-
Notifications
You must be signed in to change notification settings - Fork 0
/
pooling.py
executable file
·1410 lines (1176 loc) · 50.7 KB
/
pooling.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
'''
Copyright (C) 2014 New York University
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program. If not, see <http://www.gnu.org/licenses/>.
'''
import numpy as np
import theano
import theano.tensor as T
from theano import Op, Apply
from theano.gradient import DisconnectedType
import thutil
from thutil import test_value, Eval
if thutil.use_gpu:
import theano.sandbox.cuda
from theano.sandbox.cuda import GpuOp, gpu_from_host, host_from_gpu, \
CudaNdarrayType, CudaNdarray
def subsample2d(input, stride=(2,2), output_shape=None, transpose='n'):
(bsize, ic, ix, iy) = input.shape
(dx, dy) = stride
if transpose.lower() == 't':
if output_shape is None:
output_shape = (ix * dx, iy * dy)
out = T.zeros((bsize, ic,) + output_shape, dtype=input.dtype)
out = T.set_subtensor(out[:, :, ::dx, ::dy], input)
else:
out = input[:, :, ::dx, ::dy]
return out
def maxpool2d(input, winsize, stride=None, input_shape=None):
if input_shape is None:
input_shape = test_value(input).shape[-2:]
inds = maxinds_2d(input, winsize, stride, input_shape)
vals = index_pool_2d(input, inds, winsize, stride, input_shape)
return (vals, inds)
def maxinds_2d(input, winsize, stride=None, input_shape=None):
if input_shape is None:
input_shape = test_value(input).shape[-2:]
return MaxInds2D(input_shape, winsize, stride)(input)
def index_pool_2d(input, inds, winsize, stride=None,
input_shape=None):
if input_shape is None:
input_shape = test_value(input).shape[-2:]
return IndexPool2D(input_shape, winsize, stride)(input, inds)
def index_unpool_2d(input, inds, winsize, stride=None,
input_shape=None, output_shape=None):
if input_shape is None:
input_shape = test_value(input).shape[-2:]
return IndexUnpool2D(input_shape, winsize, stride,
output_shape=output_shape)(input, inds)
def sumpool2d(input, winsize, stride=None, input_shape=None, average=False):
if input_shape is None:
input_shape = test_value(input).shape[-2:]
return SumPool2D(input_shape, winsize, stride, average=average)(input)
def sum_unpool_2d(input, winsize, stride=None,
input_shape=None, output_shape=None, average=False):
if input_shape is None:
input_shape = test_value(input).shape[-2:]
return SumUnpool2D(input_shape, winsize, stride, output_shape,
average=average)(input)
def maxpool_features(input, winsize):
if winsize == 1:
return (input, T.zeros_like(input))
(bsize, nc, ni, nj) = input.shape
inp = input.transpose((0,2,3,1)).reshape((bsize, ni*nj, nc, 1))
(vals, inds) = maxpool2d(inp, (winsize, 1))
sz = vals.size / (bsize*ni*nj)
vals = vals.reshape((bsize, ni, nj, sz)).transpose((0,3,1,2))
return (vals, inds)
def index_unpool_features(input, inds, winsize):
if winsize == 1:
return input
(bsize, nc, ni, nj) = input.shape
inp = input.transpose((0,2,3,1)).reshape((bsize, ni*nj, nc, 1))
vals = index_unpool_2d(inp, inds, (winsize,1))
sz = vals.size / (bsize*ni*nj)
vals = vals.reshape((bsize, ni, nj, sz)).transpose((0,3,1,2))
return vals
def cmrnorm(x, winsize=5, scale=0.0001, pow=0.75, input_shape=None):
if input_shape is None:
input_shape = test_value(x.shape)[1:]
return CMRNorm(input_shape, winsize, scale, pow, x.dtype)(x)
class PoolOp(Op):
(is_pooling, is_unpooling) = (True, False)
def __init__(self, input_shape, winsize, stride,
output_shape=None, dtype=None):
if stride is None:
stride = winsize
self.input_shape = input_shape
self.winsize = winsize
self.stride = stride
if output_shape is None:
output_shape = self._infer_shape(input_shape)
self.output_shape = output_shape
self.output_dtype = dtype
self._hash_key = (type(self), self.input_shape,
self.winsize, self.stride,
self.output_shape, self.output_dtype)
def __eq__(self, other):
return (type(self) == type(other) and
self.input_shape == other.input_shape and
self.winsize == other.winsize and
self.stride == other.stride and
self.output_shape == other.output_shape and
self.output_dtype == other.output_dtype)
def __hash__(self):
return hash(self._hash_key)
def _infer_shape(self, input_shape):
(x, y) = input_shape
(wx, wy) = self.winsize
(sx, sy) = self.stride
if self.is_pooling:
return ((x-wx)//sx + 1, (y-wy)//sy + 1)
else: # unpooling
return ((x-1)*sx + wx, (y-1)*sy + wy)
def infer_shape(self, node, input_shapes):
s = input_shapes[0][:-2] + self.output_shape
return (s,)
def make_node(self, *inputs):
inputs = tuple(map(T.as_tensor_variable, inputs))
output = T.tensor4(dtype=(self.output_dtype or inputs[0].dtype))
return Apply(self, inputs, (output,))
def perform(self, node, inputs, (output,)):
(bsize, nchan, unpooli, unpoolj) = inputs[0].shape
(pooli, poolj) = self.output_shape
(wi, wj) = self.winsize
(si, sj) = self.stride
output_dtype = self.output_dtype or node.inputs[0].dtype
out = output[0]
if out is None or out.dtype != output_dtype or \
out.shape != (bsize, nchan, pooli, poolj):
out = output[0] = np.empty((bsize, nchan, pooli, poolj),
dtype=output_dtype)
inp = inputs[0]
poolfunc = self.perform_pool
for b in xrange(bsize):
for c in xrange(nchan):
x = out[b, c].flat
x[:] = [poolfunc(inp[b, c, i*si:i*si+wi, j*sj:j*sj+wj],
inputs, b, c, i, j)
for i in xrange(pooli)
for j in xrange(poolj)]
def _pool_c_code(self, node, name, input, output, body, sub,
output_type=None):
(unpooli, unpoolj) = self.input_shape
(pooli, poolj) = self.output_shape
(wi, wj) = self.winsize
(si, sj) = self.stride
fail = sub['fail']
if output_type is None:
output_type = 'PyArray_ObjectType((PyObject*) %s, 0)' % input
code = '''
#define MIN(a,b) ((a) < (b) ? (a) : (b))
int istart, jstart, iend, jend;
int ind;
int bsize = PyArray_DIMS(%(input)s)[0];
int nchan = PyArray_DIMS(%(input)s)[1];
npy_intp dims[4] = {0, 0, %(pooli)d, %(poolj)d};
dims[0] = bsize;
dims[1] = nchan;
if (PyArray_NDIM(%(input)s) != 4) {
PyErr_SetString(PyExc_ValueError, "input must be a 4d ndarray");
%(fail)s;
}
Py_XDECREF(%(output)s);
%(output)s = (PyArrayObject*) PyArray_ZEROS(
4, dims, %(output_type)s, 0);
for (int b = 0; b < bsize; ++b) {
for (int c = 0; c < nchan; ++c) {
for (int i = 0; i < %(pooli)d; ++i) {
istart = i * %(si)d;
iend = MIN(istart + %(wi)d, %(unpooli)d);
for (int j = 0; j < %(poolj)d; ++j) {
jstart = j * %(sj)d;
jend = MIN(jstart + %(wj)d, %(unpoolj)d);
%(body)s
}
}
}
}
''' % locals()
return code
class UnpoolOp(PoolOp):
(is_pooling, is_unpooling) = (False, True)
def perform(self, node, inputs, (output,)):
vals = inputs[0]
(bsize, nchan, pooli, poolj) = vals.shape
(wi, wj) = self.winsize
(si, sj) = self.stride
(unpooli, unpoolj) = self.output_shape
out = output[0] = np.zeros((bsize, nchan, unpooli, unpoolj),
dtype=vals.dtype)
for b in xrange(bsize):
for c in xrange(nchan):
for i in xrange(pooli):
for j in xrange(poolj):
x = out[b, c, i*si:i*si+wi, j*sj:j*sj+wj]
self.perform_unpool(x, vals[b,c,i,j],
inputs, b, c, i, j)
class MaxInds2D(PoolOp):
def perform_pool(self, pool_vals, inputs, b, c, i, j):
return np.argmax(pool_vals)
def make_gpu_node(self, input):
return MaxInds2D_GPU(self.input_shape, self.winsize, self.stride) \
(input)
def c_support_code(self):
code = '''
template<class T>
inline int _argmax(PyArrayObject *x, int b, int c,
int istart, int iend, int jstart, int jend)
{
int k = 0, kmax = 0;
T v, vmax;
vmax = *(T*) PyArray_GETPTR4(x, b, c, istart, jstart);
for (int i = istart; i < iend; ++i) {
for (int j = jstart; j < jend; ++j, ++k) {
v = *(T*) PyArray_GETPTR4(x, b, c, i, j);
if (v > vmax) {
vmax = v;
kmax = k;
}
}
}
return kmax;
}
'''
return code
def c_code(self, node, name, (input,), (output,), sub):
output_type = {'int32': 'NPY_INT',
'float32': 'NPY_FLOAT32',
'float64': 'NPY_FLOAT64',
}[self.output_dtype or node.inputs[0].dtype]
body = '''
int v = _argmax<dtype_%(input)s>(
%(input)s, b, c, istart, iend, jstart, jend);
*(dtype_%(output)s*) PyArray_GETPTR4(%(output)s, b, c, i, j)
= (dtype_%(output)s) v;
''' % locals()
return self._pool_c_code(node, name, input, output, body, sub)
class IndexPool2D(PoolOp):
def perform_pool(self, pool_vals, inputs, b, c, i, j):
return pool_vals.flat[int(inputs[1][b,c,i,j])]
def grad(self, (vals, inds), (dvals,)):
return (IndexUnpool2D(self.output_shape, self.winsize, self.stride,
output_shape=self.input_shape)(dvals, inds),
DisconnectedType()(),)
def make_gpu_node(self, input, inds):
return IndexPool2D_GPU(self.input_shape, self.winsize, self.stride) \
(input, inds)
def c_support_code(self):
code = '''
template<class T>
inline T _select_ind(PyArrayObject *x, int b, int c,
int istart, int iend, int jstart, int jend,
int ind)
{
int jlen = jend - jstart;
int i = istart + ind / jlen;
int j = jstart + ind % jlen;
return *(T*) PyArray_GETPTR4(x, b, c, i, j);
}
'''
return code
def c_code(self, node, name, (input, inds), (output,), sub):
body = '''
int ind = (int) *(dtype_%(inds)s*)
PyArray_GETPTR4(%(inds)s, b, c, i, j);
dtype_%(input)s v = _select_ind<dtype_%(input)s>(
%(input)s, b, c,
istart, iend, jstart, jend,
ind);
*(dtype_%(output)s*)
PyArray_GETPTR4(%(output)s, b, c, i, j) = v;
''' % locals()
return self._pool_c_code(node, name, input, output, body, sub)
class IndexUnpool2D(UnpoolOp):
def perform_unpool(self, unpool_vals, pool_val, inputs, b, c, i, j):
unpool_vals.flat[int(inputs[1][b,c,i,j])] += pool_val
def make_gpu_node(self, input, inds):
return IndexUnpool2D_GPU(self.input_shape, self.winsize, self.stride,
self.output_shape)(input, inds)
def grad(self, (vals, inds), (doutput,)):
return (IndexPool2D(self.output_shape, self.winsize, self.stride)
(doutput, inds),
DisconnectedType()(),)
def c_support_code(self):
code = '''
#define MIN(a,b) ((a) < (b) ? (a) : (b))
template<class T>
inline void _add_ind(
PyArrayObject *x, int b, int c,
int istart, int iend, int jstart, int jend,
int ind, T val)
{
int jlen = jend - jstart;
int i = istart + ind / jlen;
int j = jstart + ind % jlen;
*(T*) PyArray_GETPTR4(x, b, c, i, j) += val;
}
'''
return code
def c_code(self, node, name, (input, inds), (output,), sub):
(unpooli, unpoolj) = self.output_shape
(pooli, poolj) = self.input_shape
(wi, wj) = self.winsize
(si, sj) = self.stride
fail = sub['fail']
code = '''
int istart, jstart;
int ind;
int bsize = PyArray_DIMS(%(input)s)[0];
int nchan = PyArray_DIMS(%(input)s)[1];
npy_intp dims[4] = {0, 0, %(unpooli)d, %(unpoolj)d};
dims[0] = bsize;
dims[1] = nchan;
dtype_%(output)s v;
if (PyArray_NDIM(%(input)s) != 4) {
PyErr_SetString(PyExc_ValueError, "input must be a 4d ndarray");
%(fail)s;
}
Py_XDECREF(%(output)s);
%(output)s = (PyArrayObject*) PyArray_ZEROS(
4, dims,
PyArray_ObjectType((PyObject*) %(input)s, 0),
0);
for (int b = 0; b < bsize; ++b) {
for (int c = 0; c < nchan; ++c) {
for (int i = 0; i < %(pooli)d; ++i) {
istart = i * %(si)d;
for (int j = 0; j < %(poolj)d; ++j) {
jstart = j * %(sj)d;
ind = (int) *(dtype_%(inds)s*)
PyArray_GETPTR4(%(inds)s, b, c, i, j);
v = *(dtype_%(input)s*)
PyArray_GETPTR4(%(input)s, b, c, i, j);
_add_ind<dtype_%(output)s>(
%(output)s, b, c,
istart, MIN(istart + %(wi)d, %(unpooli)d),
jstart, MIN(jstart + %(wj)d, %(unpoolj)d),
ind, v);
}
}
}
}
''' % locals()
return code
class SumPool2D(PoolOp):
def __init__(self, *args, **kwargs):
self.average = kwargs.pop('average', False)
PoolOp.__init__(self, *args, **kwargs)
self._hash_key = self._hash_key + (self.average,)
def __eq__(self, other):
return PoolOp.__eq__(self, other) and self.average == other.average
def perform_pool(self, pool_vals, inputs, b, c, i, j):
if self.average:
return np.mean(pool_vals)
return np.sum(pool_vals)
def make_gpu_node(self, input):
return SumPool2D_GPU(self.input_shape, self.winsize, self.stride,
average=self.average)(input)
def grad(self, (vals,), (dvals,)):
return (SumUnpool2D(self.output_shape, self.winsize, self.stride,
output_shape=self.input_shape,
average=self.average)
(dvals),
)
def c_support_code(self):
code = '''
template<class T, bool average>
inline T _sum_window(PyArrayObject *x, int b, int c,
int istart, int iend, int jstart, int jend)
{
T vsum = 0;
for (int i = istart; i < iend; ++i) {
for (int j = jstart; j < jend; ++j) {
vsum += *(T*) PyArray_GETPTR4(x, b, c, i, j);
}
}
if (average)
vsum /= (iend - istart) * (jend - jstart);
return vsum;
}
'''
return code
def c_code(self, node, name, (input,), (output,), sub):
output_type = {'int32': 'NPY_INT',
'float32': 'NPY_FLOAT32',
'float64': 'NPY_FLOAT64',
}[self.output_dtype or node.inputs[0].dtype]
average = int(self.average)
body = '''
dtype_%(input)s v = _sum_window<dtype_%(input)s, %(average)d>(
%(input)s, b, c, istart, iend, jstart, jend);
*(dtype_%(output)s*) PyArray_GETPTR4(%(output)s, b, c, i, j)
= (dtype_%(output)s) v;
''' % locals()
return self._pool_c_code(node, name, input, output, body, sub)
class SumUnpool2D(UnpoolOp):
def __init__(self, *args, **kwargs):
self.average = kwargs.pop('average', False)
UnpoolOp.__init__(self, *args, **kwargs)
self._hash_key = self._hash_key + (self.average,)
def __eq__(self, other):
return UnpoolOp.__eq__(self, other) and self.average == other.average
def perform_unpool(self, unpool_vals, pool_val, inputs, b, c, i, j):
if self.average:
unpool_vals += pool_val / float(unpool_vals.size)
else:
unpool_vals += pool_val
def make_gpu_node(self, input):
return SumUnpool2D_GPU(self.input_shape, self.winsize, self.stride,
self.output_shape,
average=self.average)(input)
def grad(self, (vals,), (dvals,)):
return (SumPool2D(self.output_shape, self.winsize, self.stride,
average=self.average)
(dvals),
)
def c_support_code(self):
code = '''
#define MIN(a,b) ((a) < (b) ? (a) : (b))
template<class T, bool average>
inline void _add_val_to_window(
PyArrayObject *x, int b, int c,
int istart, int iend, int jstart, int jend,
T val)
{
if (average) {
val /= (T) ((iend - istart) * (jend - jstart));
}
for (int i = istart; i < iend; ++i) {
for (int j = jstart; j < jend; ++j) {
*(T*) PyArray_GETPTR4(x, b, c, i, j) += val;
}
}
}
'''
return code
def c_code(self, node, name, (input,), (output,), sub):
(unpooli, unpoolj) = self.output_shape
(pooli, poolj) = self.input_shape
(wi, wj) = self.winsize
(si, sj) = self.stride
average = self.average
fail = sub['fail']
code = '''
int istart, jstart;
int bsize = PyArray_DIMS(%(input)s)[0];
int nchan = PyArray_DIMS(%(input)s)[1];
npy_intp dims[4] = {0, 0, %(unpooli)d, %(unpoolj)d};
dims[0] = bsize;
dims[1] = nchan;
dtype_%(output)s v;
if (PyArray_NDIM(%(input)s) != 4) {
PyErr_SetString(PyExc_ValueError, "input must be a 4d ndarray");
%(fail)s;
}
Py_XDECREF(%(output)s);
%(output)s = (PyArrayObject*) PyArray_ZEROS(
4, dims,
PyArray_ObjectType((PyObject*) %(input)s, 0),
0);
for (int b = 0; b < bsize; ++b) {
for (int c = 0; c < nchan; ++c) {
for (int i = 0; i < %(pooli)d; ++i) {
istart = i * %(si)d;
for (int j = 0; j < %(poolj)d; ++j) {
jstart = j * %(sj)d;
v = *(dtype_%(input)s*)
PyArray_GETPTR4(%(input)s, b, c, i, j);
/* add val to all elements in this pooling window */
_add_val_to_window<dtype_%(output)s, %(average)d>(
%(output)s, b, c,
istart, MIN(istart + %(wi)d, %(unpooli)d),
jstart, MIN(jstart + %(wj)d, %(unpoolj)d),
v);
}
}
}
}
''' % locals()
return code
class CMRNorm(Op):
def __init__(self, input_shape, winsize, scale, pow, dtype=None):
self.input_shape = tuple(input_shape)
self.winsize = winsize
self.scale = scale
self.pow = pow
self.output_dtype = dtype
self.enable_grad = True
self._hash_key = (type(self), self.__class__, self.input_shape,
self.winsize, self.scale, self.pow,
self.output_dtype)
def __eq__(self, other):
return (type(self) == type(other) and
self.__class__ == other.__class__ and
self.input_shape == other.input_shape and
self.winsize == other.winsize and
self.scale == other.scale and
self.pow == other.pow and
self.output_dtype == other.output_dtype)
def __hash__(self):
return hash(self._hash_key)
def make_gpu_node(self, input):
return CMRNorm_GPU(self.input_shape,
self.winsize, self.scale, self.pow,
dtype=self.output_dtype)(input)
def infer_shape(self, node, input_shapes):
return input_shapes
def make_node(self, *inputs):
inputs = tuple(map(T.as_tensor_variable, inputs))
output = T.tensor4(dtype=(self.output_dtype or inputs[0].dtype))
return Apply(self, inputs, (output,))
def perform(self, node, (input,), (output,)):
(bsize, nchan, ni, nj) = input.shape
output_dtype = self.output_dtype or node.inputs[0].dtype
out = output[0]
if out is None or out.dtype != output_dtype or \
out.shape != (bsize, nchan, ni, ni):
out = output[0] = np.empty((bsize, nchan, ni, nj),
dtype=output_dtype)
x = input
x2 = x ** 2
sums = np.zeros_like(x)
for p in xrange(self.winsize):
d = p - (self.winsize//2)
sums[:,max(0,-d):min(nchan,nchan-d),:,:] += \
x2[:,max(0,d):min(nchan,nchan+d),:,:]
out[:] = x * ((2 + self.scale * sums) ** (-self.pow))
def grad(self, (x,), (dy,)):
if not self.enable_grad:
return [dy]
return (CMRNormGrad(self.input_shape,
self.winsize, self.scale, self.pow,
dtype=self.output_dtype)
(x, self(x), dy),
)
class CMRNormGrad(CMRNorm):
def perform(self, node, (x, y, dy), (output,)):
(bsize, nchan, ni, nj) = x.shape
output_dtype = self.output_dtype or node.inputs[0].dtype
dx = output[0]
if dx is None or dx.dtype != output_dtype or \
dx.shape != (bsize, nchan, ni, ni):
dx = output[0] = np.empty((bsize, nchan, ni, nj),
dtype=output_dtype)
x2 = x ** 2
sums = np.zeros_like(x)
for p in xrange(self.winsize):
d = p - (self.winsize//2)
sums[:,max(0,-d):min(nchan,nchan-d),:,:] += \
x2[:,max(0,d):min(nchan,nchan+d),:,:]
denom = (2 + self.scale * sums) ** (-self.pow)
a = (-2 * self.scale * self.pow) * y * denom
dx[:] = 0
x_dy = x * dy
for p in xrange(self.winsize):
d = p - (self.winsize//2)
# slices of "convolution" window sliding
lhs = slice(max(0,-d), min(nchan,nchan-d))
rhs = slice(max(0,d), min(nchan,nchan+d))
dx[:,lhs,:,:] += x_dy[:,rhs,:,:]
dx *= a
dx += dy * denom
def make_gpu_node(self, *inputs):
return CMRNormGrad_GPU(self.input_shape,
self.winsize, self.scale, self.pow,
dtype=self.output_dtype)(*inputs)
def infer_shape(self, node, input_shapes):
return (input_shapes[0],)
def grad(self, inputs, doutputs):
raise NotImplementedError
if thutil.use_gpu:
source_support_defs = '''
#define MOD %
#define MIN(a,b) ((a) < (b) ? (a) : (b))
#define IDX4(n1, n2, n3, n4, i1, i2, i3, i4) \\
((i1)*(n2)*(n3)*(n4) + (i2)*(n3)*(n4) + (i3)*(n4) + (i4))
#define IDX3(n1, n2, n3, i1, i2, i3) \\
((i1)*(n2)*(n3) + (i2)*(n3) + (i3))
#define UNRAVEL_IDX4(ind, n1, n2, n3, n4, i1, i2, i3, i4) \\
{ \\
i1 = (ind) / ((n2)*(n3)*(n4)); \\
i2 = ((ind) MOD ((n2)*(n3)*(n4))) / ((n3)*(n4)); \\
i3 = ((ind) MOD ((n3)*(n4))) / (n4); \\
i4 = ((ind) MOD (n4)); \\
}
#define SETIF(var, val, cond) \
(var = (val)*!!(cond) + (var)*!(cond))
#define DIVUP(x,y) (1 + (((x) - 1) / (y)))
static void launch_sizes(int nthreads, dim3 &grid_size, dim3 &block_size)
{
static const int min_threads = 16;
static const int max_threads = 256;
static const int max_blocks = 65535;
int ngroups = (nthreads + min_threads - 1) / min_threads;
if (ngroups == 1) {
grid_size = dim3(1);
block_size = dim3(min_threads);
} else if (nthreads < max_blocks * min_threads) {
grid_size = dim3(ngroups);
block_size = dim3(min_threads);
} else if (nthreads < max_blocks * max_threads) {
grid_size = dim3(max_blocks);
block_size = dim3((ngroups + max_blocks - 1)
/ max_blocks * min_threads);
} else {
grid_size = dim3(max_blocks);
block_size = dim3(max_threads);
}
}
'''
class PoolGpuOp(GpuOp):
def make_node(self, *inputs):
output = CudaNdarrayType((False,) * 4)()
return Apply(self, inputs, (output,))
def c_support_code(self):
if self.is_pooling:
unpooled_shape = self.input_shape
pooled_shape = self.output_shape
pooled_stride_i = 1
pooled_stride_j = 1
else:
unpooled_shape = self.output_shape
pooled_shape = self.input_shape
pooled_stride_i = 'DIVUP(wsize_i, stride_i)'
pooled_stride_j = 'DIVUP(wsize_j, stride_j)'
assert unpooled_shape[0] >= (pooled_shape[0] - 1) * self.stride[0] + self.winsize[0]
assert unpooled_shape[1] >= (pooled_shape[1] - 1) * self.stride[1] + self.winsize[1]
source = source_support_defs + '''
#define unpooled_i %(unpooled_shape[0])d
#define unpooled_j %(unpooled_shape[1])d
#define pooled_i %(pooled_shape[0])d
#define pooled_j %(pooled_shape[1])d
#define wsize_i %(self.winsize[0])d
#define wsize_j %(self.winsize[1])d
#define stride_i %(self.stride[0])d
#define stride_j %(self.stride[1])d
#define pooled_stride_i %(pooled_stride_i)s
#define pooled_stride_j %(pooled_stride_j)s
#define pooled_si DIVUP(pooled_i, pooled_stride_i)
#define pooled_sj DIVUP(pooled_j, pooled_stride_j)
#define ntiles_per_call (pooled_si * pooled_sj)
static __global__ void %(self.ker_name)s ( %(self.ker_args)s,
float *out,
int nimgs,
int pooled_start_i,
int pooled_start_j,
uint32_t randseed)
{
unsigned total_threads = gridDim.x * blockDim.x;
int tid = blockIdx.x * blockDim.x + threadIdx.x;
int t, tnum, img, tile;
int pi, pj, ui0, uj0;
%(self.ker_defs)s
for (t = tid; t < nimgs * ntiles_per_call; t += total_threads) {
tnum = t; /* tile number */
img = tnum / ntiles_per_call; /* image the tile is in */
tile = tnum MOD ntiles_per_call; /* tile in image */
pi = (tile / pooled_sj); /* pooled pixel indices */
pj = (tile MOD pooled_sj);
pi = pi * pooled_stride_i + pooled_start_i;
pj = pj * pooled_stride_j + pooled_start_j;
if (pi >= pooled_i || pj >= pooled_j)
continue;
ui0 = pi * stride_i; /* unpooled window top-left pixel */
uj0 = pj * stride_j;
%(self.ker_loop_body)s
}
}
''' % Eval()
return source
def c_code(self, node, nodename, inputs, outputs, sub):
(output,) = outputs
source = '''
const int *input_dims = CudaNdarray_HOST_DIMS(%(inputs[0])s);
const int *output_dims = %(output)s ?
CudaNdarray_HOST_DIMS(%(output)s) :
NULL;
const int dims[] = { input_dims[0],
input_dims[1],
%(self.output_shape[0])s,
%(self.output_shape[1])s };
const int nimgs = dims[0] * dims[1]; /* imgs * channels */
int ntiles = nimgs * ntiles_per_call; /* one thread per tile */
dim3 grid_size, block_size;
launch_sizes(ntiles, grid_size, block_size);
CudaNdarray %(', '.join('*%s_contig' % inp for inp in inputs))s ;
cudaError_t err;
'''
source += '''
if (%(output)s == NULL
|| !CudaNdarray_is_c_contiguous(%(output)s)
|| %(output)s->nd != 4
|| dims[0] != output_dims[0]
|| dims[1] != output_dims[1]
|| dims[2] != output_dims[2]
|| dims[3] != output_dims[3]) {
Py_XDECREF(%(output)s);
%(output)s = (CudaNdarray*)CudaNdarray_New();
if (%(output)s == NULL
|| CudaNdarray_alloc_contiguous(%(output)s, 4, dims)) {
Py_XDECREF(%(output)s);
%(output)s = NULL;
%(sub['fail'])s;
}
}
if (%(self.zero_output)d) {
if (cudaMemset(
CudaNdarray_DEV_DATA(%(output)s),
0, CudaNdarray_SIZE(%(output)s) * sizeof(float))
!= cudaSuccess) {
PyErr_Format(PyExc_MemoryError,
"%(self.ker_name)s: Error in memset");
Py_XDECREF(%(output)s);
%(output)s = NULL;
%(sub['fail'])s;
}
}
'''
for inp in inputs:
source += '''
%(inp)s_contig = %(inp)s;
if (!CudaNdarray_is_c_contiguous(%(inp)s)) {
%(inp)s_contig = (CudaNdarray*) CudaNdarray_Copy(%(inp)s);
assert(CudaNdarray_is_c_contiguous(%(inp)s_contig));
}
''' % Eval()
source += '''
/* call kernel once for each offset within the pooled stride */
for (int j = 0; j < pooled_stride_j; ++j) {
for (int i = 0; i < pooled_stride_i; ++i) {
%(self.ker_name)s <<<grid_size, block_size>>> (
%(', '.join('CudaNdarray_DEV_DATA(%s_contig)' % x
for x in inputs))s,
CudaNdarray_DEV_DATA(%(output)s),
nimgs,
i, j,
rand()
);
}
}
CNDA_THREAD_SYNC;
'''
for inp in inputs:
source += '''
if (%(inp)s_contig != %(inp)s) {
Py_DECREF(%(inp)s_contig);
}
''' % Eval()
source += '''
err = cudaGetLastError();
if (err != cudaSuccess) {
PyErr_Format(PyExc_RuntimeError,
"Cuda error: %%s: %%s",
"%(self.ker_name)s", cudaGetErrorString(err));
%(sub['fail'])s
}
'''
source = source % Eval()
return source
class MaxInds2D_GPU(PoolGpuOp, MaxInds2D):
ker_name = 'pool_maxind'
ker_args = 'float *X'
ker_defs = '''
int u, ui, uj, ismax, max_ind;
float val, max_val;
'''
zero_output = False
ker_loop_body = '''
max_val = -1000000;
for (u = 0; u < wsize_i * wsize_j; ++u) {
ui = ui0 + u / wsize_j;
uj = uj0 + u MOD wsize_j;
if ((ui < unpooled_i) && (uj < unpooled_j)) {
val = X[IDX3(nimgs, unpooled_i, unpooled_j,
img, ui, uj)];
ismax = (val > max_val);
SETIF(max_val, val, ismax);
SETIF(max_ind, u, ismax);
}
}
out[IDX3(nimgs, pooled_i, pooled_j,
img, pi, pj)]
= max_ind;
'''
def perform(self, node, (input,), (output,)):
output_host = [None]
MaxInds2D.perform(self, node, (np.array(input),), (output_host,))
output[0] = CudaNdarray(output_host[0].astype(np.float32))
class IndexPool2D_GPU(PoolGpuOp, IndexPool2D):
ker_name = 'pool_index'
ker_args = 'float *input, float *inds'
ker_defs = '''
int ui, uj, ind;
float val;
'''
zero_output = False
ker_loop_body = '''
ind = (int) inds[IDX3(nimgs, pooled_i, pooled_j,
img, pi, pj)];
ui = ui0 + ind / wsize_j;
uj = uj0 + ind MOD wsize_j;
val = input[IDX3(nimgs, unpooled_i, unpooled_j,
img, ui, uj)];
out[IDX3(nimgs, pooled_i, pooled_j,
img, pi, pj)] = val;
'''
def perform(self, node, inputs, (output,)):
output_host = [None]
IndexPool2D.perform(self, node,
map(np.array, inputs), (output_host,))
output[0] = CudaNdarray(output_host[0].astype(np.float32))
class IndexUnpool2D_GPU(PoolGpuOp, IndexUnpool2D):
ker_name = 'unpool_index'
ker_args = 'float *input, float *inds'
ker_defs = '''
int ui, uj, ind;
float val;
'''
zero_output = True
ker_loop_body = '''
ind = (int) inds[IDX3(nimgs, pooled_i, pooled_j,
img, pi, pj)];
val = input[IDX3(nimgs, pooled_i, pooled_j,
img, pi, pj)];
ui = ui0 + ind / wsize_j;
uj = uj0 + ind MOD wsize_j;
out[IDX3(nimgs, unpooled_i, unpooled_j,
img, ui, uj)] += val;
'''
def perform(self, node, inputs, (output,)):
output_host = [None]
IndexUnpool2D.perform(self, node,
map(np.array, inputs), (output_host,))
output[0] = CudaNdarray(output_host[0].astype(np.float32))
class SumPool2D_GPU(PoolGpuOp, SumPool2D):
ker_name = 'pool_sum'
ker_args = 'float *X'
ker_defs = '''
int u, ui, uj, usize_i, usize_j;
float vsum;
'''
zero_output = False
def __init__(self, *args, **kwargs):
super(SumPool2D_GPU, self).__init__(*args, **kwargs)
self.ker_loop_body = '''
vsum = 0;