forked from pawni/BayesByHypernet
-
Notifications
You must be signed in to change notification settings - Fork 0
/
layers.py
664 lines (525 loc) · 25 KB
/
layers.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
from __future__ import absolute_import, print_function, division
import tensorflow as tf
import numpy as np
from base_layers import *
class BBHDenseLayer(BBHLayer):
def _build(self, name, input_dim, output_dim, use_bias=True,
h_units=[16, 32],
h_use_bias=True, h_noise_shape=1,
num_samples=5, num_slices=1,
aligned_noise=True,
h_activation_func=lambda x: tf.maximum(0.1 * x, x)):
self.share_noise = aligned_noise
self.use_bias = use_bias
with tf.variable_scope(name):
self.w = self._get_weight(
'{}/w'.format(name), (input_dim, output_dim), units=h_units,
use_bias=h_use_bias, noise_shape=h_noise_shape,
num_samples=num_samples, num_slices=num_slices,
activation_func=h_activation_func)
if self.use_bias:
self.b = self._get_weight(
'{}/b'.format(name), (output_dim, ), units=h_units,
use_bias=h_use_bias, noise_shape=h_noise_shape,
num_samples=num_samples, num_slices=1,
activation_func=h_activation_func)
def call(self, x, sample=0):
x = tf.matmul(x, self.w[sample])
if self.use_bias:
x = x + self.b[sample]
return x
class BBHConvLayer(BBHLayer):
def _build(self, name, input_filter, output_filter, kernel_size,
padding='SAME', strides=(1, 1, 1, 1), use_bias=True,
num_samples=5, num_slices=1,
aligned_noise=True,
h_units=[16, 32], h_use_bias=True, h_noise_shape=1,
h_activation_func=lambda x: tf.maximum(0.1 * x, x)):
self.share_noise = aligned_noise
self.padding = padding
self.strides = strides
self.use_bias = use_bias
with tf.variable_scope(name):
self.w = self._get_weight(
'{}/w'.format(name), (kernel_size, kernel_size, input_filter, output_filter),
units=h_units, use_bias=h_use_bias, noise_shape=h_noise_shape,
num_samples=num_samples, num_slices=num_slices,
activation_func=h_activation_func)
if self.use_bias:
self.b = self._get_weight(
'{}/b'.format(name), (output_filter, ), units=h_units,
use_bias=h_use_bias, noise_shape=h_noise_shape,
num_samples=num_samples, num_slices=1,
activation_func=h_activation_func)
def call(self, x, sample=0):
x = tf.nn.conv2d(x, self.w[sample], self.strides, self.padding,
use_cudnn_on_gpu=True)
if self.use_bias:
x = x + self.b[sample]
return x
class BBHDynDenseLayer(BBHDynLayer):
def _build(self, name, input_dim, output_dim, use_bias=True,
h_units=[16, 32],
h_use_bias=False, h_noise_shape=1,
h_activation_func=lambda x: tf.maximum(0.1 * x, x)):
self.use_bias = use_bias
with tf.variable_scope(name):
self.w = self._get_weight(
'{}/w'.format(name), (input_dim, output_dim), units=h_units,
use_bias=h_use_bias, noise_shape=h_noise_shape,
activation_func=h_activation_func)
if self.use_bias:
self.b = self._get_weight(
'{}/b'.format(name), (output_dim, ), units=h_units,
use_bias=h_use_bias, noise_shape=h_noise_shape,
activation_func=h_activation_func)
def call(self, x, *args, **kwargs):
cond = tf.reduce_mean(x, [0])
cond = tf.concat(tf.nn.moments(x, [0]), 0)
x = tf.matmul(x, self.w(cond))
if self.use_bias:
x = x + self.b(cond)
return x
class BBHDynConvLayer(BBHDynLayer):
def _build(self, name, input_filter, output_filter, kernel_size,
padding='SAME', strides=(1, 1, 1, 1), use_bias=True,
h_units=[16, 32], h_use_bias=False, h_noise_shape=1,
h_activation_func=lambda x: tf.maximum(0.1 * x, x)):
self.padding = padding
self.strides = strides
self.use_bias = use_bias
with tf.variable_scope(name):
self.w = self._get_weight(
'{}/w'.format(name), (kernel_size, kernel_size, input_filter, output_filter),
units=h_units, use_bias=h_use_bias, noise_shape=h_noise_shape,
activation_func=h_activation_func)
if self.use_bias:
self.b = self._get_weight(
'{}/b'.format(name), (output_filter, ), units=h_units,
use_bias=h_use_bias, noise_shape=h_noise_shape,
activation_func=h_activation_func)
def call(self, x, *args, **kwargs):
cond = tf.reduce_mean(x, [0, 1, 2])
cond = tf.concat(tf.nn.moments(x, [0, 1, 2]), 0)
x = tf.nn.conv2d(x, self.w(cond), self.strides, self.padding,
use_cudnn_on_gpu=True)
if self.use_bias:
x = x + self.b(cond)
return x
class BBHNormDenseLayer(BBHLayer):
def _build(self, name, input_dim, output_dim, use_bias=True,
h_units=[16, 32],
h_use_bias=False, h_noise_shape=1,
h_activation_func=lambda x: tf.maximum(0.1 * x, x)):
self.use_bias = use_bias
with tf.variable_scope(name):
w = tf.get_variable(
'w', (input_dim, output_dim),
tf.float32,
tf.truncated_normal_initializer(0, 0.05))
w_norm = self._get_weight(
'{}/w_norm'.format(name), (1, output_dim),
units=h_units, use_bias=h_use_bias, noise_shape=h_noise_shape,
activation_func=h_activation_func)
self.w = w / tf.sqrt(tf.reduce_sum(tf.square(w), axis=[0], keep_dims=True)) * w_norm
if self.use_bias:
self.b = self._get_weight(
'{}/b'.format(name), (output_dim, ), units=h_units,
use_bias=h_use_bias, noise_shape=h_noise_shape,
activation_func=h_activation_func)
def call(self, x, *args, **kwargs):
x = tf.matmul(x, self.w)
if self.use_bias:
x = x + self.b
return x
class BBHNormConvLayer(BBHLayer):
def _build(self, name, input_filter, output_filter, kernel_size,
padding='SAME', strides=(1, 1, 1, 1), use_bias=True,
h_units=[16, 32], h_use_bias=False, h_noise_shape=1,
h_activation_func=lambda x: tf.maximum(0.1 * x, x)):
self.padding = padding
self.strides = strides
self.use_bias = use_bias
with tf.variable_scope(name):
w = tf.get_variable(
'w', (kernel_size, kernel_size, input_filter, output_filter),
tf.float32,
tf.truncated_normal_initializer(0, 0.05))
w_norm = self._get_weight(
'{}/w_norm'.format(name), (1, 1, 1, output_filter),
units=h_units, use_bias=h_use_bias, noise_shape=h_noise_shape,
activation_func=h_activation_func)
self.w = w / tf.sqrt(tf.reduce_sum(tf.square(w), axis=[0, 1, 2], keep_dims=True)) * w_norm
if self.use_bias:
self.b = self._get_weight(
'{}/b'.format(name), (output_filter, ), units=h_units,
use_bias=h_use_bias, noise_shape=h_noise_shape,
activation_func=h_activation_func)
def call(self, x, *args, **kwargs):
x = tf.nn.conv2d(x, self.w, self.strides, self.padding,
use_cudnn_on_gpu=True)
if self.use_bias:
x = x + self.b
return x
class BBBDenseLayer(BBBLayer):
def _build(self, name, input_dim, output_dim, use_bias=True, init_var=-9,
prior_scale=1., aligned_noise=False):
self.use_bias = use_bias
self.share_noise = aligned_noise
with tf.variable_scope(name):
self.w = self._get_weight(
'w', (input_dim, output_dim), init_var=init_var,
prior_scale=prior_scale)
if self.use_bias:
self.b = self._get_weight(
'b', (output_dim,), init_var=init_var,
prior_scale=prior_scale)
def call(self, x, *args, **kwargs):
x = tf.matmul(x, self.w)
if self.use_bias:
x = x + self.b
return x
class BBBConvLayer(BBBLayer):
def _build(self, name, input_filter, output_filter, kernel_size,
padding='SAME', strides=(1, 1, 1, 1), use_bias=True,
aligned_noise=False,
init_var=-9, prior_scale=1.):
self.share_noise = aligned_noise
self.padding = padding
self.strides = strides
self.use_bias = use_bias
with tf.variable_scope(name):
self.w = self._get_weight(
'w', (kernel_size, kernel_size, input_filter, output_filter),
init_var=init_var, prior_scale=prior_scale)
if self.use_bias:
self.b = self._get_weight(
'b', (output_filter,), init_var=init_var,
prior_scale=prior_scale)
def call(self, x, *args, **kwargs):
x = tf.nn.conv2d(x, self.w, self.strides, self.padding,
use_cudnn_on_gpu=True)
if self.use_bias:
x = x + self.b
return x
class VanillaDenseLayer(Layer):
def _build(self, name, input_dim, output_dim, use_bias=True):
self.use_bias = use_bias
with tf.variable_scope(name):
self.w = tf.get_variable(
'w', (input_dim, output_dim), tf.float32,
tf.variance_scaling_initializer())
tf.add_to_collection('l2', tf.reduce_sum(tf.square(self.w)))
if self.use_bias:
self.b = tf.get_variable(
'b', (output_dim, ), tf.float32,
tf.zeros_initializer())
tf.add_to_collection('l2', tf.reduce_sum(tf.square(self.b)))
def call(self, x, *args, **kwargs):
x = tf.matmul(x, self.w)
if self.use_bias:
x = x + self.b
return x
class VanillaConvLayer(Layer):
def _build(self, name, input_filter, output_filter, kernel_size,
padding='SAME', strides=(1, 1, 1, 1), use_bias=True):
self.padding = padding
self.strides = strides
self.use_bias = use_bias
with tf.variable_scope(name):
self.w = tf.get_variable(
'w', (kernel_size, kernel_size, input_filter, output_filter),
tf.float32, tf.variance_scaling_initializer())
tf.add_to_collection('l2', tf.reduce_sum(self.w ** 2))
if self.use_bias:
self.b = tf.get_variable(
'b', (output_filter,), tf.float32,
tf.zeros_initializer())
tf.add_to_collection('l2', tf.reduce_sum(self.b ** 2))
def call(self, x, *args, **kwargs):
x = tf.nn.conv2d(x, self.w, self.strides, self.padding,
use_cudnn_on_gpu=True)
if self.use_bias:
x = x + self.b
return x
########
#
# MNF layers adapted from https://github.com/AMLab-Amsterdam/MNF_VBNN
#
class MNFDenseLayer(Layer):
def _build(self, name, input_dim, output_dim, learn_p=False,
thres_var=1., init_var=-9, use_bias=True):
self.thres_var = thres_var
self.input_dim = input_dim
self.output_dim = output_dim
self.use_bias = use_bias
flow_dim_h = 50
with tf.variable_scope(name):
self.w_loc = tf.get_variable(
'w_loc', (input_dim, output_dim), tf.float32,
tf.variance_scaling_initializer())
self.w_log_scale_sq = tf.get_variable(
'w_log_scale_sq', (input_dim, output_dim), tf.float32,
tf.truncated_normal_initializer(init_var, 0.05))
self.b_loc = tf.get_variable(
'b_loc', (1, output_dim), tf.float32,
tf.truncated_normal_initializer(0, 0.05))
self.b_log_scale_sq = tf.get_variable(
'b_log_scale_sq', (1, output_dim), tf.float32,
tf.truncated_normal_initializer(init_var, 0.05))
self.qzero_mean = tf.get_variable(
'qzero_mean', (input_dim, ), tf.float32,
tf.truncated_normal_initializer(0., 0.05))
self.qzero = tf.get_variable(
'qzero', (input_dim,), tf.float32,
tf.truncated_normal_initializer(np.log(0.1), 1e-6))
self.rsr_M = tf.get_variable(
'var_r_aux', (input_dim,), tf.float32,
tf.truncated_normal_initializer(0., 0.05))
self.apvar_M = tf.get_variable(
'apvar_r_aux', (input_dim,), tf.float32,
tf.truncated_normal_initializer(0., 0.05))
self.rsri_M = tf.get_variable(
'var_r_auxi', (input_dim,), tf.float32,
tf.truncated_normal_initializer(0., 0.05))
self.pvar = tf.get_variable(
'prior_var_r_p', (input_dim,), tf.float32,
tf.truncated_normal_initializer(1., 1e-6),
trainable=learn_p)
self.pvar_bias = tf.get_variable(
'prior_var_r_p_bias', (1,), tf.float32,
tf.truncated_normal_initializer(1., 1e-6),
trainable=learn_p)
if input_dim == 1:
self.flow_r = PlanarFlow(name + '_fr', input_dim,
n_flows=2, # fixed to 2
scope=name)
else:
self.flow_r = MaskedNVPFlow(name + '_fr', input_dim,
n_flows=2, # fixed to 2
n_hidden=0,
dim_h=2 * flow_dim_h,
scope=name)
if input_dim == 1:
self.flow_q = PlanarFlow(name + '_fq', input_dim,
n_flows=2, # fixed to 2
scope=name)
else:
self.flow_q = MaskedNVPFlow(name + '_fq', input_dim,
n_flows=2, # fixed to 2
n_hidden=0,
dim_h=flow_dim_h,
scope=name)
tf.add_to_collection('mnf_kl', -1. * self.kldiv())
tf.add_to_collection('kl_term', -1. * self.kldiv())
weight_samples = tf.stack([self.get_weight() for _ in range(5)])
weight_samples = tf.reshape(weight_samples, [5, -1])
tf.add_to_collection('weight_samples', weight_samples)
def sample_z(self, size_M=1):
qm0 = tf.exp(self.qzero)
isample_M = tf.tile(tf.expand_dims(self.qzero_mean, 0), [size_M, 1])
eps = tf.random_normal(tf.stack((size_M, self.input_dim)))
sample_M = isample_M + tf.sqrt(qm0) * eps
sample_M, logdets = self.flow_q.get_output_for(sample_M)
return sample_M, logdets
def kldiv(self):
M, logdets = self.sample_z()
logdets = logdets[0]
M = tf.squeeze(M)
std_mg = tf.exp(self.w_log_scale_sq)
qm0 = tf.exp(self.qzero)
if len(M.get_shape()) == 0:
Mexp = M
else:
Mexp = tf.expand_dims(M, 1)
Mtilde = Mexp * self.w_loc
Vtilde = tf.square(std_mg)
iUp = outer(tf.exp(self.pvar), tf.ones((self.output_dim,)))
logqm = - tf.reduce_sum(.5 * (tf.log(2 * np.pi) + tf.log(qm0) + 1))
logqm -= logdets
kldiv_w = tf.reduce_sum(.5 * tf.log(iUp) - tf.log(std_mg) + (
(Vtilde + tf.square(Mtilde)) / (2 * iUp)) - .5)
kldiv_bias = tf.reduce_sum(
.5 * self.pvar_bias - .5 * self.b_log_scale_sq + (
(tf.exp(self.b_log_scale_sq) +
tf.square(self.b_loc)) / (2 * tf.exp(self.pvar_bias))) - .5)
apvar_M = self.apvar_M
# shared network for hidden layer
mw = tf.matmul(tf.expand_dims(apvar_M, 0), Mtilde)
eps = tf.expand_dims(tf.random_normal((self.output_dim,)), 0)
varw = tf.matmul(tf.square(tf.expand_dims(apvar_M, 0)), Vtilde)
a = tf.nn.tanh(mw + tf.sqrt(varw) * eps)
# split at output layer
if len(tf.squeeze(a).get_shape()) != 0:
w__ = tf.reduce_mean(outer(self.rsr_M, tf.squeeze(a)), axis=1)
wv__ = tf.reduce_mean(outer(self.rsri_M, tf.squeeze(a)), axis=1)
else:
w__ = self.rsr_M * tf.squeeze(a)
wv__ = self.rsri_M * tf.squeeze(a)
M, logrm = self.flow_r.get_output_for(tf.expand_dims(M, 0))
M = tf.squeeze(M)
logrm = logrm[0]
logrm += tf.reduce_sum(
-.5 * tf.exp(wv__) * tf.square(M - w__) - .5 * tf.log(
2 * np.pi) + .5 * wv__)
return - kldiv_w + logrm - logqm - kldiv_bias
def get_weight(self):
std_mg = tf.clip_by_value(
tf.exp(self.w_log_scale_sq), 0., self.thres_var)
sample_M, _ = self.sample_z()
w_sample = tf.transpose(sample_M) * self.w_loc
w_sample += tf.random_normal(tf.shape(w_sample)) * std_mg
return w_sample
def call(self, x, sample_shape=None):
if sample_shape is None:
sample_shape = tf.shape(x)[0]
std_mg = tf.clip_by_value(
tf.exp(self.w_log_scale_sq), 0., self.thres_var)
var_mg = tf.square(std_mg)
sample_M, _ = self.sample_z(size_M=sample_shape)
xt = x * sample_M
mu_out = tf.matmul(xt, self.w_loc)
varin = tf.matmul(tf.square(x), var_mg)
if self.use_bias:
mu_out += self.b_loc
varin += tf.clip_by_value(
tf.exp(self.b_log_scale_sq), 0., self.thres_var ** 2)
xin = tf.sqrt(varin)
sigma_out = xin * tf.random_normal(tf.shape(mu_out))
output = mu_out + sigma_out
return output
class MNFConvLayer(Layer):
def _build(self, name, input_filter, output_filter, kernel_size,
padding='SAME', strides=(1, 1, 1, 1), learn_p=False,
thres_var=1., init_var=-9, use_bias=True):
self.thres_var = thres_var
self.input_filter = input_filter
self.output_filter = output_filter
self.padding = padding
self.strides = strides
self.input_dim = kernel_size * kernel_size * input_filter
self.w_shape = (kernel_size, kernel_size, input_filter, output_filter)
flow_dim_h = 50
with tf.variable_scope(name):
self.w_loc = tf.get_variable(
'w_loc', self.w_shape, tf.float32,
tf.variance_scaling_initializer())
self.w_log_scale_sq = tf.get_variable(
'w_log_scale_sq', self.w_shape, tf.float32,
tf.truncated_normal_initializer(init_var, 0.05))
self.b_loc = tf.get_variable(
'b_loc', (output_filter, ), tf.float32,
tf.truncated_normal_initializer(0, 0.05))
self.b_log_scale_sq = tf.get_variable(
'b_log_scale_sq', (output_filter, ), tf.float32,
tf.truncated_normal_initializer(init_var, 0.05))
self.qzero_mean = tf.get_variable(
'qzero_mean', (output_filter, ), tf.float32,
tf.truncated_normal_initializer(0., 0.05))
self.qzero = tf.get_variable(
'qzero', (output_filter,), tf.float32,
tf.truncated_normal_initializer(np.log(0.1), 1e-6))
self.rsr_M = tf.get_variable(
'var_r_aux', (output_filter,), tf.float32,
tf.truncated_normal_initializer(0., 0.05))
self.apvar_M = tf.get_variable(
'apvar_r_aux', (output_filter,), tf.float32,
tf.truncated_normal_initializer(0., 0.05))
self.rsri_M = tf.get_variable(
'var_r_auxi', (output_filter,), tf.float32,
tf.truncated_normal_initializer(0., 0.05))
self.pvar = tf.get_variable(
'prior_var_r_p', (self.input_dim,), tf.float32,
tf.truncated_normal_initializer(1., 1e-6),
trainable=learn_p)
self.pvar_bias = tf.get_variable(
'prior_var_r_p_bias', (1,), tf.float32,
tf.truncated_normal_initializer(1., 1e-6),
trainable=learn_p)
self.flow_r = MaskedNVPFlow(name + '_fr', output_filter,
n_flows=2, # fixed to 2
n_hidden=0,
dim_h=2 * flow_dim_h,
scope=name)
self.flow_q = MaskedNVPFlow(name + '_fq', output_filter,
n_flows=2, # fixed to 2
n_hidden=0,
dim_h=flow_dim_h,
scope=name)
tf.add_to_collection('mnf_kl', -1. * self.kldiv())
weight_samples = tf.stack([self.get_weight() for _ in range(5)])
weight_samples = tf.reshape(weight_samples, [5, -1])
tf.add_to_collection('weight_samples', weight_samples)
def sample_z(self, size_M=1):
qm0 = tf.exp(self.qzero)
isample_M = tf.tile(tf.expand_dims(self.qzero_mean, 0), [size_M, 1])
eps = tf.random_normal(tf.stack((size_M, self.output_filter)))
sample_M = isample_M + tf.sqrt(qm0) * eps
sample_M, logdets = self.flow_q.get_output_for(sample_M)
return sample_M, logdets
def kldiv(self):
M, logdets = self.sample_z()
logdets = logdets[0]
M = tf.squeeze(M)
std_w = tf.exp(self.w_log_scale_sq)
mu = tf.reshape(self.w_loc, [-1, self.output_filter])
std_w = tf.reshape(std_w, [-1, self.output_filter])
Mtilde = mu * tf.expand_dims(M, 0)
mbias = self.b_loc * M
Vtilde = tf.square(std_w)
iUp = outer(tf.exp(self.pvar), tf.ones((self.output_filter,)))
qm0 = tf.exp(self.qzero)
logqm = - tf.reduce_sum(.5 * (tf.log(2 * np.pi)
+ tf.log(qm0 + 1e-8) +1))
logqm -= logdets
kldiv_w = tf.reduce_sum(.5 * tf.log(iUp + 1e-8) - .5 * tf.log(Vtilde)
+ ((Vtilde + tf.square(Mtilde))
/ (2 * iUp)) - .5)
kldiv_bias = tf.reduce_sum(
.5 * self.pvar_bias - .5 * self.b_log_scale_sq + (
(tf.exp(self.b_log_scale_sq) +
tf.square(mbias)) / (2 * tf.exp(self.pvar_bias))) - .5)
apvar_M = self.apvar_M
mw = tf.matmul(Mtilde, tf.expand_dims(apvar_M, 1))
vw = tf.matmul(Vtilde, tf.expand_dims(tf.square(apvar_M), 1))
eps = tf.expand_dims(tf.random_normal((self.input_dim,)), 1)
a = mw + tf.sqrt(vw) * eps
mb = tf.reduce_sum(mbias * apvar_M)
vb = tf.reduce_sum(tf.exp(self.b_log_scale_sq) * tf.square(apvar_M))
a += mb + tf.sqrt(vb) * tf.random_normal(())
w__ = tf.reduce_mean(outer(tf.squeeze(a), self.rsr_M), axis=0)
wv__ = tf.reduce_mean(outer(tf.squeeze(a), self.rsri_M), axis=0)
M, logrm = self.flow_r.get_output_for(tf.expand_dims(M, 0))
M = tf.squeeze(M)
logrm = logrm[0]
logrm += tf.reduce_sum(
-.5 * tf.exp(wv__) * tf.square(M - w__) - .5 * tf.log(
2 * np.pi) + .5 * wv__)
return - kldiv_w + logrm - logqm - kldiv_bias
def get_mean_var(self, x):
var_w = tf.clip_by_value(tf.exp(self.w_log_scale_sq), 0., self.thres_var)
var_w = tf.square(var_w)
var_b = tf.clip_by_value(tf.exp(self.b_log_scale_sq), 0.,
self.thres_var ** 2)
# formally we do cross-correlation here
muout = tf.nn.conv2d(x, self.w_loc, self.strides, self.padding,
use_cudnn_on_gpu=True) + self.b_loc
varout = tf.nn.conv2d(tf.square(x), var_w, self.strides,
self.padding, use_cudnn_on_gpu=True) + var_b
return muout, varout
def get_weight(self):
std_mg = tf.clip_by_value(
tf.exp(self.w_log_scale_sq), 0., self.thres_var)
sample_M, _ = self.sample_z()
w_sample = self.w_loc * sample_M
w_sample += tf.random_normal(tf.shape(w_sample)) * std_mg
return w_sample
def call(self, x, *args, **kwargs):
sample_M, _ = self.sample_z(size_M=tf.shape(x)[0])
sample_M = tf.expand_dims(tf.expand_dims(sample_M, 1), 2)
mean_out, var_out = self.get_mean_var(x)
mean_gout = mean_out * sample_M
var_gout = tf.sqrt(var_out) * tf.random_normal(tf.shape(mean_gout))
out = mean_gout + var_gout
output = out
return output