This repository has been archived by the owner on Aug 15, 2019. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 953
/
conv.ts
914 lines (851 loc) · 36.5 KB
/
conv.ts
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
/**
* @license
* Copyright 2018 Google Inc. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
import {ENV} from '../environment';
import {Tensor2D, Tensor3D, Tensor4D, Tensor5D} from '../tensor';
import {convertToTensor} from '../tensor_util_env';
import {TensorLike} from '../types';
import * as util from '../util';
import * as conv_util from './conv_util';
import {op} from './operation';
/**
* Computes a 1D convolution over the input x.
*
* @param x The input tensor, of rank 3 or rank 2, of shape
* `[batch, width, inChannels]`. If rank 2, batch of 1 is assumed.
* @param filter The filter, rank 3, of shape
* `[filterWidth, inDepth, outDepth]`.
* @param stride The number of entries by which the filter is moved right at
* each step.
* @param pad The type of padding algorithm.
* - `same` and stride 1: output will be of same size as input,
* regardless of filter size.
* - `valid`: output will be smaller than input if filter is larger
* than 1x1.
* - For more info, see this guide:
* [https://www.tensorflow.org/api_guides/python/nn#Convolution](
* https://www.tensorflow.org/api_guides/python/nn#Convolution)
* @param dataFormat An optional string from "NWC", "NCW". Defaults to "NWC",
* the data is stored in the order of [batch, in_width, in_channels]. Only
* "NWC" is currently supported.
* @param dilation The dilation rate in which we sample input values in
* atrous convolution. Defaults to `1`. If it is greater than 1, then
* stride must be `1`.
* @param dimRoundingMode The rounding mode used when computing output
* dimensions if pad is a number. If none is provided, it will not round
* and error if the output is of fractional size.
*/
/** @doc {heading: 'Operations', subheading: 'Convolution'} */
function conv1d_<T extends Tensor2D|Tensor3D>(
x: T|TensorLike, filter: Tensor3D|TensorLike, stride: number,
pad: 'valid'|'same'|number, dataFormat: 'NWC'|'NCW' = 'NWC', dilation = 1,
dimRoundingMode?: 'floor'|'round'|'ceil'): T {
const $x = convertToTensor(x, 'x', 'conv1d');
const $filter = convertToTensor(filter, 'filter', 'conv1d');
let x3D = $x as Tensor3D;
let reshapedTo3D = false;
if ($x.rank === 2) {
reshapedTo3D = true;
x3D = $x.as3D(1, $x.shape[0], $x.shape[1]);
}
util.assert(
x3D.rank === 3,
() => `Error in conv1d: input must be rank 3, but got rank ${x3D.rank}.`);
util.assert(
$filter.rank === 3,
() => `Error in conv1d: filter must be rank 3, but got rank ` +
`${$filter.rank}.`);
if (dimRoundingMode != null) {
util.assert(
util.isInt(pad as number),
() => `Error in conv1d: pad must be an integer when using, ` +
`dimRoundingMode ${dimRoundingMode} but got pad ${pad}.`);
}
util.assert(
x3D.shape[2] === $filter.shape[1],
() => `Error in conv1d: depth of input (${x3D.shape[2]}) must match ` +
`input depth for filter ${$filter.shape[1]}.`);
util.assert(
conv_util.eitherStridesOrDilationsAreOne(stride, dilation),
() => 'Error in conv1D: Either stride or dilation must be 1. ' +
`Got stride ${stride} and dilation '${dilation}'`);
util.assert(
dataFormat === 'NWC',
() => `Error in conv1d: got dataFormat of ${
dataFormat} but only NWC is currently supported.`);
const filter4D =
$filter.as4D(1, $filter.shape[0], $filter.shape[1], $filter.shape[2]);
const input4D = x3D.as4D(x3D.shape[0], 1, x3D.shape[1], x3D.shape[2]);
const strides: [number, number] = [1, stride];
const dilations: [number, number] = [1, dilation];
const conv2dDataFormat = 'NHWC';
const res = conv2d(
input4D, filter4D, strides, pad, conv2dDataFormat, dilations,
dimRoundingMode);
if (reshapedTo3D) {
return res.as2D(res.shape[2], res.shape[3]) as T;
}
return res.as3D(res.shape[0], res.shape[2], res.shape[3]) as T;
}
/**
* Computes a 2D convolution over the input x.
*
* @param x The input tensor, of rank 4 or rank 3, of shape
* `[batch, height, width, inChannels]`. If rank 3, batch of 1 is
* assumed.
* @param filter The filter, rank 4, of shape
* `[filterHeight, filterWidth, inDepth, outDepth]`.
* @param strides The strides of the convolution: `[strideHeight,
* strideWidth]`.
* @param pad The type of padding algorithm.
* - `same` and stride 1: output will be of same size as input,
* regardless of filter size.
* - `valid`: output will be smaller than input if filter is larger
* than 1x1.
* - For more info, see this guide:
* [https://www.tensorflow.org/api_guides/python/nn#Convolution](
* https://www.tensorflow.org/api_guides/python/nn#Convolution)
* @param dataFormat: An optional string from: "NHWC", "NCHW". Defaults to
* "NHWC". Specify the data format of the input and output data. With the
* default format "NHWC", the data is stored in the order of: [batch,
* height, width, channels]. Only "NHWC" is currently supported.
* @param dilations The dilation rates: `[dilationHeight, dilationWidth]`
* in which we sample input values across the height and width dimensions
* in atrous convolution. Defaults to `[1, 1]`. If `dilations` is a single
* number, then `dilationHeight == dilationWidth`. If it is greater than
* 1, then all values of `strides` must be 1.
* @param dimRoundingMode The rounding mode used when computing output
* dimensions if pad is a number. If none is provided, it will not round
* and error if the output is of fractional size.
*/
/** @doc {heading: 'Operations', subheading: 'Convolution'} */
function conv2d_<T extends Tensor3D|Tensor4D>(
x: T|TensorLike, filter: Tensor4D|TensorLike,
strides: [number, number]|number, pad: 'valid'|'same'|number,
dataFormat: 'NHWC'|'NCHW' = 'NHWC',
dilations: [number, number]|number = [1, 1],
dimRoundingMode?: 'floor'|'round'|'ceil'): T {
const $x = convertToTensor(x, 'x', 'conv2d');
const $filter = convertToTensor(filter, 'filter', 'conv2d');
let x4D = $x as Tensor4D;
let reshapedTo4D = false;
if ($x.rank === 3) {
reshapedTo4D = true;
x4D = $x.as4D(1, $x.shape[0], $x.shape[1], $x.shape[2]);
}
util.assert(
x4D.rank === 4,
() => `Error in conv2d: input must be rank 4, but got rank ${x4D.rank}.`);
util.assert(
$filter.rank === 4,
() => `Error in conv2d: filter must be rank 4, but got rank ` +
`${$filter.rank}.`);
if (dimRoundingMode != null) {
util.assert(
util.isInt(pad as number),
() => `Error in conv2d: pad must be an integer when using, ` +
`dimRoundingMode ${dimRoundingMode} but got pad ${pad}.`);
}
util.assert(
x4D.shape[3] === $filter.shape[2],
() => `Error in conv2d: depth of input (${x4D.shape[3]}) must match ` +
`input depth for filter ${$filter.shape[2]}.`);
util.assert(
conv_util.eitherStridesOrDilationsAreOne(strides, dilations),
() => 'Error in conv2D: Either strides or dilations must be 1. ' +
`Got strides ${strides} and dilations '${dilations}'`);
util.assert(
dataFormat === 'NHWC',
() => `Error in conv2d: got dataFormat of ${
dataFormat} but only NHWC is currently supported.`);
const convInfo = conv_util.computeConv2DInfo(
x4D.shape, $filter.shape, strides, dilations, pad, dimRoundingMode);
const grad = (dy: Tensor4D) => {
util.assert(
conv_util.tupleValuesAreOne(dilations),
() => 'Error in gradient of conv2D: dilation rates greater than 1 ' +
`are not yet supported in gradients. Got dilations '${dilations}'`);
return {
x: () => conv2dDerInput_(x4D.shape, dy, $filter, strides, pad),
$filter: () => conv2dDerFilter_(x4D, dy, $filter.shape, strides, pad)
};
};
const res = ENV.engine.runKernel(
backend => backend.conv2d(x4D, $filter, convInfo), {x: x4D, $filter},
grad);
if (reshapedTo4D) {
return res.as3D(res.shape[1], res.shape[2], res.shape[3]) as T;
}
return res as T;
}
/**
* Computes the derivative of the input of a 2D convolution.
*
* @param xShape The shape of the input: [batch, height, width, inDepth].
* If length of 3, batch of 1 is assumed.
* @param dy The derivative of the output, of rank 4 or rank 3 of shape
* `[batch, outHeight, outWidth, outDepth]`. If rank 3, batch of 1 is
* assumed.
* @param filter The filter, rank 4, of shape
* `[filterHeight, filterWidth, inDepth, outDepth]`.
* @param strides The strides of the convolution: `[strideHeight,
* strideWidth]`.
* @param pad The type of padding algorithm used:
* - `same` and stride 1: output will be of same size as input,
* regardless of filter size.
* - `valid`: output will be smaller than input if filter is larger
* than 1x1.
* @param dimRoundingMode The rounding mode used when computing output
* dimensions if pad is a number. If none is provided, it will not round
* and error if the output is of fractional size.
*/
function conv2dDerInput_<T extends Tensor3D|Tensor4D>(
xShape: [number, number, number, number]|[number, number, number], dy: T,
filter: Tensor4D, strides: [number, number]|number,
pad: 'valid'|'same'|number, dimRoundingMode?: 'floor'|'round'|'ceil'): T {
util.assert(
xShape.length === dy.rank,
() => `Length of inShape ` +
`(${xShape.length}) and rank of dy (${dy.rank}) must match`);
let xShape4D = xShape as [number, number, number, number];
let dy4D = dy as Tensor4D;
let reshapedTo4D = false;
if (dy.rank === 3) {
reshapedTo4D = true;
dy4D = dy.as4D(1, dy.shape[0], dy.shape[1], dy.shape[2]);
xShape4D = [1, xShape[0], xShape[1], xShape[2]];
}
const inDepth = xShape4D[3];
const outDepth = dy4D.shape[3];
util.assert(
xShape4D.length === 4,
() =>
`Error in conv2dDerInput: inShape must be length 4, but got length ` +
`${xShape4D.length}.`);
util.assert(
dy4D.rank === 4,
() => `Error in conv2dDerInput: dy must be rank 4, but got ` +
`rank ${dy4D.rank}`);
util.assert(
filter.rank === 4,
() => `Error in conv2dDerInput: filter must be rank 4, but got ` +
`rank ${filter.rank}`);
util.assert(
inDepth === filter.shape[2],
() => `Error in conv2dDerInput: depth of input (${inDepth}) must ` +
`match input depth for filter ${filter.shape[2]}.`);
util.assert(
outDepth === filter.shape[3],
() => `Error in conv2dDerInput: depth of output (${outDepth}) must ` +
`match output depth for filter ${filter.shape[3]}.`);
if (dimRoundingMode != null) {
util.assert(
util.isInt(pad as number),
() => `Error in conv2dDerInput: pad must be an integer when using, ` +
`dimRoundingMode ${dimRoundingMode} but got pad ${pad}.`);
}
const dilations = 1;
const grad = (ddx: Tensor4D) => {
const dataFormat = 'NHWC';
return {
dy4D: () => conv2d(
ddx, filter, strides, pad, dataFormat, dilations, dimRoundingMode),
filter: () => conv2dDerFilter(
ddx, dy4D, filter.shape, strides, pad, dimRoundingMode)
};
};
const convInfo = conv_util.computeConv2DInfo(
xShape4D, filter.shape, strides, dilations, pad, dimRoundingMode);
const res = ENV.engine.runKernel(
backend => backend.conv2dDerInput(dy4D, filter, convInfo), {dy4D, filter},
grad);
if (reshapedTo4D) {
return res.as3D(res.shape[1], res.shape[2], res.shape[3]) as T;
}
return res as T;
}
/**
* Computes the derivative of the filter of a 2D convolution.
*
* @param x The input tensor, of rank 4 or rank 3 of shape
* [batch, height, width, inChannels]. If rank 3, batch of 1 is assumed.
* @param dy The dy image, of rank 4 or rank 3, of shape
* [batch, height, width, outDepth]. If rank 3, batch of 1 is assumed.
* @param filterShape The shape of the filter, length 4,
* [filterHeight, filterWidth, inDepth, outDepth].
* @param strides The strides of the convolution: [strideHeight,
* strideWidth].
* @param pad A string from: 'same', 'valid'. The type of padding algorithm
* used in the forward prop of the op.
* @param dimRoundingMode A string from: 'ceil', 'round', 'floor'. The
* rounding mode used when computing output dimensions if pad is a
* number. If none is provided, it will not round and error if the output
* is of fractional size.
*/
function conv2dDerFilter_<T extends Tensor3D|Tensor4D>(
x: T, dy: T, filterShape: [number, number, number, number],
strides: [number, number]|number, pad: 'valid'|'same'|number,
dimRoundingMode?: 'floor'|'round'|'ceil'): Tensor4D {
let x4D = x as Tensor4D;
if (x.rank === 3) {
x4D = x.as4D(1, x.shape[0], x.shape[1], x.shape[2]);
}
let dy4D = dy as Tensor4D;
if (dy4D.rank === 3) {
dy4D = dy.as4D(1, dy.shape[0], dy.shape[1], dy.shape[2]);
}
util.assert(
x4D.rank === 4,
() => `Error in conv2dDerFilter: input must be rank 4, but got shape ` +
`${x4D.shape}.`);
util.assert(
dy4D.rank === 4,
() => `Error in conv2dDerFilter: dy must be rank 4, but got shape ` +
`${dy4D.shape}.`);
util.assert(
filterShape.length === 4,
() => `Error in conv2dDerFilter: filterShape must be length 4, but got ` +
`${filterShape}.`);
util.assert(
x4D.shape[3] === filterShape[2],
() => `Error in conv2dDerFilter: depth of input ${x4D.shape[3]}) must ` +
`match input depth in filter (${filterShape[2]}.`);
util.assert(
dy4D.shape[3] === filterShape[3],
() => `Error in conv2dDerFilter: depth of dy (${dy4D.shape[3]}) must ` +
`match output depth for filter (${filterShape[3]}).`);
if (dimRoundingMode != null) {
util.assert(
util.isInt(pad as number),
() => `Error in conv2dDerFilter: pad must be an integer when using, ` +
`dimRoundingMode ${dimRoundingMode} but got pad ${pad}.`);
}
const dilations = 1;
const convInfo = conv_util.computeConv2DInfo(
x4D.shape, filterShape, strides, dilations, pad, dimRoundingMode);
return ENV.engine.runKernel(
backend => backend.conv2dDerFilter(x4D, dy4D, convInfo), {x4D, dy4D});
}
/**
* Computes the transposed 2D convolution of an image, also known as a
* deconvolution.
*
* @param x The input image, of rank 4 or rank 3, of shape
* `[batch, height, width, inDepth]`. If rank 3, batch of 1 is assumed.
* @param filter The filter, rank 4, of shape
* `[filterHeight, filterWidth, outDepth, inDepth]`.
* `inDepth` must match `inDepth` in `x`.
* @param outputShape Output shape, of rank 4 or rank 3:
* `[batch, height, width, outDepth]`. If rank 3, batch of 1 is assumed.
* @param strides The strides of the original convolution:
* `[strideHeight, strideWidth]`.
* @param pad The type of padding algorithm used in the non-transpose version
* of the op.
* @param dimRoundingMode The rounding mode used when computing output
* dimensions if pad is a number. If none is provided, it will not round
* and error if the output is of fractional size.
*/
/** @doc {heading: 'Operations', subheading: 'Convolution'} */
function conv2dTranspose_<T extends Tensor3D|Tensor4D>(
x: T|TensorLike, filter: Tensor4D|TensorLike,
outputShape: [number, number, number, number]|[number, number, number],
strides: [number, number]|number, pad: 'valid'|'same'|number,
dimRoundingMode?: 'floor'|'round'|'ceil'): T {
const $x = convertToTensor(x, 'x', 'conv2dTranspose');
const $filter = convertToTensor(filter, 'filter', 'conv2dTranspose');
return conv2dDerInput_(
outputShape, $x, $filter, strides, pad, dimRoundingMode);
}
/**
* Depthwise 2D convolution.
*
* Given a 4D `input` array and a `filter` array of shape
* `[filterHeight, filterWidth, inChannels, channelMultiplier]` containing
* `inChannels` convolutional filters of depth 1, this op applies a
* different filter to each input channel (expanding from 1 channel to
* `channelMultiplier` channels for each), then concatenates the results
* together. The output has `inChannels * channelMultiplier` channels.
*
* See
* [https://www.tensorflow.org/api_docs/python/tf/nn/depthwise_conv2d](
* https://www.tensorflow.org/api_docs/python/tf/nn/depthwise_conv2d)
* for more details.
*
* @param x The input tensor, of rank 4 or rank 3, of shape
* `[batch, height, width, inChannels]`. If rank 3, batch of 1 is
* assumed.
* @param filter The filter tensor, rank 4, of shape
* `[filterHeight, filterWidth, inChannels, channelMultiplier]`.
* @param strides The strides of the convolution: `[strideHeight,
* strideWidth]`. If strides is a single number, then `strideHeight ==
* strideWidth`.
* @param pad The type of padding algorithm.
* - `same` and stride 1: output will be of same size as input,
* regardless of filter size.
* - `valid`: output will be smaller than input if filter is larger
* than 1x1.
* - For more info, see this guide:
* [https://www.tensorflow.org/api_guides/python/nn#Convolution](
* https://www.tensorflow.org/api_guides/python/nn#Convolution)
* @param dilations The dilation rates: `[dilationHeight, dilationWidth]`
* in which we sample input values across the height and width dimensions
* in atrous convolution. Defaults to `[1, 1]`. If `rate` is a single
* number, then `dilationHeight == dilationWidth`. If it is greater than
* 1, then all values of `strides` must be 1.
* @param dataFormat: An optional string from: "NHWC", "NCHW". Defaults to
* "NHWC". Specify the data format of the input and output data. With the
* default format "NHWC", the data is stored in the order of: [batch,
* height, width, channels]. Only "NHWC" is currently supported.
* @param dimRoundingMode The rounding mode used when computing output
* dimensions if pad is a number. If none is provided, it will not round
* and error if the output is of fractional size.
*/
/** @doc {heading: 'Operations', subheading: 'Convolution'} */
function depthwiseConv2d_<T extends Tensor3D|Tensor4D>(
x: T|TensorLike, filter: Tensor4D|TensorLike,
strides: [number, number]|number, pad: 'valid'|'same'|number,
dataFormat: 'NHWC'|'NCHW' = 'NHWC',
dilations: [number, number]|number = [1, 1],
dimRoundingMode?: 'floor'|'round'|'ceil'): T {
const $x = convertToTensor(x, 'x', 'depthwiseConv2d');
const $filter = convertToTensor(filter, 'filter', 'depthwiseConv2d');
let x4D = $x as Tensor4D;
let reshapedTo4D = false;
if ($x.rank === 3) {
reshapedTo4D = true;
x4D = $x.as4D(1, $x.shape[0], $x.shape[1], $x.shape[2]);
}
util.assert(
x4D.rank === 4,
() => `Error in depthwiseConv2d: input must be rank 4, but got ` +
`rank ${x4D.rank}.`);
util.assert(
$filter.rank === 4,
() => `Error in depthwiseConv2d: filter must be rank 4, but got rank ` +
`${$filter.rank}.`);
util.assert(
x4D.shape[3] === $filter.shape[2],
() => `Error in depthwiseConv2d: number of input channels ` +
`(${x4D.shape[3]}) must match the inChannels dimension in ` +
`filter ${$filter.shape[2]}.`);
if (dilations == null) {
dilations = [1, 1];
}
util.assert(
conv_util.eitherStridesOrDilationsAreOne(strides, dilations),
() =>
'Error in depthwiseConv2d: Either strides or dilations must be 1. ' +
`Got strides ${strides} and dilations '${dilations}'`);
if (dimRoundingMode != null) {
util.assert(
util.isInt(pad as number),
() => `Error in depthwiseConv2d: pad must be an integer when using, ` +
`dimRoundingMode ${dimRoundingMode} but got pad ${pad}.`);
}
const convInfo = conv_util.computeConv2DInfo(
x4D.shape, $filter.shape, strides, dilations, pad, dimRoundingMode,
true /* depthwise */);
const grad = (dy: Tensor4D) => {
util.assert(
conv_util.tupleValuesAreOne(dilations),
() => 'Error in gradient of depthwiseConv2d: dilation rates ' +
`greater than 1 are not yet supported. Got dilations ` +
`'${dilations}'`);
return {
x: () => depthwiseConv2dDerInput(x4D.shape, dy, $filter, convInfo),
$filter: () => depthwiseConv2dDerFilter(x4D, dy, $filter.shape, convInfo),
};
};
const res = ENV.engine.runKernel(
backend => backend.depthwiseConv2D(x4D, $filter, convInfo),
{x: x4D, $filter}, grad);
if (reshapedTo4D) {
return res.as3D(res.shape[1], res.shape[2], res.shape[3]) as T;
}
return res as T;
}
/**
* 2-D convolution with separable filters.
*
* Performs a depthwise convolution that acts separately on channels followed
* by a pointwise convolution that mixes channels. Note that this is
* separability between dimensions [1, 2] and 3, not spatial separability
* between dimensions 1 and 2.
*
* See
* [https://www.tensorflow.org/api_docs/python/tf/nn/separable_conv2d](
* https://www.tensorflow.org/api_docs/python/tf/nn/separable_conv2d)
* for more details.
*
* @param x The input tensor, of rank 4 or rank 3, of shape
* `[batch, height, width, inChannels]`. If rank 3, batch of 1 is
* assumed.
* @param depthwiseFilter The depthwise filter tensor, rank 4, of shape
* `[filterHeight, filterWidth, inChannels, channelMultiplier]`. This is
* the filter used in the first step.
* @param pointwiseFilter The pointwise filter tensor, rank 4, of shape
* `[1, 1, inChannels * channelMultiplier, outChannels]`. This is
* the filter used in the second step.
* @param strides The strides of the convolution: `[strideHeight,
* strideWidth]`. If strides is a single number, then `strideHeight ==
* strideWidth`.
* @param pad The type of padding algorithm.
* - `same` and stride 1: output will be of same size as input,
* regardless of filter size.
* - `valid`: output will be smaller than input if filter is larger
* than 1x1.
* - For more info, see this guide:
* [https://www.tensorflow.org/api_guides/python/nn#Convolution](
* https://www.tensorflow.org/api_guides/python/nn#Convolution)
* @param dilations The dilation rates: `[dilationHeight, dilationWidth]`
* in which we sample input values across the height and width dimensions
* in atrous convolution. Defaults to `[1, 1]`. If `rate` is a single
* number, then `dilationHeight == dilationWidth`. If it is greater than
* 1, then all values of `strides` must be 1.
* @param dataFormat: An optional string from: "NHWC", "NCHW". Defaults to
* "NHWC". Specify the data format of the input and output data. With the
* default format "NHWC", the data is stored in the order of: [batch,
* height, width, channels]. Only "NHWC" is currently supported.
*/
/** @doc {heading: 'Operations', subheading: 'Convolution'} */
function separableConv2d_<T extends Tensor3D|Tensor4D>(
x: T|TensorLike, depthwiseFilter: Tensor4D|TensorLike,
pointwiseFilter: Tensor4D|TensorLike, strides: [number, number]|number,
pad: 'valid'|'same', dilation: [number, number]|number = [1, 1],
dataFormat: 'NHWC'|'NCHW' = 'NHWC'): T {
const $x = convertToTensor(x, 'x', 'separableConv2d');
const $depthwiseFilter =
convertToTensor(depthwiseFilter, 'depthwiseFilter', 'separableConv2d');
const $pointwiseFilter =
convertToTensor(pointwiseFilter, 'pointwiseFilter', 'separableConv2d');
let x4D = $x as Tensor4D;
let reshapedTo4D = false;
if ($x.rank === 3) {
reshapedTo4D = true;
x4D = $x.as4D(1, $x.shape[0], $x.shape[1], $x.shape[2]);
}
if (dataFormat === 'NCHW') {
throw new Error(
'separableConv2d currently does not support dataFormat NCHW; only ' +
'NHWC is supported');
}
util.assert(
x4D.rank === 4,
() => `Error in separableConv2d: input must be rank 4, but got ` +
`rank ${x4D.rank}.`);
util.assert(
$depthwiseFilter.rank === 4,
() => `Error in separableConv2d: depthwise filter must be rank 4, but ` +
`got rank ${$depthwiseFilter.rank}.`);
util.assert(
$pointwiseFilter.rank === 4,
() => `Error in separableConv2d: pointwise filter must be rank 4, but ` +
`got rank ${$depthwiseFilter.rank}.`);
util.assert(
$pointwiseFilter.shape[0] === 1,
() =>
`Error in separableConv2d: the first dimension of pointwise filter ` +
` must be 1, but got ${$pointwiseFilter.shape[0]}.`);
util.assert(
$pointwiseFilter.shape[1] === 1,
() => `Error in separableConv2d: the second dimension of pointwise ` +
`filter must be 1, but got ${$pointwiseFilter.shape[1]}.`);
const inChannels = $depthwiseFilter.shape[2];
const channelMultiplier = $depthwiseFilter.shape[3];
util.assert(
$pointwiseFilter.shape[2] === inChannels * channelMultiplier,
() =>
`Error in separableConv2d: the third dimension of pointwise filter ` +
`must be ${inChannels * channelMultiplier}, ` +
`but got ${$pointwiseFilter.shape[2]}.`);
const depthwise = depthwiseConv2d(
x4D, $depthwiseFilter, strides, pad, dataFormat, dilation);
const pointwiseStride = 1;
const res =
conv2d(depthwise, $pointwiseFilter, pointwiseStride, 'valid', dataFormat);
if (reshapedTo4D) {
return res.as3D(res.shape[1], res.shape[2], res.shape[3]) as T;
}
return res as T;
}
function parseTupleParam(
param: number|[number, number]|[number, number, number]):
[number, number, number] {
if (typeof param === 'number') {
return [param, param, param];
}
if (param.length === 2) {
return [param[0], param[1], 1];
}
return param;
}
function tupleValuesAreOne(
param: number|[number, number]|[number, number, number]): boolean {
const [dimA, dimB, dimC] = parseTupleParam(param);
return dimA === 1 && dimB === 1 && dimC === 1;
}
function eitherStridesOrDilationsAreOne(
strides: number|[number, number]|[number, number, number],
dilations: number|[number, number]|[number, number, number]): boolean {
return tupleValuesAreOne(strides) || tupleValuesAreOne(dilations);
}
function depthwiseConv2dDerInput<T extends Tensor3D|Tensor4D>(
xShape: [number, number, number, number]|[number, number, number], dy: T,
filter: Tensor4D, convInfo: conv_util.Conv2DInfo): T {
let dy4D = dy as Tensor4D;
let reshapedTo4D = false;
if (dy.rank === 3) {
reshapedTo4D = true;
dy4D = dy.as4D(1, dy.shape[0], dy.shape[1], dy.shape[2]);
}
const res = ENV.engine.runKernel(
backend => backend.depthwiseConv2DDerInput(dy4D, filter, convInfo),
{dy4D});
if (reshapedTo4D) {
return res.as3D(res.shape[1], res.shape[2], res.shape[3]) as T;
}
return res as T;
}
function depthwiseConv2dDerFilter<T extends Tensor3D|Tensor4D>(
x: T, dy: T, filterShape: [number, number, number, number],
convInfo: conv_util.Conv2DInfo): Tensor4D {
let x4D = x as Tensor4D;
if (x.rank === 3) {
x4D = x.as4D(1, x.shape[0], x.shape[1], x.shape[2]);
}
let dy4D = dy as Tensor4D;
if (dy4D.rank === 3) {
dy4D = dy.as4D(1, dy.shape[0], dy.shape[1], dy.shape[2]);
}
return ENV.engine.runKernel(
backend => backend.depthwiseConv2DDerFilter(x4D, dy4D, convInfo),
{x4D, dy4D});
}
/**
* Computes a 3D convolution over the input x.
*
* @param x The input tensor, of rank 5 or rank 4, of shape
* `[batch, depth, height, width, channels]`. If rank 4,
* batch of 1 is assumed.
* @param filter The filter, rank 5, of shape
* `[filterDepth, filterHeight, filterWidth, inChannels, outChannels]`.
* inChannels must match between input and filter.
* @param strides The strides of the convolution: `[strideDepth, strideHeight,
* strideWidth]`.
* @param pad The type of padding algorithm.
* - `same` and stride 1: output will be of same size as input,
* regardless of filter size.
* - `valid`: output will be smaller than input if filter is larger
* than 1x1.
* - For more info, see this guide:
* [https://www.tensorflow.org/api_guides/python/nn#Convolution](
* https://www.tensorflow.org/api_guides/python/nn#Convolution)
* @param dataFormat: An optional string from: "NHWC", "NCHW". Defaults to
* "NHWC". Specify the data format of the input and output data. With the
* default format "NHWC", the data is stored in the order of: [batch,
* depth, height, width, channels]. Only "NHWC" is currently supported.
* @param dilations The dilation rates: `[dilationDepth, dilationHeight,
* dilationWidth]` in which we sample input values across the height
* and width dimensions in atrous convolution. Defaults to `[1, 1, 1]`.
* If `dilations` is a single number, then
* `dilationDepth == dilationHeight == dilationWidth`. If it is greater
* than 1, then all values of `strides` must be 1.
*/
/** @doc {heading: 'Operations', subheading: 'Convolution'} */
function conv3d_<T extends Tensor4D|Tensor5D>(
x: T|TensorLike, filter: Tensor5D|TensorLike,
strides: [number, number, number]|number, pad: 'valid'|'same',
dataFormat: 'NHWC'|'NCHW' = 'NHWC',
dilations: [number, number, number]|number = [1, 1, 1]): T {
const $x = convertToTensor(x, 'x', 'conv3d');
const $filter = convertToTensor(filter, 'filter', 'conv3d');
let x5D = $x as Tensor5D;
let reshapedTo5D = false;
if ($x.rank === 4) {
reshapedTo5D = true;
x5D = $x.as5D(1, $x.shape[0], $x.shape[1], $x.shape[2], $x.shape[3]);
}
util.assert(
x5D.rank === 5,
() => `Error in conv3d: input must be rank 5, but got rank ${x5D.rank}.`);
util.assert(
$filter.rank === 5,
() => `Error in conv3d: filter must be rank 5, but got rank ` +
`${$filter.rank}.`);
util.assert(
x5D.shape[4] === $filter.shape[3],
() => `Error in conv3d: depth of input (${x5D.shape[4]}) must match ` +
`input depth for filter ${$filter.shape[3]}.`);
util.assert(
eitherStridesOrDilationsAreOne(strides, dilations),
() => 'Error in conv3D: Either strides or dilations must be 1. ' +
`Got strides ${strides} and dilations '${dilations}'`);
util.assert(
dataFormat === 'NHWC',
() => `Error in conv3d: got dataFormat of ${
dataFormat} but only NHWC is currently supported.`);
const convInfo = conv_util.computeConv3DInfo(
x5D.shape, $filter.shape, strides, dilations, pad);
const grad = (dy: Tensor5D) => {
util.assert(
tupleValuesAreOne(dilations),
() =>
'Error in gradient of conv3D: dilation rates greater than 1 are ' +
`not yet supported in gradients. Got dilations '${dilations}'`);
return {
x: () => conv3dDerInput_(x5D.shape, dy, $filter, strides, pad),
$filter: () => conv3dDerFilter_(x5D, dy, $filter.shape, strides, pad)
};
};
const res = ENV.engine.runKernel(
backend => backend.conv3d(x5D, $filter, convInfo), {x: x5D, $filter},
grad);
if (reshapedTo5D) {
return res.as4D(res.shape[1], res.shape[2], res.shape[3], res.shape[4]) as
T;
}
return res as T;
}
/**
* Computes the derivative of the input of a 3D convolution.
*
* @param xShape The shape of the input: [batch, depth, height, width,
* in_channels]. If length of 4, batch of 1 is assumed.
* @param dy The derivative of the output, of rank 5 or rank 4 of shape
* `[batch, outDepth, outHeight, outWidth, in_channels]`.
* If rank 4, batch of 1 is assumed.
* @param filter The filter, rank 5, of shape
* `[filterDepth, filterHeight, filterWidth, inDepth, outDepth]`.
* @param strides The strides of the convolution: `[strideDepth, strideHeight,
* strideWidth]`.
* @param pad The type of padding algorithm used:
* - `same` and stride 1: output will be of same size as input,
* regardless of filter size.
* - `valid`: output will be smaller than input if filter is larger
* than 1x1.
*/
function conv3dDerInput_<T extends Tensor4D|Tensor5D>(
xShape:
[number, number, number, number,
number]|[number, number, number, number],
dy: T, filter: Tensor5D, strides: [number, number, number]|number,
pad: 'valid'|'same'): T {
util.assert(
xShape.length === dy.rank,
() => `Length of inShape ` +
`(${xShape.length}) and rank of dy (${dy.rank}) must match`);
let xShape5D = xShape as [number, number, number, number, number];
let dy5D = dy as Tensor5D;
let reshapedTo5D = false;
if (dy.rank === 4) {
reshapedTo5D = true;
dy5D = dy.as5D(1, dy.shape[0], dy.shape[1], dy.shape[2], dy.shape[3]);
xShape5D = [1, xShape[0], xShape[1], xShape[2], xShape[3]];
}
const inDepth = xShape5D[4];
const outDepth = dy5D.shape[4];
util.assert(
xShape5D.length === 5,
() =>
`Error in conv3dDerInput: inShape must be length 5, but got length ` +
`${xShape5D.length}.`);
util.assert(
dy5D.rank === 5,
() => `Error in conv3dDerInput: dy must be rank 5, but got ` +
`rank ${dy5D.rank}`);
util.assert(
filter.rank === 5,
() => `Error in conv3dDerInput: filter must be rank 5, but got ` +
`rank ${filter.rank}`);
util.assert(
inDepth === filter.shape[3],
() => `Error in conv3dDerInput: depth of input (${inDepth}) must ` +
`match input depth for filter ${filter.shape[3]}.`);
util.assert(
outDepth === filter.shape[4],
() => `Error in conv3dDerInput: depth of output (${outDepth}) must ` +
`match output depth for filter ${filter.shape[4]}.`);
const dilations = 1;
const convInfo = conv_util.computeConv3DInfo(
xShape5D, filter.shape, strides, dilations, pad);
const res = ENV.engine.runKernel(
backend => backend.conv3dDerInput(dy5D, filter, convInfo), {dy5D});
if (reshapedTo5D) {
return res.as4D(res.shape[1], res.shape[2], res.shape[3], res.shape[4]) as
T;
}
return res as T;
}
/**
* Computes the derivative of the filter of a 3D convolution.
*
* @param x The input tensor, of rank 5 or rank 4 of shape
* [batch, depth, height, width, inChannels]. If rank 4, batch of 1 is
* assumed.
* @param dy The dy image, of rank 5 or rank 4, of shape
* [batch, depth, height, width, outDepth]. If rank 4, batch of 1 is
* assumed.
* @param filterShape The shape of the filter, length 5,
* [filterDepth, filterHeight, filterWidth, inDepth, outDepth].
* @param strides The strides of the convolution: [strideDepth, strideHeight,
* strideWidth].
* @param pad A string from: 'same', 'valid'. The type of padding algorithm
* used in the forward prop of the op.
*/
function conv3dDerFilter_<T extends Tensor4D|Tensor5D>(
x: T, dy: T, filterShape: [number, number, number, number, number],
strides: [number, number, number]|number, pad: 'valid'|'same'): Tensor5D {
let x5D = x as Tensor5D;
if (x.rank === 4) {
x5D = x.as5D(1, x.shape[0], x.shape[1], x.shape[2], x.shape[3]);
}
let dy5D = dy as Tensor5D;
if (dy5D.rank === 4) {
dy5D = dy.as5D(1, dy.shape[0], dy.shape[1], dy.shape[2], dy.shape[3]);
}
util.assert(
x5D.rank === 5,
() => `Error in conv3dDerFilter: input must be rank 5, but got shape ` +
`${x5D.shape}.`);
util.assert(
dy5D.rank === 5,
() => `Error in conv3dDerFilter: dy must be rank 5, but got shape ` +
`${dy5D.shape}.`);
util.assert(
filterShape.length === 5,
() => `Error in conv3dDerFilter: filterShape must be length 5, but got ` +
`${filterShape}.`);
util.assert(
x5D.shape[4] === filterShape[3],
() => `Error in conv3dDerFilter: depth of input ${x5D.shape[4]}) must ` +
`match input depth in filter (${filterShape[3]}.`);
util.assert(
dy5D.shape[4] === filterShape[4],
() => `Error in conv3dDerFilter: depth of dy (${dy5D.shape[4]}) must ` +
`match output depth for filter (${filterShape[4]}).`);
const dilations = 1;
const convInfo = conv_util.computeConv3DInfo(
x5D.shape, filterShape, strides, dilations, pad);
return ENV.engine.runKernel(
backend => backend.conv3dDerFilter(x5D, dy5D, convInfo), {x5D, dy5D});
}
export const conv1d = op({conv1d_});
export const conv2d = op({conv2d_});
export const conv3d = op({conv3d_});
export const conv2dDerFilter = op({conv2dDerFilter_});
export const depthwiseConv2d = op({depthwiseConv2d_});
export const separableConv2d = op({separableConv2d_});
export const conv2dTranspose = op({conv2dTranspose_});