1 | /**
|
2 | * @license
|
3 | * Copyright 2020 Google LLC. All Rights Reserved.
|
4 | * Licensed under the Apache License, Version 2.0 (the "License");
|
5 | * you may not use this file except in compliance with the License.
|
6 | * You may obtain a copy of the License at
|
7 | *
|
8 | * http://www.apache.org/licenses/LICENSE-2.0
|
9 | *
|
10 | * Unless required by applicable law or agreed to in writing, software
|
11 | * distributed under the License is distributed on an "AS IS" BASIS,
|
12 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13 | * See the License for the specific language governing permissions and
|
14 | * limitations under the License.
|
15 | * =============================================================================
|
16 | */
|
17 | import { ENGINE } from '../engine';
|
18 | import { Conv2DBackpropFilter } from '../kernel_names';
|
19 | import * as util from '../util';
|
20 | import { op } from './operation';
|
21 | import { reshape } from './reshape';
|
22 | /**
|
23 | * Computes the derivative of the filter of a 2D convolution.
|
24 | *
|
25 | * @param x The input tensor, of rank 4 or rank 3 of shape
|
26 | * [batch, height, width, inChannels]. If rank 3, batch of 1 is assumed.
|
27 | * @param dy The dy image, of rank 4 or rank 3, of shape
|
28 | * [batch, height, width, outDepth]. If rank 3, batch of 1 is assumed.
|
29 | * @param filterShape The shape of the filter, length 4,
|
30 | * [filterHeight, filterWidth, inDepth, outDepth].
|
31 | * @param strides The strides of the convolution: [strideHeight,
|
32 | * strideWidth].
|
33 | * @param pad A string from: 'same', 'valid'. The type of padding algorithm
|
34 | * used in the forward prop of the op.
|
35 | * @param dataFormat: An optional string from: "NHWC", "NCHW". Defaults to
|
36 | * "NHWC". Specify the data format of the input and output data. With the
|
37 | * default format "NHWC", the data is stored in the order of: [batch,
|
38 | * height, width, channels].
|
39 | * @param dimRoundingMode A string from: 'ceil', 'round', 'floor'. If none is
|
40 | * provided, it will default to truncate.
|
41 | */
|
42 | function conv2DBackpropFilter_(x, dy, filterShape, strides, pad, dataFormat = 'NHWC', dimRoundingMode) {
|
43 | let x4D = x;
|
44 | if (x.rank === 3) {
|
45 | x4D = reshape(x, [1, x.shape[0], x.shape[1], x.shape[2]]);
|
46 | }
|
47 | let dy4D = dy;
|
48 | if (dy4D.rank === 3) {
|
49 | dy4D = reshape(dy, [1, dy.shape[0], dy.shape[1], dy.shape[2]]);
|
50 | }
|
51 | util.assert(x4D.rank === 4, () => `Error in conv2dDerFilter: input must be rank 4, but got shape ` +
|
52 | `${x4D.shape}.`);
|
53 | util.assert(dy4D.rank === 4, () => `Error in conv2dDerFilter: dy must be rank 4, but got shape ` +
|
54 | `${dy4D.shape}.`);
|
55 | util.assert(filterShape.length === 4, () => `Error in conv2dDerFilter: filterShape must be length 4, but got ` +
|
56 | `${filterShape}.`);
|
57 | const inDepth = dataFormat === 'NHWC' ? x4D.shape[3] : x4D.shape[1];
|
58 | const outDepth = dataFormat === 'NHWC' ? dy4D.shape[3] : dy4D.shape[1];
|
59 | util.assert(inDepth === filterShape[2], () => `Error in conv2dDerFilter: depth of input ${inDepth}) must ` +
|
60 | `match input depth in filter (${filterShape[2]}.`);
|
61 | util.assert(outDepth === filterShape[3], () => `Error in conv2dDerFilter: depth of dy (${outDepth}) must ` +
|
62 | `match output depth for filter (${filterShape[3]}).`);
|
63 | if (dimRoundingMode != null) {
|
64 | util.assert(util.isInt(pad), () => `Error in conv2dDerFilter: pad must be an integer when using, ` +
|
65 | `dimRoundingMode ${dimRoundingMode} but got pad ${pad}.`);
|
66 | }
|
67 | const inputs = { x: x4D, dy: dy4D };
|
68 | const attrs = { strides, pad, dataFormat, dimRoundingMode, filterShape };
|
69 | // tslint:disable-next-line: no-unnecessary-type-assertion
|
70 | return ENGINE.runKernel(Conv2DBackpropFilter, inputs, attrs);
|
71 | }
|
72 | export const conv2DBackpropFilter = op({ conv2DBackpropFilter_ });
|
73 | //# sourceMappingURL=conv2d_backprop_filter.js.map |
\ | No newline at end of file |