-
Notifications
You must be signed in to change notification settings - Fork 2k
/
Copy pathpool.ts
167 lines (149 loc) · 6.91 KB
/
pool.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
/**
* @license
* Copyright 2018 Google LLC. 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 {Tensor3D, Tensor4D} from '../tensor';
import {convertToTensor} from '../tensor_util_env';
import {TensorLike} from '../types';
import * as util from '../util';
import {avgPool} from './avg_pool';
import {batchToSpaceND} from './batch_to_space_nd';
import * as conv_util from './conv_util';
import {maxPool} from './max_pool';
import {op} from './operation';
import {reshape} from './reshape';
import {spaceToBatchND} from './space_to_batch_nd';
/**
* Performs an N-D pooling operation
*
* @param input The input tensor, of rank 4 or rank 3 of shape
* `[batch, height, width, inChannels]`. If rank 3, batch of 1 is assumed.
* @param windowShape The filter size: `[filterHeight, filterWidth]`. If
* `filterSize` is a single number, then `filterHeight == filterWidth`.
* @param poolingType The type of pooling, either 'max' or 'avg'.
* @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 dilated pooling. Defaults to `[1, 1]`. If `dilationRate` is a single
* number, then `dilationHeight == dilationWidth`. If it is greater than
* 1, then all values of `strides` must be 1.
* @param strides The strides of the pooling: `[strideHeight, strideWidth]`. If
* `strides` is a single number, then `strideHeight == strideWidth`.
* @param dimRoundingMode A string from: 'ceil', 'round', 'floor'. If none is
* provided, it will default to truncate.
*
* @doc {heading: 'Operations', subheading: 'Convolution'}
*/
function pool_<T extends Tensor3D|Tensor4D>(
input: T|TensorLike, windowShape: [number, number]|number,
poolingType: 'avg'|'max',
pad: 'valid'|'same'|number|conv_util.ExplicitPadding,
dilations?: [number, number]|number, strides?: [number, number]|number,
dimRoundingMode?: 'floor'|'round'|'ceil') {
if (dilations == null) {
dilations = [1, 1];
}
if (strides == null) {
strides = 1;
}
if (pad === 0) {
pad = 'valid';
}
const $x = convertToTensor(input, 'x', 'maxPool');
let x4D = $x as Tensor4D;
let reshapedTo4D = false;
if ($x.rank === 3) {
reshapedTo4D = true;
x4D = reshape($x, [1, $x.shape[0], $x.shape[1], $x.shape[2]]);
}
util.assert(
conv_util.eitherStridesOrDilationsAreOne(strides, dilations),
() => 'Error in pool: Either strides or dilations must be 1. ' +
`Got strides ${strides} and dilations '${dilations}'`);
const convInfo = conv_util.computePool2DInfo(
x4D.shape, windowShape, strides, dilations, pad);
const dilation: [number, number] =
[convInfo.dilationHeight, convInfo.dilationWidth];
// The following implementation does batchToSpace(pool(spaceToBatch(x)))
// whenever dilation > 1 since the TF kernels do not support dilation > 1.
// tslint:disable-next-line:max-line-length
// https://github.com/tensorflow/tensorflow/blob/50f6bb67dc98c9b74630b6047aae7a4f8a40fd02/tensorflow/python/ops/nn_ops.py#L1037
let basePadding: number[][];
if (pad === 'same') {
basePadding = withSpaceToBatchBasePaddings(
[convInfo.filterHeight, convInfo.filterWidth], dilation);
} else {
basePadding = [[0, 0], [0, 0]];
}
const isDilationOne = dilation[0] === 1 && dilation[1] === 1;
const [adjustedPadding, adjustedCrops] = requiredSpaceToBatchPaddings(
[convInfo.inHeight, convInfo.inWidth], dilation, basePadding);
const convertedPad = isDilationOne ? pad : 'valid';
const convertedX =
isDilationOne ? x4D : spaceToBatchND(x4D, dilation, adjustedPadding);
const forwardOp = poolingType === 'avg' ?
() => avgPool(convertedX, windowShape, strides, convertedPad,
dimRoundingMode) :
() => maxPool(convertedX, windowShape, strides, convertedPad,
dimRoundingMode);
const y = forwardOp();
const res = isDilationOne ? y : batchToSpaceND(y, dilation, adjustedCrops);
if (reshapedTo4D) {
return reshape(res, [res.shape[1], res.shape[2], res.shape[3]]) as T;
}
return res as T;
}
// Helper function to compute crops and paddings for pool with dilation > 1.
// tslint:disable-next-line:max-line-length
// https://github.com/tensorflow/tensorflow/blob/50f6bb67dc98c9b74630b6047aae7a4f8a40fd02/tensorflow/python/ops/array_ops.py#L2184
function requiredSpaceToBatchPaddings(
inputShape: [number, number], blockShape: [number, number],
basePadding: number[][]) {
const padStart = basePadding.map(b => b[0]);
const origPadEnd = basePadding.map(b => b[1]);
const fullInputShape = inputShape.concat(padStart, origPadEnd);
const padEndExtra = blockShape.map((b, i) => (b - fullInputShape[i] % b) % b);
const padEnd = origPadEnd.map((s, i) => s + padEndExtra[i]);
const paddings = blockShape.map((_, i) => [padStart[i], padEnd[i]]);
const crops = blockShape.map((_, i) => [0, padEndExtra[i]]);
return [paddings, crops];
}
// Helper function to compute base paddings for pool with dilation > 1.
// tslint:disable-next-line:max-line-length
// https://github.com/tensorflow/tensorflow/blob/50f6bb67dc98c9b74630b6047aae7a4f8a40fd02/tensorflow/python/ops/nn_ops.py#L524
function withSpaceToBatchBasePaddings(
filterShape: [number, number], dilation: [number, number]) {
// Spatial dimensions of the filters and the upsampled filters in which we
// introduce (rate - 1) zeros between consecutive filter values.
const dilatedFilterShape = filterShape.map((s, i) => {
return s + (s - 1) * (dilation[i] - 1);
});
const padExtraShape = dilatedFilterShape.map(s => s - 1);
// When padding is odd, we pad more at end, following the same
// convention as conv2d.
const padExtraStart = padExtraShape.map(s => Math.floor(s / 2));
const padExtraEnd = padExtraShape.map((s, i) => s - padExtraStart[i]);
return padExtraShape.map((_, i) => {
return [padExtraStart[i], padExtraEnd[i]];
});
}
export const pool = /* @__PURE__ */ op({pool_});