-
Notifications
You must be signed in to change notification settings - Fork 9
/
Copy pathmodules.py
1302 lines (1041 loc) · 46.6 KB
/
modules.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
# -*- coding: utf-8 -*-
"""
Created on Wed May 10 17:29:14 2017
@author: Chin-Wei
"""
import theano
import theano.tensor as T
import numpy as np
floatX = theano.config.floatX
#import matplotlib.pyplot as plt
import lasagne
from lasagne import init
from lasagne import nonlinearities
from lasagne.layers import get_output
from lasagne.layers import Conv2DLayer, DenseLayer
from lasagne.nonlinearities import rectify
from lasagne.objectives import categorical_crossentropy as cc
from lasagne.objectives import squared_error as se
from theano.tensor.var import TensorVariable as tv
from utils import log_normal
from externals_modules.made_modules import MADE
conv = lasagne.theano_extensions.conv
softplus = lambda x: T.nnet.softplus(x) + delta
exp = lambda x: T.exp(x) + delta
delta = 0.001
# DK
from theano.tensor.shared_randomstreams import RandomStreams
srng = RandomStreams(seed=427)
# TODO: separate primary/hnet layers!
class CoupledDenseLayer(lasagne.layers.base.Layer):
def __init__(self, incoming, num_units, W=init.Normal(0.0001),
b=init.Constant(0.), nonlinearity=nonlinearities.rectify,
**kwargs):
super(CoupledDenseLayer, self).__init__(incoming, **kwargs)
self.nonlinearity = (nonlinearities.identity if nonlinearity is None
else nonlinearity)
self.num_units = num_units
num_inputs1 = int(self.input_shape[1]/2)
num_inputs2 = self.input_shape[1] - num_inputs1
self.W1 = self.add_param(W, (num_inputs1, num_units), name="cpds_W1")
self.W21 = self.add_param(W, (num_units, num_inputs2), name="cpds_W21")
self.W22 = self.add_param(W, (num_units, num_inputs2), name="cdds_W22")
if b is None:
self.b1 = None
self.b21 = None
self.b22 = None
else:
self.b1 = self.add_param(b, (num_units,), name="cpds_b1",
regularizable=False)
self.b21 = self.add_param(b, (num_inputs2,), name="cpds_b21",
regularizable=False)
self.b22 = self.add_param(b, (num_inputs2,), name="cpds_b22",
regularizable=False)
def get_output_shape_for(self, input_shape):
return input_shape
def get_output_for(self, input, **kwargs):
num_inputs = input.shape[1]
input1 = input[:,:num_inputs/2]
input2 = input[:,num_inputs/2:]
output1 = input1
a = T.dot(input1,self.W1)
if self.b1 is not None:
a = a + self.b1
h = self.nonlinearity(a)
s_ = T.dot(h,self.W21)
if self.b21 is not None:
s_ = s_ + self.b21
s = T.exp(s_) + 0.001
ls = T.log(s)
m = T.dot(h,self.W22)
if self.b22 is not None:
m = m + self.b22
output2 = s * input2 + m
output = T.concatenate([output1,output2],1)
return output, ls.sum(1)
class CoupledWNDenseLayer(lasagne.layers.base.Layer):
def __init__(self, incoming, num_units,
W=init.Normal(0.0001),
r=init.Normal(0.0001),
b=init.Constant(0.), nonlinearity=nonlinearities.rectify,
**kwargs):
super(CoupledWNDenseLayer, self).__init__(incoming, **kwargs)
self.nonlinearity = (nonlinearities.identity if nonlinearity is None
else nonlinearity)
self.num_units = num_units
num_inputs1 = int(self.input_shape[1]/2)
num_inputs2 = self.input_shape[1] - num_inputs1
self.W1 = self.add_param(W, (num_inputs1, num_units), name="cpds_W1")
self.W21 = self.add_param(W, (num_units, num_inputs2), name="cpds_W21")
self.W22 = self.add_param(W, (num_units, num_inputs2), name="cdds_W22")
self.r1 = self.add_param(r, (num_units,), name='cpds_r1')
self.r21 = self.add_param(r, (num_inputs2,), name='cpds_r21')
self.r22 = self.add_param(r, (num_inputs2,), name='cpds_r22')
if b is None:
self.b1 = None
self.b21 = None
self.b22 = None
else:
self.b1 = self.add_param(b, (num_units,), name="cpds_b1",
regularizable=False)
self.b21 = self.add_param(b, (num_inputs2,), name="cpds_b21",
regularizable=False)
self.b22 = self.add_param(b, (num_inputs2,), name="cpds_b22",
regularizable=False)
def get_output_shape_for(self, input_shape):
return input_shape
def get_output_for(self, input, **kwargs):
num_inputs = input.shape[1]
input1 = input[:,:num_inputs/2]
input2 = input[:,num_inputs/2:]
output1 = input1
norm1 = T.sqrt(T.sum(T.square(self.W1),axis=0,keepdims=True))
W1 = self.W1 / norm1
norm21 = T.sqrt(T.sum(T.square(self.W21),axis=0,keepdims=True))
W21 = self.W21 / norm21
norm22 = T.sqrt(T.sum(T.square(self.W22),axis=0,keepdims=True))
W22 = self.W22 / norm22
a = self.r1 * T.dot(input1,W1)
if self.b1 is not None:
a = a + self.b1
h = self.nonlinearity(a)
s_ = self.r21 * T.dot(h,W21)
if self.b21 is not None:
s_ = s_ + self.b21
s = T.exp(s_) + 0.001
ls = T.log(s)
m = self.r22 * T.dot(h,W22)
if self.b22 is not None:
m = m + self.b22
output2 = s * input2 + m
output = T.concatenate([output1,output2],1)
return output, ls.sum(1)
def get_wn_params(P,add_param,specs,name,d1,d2):
u,g,b = specs
P['u_{}'.format(name)] = add_param(u,(d1,d2))
P['g_{}'.format(name)] = add_param(g,(d2,))
P['b_{}'.format(name)] = add_param(b,(d2,),
regularizable=False)
def weightnorm(W_,g):
if W_.ndim==4:
W_axes_to_sum = (1,2,3)
W_dimshuffle_args = [0,'x','x','x']
else:
W_axes_to_sum = 0
W_dimshuffle_args = ['x',0]
norm = T.sqrt(T.sum(T.square(W_),axis=W_axes_to_sum,keepdims=True))
W_normed = W_ / norm
return W_normed * g.dimshuffle(*W_dimshuffle_args)
def dot_():
dot = lambda X,W,b: T.dot(X,W) + b.dimshuffle('x',0)
return dot
def weightnormdot(X,W,g,b,operator=dot_(),nonl=None):
h = operator(X,weightnorm(W,g),b)
if nonl is not None:
return nonl(h)
else:
return h
class IAFDenseLayer(lasagne.layers.base.Layer):
def __init__(self, incoming, num_units,
num_hids=1, L=1, cond_bias=False,
W=init.Normal(0.0001),
r=init.Normal(0.0001),
b=init.Constant(0.), nonlinearity=nonlinearities.rectify,
**kwargs):
super(IAFDenseLayer, self).__init__(incoming, **kwargs)
self.nonlinearity = (nonlinearities.identity if nonlinearity is None
else nonlinearity)
self.num_units = num_units
aux_input = T.matrix()
masks = list()
P = dict()
name = 'iaf'
for l in range(L):
made = MADE(aux_input,incoming.input_shape[1],
hidden_sizes=[num_units,]*num_hids,
hidden_activation=None,
random_seed=1234+l)
made.shuffle('Once')
made.shuffle('Full')
# saving the masks generated by `MADE`
masks.append([m.get_value() for m in made.masks])
# initializing parameters,
# # last one is mean
for h,m in enumerate(masks[l]):
d1,d2 = m.shape
get_wn_params(P,self.add_param,[W,r,b],
'{}_l{}h{}'.format(name,l,h),d1,d2)
# # std
get_wn_params(P,self.add_param,[W,r,b],
'{}_l{}h{}s'.format(name,l,h),d1,d2)
# # additional conditional bias
if cond_bias:
d1,d2 = masks[l][0]
get_wn_params(P,self.add_param,[W,r,b],
'{}_l{}cb1'.format(name,l),d1,d2)
if l != L-1:
get_wn_params(P,self.add_param,[W,r,b],
'{}_l{}cb2'.format(name,l),d1,d2)
self.P = P
self.cond_bias = cond_bias
self.num_hids = num_hids
self.L = L
self.masks = masks
def get_output_shape_for(self, input_shape):
return input_shape
def get_output_for(self, input, cond_bias=None, **kwargs):
z = input
zs = list()
zs.append(input)
ss = T.zeros((input.shape[0],)) # logdet jacobian
masks = self.masks
P = self.P
L = self.L
num_hids = self.num_hids
#cond_bias = self.cond_bias
nonl = self.nonlinearity
name = 'iaf'
for l in range(L):
hidden = zs[l]
for h in range(num_hids+1):
mask = masks[l][h]
u = P['u_{}_l{}h{}'.format(name,l,h)]
g = P['g_{}_l{}h{}'.format(name,l,h)]
b = P['b_{}_l{}h{}'.format(name,l,h)]
u_ = T.switch(mask,u,0)
if h != num_hids:
hidden = weightnormdot(hidden,u_,g,b,nonl=nonl)
else:
mean = weightnormdot(hidden,u_,g,b,nonl=None)
if h == 0 and cond_bias is not None:
u = P['u_{}_l{}cb1'.format(name,l,h)]
g = P['g_{}_l{}cb1'.format(name,l,h)]
b = P['b_{}_l{}cb1'.format(name,l,h)]
hidden_cb = weightnormdot(cond_bias,u,g,b,nonl=nonl)
hidden += hidden_cb
if l != L-1:
u = P['u_{}_l{}cb2'.format(name,l,h)]
g = P['g_{}_l{}cb2'.format(name,l,h)]
b = P['b_{}_l{}cb2'.format(name,l,h)]
cond_bias = weightnormdot(hidden_cb,u,g,b,nonl=nonl)
u = P['u_{}_l{}h{}s'.format(name,l,h)]
g = P['g_{}_l{}h{}s'.format(name,l,h)]
b = P['b_{}_l{}h{}s'.format(name,l,h)]
u_ = T.switch(mask,u,0)
std = weightnormdot(hidden,u_,g,b,nonl=exp)
z = mean + std * zs[l]
zs.append(z)
ss += T.sum(T.log(std),1)
return z, ss
class CoupledConv1DLayer(lasagne.layers.base.Layer):
"""
shape[1] should be even number
"""
def __init__(self, incoming, num_units, filter_size,
W=init.GlorotUniform(),
b=init.Constant(0.), nonlinearity=nonlinearities.rectify,
flip_filters=True, convolution=conv.conv1d_mc0, **kwargs):
super(CoupledConv1DLayer, self).__init__(incoming, **kwargs)
self.nonlinearity = (nonlinearities.identity if nonlinearity is None
else nonlinearity)
self.filter_size = filter_size
self.num_units = num_units
self.flip_filters = flip_filters
self.convolution = convolution
W1_shape = (num_units,1,filter_size)
W21_shape = (1,num_units,filter_size)
W22_shape = (1,num_units,filter_size)
self.W1 = self.add_param(W, W1_shape, name="W1")
self.W21 = self.add_param(W, W21_shape, name="W21")
self.W22 = self.add_param(W, W22_shape, name="W22")
if b is None:
self.b1 = None
self.b21 = None
self.b22 = None
else:
self.b1 = self.add_param(b, (num_units,), name="b1",
regularizable=False)
self.b21 = self.add_param(b, (1,), name="b21",
regularizable=False)
self.b22 = self.add_param(b, (1,), name="b22",
regularizable=False)
def get_output_shape_for(self, input_shape):
return input_shape
def get_output_for(self, input, **kwargs):
border_mode = 'half'
num_units = self.num_units
filter_size = self.filter_size
W1_shape = (num_units,1,filter_size)
W21_shape = (1,num_units,filter_size)
W22_shape = (1,num_units,filter_size)
num_inputs = self.input_shape[1]
input1 = input[:,:num_inputs/2]
input2 = input[:,num_inputs/2:]
output1 = input1
input_shape = self.input_shape
input1_shape = (input_shape[0], 1, num_inputs/2)
h_shape = (input_shape[0], num_units, num_inputs/2)
conved = self.convolution(input1.reshape((-1,1,num_inputs/2)), self.W1,
input1_shape, W1_shape,
subsample=(1,),
border_mode=border_mode,
filter_flip=self.flip_filters)
if self.b1 is not None:
a = conved + self.b1.dimshuffle('x',0,'x')
h = self.nonlinearity(a)
conved = self.convolution(h, self.W21,
h_shape, W21_shape,
subsample=(1,),
border_mode=border_mode,
filter_flip=self.flip_filters)
if self.b21 is not None:
s_ = conved + self.b21.dimshuffle('x',0,'x')
s = T.nnet.softplus(s_).reshape((-1,num_inputs/2)) + delta
ls = T.log(s)
conved = self.convolution(h, self.W22,
h_shape, W22_shape,
subsample=(1,),
border_mode=border_mode,
filter_flip=self.flip_filters)
if self.b22 is not None:
m = conved + self.b22.dimshuffle('x',0,'x')
m = m.reshape((-1,num_inputs/2))
output2 = s * input2 + m
output = T.concatenate([output1,output2],1)
return output, ls.sum(1)
class LinearFlowLayer(lasagne.layers.base.Layer):
"""
Scale and shift inputs, elementwise
"""
def __init__(self, incoming, W=init.Normal(0.01,-7),
b=init.Normal(0.01,0),
**kwargs):
super(LinearFlowLayer, self).__init__(incoming, **kwargs)
num_inputs = int(np.prod(self.input_shape[1]))
self.W = self.add_param(W, (num_inputs,), name="lf_W")
if b is None:
self.b = None
else:
self.b = self.add_param(b, (num_inputs,), name="lf_b",
regularizable=False)
def get_output_shape_for(self, input_shape):
return input_shape
def get_output_for(self, input, **kwargs):
s = T.exp(self.W) + delta
output = input * s
if self.b is not None:
output = output + self.b
return output, (T.ones_like(input)*T.log(s)).sum(1)
class ConvexBiasLayer(lasagne.layers.base.Layer):
"""
Scale and shift inputs, elementwise
"""
def __init__(self, incoming, W=init.Normal(0.01,-7),
b=init.Normal(0.01,0),
**kwargs):
super(ConvexBiasLayer, self).__init__(incoming, **kwargs)
num_inputs = int(np.prod(self.input_shape[1]))
self.W = self.add_param(W, (1,), name="lf_W")
self.b = self.add_param(b, (num_inputs,), name="lf_b",
regularizable=False)
def get_output_shape_for(self, input_shape):
return input_shape
def get_output_for(self, input, **kwargs):
s = T.nnet.sigmoid(self.W) + delta
output = input * s + (1-s) * self.b
return output, (T.ones_like(input)*T.log(s)).sum(1)
class IndexLayer(lasagne.layers.Layer):
"""
Return the given index of input tuple
"""
def __init__(self, incoming, index, output_shape=None, **kwargs):
super(IndexLayer, self).__init__(incoming, **kwargs)
self.index = index
self.output_shape_ = output_shape
def get_output_shape_for(self, input_shape):
if self.output_shape_ is not None:
return self.output_shape_
else:
return super(IndexLayer, self).get_output_shape_for(input_shape)
def get_output_for(self, input, **kwargs):
return input[self.index]
class ReverseLayer(lasagne.layers.Layer):
"""
Reverse the order of features
"""
def __init__(self, incoming, num_units, axis=-1, **kwargs):
super(ReverseLayer, self).__init__(incoming, **kwargs)
indices = (np.arange(num_units))[::-1]
self.indices = indices
self.axis = axis
def get_output_shape_for(self, input_shape):
return input_shape
def get_output_for(self, input, **kwargs):
slc = [slice(None)] * input.ndim
slc[self.axis] = self.indices
return input[slc]
class PermuteLayer(lasagne.layers.Layer):
def __init__(self, incoming, num_units, axis=-1, **kwargs):
super(PermuteLayer, self).__init__(incoming, **kwargs)
indices = np.random.permutation(np.arange(num_units))
while np.all(indices == np.arange(num_units)):
indices = np.random.permutation(np.arange(num_units))
self.indices = indices
self.axis = axis
def get_output_shape_for(self, input_shape):
return input_shape
def get_output_for(self, input, **kwargs):
slc = [slice(None)] * input.ndim
slc[self.axis] = self.indices
return input[slc]
class SplitLayer(lasagne.layers.Layer):
def __init__(self,incoming, index, axis=-1, **kwargs):
super(SplitLayer, self).__init__(incoming, **kwargs)
self.index = index
self.axis = axis
def get_output_shape_for(self, input_shape):
index = self.index
axis = self.axis
output_shape1 = input_shape[:axis] + \
(index,) + input_shape[axis+1:]
output_shape2 = input_shape[:axis] + \
(input_shape[axis]-index,) + input_shape[axis+1:]
return output_shape1, output_shape2
def get_output_for(self, input, **kwargs):
index = self.index
axis = self.axis
slc1 = [slice(None)] * input.ndim
slc1[axis] = np.arange(index)
slc2 = [slice(None)] * input.ndim
slc2[axis] = np.arange(index,self.input_shape[axis])
output1 = input[slc1]
output2 = input[slc2]
return output1, output2
class stochasticDenseLayer(lasagne.layers.base.MergeLayer):
def __init__(self, incomings, num_units,
b=init.Constant(0.), nonlinearity=nonlinearities.rectify,
num_leading_axes=1, **kwargs):
super(stochasticDenseLayer, self).__init__(incomings, **kwargs)
self.nonlinearity = (nonlinearities.identity if nonlinearity is None
else nonlinearity)
self.num_units = num_units
if b is None:
self.b = None
else:
self.b = self.add_param(b, (num_units,), name="stocds_b",
regularizable=False)
def get_output_shape_for(self,input_shapes):
input_shape = input_shapes[0]
weight_shape = input_shapes[1]
try:
return (input_shape[0], weight_shape[2])
except:
return (input_shape[0], weight_shape[1])
def get_output_for(self, inputs, **kwargs):
"""
inputs[0].shape = (None, num_inputs)
inputs[1].shape = (None/1, num_inputs, num_units)
"""
input = inputs[0]
W = inputs[1]
activation = T.sum(input.dimshuffle(0,1,'x') * W, axis = 1)
if self.b is not None:
activation = activation + self.b
return self.nonlinearity(activation)
class stochasticDenseLayerWithBias(lasagne.layers.base.MergeLayer):
def __init__(self, incomings, num_units,
nonlinearity=nonlinearities.rectify,
num_leading_axes=1, **kwargs):
super(stochasticDenseLayerWithBias, self).__init__(incomings, **kwargs)
self.nonlinearity = (nonlinearities.identity if nonlinearity is None
else nonlinearity)
self.num_units = num_units
def get_output_shape_for(self,input_shapes):
input_shape = input_shapes[0]
weight_shape = input_shapes[1]
try:
return (input_shape[0], weight_shape[2])
except:
return (input_shape[0], weight_shape[1])
def get_output_for(self, inputs, **kwargs):
"""
inputs[0].shape = (None, num_inputs)
inputs[1].shape = (None/1, num_inputs, num_units)
"""
input = inputs[0]
W = inputs[1]
b = inputs[2]
activation = T.sum(input.dimshuffle(0,1,'x') * W, axis = 1) + b
return self.nonlinearity(activation)
class stochasticDenseLayer2(lasagne.layers.base.MergeLayer):
"""
stochastic dense layer with weightnorm reparameterization
noise on preactivation rescaling parameters
"""
def __init__(self, incomings, num_units, W=init.GlorotUniform(),
b=init.Constant(0.), nonlinearity=nonlinearities.rectify,
num_leading_axes=1, **kwargs):
super(stochasticDenseLayer2, self).__init__(incomings, **kwargs)
self.nonlinearity = (nonlinearities.identity if nonlinearity is None
else nonlinearity)
self.num_units = num_units
input_shape = incomings[0].output_shape
if num_leading_axes >= len(input_shape):
raise ValueError(
"Got num_leading_axes=%d for a %d-dimensional input, "
"leaving no trailing axes for the dot product." %
(num_leading_axes, len(input_shape)))
elif num_leading_axes < -len(input_shape):
raise ValueError(
"Got num_leading_axes=%d for a %d-dimensional input, "
"requesting more trailing axes than there are input "
"dimensions." % (num_leading_axes, len(input_shape)))
self.num_leading_axes = num_leading_axes
if any(s is None for s in input_shape[num_leading_axes:]):
raise ValueError(
"A DenseLayer requires a fixed input shape (except for "
"the leading axes). Got %r for num_leading_axes=%d." %
(input_shape, self.num_leading_axes))
num_inputs = int(np.prod(input_shape[num_leading_axes:]))
self.W = self.add_param(W, (num_inputs, num_units), name="W")
if b is None:
self.b = None
else:
self.b = self.add_param(b, (num_units,), name="b",
regularizable=False)
def get_output_shape_for(self,input_shapes):
return input_shapes[0][:self.num_leading_axes] + (self.num_units,)
input_shape = input_shapes[0]
weight_shape = input_shapes[1]
return (input_shape[0], weight_shape[2])
def get_output_for(self, inputs, **kwargs):
"""
inputs[0].shape = (None, num_inputs)
inputs[1].shape = (None/1, num_units)
"""
input = inputs[0]
r = inputs[1]
norm = T.sqrt(T.sum(T.square(self.W),axis=0,keepdims=True))
W = self.W / norm
num_leading_axes = self.num_leading_axes
if num_leading_axes < 0:
num_leading_axes += input.ndim
if input.ndim > num_leading_axes + 1:
# flatten trailing axes (into (n+1)-tensor for num_leading_axes=n)
input = input.flatten(num_leading_axes + 1)
activation = T.dot(input, W) * r
if self.b is not None:
activation = activation + self.b
return self.nonlinearity(activation)
def stochasticConv2DLayer(incomings, num_filters, filter_size, stride=(1, 1),
pad=0, untie_biases=False, W=init.GlorotUniform(),
b=init.Constant(0.),
nonlinearity=nonlinearities.rectify,
flip_filters=True,
convolution=T.nnet.conv2d, **kwargs):
incoming = incomings[0]
W = incomings[1]
layer = lasagne.layers.Conv2DLayer(
incoming, num_filters, filter_size, stride,
pad, untie_biases, W, b,
nonlinearity, flip_filters,
convolution
)
layer.W = W
return layer
class MNFLayer(lasagne.layers.base.MergeLayer):
def __init__(self, incomings, num_units, num_inputs,
nonlinearity=nonlinearities.rectify,
srgn=srng,
W_mu=lasagne.init.Normal(0.05),
b_mu=init.Constant(0.),
W_logsig=lasagne.init.Normal(0.05 * 1e-6, -9),
b_logsig=lasagne.init.Normal(0.05 * 1e-6, -9),
num_leading_axes=1, **kwargs):
super(MNFLayer, self).__init__(incomings, **kwargs)
self.nonlinearity = (nonlinearities.identity if nonlinearity is None
else nonlinearity)
self.num_units = num_units
self.srng = srng
# add params
self.W_mu = self.add_param(W_mu, (num_inputs, num_units,), name="W_mu", regularizable=False)
self.W_logsig = self.add_param(W_logsig, (num_inputs, num_units,), name="W_logsig", regularizable=False)
self.b_mu = self.add_param(b_mu, (num_units,), name="b_mu", regularizable=False)
self.b_logsig = self.add_param(b_logsig, (num_units,), name="b_logsig", regularizable=False)
def get_output_shape_for(self,input_shapes):
input_shape = input_shapes[0]
weight_shape = input_shapes[1]
try:
return (input_shape[0], weight_shape[2])
except:
return (input_shape[0], weight_shape[1])
def get_output_for(self, inputs, **kwargs):
"""
inputs[0].shape = (None, num_inputs)
inputs[1].shape = (None, num_inputs)
"""
input = inputs[0]
Z = inputs[1] # Z_{T_f}
activation = input * Z
# TODO: here (what is going on?)
mu = T.sum(activation.dimshuffle(0,1,'x') * self.W_mu, axis = 1) + self.b_mu
sig = T.sum((input**2).dimshuffle(0,1,'x') * T.exp(self.W_logsig), axis = 1)**.5 + T.exp(self.b_logsig)
ep = self.srng.normal(size=sig.shape,dtype=floatX)
preactivation = mu + ep * sig
return self.nonlinearity(preactivation)#, mu, sig
class WeightNormLayer(lasagne.layers.Layer):
"""
bias is a new feature I added which allows us to put a posterior over the biases of the model - DK
"""
def __init__(self, incoming, b=lasagne.init.Constant(0.),
g=lasagne.init.Constant(1.),
W=lasagne.init.Normal(0.05),
nonlinearity=nonlinearities.rectify,
bias=None,
**kwargs):
super(WeightNormLayer, self).__init__(incoming, **kwargs)
self.bias = bias
self.nonlinearity = nonlinearity
k = self.input_shape[1]
if b is not None:
self.b = self.add_param(b, (k,), name="b", regularizable=False)
if bias is not None:
self.b = bias
if type(g) == tv:
self.g = g
elif g is not None:
self.g = self.add_param(g, (k,), name="g")
if len(self.input_shape)==4:
self.axes_to_sum = (0,2,3)
self.dimshuffle_args = ['x',0,'x','x']
else:
self.axes_to_sum = 0
self.dimshuffle_args = ['x',0]
# scale weights in layer below
incoming.W_param = incoming.W
incoming.W_param.set_value(W.sample(incoming.W_param.get_value().shape))
if incoming.W_param.ndim==4:
W_axes_to_sum = (1,2,3)
W_dimshuffle_args = [0,'x','x','x']
else:
W_axes_to_sum = 0
W_dimshuffle_args = ['x',0]
if g is not None:
incoming.W = self.g * incoming.W_param / T.sqrt(T.sum(T.square(incoming.W_param),axis=W_axes_to_sum)).dimshuffle(*W_dimshuffle_args)
else:
incoming.W = incoming.W_param / T.sqrt(T.sum(T.square(incoming.W_param),axis=W_axes_to_sum,keepdims=True))
def get_output_for(self, input, init=False, **kwargs):
if init:
m = T.mean(input, self.axes_to_sum)
input -= m.dimshuffle(*self.dimshuffle_args)
stdv = T.sqrt(T.mean(T.square(input),axis=self.axes_to_sum))
input /= stdv.dimshuffle(*self.dimshuffle_args)
self.init_updates = [(self.b, -m/stdv), (self.g, self.g/stdv)]
elif hasattr(self,'b'):
# FIXME (?): we should do this even if self.bias is not None!? (else we aren't using the random biases!?)
if self.bias is None: # else we sample a different bias for each mini-batch (??)
input += self.b.dimshuffle(*self.dimshuffle_args)
return self.nonlinearity(input)
def stochastic_weight_norm(layer, weight, bias=None, **kwargs):
nonlinearity = getattr(layer, 'nonlinearity', None)
if nonlinearity is not None:
layer.nonlinearity = lasagne.nonlinearities.identity
if hasattr(layer, 'b'):
del layer.params[layer.b]
layer.b = None
layer_out = WeightNormLayer(layer, g = weight,
nonlinearity=nonlinearity,
bias=bias,
**kwargs)
return layer_out
# new BHN with WN/BN
from lasagne.layers import InputLayer, ElemwiseSumLayer, BatchNormLayer, \
MergeLayer, get_all_layers, Layer
from lasagne import utils
from difflib import get_close_matches
from warnings import warn
def NVP_dense_layer(incoming,
num_units=200,
L=2,
W=init.Normal(0.0001),
r=init.Normal(0.0001),
b=init.Constant(0.),
nonlinearity=nonlinearities.rectify,):
layer = incoming
shape = layer.output_shape[1]
logdets_layers = list()
for c in range(L):
layer = PermuteLayer(layer,shape)
layer_temp = CoupledWNDenseLayer(layer,num_units,
W=W,r=r,b=b,nonlinearity=nonlinearity)
layer = IndexLayer(layer_temp,0)
logdets_layers.append(IndexLayer(layer_temp,1))
return layer, logdets_layers
def IAF_dense_layer(incoming,
num_units=200,
L=2,
num_hids=1,
W=init.Normal(0.0001),
r=init.Normal(0.0001),
b=init.Constant(0.),
nonlinearity=nonlinearities.rectify):
layer_temp = IAFDenseLayer(incoming,num_units,num_hids,L=L,cond_bias=False)
layer = IndexLayer(layer_temp,0)
logdets_layers = IndexLayer(layer_temp,1)
return layer, [logdets_layers,]
normalizable_layers = [DenseLayer,Conv2DLayer]
nlb = lambda layer: any([isinstance(layer,nl) for nl in normalizable_layers])
def hypernet(net,
hidden_size=512,
layers=0,
flow='IAF',
output_size=None,
nlb = nlb,
copies = 1, # 2 for bias + scale; 1 for scale only
**kargs):
if output_size is None:
all_layers = lasagne.layers.get_all_layers(net)
output_size = sum([layer.output_shape[1] for layer in all_layers
if not isinstance(layer,InputLayer) and
not isinstance(layer,ElemwiseSumLayer) and
nlb(layer)]) * copies
logdets_layers = []
layer = InputLayer(shape=(None,output_size))
layer_temp = LinearFlowLayer(layer)
layer = IndexLayer(layer_temp,0)
logdets_layers.append(IndexLayer(layer_temp,1))
if layers > 0:
if flow == 'RealNVP':
layer, ld_layers = NVP_dense_layer(layer, hidden_size,
layers, **kargs)
elif flow == 'IAF':
layer, ld_layers = IAF_dense_layer(layer, hidden_size,
layers, **kargs)
logdets_layers = logdets_layers + ld_layers
return layer, ElemwiseSumLayer(logdets_layers), output_size
def slicing(params, start_index, size):
end_index = start_index + size
return params[:,start_index:end_index], end_index
def N_get_output(layer_or_layers, inputs, hnet, input_h,
deterministic=False, norm_type='BN',
static_bias=None, nlb=nlb,
test_time=False,
**kwargs):
# check if the keys of the dictionary are valid
if isinstance(inputs, dict):
for input_key in inputs.keys():
if (input_key is not None) and (not isinstance(input_key, Layer)):
raise TypeError("The inputs dictionary keys must be"
" lasagne layers not %s." %
type(input_key))
# track accepted kwargs used by get_output_for
accepted_kwargs = {'deterministic'}
# obtain topological ordering of all layers the output layer(s) depend on
treat_as_input = inputs.keys() if isinstance(inputs, dict) else []
all_layers = get_all_layers(layer_or_layers, treat_as_input)
# initialize layer-to-expression mapping from all input layers
all_outputs = dict((layer, layer.input_var)
for layer in all_layers
if isinstance(layer, InputLayer) and
layer not in treat_as_input)
# update layer-to-expression mapping from given input(s), if any
if isinstance(inputs, dict):
all_outputs.update((layer, utils.as_theano_expression(expr))
for layer, expr in inputs.items())
elif inputs is not None:
if len(all_outputs) > 1:
raise ValueError("get_output() was called with a single input "
"expression on a network with multiple input "
"layers. Please call it with a dictionary of "
"input expressions instead.")
for input_layer in all_outputs:
all_outputs[input_layer] = utils.as_theano_expression(inputs)
N_params = lasagne.layers.get_output(hnet,input_h)
index = 0
if static_bias is not None:
index_b = 0
if static_bias.ndim == 1:
static_bias = static_bias.dimshuffle('x',0)
# update layer-to-expression mapping by propagating the inputs
#last_layer = all_layers[-1]
for layer in all_layers:
if layer not in all_outputs:
try:
if isinstance(layer, MergeLayer):
layer_inputs = [all_outputs[input_layer]
for input_layer in layer.input_layers]
else:
layer_inputs = all_outputs[layer.input_layer]
except KeyError:
# one of the input_layer attributes must have been `None`
raise ValueError("get_output() was called without giving an "
"input expression for the free-floating "
"layer %r. Please call it with a dictionary "
"mapping this layer to an input expression."
% layer)
#if layer is not last_layer and not isinstance(layer,
# ElemwiseSumLayer):
if not isinstance(layer, ElemwiseSumLayer) and \
nlb(layer):
nonlinearity = getattr(layer, 'nonlinearity', None)
if nonlinearity is not None:
layer.nonlinearity = lambda x: x