeleton, P_Pred_tar[i]))
B_pred_loss = tf.reduce_mean(tf.losses.absolute_difference(B_pred_skeleton, B_Pred_tar[i]))
elif FLAGS.loss == 'MSE':
P_pred_loss = tf.reduce_mean(tf.losses.mean_squared_error(P_pred_skeleton, P_Pred_tar[i]))
B_pred_loss = tf.reduce_mean(tf.losses.mean_squared_error(B_pred_skeleton, B_Pred_tar[i]))
elif FLAGS.loss == 'l2':
P_pred_loss = tf.reduce_mean(tf.nn.l2_loss(P_pred_skeleton - P_Pred_tar[i]))
B_pred_loss = tf.reduce_mean(tf.nn.l2_loss(B_pred_skeleton - B_Pred_tar[i]))
if no_multi_pred:
P_all_pred_loss.append(P_pred_loss)
else:
P_all_pred_loss.append(P_pred_loss+B_pred_loss)
P_encoder_output = tf.reshape(P_encoder_output, [batch_size, k_values[i % levels], -1])
# Average
P_en_outs.append(P_encoder_output[:, -1, :])
for i in range(levels + 1):
P_test_seq_ftr = P_Test_ftr[i]
P_encoder_output, P_encoder_state = tf.nn.dynamic_rnn(P_cell, P_test_seq_ftr, dtype=tf.float32)
P_encoder_output = tf.reshape(P_encoder_output, [-1, LSTM_embed])
P_encoder_output = tf.reshape(P_encoder_output, [batch_size, i+1, -1])
P_en_outs_test.append(P_encoder_output[:, -1, :])
with tf.name_scope("B_pred"), tf.variable_scope("B_pred", reuse=tf.AUTO_REUSE):
B_W1_pred = tf.Variable(tf.random_normal([LSTM_embed, LSTM_embed]))
B_b1_pred = tf.Variable(tf.zeros(shape=[LSTM_embed, ]))
B_W2_pred = tf.Variable(tf.random_normal([LSTM_embed, 5 * 3]))
B_b2_pred = tf.Variable(tf.zeros(shape=[5 * 3, ]))
for i in range((levels - int(pred_margin)) * sample_num):
B_sampled_seq_ftr = B_T_m_ftr[i]
B_encoder_output, B_encoder_state = tf.nn.dynamic_rnn(B_cell, B_sampled_seq_ftr, dtype=tf.float32)
B_encoder_output = tf.reshape(B_encoder_output, [-1, LSTM_embed])
# Skeleton-level Prediction
B_pred_embedding_B = tf.nn.relu(tf.matmul(B_encoder_output, B_W1_pred) + B_b1_pred)
B_pred_skeleton = tf.matmul(B_pred_embedding_B, B_W2_pred) + B_b2_pred
B_pred_skeleton = tf.reshape(B_pred_skeleton, [batch_size, k_values[i % levels], 5 * 3])
# tf.nn.l1_loss
if k_filter == -1 or (k_filter !=-1 and i == 0):
if FLAGS.loss == 'l1':
B_pred_loss = tf.reduce_mean(tf.losses.absolute_difference(B_pred_skeleton, B_Pred_tar[i]))
elif FLAGS.loss == 'MSE':
B_pred_loss = tf.reduce_mean(tf.losses.mean_squared_error(B_pred_skeleton, B_Pred_tar[i]))
elif FLAGS.loss == 'l2':
B_pred_loss = tf.reduce_mean(tf.nn.l2_loss(B_pred_skeleton - B_Pred_tar[i]))
B_all_pred_loss.append(B_pred_loss)
B_encoder_output = tf.reshape(B_encoder_output, [batch_size, k_values[i % levels], -1])
# Average
B_en_outs.append(B_encoder_output[:, -1, :])
for i in range(levels + 1):
B_test_seq_ftr = B_Test_ftr[i]
B_encoder_output, B_encoder_state = tf.nn.dynamic_rnn(B_cell, B_test_seq_ftr, dtype=tf.float32)
B_encoder_output = tf.reshape(B_encoder_output, [-1, LSTM_embed])
B_encoder_output = tf.reshape(B_encoder_output, [batch_size, i+1, -1])
B_en_outs_test.append(B_encoder_output[:, -1, :])
with tf.name_scope("I_pred"), tf.variable_scope("I_pred", reuse=tf.AUTO_REUSE):
J_W1_pred = tf.Variable(tf.random_normal([LSTM_embed, LSTM_embed]))
J_b1_pred = tf.Variable(tf.zeros(shape=[LSTM_embed, ]))
J_W2_pred = tf.Variable(tf.random_normal([LSTM_embed, nb_nodes * 3]))
J_b2_pred = tf.Variable(tf.zeros(shape=[nb_nodes * 3, ]))
P_W1_pred = tf.Variable(tf.random_normal([LSTM_embed, LSTM_embed]))
P_b1_pred = tf.Variable(tf.zeros(shape=[LSTM_embed, ]))
P_W2_pred = tf.Variable(tf.random_normal([LSTM_embed, 10 * 3]))
P_b2_pred = tf.Variable(tf.zeros(shape=[10 * 3, ]))
B_W1_pred = tf.Variable(tf.random_normal([LSTM_embed, LSTM_embed]))
B_b1_pred = tf.Variable(tf.zeros(shape=[LSTM_embed, ]))
B_W2_pred = tf.Variable(tf.random_normal([LSTM_embed, 5 * 3]))
B_b2_pred = tf.Variable(tf.zeros(shape=[5 * 3, ]))
I_W1_pred = tf.Variable(tf.random_normal([LSTM_embed, LSTM_embed]))
I_b1_pred = tf.Variable(tf.zeros(shape=[LSTM_embed, ]))
I_W2_pred = tf.Variable(tf.random_normal([LSTM_embed, I_nodes * 3]))
I_b2_pred = tf.Variable(tf.zeros(shape=[I_nodes * 3, ]))
for i in range((levels - int(pred_margin)) * sample_num):
# if k_filter == time_step and i != 0:
# continue
I_sampled_seq_ftr = I_T_m_ftr[i]
I_encoder_output, I_encoder_state = tf.nn.dynamic_rnn(I_cell, I_sampled_seq_ftr, dtype=tf.float32)
I_encoder_output = tf.reshape(I_encoder_output, [-1, LSTM_embed])
# Skeleton-level Prediction
I_pred_embedding_I = tf.nn.relu(tf.matmul(I_encoder_output, I_W1_pred) + I_b1_pred)
I_pred_skeleton = tf.matmul(I_pred_embedding_I, I_W2_pred) + I_b2_pred
I_pred_skeleton = tf.reshape(I_pred_skeleton, [batch_size, k_values[i % levels], I_nodes * 3])
# tf.nn.l1_loss
J_pred_embedding_I = tf.nn.relu(tf.matmul(I_encoder_output, J_W1_pred) + J_b1_pred)
J_pred_skeleton = tf.matmul(J_pred_embedding_I, J_W2_pred) + J_b2_pred
J_pred_skeleton = tf.reshape(J_pred_skeleton, [batch_size, k_values[i % levels], nb_nodes * 3])
B_pred_embedding_I = tf.nn.relu(tf.matmul(I_encoder_output, B_W1_pred) + B_b1_pred)
B_pred_skeleton = tf.matmul(B_pred_embedding_I, B_W2_pred) + B_b2_pred
B_pred_skeleton = tf.reshape(B_pred_skeleton, [batch_size, k_values[i % levels], 5 * 3])
P_pred_embedding_I = tf.nn.relu(tf.matmul(I_encoder_output, P_W1_pred) + P_b1_pred)
P_pred_skeleton = tf.matmul(P_pred_embedding_I, P_W2_pred) + P_b2_pred
P_pred_skeleton = tf.reshape(P_pred_skeleton, [batch_size, k_values[i % levels], 10 * 3])
#
if k_filter == -1 or (k_filter !=-1 and i == 0):
if FLAGS.loss == 'l1':
I_pred_loss = tf.reduce_mean(tf.losses.absolute_difference(I_pred_skeleton, I_Pred_tar[i]))
J_pred_loss = tf.reduce_mean(tf.losses.absolute_difference(J_pred_skeleton, J_Pred_tar[i]))
P_pred_loss = tf.reduce_mean(tf.losses.absolute_difference(P_pred_skeleton, P_Pred_tar[i]))
B_pred_loss = tf.reduce_mean(tf.losses.absolute_difference(B_pred_skeleton, B_Pred_tar[i]))
elif FLAGS.loss == 'MSE':
I_pred_loss = tf.reduce_mean(tf.losses.mean_squared_error(I_pred_skeleton, I_Pred_tar[i]))
J_pred_loss = tf.reduce_mean(tf.losses.mean_squared_error(J_pred_skeleton, J_Pred_tar[i]))
P_pred_loss = tf.reduce_mean(tf.losses.mean_squared_error(P_pred_skeleton, P_Pred_tar[i]))
B_pred_loss = tf.reduce_mean(tf.losses.mean_squared_error(B_pred_skeleton, B_Pred_tar[i]))
elif FLAGS.loss == 'l2':
I_pred_loss = tf.reduce_mean(tf.nn.l2_loss(I_pred_skeleton - I_Pred_tar[i]))
J_pred_loss = tf.reduce_mean(tf.nn.l2_loss(J_pred_skeleton - J_Pred_tar[i]))
P_pred_loss = tf.reduce_mean(tf.nn.l2_loss(P_pred_skeleton - P_Pred_tar[i]))
B_pred_loss = tf.reduce_mean(tf.nn.l2_loss(B_pred_skeleton - B_Pred_tar[i]))
if no_multi_pred:
I_all_pred_loss.append(I_pred_loss)
else:
I_all_pred_loss.append(P_pred_loss + J_pred_loss + I_pred_loss + B_pred_loss)
I_encoder_output = tf.reshape(I_encoder_output, [batch_size, k_values[i % levels], -1])
# Average
I_en_outs.append(I_encoder_output[:, -1, :])
for i in range(levels + 1):
I_test_seq_ftr = I_Test_ftr[i]
I_encoder_output, I_encoder_state = tf.nn.dynamic_rnn(I_cell, I_test_seq_ftr, dtype=tf.float32)
I_encoder_output = tf.reshape(I_encoder_output, [-1, LSTM_embed])
I_encoder_output = tf.reshape(I_encoder_output, [batch_size, i+1, -1])
I_en_outs_test.append(I_encoder_output[:, -1, :])
# en_outs.append(tf.reshape(encoder_output, [batch_size, k_values[i], -1]))
J_pred_opt = tf.train.AdamOptimizer(learning_rate=lr)
J_pred_train_op = J_pred_opt.minimize(tf.reduce_mean(J_all_pred_loss))
P_pred_opt = tf.train.AdamOptimizer(learning_rate=lr)
P_pred_train_op = P_pred_opt.minimize(tf.reduce_mean(P_all_pred_loss))
B_pred_opt = tf.train.AdamOptimizer(learning_rate=lr)
B_pred_train_op = B_pred_opt.minimize(tf.reduce_mean(B_all_pred_loss))
I_pred_opt = tf.train.AdamOptimizer(learning_rate=lr)
I_pred_train_op = I_pred_opt.minimize(tf.reduce_mean(I_all_pred_loss))
with tf.name_scope("Recognition"), tf.variable_scope("Recognition", reuse=tf.AUTO_REUSE):
#
if no_interp:
en_to_pred = tf.concat([J_en_outs_test[0], P_en_outs_test[0],
B_en_outs_test[0]], axis=-1)
for i in range(1, levels + 1):
temp = tf.concat([J_en_outs_test[i], P_en_outs_test[i],
B_en_outs_test[i]], axis=-1)
en_to_pred = tf.concat([en_to_pred, temp], axis=0)
elif not only_ske_embed:
en_to_pred = tf.concat([I_en_outs_test[0], J_en_outs_test[0], P_en_outs_test[0],
B_en_outs_test[0]], axis=-1)
for i in range(1, levels+1):
temp = tf.concat([I_en_outs_test[i], J_en_outs_test[i], P_en_outs_test[i],
B_en_outs_test[i]], axis=-1)
en_to_pred = tf.concat([en_to_pred, temp], axis=0)
else:
en_to_pred = tf.concat([J_en_outs_test[0]], axis=-1)
for i in range(1, levels + 1):
temp = tf.concat([J_en_outs_test[i]], axis=-1)
en_to_pred = tf.concat([en_to_pred, temp], axis=0)
# Frozen
if frozen == '1':
en_to_pred = tf.stop_gradient(en_to_pred)
if no_interp:
# original
W_1 = tf.Variable(tf.random_normal([LSTM_embed * 3, LSTM_embed * 3]))
b_1 = tf.Variable(tf.zeros(shape=[LSTM_embed * 3, ]))
W_2 = tf.Variable(tf.random_normal([LSTM_embed * 3, nb_classes]))
b_2 = tf.Variable(tf.zeros(shape=[nb_classes, ]))
elif not only_ske_embed:
# original
W_1 = tf.Variable(tf.random_normal([LSTM_embed * 4, LSTM_embed * 4]))
b_1 = tf.Variable(tf.zeros(shape=[LSTM_embed * 4, ]))
W_2 = tf.Variable(tf.random_normal([LSTM_embed * 4, nb_classes]))
b_2 = tf.Variable(tf.zeros(shape=[nb_classes, ]))
else:
W_1 = tf.Variable(tf.random_normal([LSTM_embed, LSTM_embed]))
b_1 = tf.Variable(tf.zeros(shape=[LSTM_embed, ]))
W_2 = tf.Variable(tf.random_normal([LSTM_embed, nb_classes]))
b_2 = tf.Variable(tf.zeros(shape=[nb_classes, ]))
# original
logits = tf.matmul(tf.nn.relu(tf.matmul(en_to_pred, W_1) + b_1), W_2) + b_2
logits_pred = tf.matmul(tf.nn.relu(tf.matmul(en_to_pred, W_1) + b_1), W_2) + b_2
log_resh = tf.reshape(logits, [-1, nb_classes])
lab_resh = tf.reshape(lbl_in, [-1, nb_classes])
if not last_pre:
aver_pred = logits[:batch_size]
aver_final_pred = logits_pred[:batch_size]
for i in range(1, levels+1):
aver_pred += logits[batch_size*i:batch_size*(i+1)]
aver_final_pred += logits_pred[batch_size * i:batch_size * (i + 1)]
else:
aver_pred = logits[-batch_size:]
aver_final_pred = logits_pred[-batch_size:]
correct_pred = tf.equal(tf.argmax(aver_pred, -1), tf.argmax(lab_resh, -1))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
correct_final_pred = tf.equal(tf.argmax(aver_final_pred, -1), tf.argmax(lab_resh, -1))
accuracy_final = tf.reduce_mean(tf.cast(correct_final_pred, tf.float32))
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=aver_pred, labels=lab_resh))
# frozen
opt = tf.train.AdamOptimizer(learning_rate=lr)
train_op = opt.minimize(loss)
saver = tf.train.Saver()
init_op = tf.group(tf.global_variables_initializer(), tf.local_variables_initializer())
vlss_mn = np.inf
vacc_mx = 0.0
vnAUC_mx = 0.0
curr_step = 0
X_train = X_train_J
X_test = X_test_J
with tf.Session(config=config) as sess:
sess.run(init_op)
train_loss_avg = 0
train_acc_avg = 0
val_loss_avg = 0
val_acc_avg = 0
for epoch in range(pre_epochs):
tr_step = 0
tr_size = X_train.shape[0]
while tr_step * batch_size < tr_size:
if (tr_step + 1) * batch_size > tr_size:
break
X_input_J = X_train_J[tr_step * batch_size:(tr_step + 1) * batch_size]
X_input_J = X_input_J.reshape([-1, nb_nodes, 3])
X_input_P = X_train_P[tr_step * batch_size:(tr_step + 1) * batch_size]
X_input_P = X_input_P.reshape([-1, 10, 3])
X_input_B = X_train_B[tr_step * batch_size:(tr_step + 1) * batch_size]
X_input_B = X_input_B.reshape([-1, 5, 3])
# interpolation
X_input_I = X_train_I[tr_step * batch_size:(tr_step + 1) * batch_size]
X_input_I = X_input_I.reshape([-1, I_nodes, 3])
loss_rec, loss_attr, acc_attr, loss_pred = 0, 0, 0, 0
if no_interp:
_, _, _, loss_pred, P_loss_pred, B_loss_pred, I_loss_pred, \
= sess.run([J_pred_train_op, P_pred_train_op,
B_pred_train_op,
J_pred_loss, P_pred_loss, B_pred_loss, I_pred_loss,
],
feed_dict={
J_in: X_input_J,
P_in: X_input_P,