@tensorflow/tfjs-core
Version:
Hardware-accelerated JavaScript library for machine intelligence
366 lines • 17.8 kB
JavaScript
/**
* @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 * as tf from '../index';
import { ALL_ENVS, describeWithFlags } from '../jasmine_util';
import { expectArraysClose } from '../test_util';
describeWithFlags('separableConv2d', ALL_ENVS, () => {
it('input=1x3x3x1,f=2,s=1,d=1,p=valid,chMul=1,outDepth=2', async () => {
const fSize = 2;
const pad = 'valid';
const stride = 1;
const chMul = 1;
const inDepth = 1;
const outDepth = 2;
const x = tf.tensor4d([
0.230664, 0.987388, 0.0685208, 0.419224, 0.887861, 0.731641,
0.0741907, 0.409265, 0.351377
], [1, 3, 3, inDepth]);
const depthwiseFilter = tf.tensor4d([0.303873, 0.229223, 0.144333, 0.803373], [fSize, fSize, inDepth, chMul]);
const pointwiseFilter = tf.tensor4d([0.1, -0.2], [1, 1, inDepth * chMul, outDepth]);
const result = tf.separableConv2d(x, depthwiseFilter, pointwiseFilter, stride, pad);
expectArraysClose(await result.data(), [
0.10702161, -0.21404321, 0.10316753, -0.20633507, 0.06704096, -0.13408193,
0.07788632, -0.15577264
]);
expect(result.shape).toEqual([1, 2, 2, outDepth]);
});
it('input=1x3x3x1,f=2,s=1,d=1,p=valid,chMul=1,outDepth=2 in tensor', async () => {
const fSize = 2;
const pad = 'valid';
const stride = 1;
const chMul = 1;
const inDepth = 1;
const outDepth = 2;
const x = tf.tensor4d([
0.230664, 0.987388, 0.0685208, 0.419224, 0.887861, 0.731641,
0.0741907, 0.409265, 0.351377
], [1, 3, 3, inDepth]);
const depthwiseFilter = tf.tensor4d([0.303873, 0.229223, 0.144333, 0.803373], [fSize, fSize, inDepth, chMul]);
const pointwiseFilter = tf.tensor4d([0.1, -0.2], [1, 1, inDepth * chMul, outDepth]);
const result = x.separableConv2d(depthwiseFilter, pointwiseFilter, stride, pad);
expectArraysClose(await result.data(), [
0.10702161, -0.21404321, 0.10316753, -0.20633507, 0.06704096,
-0.13408193, 0.07788632, -0.15577264
]);
expect(result.shape).toEqual([1, 2, 2, outDepth]);
});
it('input=1x3x3x1,f=2,s=1,d=1,p=valid,chMul=2,outDepth=2', async () => {
const fSize = 2;
const pad = 'valid';
const stride = 1;
const chMul = 2;
const inDepth = 1;
const outDepth = 3;
const x = tf.tensor4d([
0.230664, 0.987388, 0.0685208, 0.419224, 0.887861, 0.731641,
0.0741907, 0.409265, 0.351377
], [1, 3, 3, inDepth]);
const depthwiseFilter = tf.tensor4d([
0.303873, 0.229223, 0.144333, 0.803373, -0.303873, -0.229223,
-0.144333, -0.803373
], [fSize, fSize, inDepth, chMul]);
const pointwiseFilter = tf.tensor4d([0.1, -0.2, -0.1, 0.2, 0.15, 0.15], [1, 1, inDepth * chMul, outDepth]);
const result = tf.separableConv2d(x, depthwiseFilter, pointwiseFilter, stride, pad);
expectArraysClose(await result.data(), [
0.00305368, 0.0140969, 0.00980358, -0.10853045, -0.06339455, -0.0699412,
0.11010849, 0.0347524, 0.05214475, 0.10307151, 0.02221644, 0.04224815
]);
expect(result.shape).toEqual([1, 2, 2, outDepth]);
});
it('input=1x3x3x1,f=2,s=1,d=1,p=valid,chMul=1,outDepth=2,3D input', async () => {
const fSize = 2;
const pad = 'valid';
const stride = 1;
const chMul = 1;
const inDepth = 1;
const outDepth = 2;
const x = tf.tensor3d([
0.230664, 0.987388, 0.0685208, 0.419224, 0.887861, 0.731641,
0.0741907, 0.409265, 0.351377
], [3, 3, inDepth]);
const depthwiseFilter = tf.tensor4d([0.303873, 0.229223, 0.144333, 0.803373], [fSize, fSize, inDepth, chMul]);
const pointwiseFilter = tf.tensor4d([0.1, -0.2], [1, 1, inDepth * chMul, outDepth]);
const result = tf.separableConv2d(x, depthwiseFilter, pointwiseFilter, stride, pad);
expectArraysClose(await result.data(), [
0.10702161, -0.21404321, 0.10316753, -0.20633507, 0.06704096,
-0.13408193, 0.07788632, -0.15577264
]);
expect(result.shape).toEqual([2, 2, outDepth]);
});
it('input=1x4x4x1,f=2,s=2,d=1,p=valid,chMul=1,outDepth=2', async () => {
const fSize = 2;
const pad = 'valid';
const stride = [2, 2];
const chMul = 1;
const inDepth = 1;
const outDepth = 2;
const x = tf.tensor4d([
0.675707, 0.758567, 0.413529, 0.963967, 0.217291, 0.101335, 0.804231,
0.329673, 0.924503, 0.728742, 0.180217, 0.210459, 0.133869, 0.650827,
0.047613, 0.554795
], [1, 4, 4, inDepth]);
const depthwiseFilter = tf.tensor4d([0.303873, 0.229223, 0.144333, 0.803373], [fSize, fSize, inDepth, chMul]);
const pointwiseFilter = tf.tensor4d([0.1, -0.2], [1, 1, inDepth * chMul, outDepth]);
const result = tf.separableConv2d(x, depthwiseFilter, pointwiseFilter, stride, pad);
expectArraysClose(await result.data(), [
0.04919822, -0.09839644, 0.07275512, -0.14551024, 0.09901544, -0.19803089,
0.05555845, -0.11111691
]);
expect(result.shape).toEqual([1, 2, 2, outDepth]);
});
it('input=2x4x4x1,f=2,s=2,d=1,p=valid,chMul=1,outDepth=2', async () => {
const fSize = 2;
const pad = 'valid';
const stride = [2, 2];
const chMul = 1;
const inDepth = 1;
const outDepth = 2;
const x = tf.tensor4d([
0.675707, 0.758567, 0.413529, 0.963967, 0.217291, 0.101335,
0.804231, 0.329673, 0.924503, 0.728742, 0.180217, 0.210459,
0.133869, 0.650827, 0.047613, 0.554795, -0.675707, -0.758567,
-0.413529, -0.963967, -0.217291, -0.101335, -0.804231, -0.329673,
-0.924503, -0.728742, -0.180217, -0.210459, -0.133869, -0.650827,
-0.047613, -0.554795
], [2, 4, 4, inDepth]);
const depthwiseFilter = tf.tensor4d([0.303873, 0.229223, 0.144333, 0.803373], [fSize, fSize, inDepth, chMul]);
const pointwiseFilter = tf.tensor4d([0.1, -0.2], [1, 1, inDepth * chMul, outDepth]);
const result = tf.separableConv2d(x, depthwiseFilter, pointwiseFilter, stride, pad);
expectArraysClose(await result.data(), [
0.04919822, -0.09839644, 0.07275512, -0.14551024, 0.09901544, -0.19803089,
0.05555845, -0.11111691, -0.04919822, 0.09839644, -0.07275512, 0.14551024,
-0.09901544, 0.19803089, -0.05555845, 0.11111691
]);
expect(result.shape).toEqual([2, 2, 2, outDepth]);
});
it('input=1x4x4x2,f=2,s=2,d=1,p=valid,chMul=1,outDepth=2', async () => {
const fSize = 2;
const pad = 'valid';
const stride = [2, 2];
const chMul = 1;
const inDepth = 2;
const outDepth = 2;
const x = tf.tensor4d([
0.675707, 0.758567, 0.413529, 0.963967, 0.217291, 0.101335,
0.804231, 0.329673, 0.924503, 0.728742, 0.180217, 0.210459,
0.133869, 0.650827, 0.047613, 0.554795, -0.675707, -0.758567,
-0.413529, -0.963967, -0.217291, -0.101335, -0.804231, -0.329673,
-0.924503, -0.728742, -0.180217, -0.210459, -0.133869, -0.650827,
-0.047613, -0.554795
], [1, 4, 4, inDepth]);
const depthwiseFilter = tf.tensor4d([
0.303873, 0.229223, 0.144333, 0.803373, 0.98976838, 0.56597068,
0.42654137, 0.66445535
], [fSize, fSize, inDepth, chMul]);
const pointwiseFilter = tf.tensor4d([0.1, -0.2, 0.05, -0.05], [1, 1, inDepth * chMul, outDepth]);
const result = tf.separableConv2d(x, depthwiseFilter, pointwiseFilter, stride, pad);
expectArraysClose(await result.data(), [
0.20072255, -0.32641545, 0.08474462, -0.11823604, -0.20072255, 0.32641545,
-0.08474462, 0.11823604
]);
expect(result.shape).toEqual([1, 2, 2, outDepth]);
});
it('input=1x4x4x1,f=2,s=1,d=2,p=valid,chMul=1,outDepth=2', async () => {
const fSize = 2;
const pad = 'valid';
const stride = 1;
const chMul = 1;
const inDepth = 1;
const outDepth = 2;
const dilationRate = 2;
const x = tf.tensor4d([
0.675707, 0.758567, 0.413529, 0.963967, 0.217291, 0.101335, 0.804231,
0.329673, 0.924503, 0.728742, 0.180217, 0.210459, 0.133869, 0.650827,
0.047613, 0.554795
], [1, 4, 4, inDepth]);
const depthwiseFilter = tf.tensor4d([0.303873, 0.229223, 0.144333, 0.803373], [fSize, fSize, inDepth, chMul]);
const pointwiseFilter = tf.tensor4d([0.1, -0.2], [1, 1, inDepth * chMul, outDepth]);
const result = tf.separableConv2d(x, depthwiseFilter, pointwiseFilter, stride, pad, dilationRate);
expectArraysClose(await result.data(), [
0.05783373, -0.11566745, 0.07257301, -0.14514601, 0.03079498, -0.06158997,
0.06460048, -0.12920095
]);
expect(result.shape).toEqual([1, 2, 2, outDepth]);
});
it('input=1x4x4x1,f=2,s=1,d=1,p=same,chMul=1,outDepth=2', async () => {
const fSize = 2;
const pad = 'same';
const stride = 1;
const chMul = 1;
const inDepth = 1;
const outDepth = 2;
const x = tf.tensor4d([
0.675707, 0.758567, 0.413529, 0.963967, 0.217291, 0.101335, 0.804231,
0.329673, 0.924503, 0.728742, 0.180217, 0.210459, 0.133869, 0.650827,
0.047613, 0.554795
], [1, 4, 4, inDepth]);
const depthwiseFilter = tf.tensor4d([0.303873, 0.229223, 0.144333, 0.803373], [fSize, fSize, inDepth, chMul]);
const pointwiseFilter = tf.tensor4d([0.1, -0.2], [1, 1, inDepth * chMul, outDepth]);
const result = tf.separableConv2d(x, depthwiseFilter, pointwiseFilter, stride, pad);
expectArraysClose(await result.data(), [
0.04919822, -0.09839644, 0.09860218, -0.19720435, 0.07275512, -0.14551024,
0.03405062, -0.06810125, 0.08081452, -0.16162904, 0.04651042, -0.09302084,
0.05150411, -0.10300821, 0.01305549, -0.02611098, 0.09901544, -0.19803089,
0.03949417, -0.07898834, 0.05555845, -0.11111691, 0.0144028, -0.02880561,
0.01898637, -0.03797274, 0.02086828, -0.04173655, 0.01416401, -0.02832802,
0.01685872, -0.03371745
]);
expect(result.shape).toEqual([1, 4, 4, outDepth]);
});
it('TensorLike', async () => {
const pad = 'valid';
const stride = 1;
const outDepth = 2;
const x = [[
[[0.230664], [0.987388], [0.0685208]],
[[0.419224], [0.887861], [0.731641]],
[[0.0741907], [0.409265], [0.351377]]
]];
const depthwiseFilter = [[[[0.303873]], [[0.229223]]], [[[0.144333]], [[0.803373]]]];
const pointwiseFilter = [[[[0.1, -0.2]]]];
const result = tf.separableConv2d(x, depthwiseFilter, pointwiseFilter, stride, pad);
expectArraysClose(await result.data(), [
0.10702161, -0.21404321, 0.10316753, -0.20633507, 0.06704096, -0.13408193,
0.07788632, -0.15577264
]);
expect(result.shape).toEqual([1, 2, 2, outDepth]);
});
it('TensorLike Chained', async () => {
const pad = 'valid';
const stride = 1;
const outDepth = 2;
const inDepth = 1;
const x = tf.tensor4d([
0.230664, 0.987388, 0.0685208, 0.419224, 0.887861, 0.731641,
0.0741907, 0.409265, 0.351377
], [1, 3, 3, inDepth]);
const depthwiseFilter = [[[[0.303873]], [[0.229223]]], [[[0.144333]], [[0.803373]]]];
const pointwiseFilter = [[[[0.1, -0.2]]]];
const result = x.separableConv2d(depthwiseFilter, pointwiseFilter, stride, pad);
expectArraysClose(await result.data(), [
0.10702161, -0.21404321, 0.10316753, -0.20633507, 0.06704096, -0.13408193,
0.07788632, -0.15577264
]);
expect(result.shape).toEqual([1, 2, 2, outDepth]);
});
it('Incorrect input rank raises error', () => {
// tslint:disable-next-line:no-any
const x = tf.zeros([4, 4]);
const depthwiseFilter = tf.zeros([2, 2, 1, 3]);
const pointwiseFilter = tf.zeros([1, 1, 2, 4]);
expect(() => tf.separableConv2d(x, depthwiseFilter, pointwiseFilter, 1, 'valid'))
.toThrowError(/rank 4/);
});
it('Incorrect depthwise filter rank raises error', () => {
const x = tf.zeros([1, 4, 4, 1]);
// tslint:disable-next-line:no-any
const depthwiseFilter = tf.zeros([2, 2, 1]);
const pointwiseFilter = tf.zeros([1, 1, 2, 4]);
expect(() => tf.separableConv2d(x, depthwiseFilter, pointwiseFilter, 1, 'valid'))
.toThrowError(/rank 4/);
});
it('Incorrect depthwise filter rank raises error', () => {
const x = tf.zeros([1, 4, 4, 1]);
const depthwiseFilter = tf.zeros([2, 2, 1, 3]);
// tslint:disable-next-line:no-any
const pointwiseFilter = tf.zeros([1, 1, 2]);
expect(() => tf.separableConv2d(x, depthwiseFilter, pointwiseFilter, 1, 'valid'))
.toThrowError(/rank 4/);
});
it('Incorrect point filter 1st dimension raises error', () => {
const x = tf.zeros([1, 4, 4, 1]);
const depthwiseFilter = tf.zeros([2, 2, 1, 3]);
const pointwiseFilter = tf.zeros([2, 1, 3, 6]);
expect(() => tf.separableConv2d(x, depthwiseFilter, pointwiseFilter, 1, 'valid'))
.toThrowError(/must be 1, but got 2/);
});
it('Incorrect point filter 2nd dimension raises error', () => {
const x = tf.zeros([1, 4, 4, 1]);
const depthwiseFilter = tf.zeros([2, 2, 1, 3]);
const pointwiseFilter = tf.zeros([1, 5, 3, 6]);
expect(() => tf.separableConv2d(x, depthwiseFilter, pointwiseFilter, 1, 'valid'))
.toThrowError(/must be 1, but got 5/);
});
it('Incorrect pointwise filter 3rd dimension raises error', () => {
const x = tf.zeros([1, 4, 4, 1]);
const depthwiseFilter = tf.zeros([2, 2, 1, 3]);
const pointwiseFilter = tf.zeros([1, 1, 4, 6]);
expect(() => tf.separableConv2d(x, depthwiseFilter, pointwiseFilter, 1, 'valid'))
.toThrowError(/must be 3, but got 4/);
});
it('throws when passed x as a non-tensor', () => {
const fSize = 2;
const pad = 'valid';
const stride = 1;
const chMul = 1;
const inDepth = 1;
const outDepth = 2;
const depthwiseFilter = tf.zeros([fSize, fSize, inDepth, chMul]);
const pointwiseFilter = tf.zeros([1, 1, inDepth * chMul, outDepth]);
const e = /Argument 'x' passed to 'separableConv2d' must be a Tensor/;
expect(() => tf.separableConv2d({}, depthwiseFilter, pointwiseFilter, stride, pad))
.toThrowError(e);
});
it('throws when passed depthwiseFilter as a non-tensor', () => {
const pad = 'valid';
const stride = 1;
const chMul = 1;
const inDepth = 1;
const outDepth = 2;
const x = tf.zeros([1, 3, 3, inDepth]);
const pointwiseFilter = tf.zeros([1, 1, inDepth * chMul, outDepth]);
const e = new RegExp('Argument \'depthwiseFilter\' passed to \'separableConv2d\'' +
' must be a Tensor');
expect(() => tf.separableConv2d(x, {}, pointwiseFilter, stride, pad))
.toThrowError(e);
});
it('throws when passed pointwiseFilter as a non-tensor', () => {
const fSize = 2;
const pad = 'valid';
const stride = 1;
const chMul = 1;
const inDepth = 1;
const x = tf.zeros([1, 3, 3, inDepth]);
const depthwiseFilter = tf.zeros([fSize, fSize, inDepth, chMul]);
const e = new RegExp('Argument \'pointwiseFilter\' passed to \'separableConv2d\'' +
' must be a Tensor');
expect(() => tf.separableConv2d(x, depthwiseFilter, {}, stride, pad))
.toThrowError(e);
});
it('accepts a tensor-like object', async () => {
const pad = 'valid';
const stride = 1;
const outDepth = 2;
// 3x3x1
const x = [
[[0.230664], [0.987388], [0.0685208]],
[[0.419224], [0.887861], [0.731641]],
[[0.0741907], [0.409265], [0.351377]]
];
// 2x2x1x1
const depthwiseFilter = [[[[0.303873]], [[0.229223]]], [[[0.144333]], [[0.803373]]]];
// 1x1x1x2
const pointwiseFilter = [[[[0.1, -0.2]]]];
const result = tf.separableConv2d(x, depthwiseFilter, pointwiseFilter, stride, pad);
expectArraysClose(await result.data(), [
0.10702161, -0.21404321, 0.10316753, -0.20633507, 0.06704096, -0.13408193,
0.07788632, -0.15577264
]);
expect(result.shape).toEqual([2, 2, outDepth]);
});
});
//# sourceMappingURL=conv2d_separable_test.js.map