|
| 1 | +/** |
| 2 | + * @license |
| 3 | + * Copyright 2017 Google Inc. 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 | + |
| 18 | +import * as test_util from '../test_util'; |
| 19 | +import {MathTests} from '../test_util'; |
| 20 | + |
| 21 | +import {Array1D, Array3D} from './ndarray'; |
| 22 | + |
| 23 | +// math.batchNormalization3D |
| 24 | +{ |
| 25 | + // TODO(nsthorat): Fix the precision for byte-packed batchnorm. |
| 26 | + const epsilon = 1e-1; |
| 27 | + const tests: MathTests = it => { |
| 28 | + it('simple batchnorm, no offset or scale, 2x1x2', math => { |
| 29 | + const x = Array3D.new([2, 1, 2], new Float32Array([2, 100, 4, 400])); |
| 30 | + const mean = Array1D.new([1, 2]); |
| 31 | + const variance = Array1D.new([2, 3]); |
| 32 | + const varianceEpsilon = .001; |
| 33 | + |
| 34 | + const result = math.batchNormalization3D( |
| 35 | + x, mean, variance, varianceEpsilon, undefined, undefined); |
| 36 | + |
| 37 | + test_util.expectArraysClose( |
| 38 | + result.getValues(), new Float32Array([ |
| 39 | + (x.get(0, 0, 0) - mean.get(0)) * 1 / |
| 40 | + Math.sqrt(variance.get(0) + varianceEpsilon), |
| 41 | + (x.get(0, 0, 1) - mean.get(1)) * 1 / |
| 42 | + Math.sqrt(variance.get(1) + varianceEpsilon), |
| 43 | + (x.get(1, 0, 0) - mean.get(0)) * 1 / |
| 44 | + Math.sqrt(variance.get(0) + varianceEpsilon), |
| 45 | + (x.get(1, 0, 1) - mean.get(1)) * 1 / |
| 46 | + Math.sqrt(variance.get(1) + varianceEpsilon) |
| 47 | + ]), |
| 48 | + epsilon); |
| 49 | + |
| 50 | + x.dispose(); |
| 51 | + mean.dispose(); |
| 52 | + variance.dispose(); |
| 53 | + }); |
| 54 | + |
| 55 | + it('simple batchnorm, no offset, 2x1x2', math => { |
| 56 | + const x = Array3D.new([2, 1, 2], new Float32Array([2, 100, 4, 400])); |
| 57 | + const mean = Array1D.new([1, 2]); |
| 58 | + const variance = Array1D.new([2, 3]); |
| 59 | + const scale = Array1D.new([4, 5]); |
| 60 | + const varianceEpsilon = .001; |
| 61 | + |
| 62 | + const result = math.batchNormalization3D( |
| 63 | + x, mean, variance, varianceEpsilon, scale, undefined); |
| 64 | + |
| 65 | + test_util.expectArraysClose( |
| 66 | + result.getValues(), new Float32Array([ |
| 67 | + (x.get(0, 0, 0) - mean.get(0)) * scale.get(0) / |
| 68 | + Math.sqrt(variance.get(0) + varianceEpsilon), |
| 69 | + (x.get(0, 0, 1) - mean.get(1)) * scale.get(1) / |
| 70 | + Math.sqrt(variance.get(1) + varianceEpsilon), |
| 71 | + (x.get(1, 0, 0) - mean.get(0)) * scale.get(0) / |
| 72 | + Math.sqrt(variance.get(0) + varianceEpsilon), |
| 73 | + (x.get(1, 0, 1) - mean.get(1)) * scale.get(1) / |
| 74 | + Math.sqrt(variance.get(1) + varianceEpsilon) |
| 75 | + ]), |
| 76 | + epsilon); |
| 77 | + |
| 78 | + x.dispose(); |
| 79 | + mean.dispose(); |
| 80 | + variance.dispose(); |
| 81 | + scale.dispose(); |
| 82 | + }); |
| 83 | + |
| 84 | + it('simple batchnorm, no scale, 2x1x2', math => { |
| 85 | + const x = Array3D.new([2, 1, 2], new Float32Array([2, 100, 4, 400])); |
| 86 | + const mean = Array1D.new([1, 2]); |
| 87 | + const variance = Array1D.new([2, 3]); |
| 88 | + const offset = Array1D.new([4, 5]); |
| 89 | + |
| 90 | + const varianceEpsilon = .001; |
| 91 | + |
| 92 | + const result = math.batchNormalization3D( |
| 93 | + x, mean, variance, varianceEpsilon, undefined, offset); |
| 94 | + |
| 95 | + test_util.expectArraysClose( |
| 96 | + result.getValues(), new Float32Array([ |
| 97 | + offset.get(0) + |
| 98 | + (x.get(0, 0, 0) - mean.get(0)) * 1 / |
| 99 | + Math.sqrt(variance.get(0) + varianceEpsilon), |
| 100 | + offset.get(1) + |
| 101 | + (x.get(0, 0, 1) - mean.get(1)) * 1 / |
| 102 | + Math.sqrt(variance.get(1) + varianceEpsilon), |
| 103 | + offset.get(0) + |
| 104 | + (x.get(1, 0, 0) - mean.get(0)) * 1 / |
| 105 | + Math.sqrt(variance.get(0) + varianceEpsilon), |
| 106 | + offset.get(1) + |
| 107 | + (x.get(1, 0, 1) - mean.get(1)) * 1 / |
| 108 | + Math.sqrt(variance.get(1) + varianceEpsilon) |
| 109 | + ]), |
| 110 | + epsilon); |
| 111 | + x.dispose(); |
| 112 | + mean.dispose(); |
| 113 | + variance.dispose(); |
| 114 | + offset.dispose(); |
| 115 | + }); |
| 116 | + |
| 117 | + it('simple batchnorm, 2x1x2', math => { |
| 118 | + const x = Array3D.new([2, 1, 2], new Float32Array([2, 100, 4, 400])); |
| 119 | + const mean = Array1D.new([1, 2]); |
| 120 | + const variance = Array1D.new([2, 3]); |
| 121 | + const offset = Array1D.new([3, 4]); |
| 122 | + const scale = Array1D.new([4, 5]); |
| 123 | + |
| 124 | + const varianceEpsilon = .001; |
| 125 | + |
| 126 | + const result = math.batchNormalization3D( |
| 127 | + x, mean, variance, varianceEpsilon, scale, offset); |
| 128 | + |
| 129 | + test_util.expectArraysClose( |
| 130 | + result.getValues(), new Float32Array([ |
| 131 | + offset.get(0) + |
| 132 | + (x.get(0, 0, 0) - mean.get(0)) * scale.get(0) / |
| 133 | + Math.sqrt(variance.get(0) + varianceEpsilon), |
| 134 | + offset.get(1) + |
| 135 | + (x.get(0, 0, 1) - mean.get(1)) * scale.get(1) / |
| 136 | + Math.sqrt(variance.get(1) + varianceEpsilon), |
| 137 | + offset.get(0) + |
| 138 | + (x.get(1, 0, 0) - mean.get(0)) * scale.get(0) / |
| 139 | + Math.sqrt(variance.get(0) + varianceEpsilon), |
| 140 | + offset.get(1) + |
| 141 | + (x.get(1, 0, 1) - mean.get(1)) * scale.get(1) / |
| 142 | + Math.sqrt(variance.get(1) + varianceEpsilon) |
| 143 | + ]), |
| 144 | + epsilon); |
| 145 | + x.dispose(); |
| 146 | + mean.dispose(); |
| 147 | + variance.dispose(); |
| 148 | + scale.dispose(); |
| 149 | + offset.dispose(); |
| 150 | + }); |
| 151 | + |
| 152 | + it('batchnorm matches tensorflow, 2x3x3', math => { |
| 153 | + const x = Array3D.new( |
| 154 | + [2, 3, 3], new Float32Array([ |
| 155 | + 0.49955603, 0.04158615, -1.09440524, 2.03854165, -0.61578344, |
| 156 | + 2.87533573, 1.18105987, 0.807462, 1.87888837, 2.26563962, |
| 157 | + -0.37040935, 1.35848753, -0.75347094, 0.15683117, 0.91925946, |
| 158 | + 0.34121279, 0.92717143, 1.89683965 |
| 159 | + ])); |
| 160 | + const mean = Array1D.new([0.39745062, -0.48062894, 0.4847822]); |
| 161 | + const variance = Array1D.new([0.32375343, 0.67117643, 1.08334653]); |
| 162 | + const offset = Array1D.new([0.69398749, -1.29056387, 0.9429723]); |
| 163 | + const scale = Array1D.new([-0.5607271, 0.9878457, 0.25181573]); |
| 164 | + const varianceEpsilon = .001; |
| 165 | + |
| 166 | + const result = math.batchNormalization3D( |
| 167 | + x, mean, variance, varianceEpsilon, scale, offset); |
| 168 | + |
| 169 | + test_util.expectArraysClose( |
| 170 | + result.getValues(), new Float32Array([ |
| 171 | + 0.59352049, -0.66135202, 0.5610874, -0.92077015, -1.45341019, |
| 172 | + 1.52106473, -0.07704776, 0.26144429, 1.28010017, -1.14422404, |
| 173 | + -1.15776136, 1.15425493, 1.82644104, -0.52249442, 1.04803919, |
| 174 | + 0.74932291, 0.40568101, 1.2844412 |
| 175 | + ])); |
| 176 | + |
| 177 | + x.dispose(); |
| 178 | + mean.dispose(); |
| 179 | + variance.dispose(); |
| 180 | + scale.dispose(); |
| 181 | + offset.dispose(); |
| 182 | + }); |
| 183 | + }; |
| 184 | + |
| 185 | + test_util.describeMathCPU('batchNormalization3D', [tests]); |
| 186 | + test_util.describeMathGPU('batchNormalization3D', [tests], [ |
| 187 | + {'WEBGL_FLOAT_TEXTURE_ENABLED': true, 'WEBGL_VERSION': 1}, |
| 188 | + {'WEBGL_FLOAT_TEXTURE_ENABLED': true, 'WEBGL_VERSION': 2}, |
| 189 | + {'WEBGL_FLOAT_TEXTURE_ENABLED': false, 'WEBGL_VERSION': 1} |
| 190 | + ]); |
| 191 | +} |
0 commit comments