/
conv.py
2365 lines (2102 loc) · 92.8 KB
/
conv.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
"""
Contains an Op for convolving input images with a set of filters. This was
developed especially for Convolutional Neural Networks.
For related ops, including downsampling and subsampling, see
tensor.signal and tensor.signal.pool.
See especially conv2d().
"""
from __future__ import absolute_import, print_function, division
import logging
import numpy as np
from six.moves import xrange
import warnings
import theano
from theano import OpenMPOp
from theano.tensor import (as_tensor_variable, blas, get_scalar_constant_value,
patternbroadcast, NotScalarConstantError)
from theano.gof import Apply
from theano.tensor.nnet.abstract_conv import (get_conv_output_shape,
get_conv_shape_1axis)
try:
# TODO: move these back out to global scope when they no longer
# cause an atexit error
from scipy.signal.signaltools import _valfrommode, _bvalfromboundary
from scipy.signal.sigtools import _convolve2d
imported_scipy_signal = True
except ImportError:
imported_scipy_signal = False
__docformat__ = "restructuredtext en"
_logger = logging.getLogger("theano.tensor.nnet.conv")
def conv2d(input, filters, image_shape=None, filter_shape=None,
border_mode='valid', subsample=(1, 1), **kargs):
"""
Deprecated, old conv2d interface.
This function will build the symbolic graph for convolving a stack of
input images with a set of filters. The implementation is modelled after
Convolutional Neural Networks (CNN). It is simply a wrapper to the ConvOp
but provides a much cleaner interface.
Parameters
----------
input : symbolic 4D tensor
Mini-batch of feature map stacks, of shape
(batch size, stack size, nb row, nb col)
see the optional parameter image_shape
filters: symbolic 4D tensor
Set of filters used in CNN layer of shape
(nb filters, stack size, nb row, nb col)
see the optional parameter filter_shape
border_mode : {'valid', 'full'}
'valid'only apply filter to complete patches of the image. Generates
output of shape: image_shape - filter_shape + 1.
'full' zero-pads image to multiple of filter shape to generate output
of shape: image_shape + filter_shape - 1.
subsample: tuple of len 2
Factor by which to subsample the output. Also called strides elsewhere.
image_shape: None, tuple/list of len 4 of int, None or Constant variable
The shape of the input parameter.
Optional, used for optimization like loop unrolling
You can put None for any element of the list to tell that this element
is not constant.
filter_shape : None, tuple/list of len 4 of int, None or Constant variable
Optional, used for optimization like loop unrolling
You can put None for any element of the list
to tell that this element is not constant.
kwargs
Kwargs are passed onto ConvOp. Can be used to set the following:
unroll_batch, unroll_kern, unroll_patch, openmp (see ConvOp doc).
openmp: By default have the same value as
config.openmp. For small image, filter,
batch size, nkern and stack size, it can be
faster to disable manually openmp. A fast and
incomplete test show that with image size
6x6, filter size 4x4, batch size==1,
n kern==1 and stack size==1, it is faster
to disable it in valid mode. But if we
grow the batch size to 10, it is faster
with openmp on a core 2 duo.
Returns
-------
symbolic 4D tensor
Set of feature maps generated by convolutional layer. Tensor is
of shape (batch size, nb filters, output row, output col).
"""
# accept Constant value for image_shape and filter_shape.
if image_shape is not None:
image_shape = list(image_shape)
for i in xrange(len(image_shape)):
if image_shape[i] is not None:
try:
image_shape[i] = get_scalar_constant_value(
as_tensor_variable(image_shape[i]))
except NotScalarConstantError:
raise NotScalarConstantError(
"The convolution need that the shape"
" information are constant values. We got"
" %s for the image_shape parameter" %
image_shape[i])
assert image_shape[i].dtype in theano.tensor.discrete_dtypes
image_shape[i] = int(image_shape[i])
if filter_shape is not None:
filter_shape = list(filter_shape)
for i in xrange(len(filter_shape)):
if filter_shape[i] is not None:
try:
filter_shape[i] = get_scalar_constant_value(
as_tensor_variable(filter_shape[i]))
except NotScalarConstantError:
raise NotScalarConstantError(
"The convolution need that the shape"
" information are constant values. We got"
" %s for the filter_shape "
"parameter" % filter_shape[i])
assert filter_shape[i].dtype in theano.tensor.discrete_dtypes
filter_shape[i] = int(filter_shape[i])
if image_shape and filter_shape:
try:
if image_shape[1] is not None and filter_shape[1] is not None:
assert image_shape[1] == filter_shape[1]
except Exception:
print('image ', image_shape, ' filters ', filter_shape)
raise
if filter_shape is not None:
nkern = filter_shape[0]
kshp = filter_shape[2:]
else:
nkern, kshp = None, None
if image_shape is not None:
bsize = image_shape[0]
imshp = image_shape[1:]
else:
bsize, imshp = None, None
op = ConvOp(output_mode=border_mode, dx=subsample[0], dy=subsample[1],
imshp=imshp, kshp=kshp, nkern=nkern, bsize=bsize, **kargs)
return op(input, filters)
class ConvOp(OpenMPOp):
"""
This Op serves a dual purpose: it can implement a vanilla 2D convolution
(as taught in any signal processing class) or implement the
convolutional layers found in Convolutional Neural Networks.
In this setting, a set of 3D images is convolved with a set of 3D kernels,
with the particularity that their leading dimensions are of equal length.
Vanilla 2D convolution is treated as a special case of this.
The input parameter represents a mini-batch of multiple images. Its shape is:
batch size x num. input feature maps x image height x image width
The kernel parameter represents a set of 3D kernels. Its shape is:
number of filters x num. input images x filter height x filter width
The output of ConvOp is a 4D tensor, generated as follows:
output[b,k,:,:] = \sum_i input[b,i,:,:] * filter[k,i,:,:] \forall b,k
where b is the mini-batch index, k the filter index and * is the
convolution operator.
The constructor initializes a ConvOp with given output_mode (full/valid).
All other parameters are optional and are only used to generate more
optimized c code, or to enable graph optimizers to optimally replace the
ConvOp.
NOTES ON OPTIMIZATION:
There are two types of optimization. The first is the selection of the
fastest algo when bsize and nkern are provided with imshp and kshp.
By default we try to select the fastest version. You can specify it
with the unroll_batch, unroll_kern, and unroll_patch parameter.
The second type of optimization is hardcoding some dimensions into the
code when all shape are know.
This make a significant difference for the 'full' output_mode.
Sometimes, the fastest implementation on x86-64 uses
{unroll_batch=4, unroll_kern=4, unroll_patch=False}
with all other shape parameters being provided.
For optimizing other architectures, see:
Kazushige Goto and Robert A. Van De Geijn, Anatomy of High-Performance
Matrix Multiplication, (mr x nr). ACM Transactions on Mathematical
Software, May 2008.
Figure 12: (mr x nr). For x86 use 2x4, itanium 8x8, etc.
Parameters
----------
output_mode : {'valid', 'full'}
'valid' gives an output smaller then the image.
'full' gives an output bigger then the image.
See 'border_mode' in conv2d's doc.
Optional parameters: (will generate more optimal c code)
imshp : tuple of len 2 or 3: 2 for 2d image, 3 for a stack of 2d images.
Stacksize, nb image row, nb image col.
kshp : tuple of len 2
Nb kernel row, nb kernel col.
nkern : int
The number of kernel.
bsize : int
The size of the minibatch.
dx : int
Patch stride rows.
dy : int
Patch stride cols
Params which select the version of code used:
unroll_patch : bool
Use a version of c_code that unroll the patch loop that don't
request all shape information to work, but if all shape information
are present, will use it to hardcode the value in the code for
faster code.
unroll_batch : int
Use a version of c_code that unroll the batch (by unroll_batch)
and the nkern (by unroll_kern) loop. The size must by a multiple
of bsize or nkern respectively.
unroll_kern : int
Use a version of c_code that unroll the batch
(by unroll_batch) and the nkern(by unroll_kern) loop. The size
must by a multiple of bsize or nkern respectively.
verbose : int
Passed to GpuConv.
version: int or str
Passed to GpuConv, if version='no_fft', fft
optimization will be desactivated at the op level.
direction_hint: {'forward', 'bprop weights', 'bprop inputs'}
Passed to GpuConv, used by graph optimizers to aid algorithm choice.
The 3 following parameters are used internally when we generate
the gradient when dx!=1 or dy!=1.
imshp_logical
Default None. None value is equivalent to imshp value.
When imshp_logical != imshp, it tell we need to insert 0 in
the image before we do the convolution. For example, when dx==dy==2
and the image is [[1, 2], [3, 4]], we should make as if the image
was [[1, 0, 2, 0], [0, 0, 0, 0], [3, 0, 4, 0], [0, 0, 0, 0]].
Our python code insert the zero, but the c code optimize it.
imshp_logical != imshp when taking the grad again the weights or
the image when the output_mode is full and `dx != 1` or `dy != 1`.
kshp_logical
Idem but for kshp and used for the grad again the
weights when the output_mode is valid and `dx != 1` or `dy != 1`.
kshp_logical_top_aligned
Used in the same case. Default to True.
Set to False in the grad again the weight when the
output_mode is full.
"""
__attrnames = ['imshp', 'kshp', 'nkern', 'bsize', 'dx', 'dy', 'out_mode',
'unroll_batch', 'unroll_kern', 'unroll_patch',
'imshp_logical', 'kshp_logical', 'kshp_logical_top_aligned']
"""These attributes uniquely identify the behaviour of this op for
given inputs. Do not set openmp here.
"""
# the value of speed_unroll_batch_kern,speed_unroll_patch_noshape,speed_unroll_patch_shape
# have bean calculated on maggie36 when their is only 1 session logged on and only this was running.
# It is an Intel(R) Xeon(R) CPU E5430 @ 2.66GHz. It is computer with theano/tensor/nnet/tests/speed_test_conv.py
# and took 5 minutes to run.
# TODO: we should compute this table for each computer/os as this can change.
# I saw on one computer that the speed with the shape can be slower than without!
# using the real shape and the same dtype could also help.
# unroll_batch, unroll_kern, valid time, full time
speed_unroll_batch_kern = [(1, 1, 2.4661250114440918, 6.5472931861877441),
(1, 2, 1.5869178771972656, 5.1499760150909424),
(1, 3, 1.4270510673522949, 3.6593470573425293),
(1, 4, 1.3373479843139648, 3.3451821804046631),
(1, 5, 1.2818830013275146, 3.1444568634033203),
(1, 6, 1.2521560192108154, 3.0256359577178955),
(1, 10, 1.2134110927581787, 2.9174180030822754),
(2, 1, 1.657214879989624, 4.5261678695678711),
(2, 2, 1.2123160362243652, 2.9747390747070312),
(2, 3, 1.0758891105651855, 2.5690360069274902),
(2, 4, 1.0683329105377197, 2.4233770370483398),
(2, 5, 1.0955719947814941, 2.3999948501586914),
(2, 6, 1.5935721397399902, 2.6878271102905273),
(2, 10, 1.8511250019073486, 3.2417428493499756),
(3, 1, 1.5948119163513184, 3.631148099899292),
(3, 2, 1.0761330127716064, 2.6011371612548828),
(3, 3, 1.0551531314849854, 2.4200370311737061),
(3, 4, 1.3930759429931641, 2.5211219787597656),
(3, 5, 1.4330689907073975, 2.5704989433288574),
(3, 6, 1.362138032913208, 2.5964410305023193),
(3, 10, 1.6582000255584717, 2.9907989501953125),
(4, 1, 1.4793620109558105, 3.3473429679870605),
(4, 2, 1.0671560764312744, 2.4171769618988037),
(4, 3, 1.2569692134857178, 2.2807950973510742),
(4, 4, 1.3456289768218994, 2.6219108104705811),
(4, 5, 1.4055080413818359, 2.4606490135192871),
(4, 6, 1.372107982635498, 2.551663875579834),
(4, 10, 1.599470853805542, 2.9172940254211426),
(5, 1, 1.4115700721740723, 3.2077109813690186),
(5, 2, 1.0635769367218018, 2.2648060321807861),
(5, 3, 1.3842809200286865, 2.6135518550872803),
(5, 4, 1.3470511436462402, 2.3852400779724121),
(5, 5, 1.3539440631866455, 2.5245928764343262),
(5, 6, 1.4037849903106689, 2.5985310077667236),
(5, 10, 1.6120610237121582, 2.8127608299255371),
(6, 1, 1.3623628616333008, 3.021122932434082),
(6, 2, 1.1697649955749512, 2.6285450458526611),
(6, 3, 1.2980999946594238, 2.4746189117431641),
(6, 4, 1.3739941120147705, 2.5579929351806641),
(6, 5, 1.3967819213867188, 2.5522029399871826),
(6, 6, 1.4279270172119141, 2.6127138137817383),
(6, 10, 1.605496883392334, 2.864037036895752),
(10, 1, 1.6401121616363525, 2.970099925994873),
(10, 2, 1.46710205078125, 2.7231831550598145),
(10, 3, 1.4193780422210693, 2.6087639331817627),
(10, 4, 1.4657118320465088, 2.6246678829193115),
(10, 5, 1.5052611827850342, 2.6542458534240723),
(10, 6, 1.5214400291442871, 2.7243161201477051),
(10, 10, 1.6116268634796143, 2.956165075302124)]
# valid time, full time
speed_unroll_patch_noshape = [2.0109100341796875, 5.8175678253173828]
# valid time, full time
speed_unroll_patch_shape = [1.2967290878295898, 5.5283889770507812]
@staticmethod
def has_all_shape(imshp, kshp, nkern=1, bsize=1):
return (nkern is not None and bsize is not None and
all(shp is not None for shp in imshp) and
all(shp is not None for shp in kshp))
@staticmethod
def getOutputShape(inshp, kshp, stride=(1, 1), mode='valid'):
"""
Computes the output dimensions of convolving an image of shape "inshp"
with kernels of shape "kshp". Accepts symbolic or integer shapes.
Propagates `None`s (for unknown shapes).
Parameters
----------
inshp
(rows,cols) of input image.
kshp
(rows,cols) of filters.
mode: {'valid', 'full'}
See 'border_mode' in conv2d's doc.
Returns
-------
object
(rows,cols) of output image.
"""
# The formula would be ceil((i + s * k - s * 1) / float(d)),
# with s=1 for mode=='full' and s=-1 for mode=='valid'.
# To support symbolic shapes, we express this with integer arithmetics.
warnings.warn("The method `getOutputShape` is deprecated use"
"`get_conv_output_shape` instead.", stacklevel=2)
return tuple(get_conv_shape_1axis(i, k, mode, d)
for i, k, d in zip(inshp, kshp, stride))
def __init__(self, imshp=None, kshp=None, nkern=None, bsize=None,
dx=1, dy=1,
output_mode='valid',
unroll_batch=None,
unroll_kern=None,
unroll_patch=None,
imshp_logical=None,
kshp_logical=None,
kshp_logical_top_aligned=True,
verbose=0,
version=-1,
direction_hint='forward',
openmp=None):
# Deactivate fft_optimization at the op level if specified
if version == "no_fft":
self.fft_opt = False
version = -1
else:
self.fft_opt = True
# Expand unknown image / kernel shapes into tuples of Nones
if imshp is None:
imshp = (None, None, None)
else:
imshp = tuple(imshp)
if kshp is None:
kshp = (None, None)
else:
kshp = tuple(kshp)
# Check imshp and kshp dimensionality
if len(imshp) == 2:
imshp = (1,) + imshp
elif len(imshp) != 3:
raise ValueError("len(imshp) must be 2 or 3, got %d" % len(imshp))
if len(kshp) != 2:
raise ValueError("len(kshp) must be 2, got %d" % len(kshp))
# We must continue to consider None as 1 for backward compatibility.
if dx is None:
dx = 1
if dy is None:
dy = 1
if int(dx) != dx:
raise TypeError('ConvOp.__init__ param dx must be an int', dx)
dx = int(dx)
if int(dy) != dy:
raise TypeError('ConvOp.__init__ param dy must be an int', dy)
dy = int(dy)
all_shape = self.has_all_shape(imshp, kshp, nkern, bsize)
if (unroll_batch or unroll_kern) and not all_shape:
raise Exception("In ConvOp, when using unroll_batch and"
" unroll_nkern, all shape are needed")
# Init the openmp attribute
super(ConvOp, self).__init__(openmp=openmp)
if not all_shape or self.openmp:
# Only this version is parallelized
unroll_patch = True
self.imshp = imshp
self.kshp = kshp
self.nkern = nkern
self.bsize = bsize
self.dx = dx
self.dy = dy
self.verbose = verbose
self.version = version
self.direction_hint = direction_hint
# a triple
if imshp_logical is None:
self.imshp_logical = self.imshp
else:
imshp_logical = tuple(imshp_logical)
if len(imshp_logical) != 3:
raise ValueError("len(imshp_logical) must be 3, got %d" % len(imshp_logical))
self.imshp_logical = imshp_logical
# a pair
if kshp_logical is None:
self.kshp_logical = self.kshp
else:
kshp_logical = tuple(kshp_logical)
if len(kshp_logical) != 2:
raise ValueError("len(kshp_logical) must be 2, got %d" % len(kshp_logical))
self.kshp_logical = kshp_logical
# a bool
self.kshp_logical_top_aligned = kshp_logical_top_aligned
self.unroll_batch = unroll_batch
self.unroll_kern = unroll_kern
self.unroll_patch = unroll_patch
if self.unroll_batch and not self.unroll_kern:
self.unroll_kern = 1
if self.unroll_kern and not self.unroll_batch:
self.unroll_batch = 1
# downcast unroll_batch if not a divisor of batch size
if self.unroll_batch is not None and self.unroll_batch > 0 and self.bsize % self.unroll_batch != 0:
if self.bsize <= self.unroll_batch:
self.unroll_batch = self.bsize
else:
# find the maximum value under unroll_batch that would work
new = self.unroll_batch
assert(new >= 1)
while self.bsize % new != 0:
new -= 1
warnstr = ("OPTIMISATION WARNING: in ConvOp.__init__() "
"unroll_batch(%i) must be 0 or a divisor of"
" bsize(%i). We revert it to %i. This"
" won't change the result, but may make it slower.")
_logger.warn(warnstr, self.unroll_batch, self.bsize, new)
self.unroll_batch = new
# downcast unroll_kern if not a divisor of nb of kernel
if self.unroll_kern is not None and self.unroll_kern > 0 and self.nkern % self.unroll_kern != 0:
if self.nkern <= self.unroll_kern:
self.unroll_kern = self.nkern
else:
# find the maximum value under unroll_kern that would work
new = self.unroll_kern
assert(new >= 1)
while self.nkern % new != 0:
new -= 1
warnstr = ("OPTIMISATION WARNING: in ConvOp.__init__()"
" unroll_kern(%i) should be 0 or a divisor of"
" nkern(%i). We revert it to %i. This"
" won't change the result, but may make it slower.")
_logger.warn(warnstr, self.unroll_kern, self.nkern, new)
self.unroll_kern = new
self.outshp = get_conv_output_shape(
(None,) + self.imshp_logical,
(None, None,) + self.kshp_logical,
output_mode,
(dx, dy))[2:]
self.fulloutshp = get_conv_output_shape(
(None,) + self.imshp_logical,
(None, None,) + self.kshp_logical,
output_mode,
(1, 1))[2:]
self.out_mode = output_mode
if self.out_mode not in ["valid", "full"]:
raise Exception("Mode %s not implemented" % str(self.out_mode))
if any((shp is not None) and (shp <= 0) for shp in self.outshp):
raise Exception("Bad size for the output shape. Verify that [post-"
"supersampling] input shape (%s) and kern"
" shape(%s) are ok. (Hint: kerns must fit inside"
" image in valid mode)" %
(self.imshp_logical, self.kshp_logical))
if (self.unroll_kern is None and
self.unroll_batch is None and
self.unroll_patch is None):
# no version specified. Find the faster we have
if self.bsize is None and self.nkern is None:
self.unroll_patch = True
elif self.bsize is not None and self.nkern is not None:
bsize = self.bsize
nkern = self.nkern
mode_idx = 0
if self.out_mode != "valid":
mode_idx = 1
if self.has_all_shape(self.imshp, self.kshp):
time_unroll_patch = self.speed_unroll_patch_shape[mode_idx]
else:
time_unroll_patch = self.speed_unroll_patch_noshape[
mode_idx]
time_unroll_batch_kern = 9999999
for i in xrange(len(self.speed_unroll_batch_kern)):
if (bsize % self.speed_unroll_batch_kern[i][0] == 0 and
nkern % self.speed_unroll_batch_kern[i][1] == 0):
if self.speed_unroll_batch_kern[i][2 + mode_idx] < time_unroll_batch_kern:
time_unroll_batch_kern = self.speed_unroll_batch_kern[i][2 + mode_idx]
time_unroll_batch_kern_idx = i
if time_unroll_patch < time_unroll_batch_kern:
self.unroll_patch = True
else:
self.unroll_batch = self.speed_unroll_batch_kern[
time_unroll_batch_kern_idx][0]
self.unroll_kern = self.speed_unroll_batch_kern[
time_unroll_batch_kern_idx][1]
self.unroll_patch = False
_logger.debug("AUTO FIND VERSION OF C_CODE OF CONV OP "
"%s %s %s %s %s %s %s",
self.unroll_batch, self.unroll_kern,
self.unroll_patch,
self.bsize, self.nkern, time_unroll_patch,
time_unroll_batch_kern)
self._rehash()
def __eq__(self, other):
if type(self) != type(other):
return False
for a in self.__attrnames:
if getattr(self, a) != getattr(other, a):
return False
return True
def __setstate__(self, d):
super(ConvOp, self).__setstate__(d)
self.direction_hint = d.get("direction_hint", None)
self._rehash()
def _rehash(self):
hashval = hash(type(self))
for a in self.__attrnames:
hashval = hashval ^ hash(getattr(self, a))
self.__hashval = hashval
def __hash__(self):
return self.__hashval
def __str__(self):
return "ConvOp{" + ",".join(str((a, getattr(self, a)))
for a in self.__attrnames) + "}"
def flops(self, inputs, outputs):
"""
Useful with the hack in profiling to print the MFlops.
"""
images, kerns = inputs
out, = outputs
assert images[1] == kerns[1]
flops = 0
if self.out_mode == "valid":
# nb mul and add by output pixel
flops = kerns[2] * kerns[3] * 2
# nb flops by output image
flops *= out[2] * out[3]
# nb patch multiplied
flops *= images[1] * kerns[0] * images[0]
else:
flops = (images[0] * kerns[0] * images[1] *
kerns[2] * kerns[3] *
images[2] * images[3] * 2)
return flops
def make_node(self, inputs, kerns):
# TODO: find a way to make ConvOp work for N-D (after NIPS09)
"""
Parameters
----------
inputs
4 dim: batches x stacksize x rows x cols.
kerns
4 dim: nkern x stackidx x rows x cols.
"""
_inputs = as_tensor_variable(inputs)
_kerns = as_tensor_variable(kerns)
# TODO: lift this restriction by upcasting either inputs or kerns
if _inputs.ndim != 4:
raise TypeError('ConvOp (make_node) requires input be a 4D tensor;'
' received "%s" (%i dims)' %
(inputs, _inputs.ndim))
if _kerns.ndim != 4:
raise TypeError('make_node requires 4D tensor of kernels')
if _inputs.type.dtype != _kerns.type.dtype:
raise NotImplementedError(
"The image and the kernel must have the same type."
"inputs(%s), kerns(%s)" % (_inputs.dtype, _kerns.dtype))
bcastable23 = [self.outshp[0] == 1, self.outshp[1] == 1]
output = theano.tensor.tensor(dtype=_inputs.type.dtype,
broadcastable=[_inputs.broadcastable[0],
_kerns.broadcastable[0]] +
bcastable23)
return Apply(self, [_inputs, _kerns], [output])
def infer_shape(self, node, input_shapes):
imshp = input_shapes[0] # 4D image shape
kshp = input_shapes[1] # 4D filter shape
bsize, imshp = imshp[0], list(imshp[1:])
nkern, kshp = kshp[0], list(kshp[2:])
# replace symbolic shapes with known shapes
if self.bsize is not None:
bsize = self.bsize
for i in [0, 1, 2]:
if self.imshp_logical[i] is not None:
imshp[i] = self.imshp_logical[i]
if self.nkern is not None:
nkern = self.nkern
for i in [0, 1]:
if self.kshp_logical[i] is not None:
kshp[i] = self.kshp_logical[i]
# infer output shape from what we have
res = get_conv_output_shape(
(bsize,) + tuple(imshp),
(nkern, None,) + tuple(kshp),
self.out_mode,
(self.dx, self.dy))
return [res]
def perform(self, node, inp, out):
"""
By default if len(img2d.shape)==3, we TODO
"""
img2d, filtersflipped = inp
z, = out
if not imported_scipy_signal:
raise theano.gof.utils.MethodNotDefined(
"c_headers", type(self), self.__class__.__name__,
"Need the python package for scipy.signal to be installed "
"for the python implementation. You can use the C"
" implementation instead.")
# TODO: move these back out to global scope when they no longer
# cause an atexit error
imshp = self.imshp
if any(x is None for x in imshp):
imshp = tuple(img2d.shape[1:])
if imshp != img2d.shape[1:]:
raise ValueError("The image shape provided at build time "
"is different from the one passed at run time",
imshp, img2d.shape[1:])
kshp = self.kshp
if any(x is None for x in kshp):
kshp = tuple(filtersflipped.shape[2:])
if kshp != filtersflipped.shape[2:]:
raise ValueError("The filter shape provided at build time "
"is different from the one passed at run time",
kshp, filtersflipped.shape[2:])
bsize = self.bsize
if bsize is None:
bsize = img2d.shape[0]
elif bsize != img2d.shape[0]:
raise ValueError("The batch size provided at build time "
"is different from the one passed at run time",
bsize, img2d.shape[0])
nkern = self.nkern
if nkern is None:
nkern = filtersflipped.shape[0]
elif nkern != filtersflipped.shape[0]:
raise ValueError("The number of filters provided at build time "
"is different from the one passed at run time",
nkern, filtersflipped.shape[0])
imshp_logical = self.imshp_logical
if imshp_logical[0] is None:
imshp_logical = (imshp[0],) + imshp_logical[1:]
if imshp_logical[1] is None:
imshp_logical = (imshp_logical[0], imshp[1], imshp_logical[2])
if imshp_logical[2] is None:
imshp_logical = imshp_logical[:2] + (imshp[2],)
assert all(x is not None for x in imshp_logical)
kshp_logical = self.kshp_logical
if kshp_logical[0] is None:
kshp_logical = (kshp[0], kshp_logical[1])
if kshp_logical[1] is None:
kshp_logical = (kshp_logical[0], kshp[1])
assert all(x is not None for x in kshp_logical)
if all(shp is not None for shp in self.fulloutshp):
fulloutshp = tuple(self.fulloutshp)
else:
fulloutshp = get_conv_output_shape(
(None,) + imshp_logical,
(None, None,) + kshp_logical,
self.out_mode,
(1, 1))[2:]
if z[0] is None or z[0].shape != (bsize, nkern,) + fulloutshp:
z[0] = np.zeros((bsize, nkern,) + fulloutshp,
dtype=img2d.dtype)
zz = z[0]
stacklen = imshp[0]
img2d = img2d.reshape((bsize,) + imshp)
filtersflipped = filtersflipped.reshape((nkern, stacklen) + kshp)
if self.imshp != self.imshp_logical:
# assuming that to get from imshp to imshp logical we insert zeros in missing spots
rstride = int(np.ceil(imshp_logical[1] / float(imshp[1])))
cstride = int(np.ceil(imshp_logical[2] / float(imshp[2])))
buf = np.zeros((bsize,) + imshp_logical, dtype=img2d.dtype)
buf[:, :, ::rstride, ::cstride] = img2d
img2d = buf
del buf, rstride, cstride
if kshp != kshp_logical:
rstride = int(np.ceil(kshp_logical[0] / float(kshp[0])))
cstride = int(np.ceil(kshp_logical[1] / float(kshp[1])))
buf = np.zeros((nkern, stacklen) +
self.kshp_logical, dtype=filtersflipped.dtype)
if self.kshp_logical_top_aligned:
roffset = coffset = 0
else:
roffset = (kshp_logical[0] - (kshp[0] *
rstride) - 1 + rstride) % rstride
coffset = (kshp_logical[1] - (kshp[1] *
cstride) - 1 + cstride) % cstride
assert roffset >= 0
assert coffset >= 0
buf[:, :, roffset::rstride, coffset::cstride] = filtersflipped
filtersflipped = buf
del buf, rstride, cstride
val = _valfrommode(self.out_mode)
bval = _bvalfromboundary('fill')
with warnings.catch_warnings():
warnings.simplefilter('ignore', np.ComplexWarning)
for b in xrange(bsize):
for n in xrange(nkern):
zz[b, n, ...].fill(0)
for im0 in xrange(stacklen):
# some cast generates a warning here
zz[b, n, ...] += _convolve2d(img2d[b, im0, ...],
filtersflipped[n, im0, ...],
1, val, bval, 0)
if False:
if False and self.out_mode == "full":
img2d2 = np.zeros((bsize, stacklen,
imshp[1] + 2 * kshp[0] - 2,
imshp[2] + 2 * kshp[1] - 2))
img2d2[:, :, kshp[0] - 1:kshp[0] - 1 + imshp[1],
kshp[1] - 1:kshp[1] - 1 + imshp[2]] = img2d
img2d = img2d2
# N_image_shape = image_data.shape
for b in xrange(bsize):
for n in xrange(nkern):
zz[b, n, ...].fill(0)
for im0 in xrange(stacklen):
for row in xrange(0, zz.shape[2], self.dx):
for col in xrange(0, zz.shape[3], self.dy):
zz[b, n, row, col] += (
img2d[b, im0, row:row + kshp[0],
col:col + kshp[1]] *
filtersflipped[n, im0, ::-1, ::-1]).sum()
# We copy it to remove the Stride mismatch warning from DEBUG_MODE.
# The copy make that we return an object with the same stride as the c version.
# The copy don't affect the performence during our experience as in that case we
# execute the c version which is much faster.
if self.dx > 1 or self.dy > 1:
zz = zz[:, :, 0::self.dx, 0::self.dy].copy()
z[0] = zz
def R_op(self, inputs, eval_points):
rval = None
if eval_points[0] is not None:
rval = self.make_node(eval_points[0], inputs[1]).outputs[0]
if eval_points[1] is not None:
if rval is None:
rval = self.make_node(inputs[0], eval_points[1]).outputs[0]
else:
rval += self.make_node(inputs[0], eval_points[1]).outputs[0]
return [rval]
def grad(self, inp, grads):
inputs, kerns = inp
gz, = grads
if self.imshp != self.imshp_logical or self.kshp != self.kshp_logical:
raise NotImplementedError('todo')
if self.out_mode == 'valid' and (self.dx, self.dy) != (1, 1):
# Use the gradient as defined in conv3D, because the implementation
# by Conv is slow (about 3x slower than conv3D, and probably 10x
# slower than it could be), and incorrect when dx or dy > 2.
# build a "node", that should be equivalent to the one given by
# self.make_node, but using conv3D instead of self.
shuffled_inputs = inputs.dimshuffle(0, 2, 3, 'x', 1)
if inputs.name is not None:
shuffled_inputs.name = 'shuffle_for_conv3D(%s)' % inputs.name
flipped_kerns = kerns[:, :, ::-1, ::-1]
if kerns.name is not None:
flipped_kerns.name = 'flipped(%s)' % kerns.name
shuffled_kerns = flipped_kerns.dimshuffle(0, 2, 3, 'x', 1)
if flipped_kerns.name is not None:
shuffled_kerns.name = 'shuffled_for_conv3D(%s)' % flipped_kerns.name
tmp_node = theano.tensor.nnet.conv3D(
V=shuffled_inputs,
W=shuffled_kerns,
b=theano.tensor.alloc(np.asarray(0, dtype=kerns.dtype),
kerns.shape[0]),
d=(self.dx, self.dy, 1))
node = theano.tensor.addbroadcast(
tmp_node, 3).dimshuffle(0, 4, 1, 2)
# mimic what happens inside theano.grad: get the input gradient
# of the final cost wrt all variables involved.
return theano.gradient.grad(cost=None, known_grads={node: gz},
wrt=[inputs, kerns])
if self.dx not in (1, 2) or self.dy not in (1, 2):
raise NotImplementedError(
"ERROR: We disable ConvOp.grad now when output_mode is not"
" 'valid' and dx or dy are greater than 2, as there is a bug"
" in it. See `abstract_conv2d <>`_ for a version that support this.")
all_shape = self.has_all_shape(self.imshp, self.kshp,
self.nkern, self.bsize)
if not all_shape and (self.dx != 1 or self.dy != 1):
raise Exception("ConvOp.grad when dx!=1 or dy!=1 we must have all "
"the optional shape information")
# Determine gradient on kernels ########
assert inputs.ndim == 4 and kerns.ndim == 4
newin = inputs.dimshuffle((1, 0, 2, 3))
newgz = gz.dimshuffle((1, 0, 2, 3))
if self.out_mode == 'valid':
(img, filters) = (newin, newgz)
kshp_logical = self.fulloutshp
kshp_logical_top_aligned = False
imshp_logical = None
(bsize, nkern) = (self.imshp[0], self.nkern)
imshp = (self.bsize, self.imshp[1], self.imshp[2])
kshp = self.outshp
elif self.out_mode == 'full':
(img, filters) = (newgz, newin)
kshp_logical = None
kshp_logical_top_aligned = True
imshp_logical = (self.bsize,
self.fulloutshp[0],
self.fulloutshp[1])
(bsize, nkern) = (self.nkern, self.imshp[0])
imshp = (self.bsize, self.outshp[0], self.outshp[1])
kshp = self.imshp[1:]
else:
raise NotImplementedError(
'Only [full,valid] modes are currently supported.')
filters = filters[:, :, ::-1, ::-1] # flip them
dw = ConvOp(imshp, kshp, nkern, bsize, 1, 1, output_mode='valid',
unroll_batch=None, unroll_kern=None, unroll_patch=None,
imshp_logical=imshp_logical,
kshp_logical=kshp_logical,
kshp_logical_top_aligned=kshp_logical_top_aligned,
version=self.version,
direction_hint='bprop weights',
verbose=self.verbose)
dw = dw(img, filters)
if all_shape:
assert all(o == k for o, k in zip(dw.owner.op.outshp, self.kshp))
if self.out_mode == 'valid':
# before DimShuffle, dw is of shape visdim x nkern x kshp[0] x kshp[1]
dw = dw.dimshuffle((1, 0, 2, 3))
dw = dw[:, :, ::-1, ::-1]
# Determine gradient on inputs ########
mode = 'valid'
if not self.out_mode == 'full':
mode = 'full'
filters = kerns.dimshuffle((1, 0, 2, 3))
filters = filters[:, :, ::-1, ::-1]
nkern = self.imshp[0]
imshp = (self.nkern, self.outshp[0], self.outshp[1])
imshp_logical = (self.nkern, self.fulloutshp[0],
self.fulloutshp[1])
din = ConvOp(imshp, self.kshp, nkern, self.bsize,
1, 1, output_mode=mode,
unroll_batch=None, unroll_kern=None,
unroll_patch=None,
imshp_logical=imshp_logical,
kshp_logical=None,
version=-1, # we we change the mode, we don't forward the version.
direction_hint='bprop inputs',
verbose=self.verbose)
din = din(gz, filters)
assert all(o is None or o == i
for o, i in zip(din.owner.op.outshp, self.imshp[1:]))
# din and dw should have the same broadcasting pattern as the
# parameters they are the gradient of (resp. inputs and kerns).
din = patternbroadcast(din, inputs.broadcastable)
dw = patternbroadcast(dw, kerns.broadcastable)
return [din, dw]
def c_headers(self):
return ['<numpy/noprefix.h>', '<iostream>', '<sstream>']
def c_code_cache_version(self):
return (15, self.openmp, blas.blas_header_version())
def c_support_code(self):
return """
#define STRIDES(arr) (PyArray_STRIDES(arr))
#define FULL 2
#define SAME 1
#define VALID 0
#define MOD %
using namespace std;
""" + blas.blas_header_text()
def use_blas(self):
""" Return True if we will generate code that use gemm.
"""