UNPKG

@tensorflow/tfjs-core

Version:

Hardware-accelerated JavaScript library for machine intelligence

374 lines 16.8 kB
/** * @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'; import { sizeFromShape } from '../util'; // Generates small floating point inputs to avoid overflows function generateCaseInputs(totalSizeTensor, totalSizeFilter) { const inp = new Array(totalSizeTensor); const filt = new Array(totalSizeFilter); for (let i = 0; i < totalSizeTensor; i++) { inp[i] = (i + 1) / totalSizeTensor; } for (let i = 0; i < totalSizeFilter; i++) { filt[i] = (i + 1) / totalSizeFilter; } return { input: inp, filter: filt }; } function generateGradientCaseInputs(totalSizeTensor, totalSizeFilter) { const inp = new Array(totalSizeTensor); const filt = new Array(totalSizeFilter); for (let i = 0; i < totalSizeTensor; i++) { inp[i] = i + 1; } for (let i = 0; i < totalSizeFilter; i++) { filt[i] = i + 1; } return { input: inp, filter: filt }; } function runConv3DTestCase(batch, inDepth, inHeight, inWidth, inChannels, outChannels, fDepth, fHeight, fWidth, pad, stride) { const inputShape = [batch, inDepth, inHeight, inWidth, inChannels]; const filterShape = [fDepth, fHeight, fWidth, inChannels, outChannels]; const totalSizeTensor = sizeFromShape(inputShape); const totalSizeFilter = sizeFromShape(filterShape); const inputs = generateCaseInputs(totalSizeTensor, totalSizeFilter); const x = tf.tensor5d(inputs.input, inputShape); const w = tf.tensor5d(inputs.filter, filterShape); const result = tf.conv3d(x, w, stride, pad); return result; } function runGradientConv3DTestCase(batch, inDepth, inHeight, inWidth, inChannels, outChannels, fDepth, fHeight, fWidth, pad, stride) { const inputShape = [batch, inDepth, inHeight, inWidth, inChannels]; const filterShape = [fDepth, fHeight, fWidth, inChannels, outChannels]; const totalSizeTensor = sizeFromShape(inputShape); const totalSizeFilter = sizeFromShape(filterShape); const inputs = generateGradientCaseInputs(totalSizeTensor, totalSizeFilter); const x = tf.tensor5d(inputs.input, inputShape); const w = tf.tensor5d(inputs.filter, filterShape); const grads = tf.grads((x, filter) => tf.conv3d(x.clone(), filter.clone(), stride, pad).clone()); const [dx, dfilter] = grads([x, w]); expect(dx.shape).toEqual(x.shape); expect(dfilter.shape).toEqual(w.shape); return [dx, dfilter]; } describeWithFlags('conv3d', ALL_ENVS, () => { it('x=[1, 2, 3, 1, 3] f=[1, 1, 1, 3, 3] s=1 d=1 p=valid', async () => { const batch = 1; const inDepth = 2; const inHeight = 3; const inWidth = 1; const inChannels = 3; const outChannels = 3; const fSize = 1; const pad = 'valid'; const stride = 1; const result = runConv3DTestCase(batch, inDepth, inHeight, inWidth, inChannels, outChannels, fSize, fSize, fSize, pad, stride); const expectedOutput = [ 0.18518519, 0.22222222, 0.25925926, 0.40740741, 0.5, 0.59259259, 0.62962963, 0.77777778, 0.92592593, 0.85185185, 1.05555556, 1.25925926, 1.07407407, 1.33333333, 1.59259259, 1.2962963, 1.61111111, 1.92592593 ]; expectArraysClose(await result.data(), expectedOutput); }); it('x=[1, 2, 1, 3, 3] f=[1, 1, 1, 3, 3] s=1 d=1 p=valid', async () => { const batch = 1; const inDepth = 2; const inHeight = 1; const inWidth = 3; const inChannels = 3; const outChannels = 3; const fSize = 1; const pad = 'valid'; const stride = 1; const result = runConv3DTestCase(batch, inDepth, inHeight, inWidth, inChannels, outChannels, fSize, fSize, fSize, pad, stride); const expectedOutput = [ 0.18518519, 0.22222222, 0.25925926, 0.40740741, 0.5, 0.59259259, 0.62962963, 0.77777778, 0.92592593, 0.85185185, 1.05555556, 1.25925926, 1.07407407, 1.33333333, 1.59259259, 1.2962963, 1.61111111, 1.92592593 ]; expectArraysClose(await result.data(), expectedOutput); }); it('x=[1, 1, 2, 3, 3] f=[1, 1, 1, 3, 3] s=1 d=1 p=valid', async () => { const batch = 1; const inDepth = 1; const inHeight = 2; const inWidth = 3; const inChannels = 3; const outChannels = 3; const fSize = 1; const pad = 'valid'; const stride = 1; const result = runConv3DTestCase(batch, inDepth, inHeight, inWidth, inChannels, outChannels, fSize, fSize, fSize, pad, stride); const expectedOutput = [ 0.18518519, 0.22222222, 0.25925926, 0.40740741, 0.5, 0.59259259, 0.62962963, 0.77777778, 0.92592593, 0.85185185, 1.05555556, 1.25925926, 1.07407407, 1.33333333, 1.59259259, 1.2962963, 1.61111111, 1.92592593 ]; expectArraysClose(await result.data(), expectedOutput); }); it('x=[1, 4, 2, 3, 3] f=[2, 2, 2, 3, 3] s=1 d=1 p=valid', async () => { const batch = 1; const inDepth = 4; const inHeight = 2; const inWidth = 3; const inChannels = 3; const outChannels = 3; const fSize = 2; const pad = 'valid'; const stride = 1; const result = runConv3DTestCase(batch, inDepth, inHeight, inWidth, inChannels, outChannels, fSize, fSize, fSize, pad, stride); const expectedOutput = [ 3.77199074, 3.85069444, 3.92939815, 4.2650463, 4.35763889, 4.45023148, 6.73032407, 6.89236111, 7.05439815, 7.22337963, 7.39930556, 7.57523148, 9.68865741, 9.93402778, 10.17939815, 10.18171296, 10.44097222, 10.70023148 ]; expectArraysClose(await result.data(), expectedOutput); }); it('x=[1, 5, 8, 7, 1] f=[1, 2, 3, 1, 1] s=[2, 3, 1] d=1 p=same', async () => { const batch = 1; const inDepth = 5; const inHeight = 8; const inWidth = 7; const inChannels = 1; const outChannels = 1; const fDepth = 1; const fHeight = 2; const fWidth = 3; const pad = 'same'; const stride = [2, 3, 1]; const result = runConv3DTestCase(batch, inDepth, inHeight, inWidth, inChannels, outChannels, fDepth, fHeight, fWidth, pad, stride); const expectedOutput = [ 0.06071429, 0.08988095, 0.10238095, 0.11488095, 0.12738095, 0.13988095, 0.08452381, 0.26071429, 0.35238095, 0.36488095, 0.37738095, 0.38988095, 0.40238095, 0.23452381, 0.46071429, 0.61488095, 0.62738095, 0.63988095, 0.65238095, 0.66488095, 0.38452381, 1.12738095, 1.48988095, 1.50238095, 1.51488095, 1.52738095, 1.53988095, 0.88452381, 1.32738095, 1.75238095, 1.76488095, 1.77738095, 1.78988095, 1.80238095, 1.03452381, 1.52738095, 2.01488095, 2.02738095, 2.03988095, 2.05238095, 2.06488095, 1.18452381, 2.19404762, 2.88988095, 2.90238095, 2.91488095, 2.92738095, 2.93988095, 1.68452381, 2.39404762, 3.15238095, 3.16488095, 3.17738095, 3.18988095, 3.20238095, 1.83452381, 2.59404762, 3.41488095, 3.42738095, 3.43988095, 3.45238095, 3.46488095, 1.98452381 ]; expectArraysClose(await result.data(), expectedOutput); }); it('x=[1, 4, 2, 3, 3] f=[2, 2, 2, 3, 3] s=2 d=1 p=valid', async () => { const batch = 1; const inDepth = 4; const inHeight = 2; const inWidth = 3; const inChannels = 3; const outChannels = 3; const fSize = 2; const pad = 'valid'; const stride = 2; const result = runConv3DTestCase(batch, inDepth, inHeight, inWidth, inChannels, outChannels, fSize, fSize, fSize, pad, stride); const expectedOutput = [ 3.77199074, 3.85069444, 3.92939815, 9.68865741, 9.93402778, 10.17939815 ]; expectArraysClose(await result.data(), expectedOutput); }); it('x=[1, 6, 7, 8, 2] f=[3, 2, 1, 2, 3] s=3 d=1 p=valid', async () => { const batch = 1; const inDepth = 6; const inHeight = 7; const inWidth = 8; const inChannels = 2; const outChannels = 3; const fDepth = 3; const fHeight = 2; const fWidth = 1; const pad = 'valid'; const stride = 3; const result = runConv3DTestCase(batch, inDepth, inHeight, inWidth, inChannels, outChannels, fDepth, fHeight, fWidth, pad, stride); const expectedOutput = [ 1.51140873, 1.57167659, 1.63194444, 1.56349206, 1.62673611, 1.68998016, 1.6155754, 1.68179563, 1.74801587, 1.9280754, 2.01215278, 2.09623016, 1.98015873, 2.0672123, 2.15426587, 2.03224206, 2.12227183, 2.21230159, 4.4280754, 4.65500992, 4.88194444, 4.48015873, 4.71006944, 4.93998016, 4.53224206, 4.76512897, 4.99801587, 4.84474206, 5.09548611, 5.34623016, 4.8968254, 5.15054563, 5.40426587, 4.94890873, 5.20560516, 5.46230159 ]; expectArraysClose(await result.data(), expectedOutput); }); it('x=[1, 4, 2, 3, 3] f=[2, 2, 2, 3, 3] s=2 d=1 p=same', async () => { const batch = 1; const inDepth = 4; const inHeight = 2; const inWidth = 3; const inChannels = 3; const outChannels = 3; const fSize = 2; const pad = 'same'; const stride = 2; const result = runConv3DTestCase(batch, inDepth, inHeight, inWidth, inChannels, outChannels, fSize, fSize, fSize, pad, stride); const expectedOutput = [ 3.77199074, 3.85069444, 3.92939815, 2.0162037, 2.06597222, 2.11574074, 9.68865741, 9.93402778, 10.17939815, 4.59953704, 4.73263889, 4.86574074 ]; expectArraysClose(await result.data(), expectedOutput); }); it('x=[1, 3, 3, 3, 1] f=[1, 1, 1, 1, 1] s=2 d=1 p=same', async () => { const batch = 1; const inDepth = 3; const inHeight = 3; const inWidth = 3; const inChannels = 1; const outChannels = 1; const fSize = 1; const pad = 'same'; const stride = 2; const result = runConv3DTestCase(batch, inDepth, inHeight, inWidth, inChannels, outChannels, fSize, fSize, fSize, pad, stride); const expectedOutput = [ 0.03703704, 0.11111111, 0.25925926, 0.33333333, 0.7037037, 0.77777778, 0.92592593, 1. ]; expectArraysClose(await result.data(), expectedOutput); }); it('x=[1, 3, 3, 3, 1] f=[1, 1, 1, 1, 1] s=2 d=1 p=valid', async () => { const batch = 1; const inDepth = 3; const inHeight = 3; const inWidth = 3; const inChannels = 1; const outChannels = 1; const fSize = 1; const pad = 'valid'; const stride = 2; const result = runConv3DTestCase(batch, inDepth, inHeight, inWidth, inChannels, outChannels, fSize, fSize, fSize, pad, stride); const expectedOutput = [ 0.03703704, 0.11111111, 0.25925926, 0.33333333, 0.7037037, 0.77777778, 0.92592593, 1. ]; expectArraysClose(await result.data(), expectedOutput); }); it('x=[1, 7, 7, 7, 1] f=[2, 2, 2, 1, 1] s=3 d=1 p=same', async () => { const batch = 1; const inDepth = 7; const inHeight = 7; const inWidth = 7; const inChannels = 1; const outChannels = 1; const fSize = 2; const pad = 'same'; const stride = 3; const result = runConv3DTestCase(batch, inDepth, inHeight, inWidth, inChannels, outChannels, fSize, fSize, fSize, pad, stride); const expectedOutput = [ 0.54081633, 0.58017493, 0.28061224, 0.81632653, 0.85568513, 0.40306122, 0.41873178, 0.4340379, 0.19642857, 2.46938776, 2.50874636, 1.1377551, 2.74489796, 2.78425656, 1.26020408, 1.16873178, 1.1840379, 0.51785714, 1.09511662, 1.10604956, 0.44642857, 1.17164723, 1.18258017, 0.47704082, 0.3691691, 0.37244898, 0.125 ]; expectArraysClose(await result.data(), expectedOutput); }); it('x=[1, 7, 7, 7, 1] f=[2, 2, 2, 1, 1] s=3 d=1 p=valid', async () => { const batch = 1; const inDepth = 7; const inHeight = 7; const inWidth = 7; const inChannels = 1; const outChannels = 1; const fSize = 2; const pad = 'valid'; const stride = 3; const result = runConv3DTestCase(batch, inDepth, inHeight, inWidth, inChannels, outChannels, fSize, fSize, fSize, pad, stride); const expectedOutput = [ 0.540816, 0.580175, 0.816327, 0.855685, 2.469388, 2.508746, 2.744898, 2.784257 ]; expectArraysClose(await result.data(), expectedOutput); }); it('x=[1, 2, 1, 2, 1] f=[2, 1, 2, 1, 2] s=1 d=1 p=valid', async () => { const batch = 1; const inDepth = 2; const inHeight = 1; const inWidth = 2; const inChannels = 1; const outChannels = 2; const fDepth = 2; const fHeight = 1; const fWidth = 2; const pad = 'valid'; const stride = 1; const result = runConv3DTestCase(batch, inDepth, inHeight, inWidth, inChannels, outChannels, fDepth, fHeight, fWidth, pad, stride); const expectedOutput = [1.5625, 1.875]; expectArraysClose(await result.data(), expectedOutput); }); it('gradient with clones, x=[1,3,6,1,1] filter=[2,2,1,1,1] s=1 d=1 p=valid', async () => { const batch = 1; const inDepth = 3; const inHeight = 6; const inWidth = 1; const inChannels = 1; const outChannels = 1; const fDepth = 2; const fHeight = 2; const fWidth = 1; const pad = 'valid'; const stride = 1; const [dx, dfilter] = runGradientConv3DTestCase(batch, inDepth, inHeight, inWidth, inChannels, outChannels, fDepth, fHeight, fWidth, pad, stride); const expectedFilterOutput = [60.0, 70.0, 120.0, 130.0]; const expectedOutput = [ 1.0, 3.0, 3.0, 3.0, 3.0, 2.0, 4.0, 10.0, 10.0, 10.0, 10.0, 6.0, 3.0, 7.0, 7.0, 7.0, 7.0, 4.0 ]; expectArraysClose(await dx.data(), expectedOutput); expectArraysClose(await dfilter.data(), expectedFilterOutput); }); it('throws when passed x as a non-tensor', () => { const inputDepth = 1; const outputDepth = 1; const fSize = 1; const pad = 'valid'; const stride = 1; const w = tf.tensor5d([2], [fSize, fSize, fSize, inputDepth, outputDepth]); expect(() => tf.conv3d({}, w, stride, pad)) .toThrowError(/Argument 'x' passed to 'conv3d' must be a Tensor/); }); it('throws when passed filter as a non-tensor', () => { const inputDepth = 1; const inputShape = [2, 2, 1, inputDepth]; const pad = 'valid'; const stride = 1; const x = tf.tensor4d([1, 2, 3, 4], inputShape); expect(() => tf.conv3d(x, {}, stride, pad)) .toThrowError(/Argument 'filter' passed to 'conv3d' must be a Tensor/); }); it('accepts a tensor-like object', async () => { const pad = 'valid'; const stride = 1; const x = [[[[1], [2]], [[3], [4]]]]; // 2x2x1x1 const w = [[[[[2]]]]]; // 1x1x1x1x1 const result = tf.conv3d(x, w, stride, pad); expectArraysClose(await result.data(), [2, 4, 6, 8]); }); it('throws when data format not NDHWC', () => { const inputDepth = 1; const outputDepth = 1; const inputShape = [2, 2, 1, inputDepth]; const pad = 'valid'; const fSize = 1; const stride = 1; const dataFormat = 'NCDHW'; const x = tf.tensor4d([1, 2, 3, 4], inputShape); const w = tf.tensor5d([2], [fSize, fSize, fSize, inputDepth, outputDepth]); expect(() => tf.conv3d(x, w, stride, pad, dataFormat)).toThrowError(); }); }); //# sourceMappingURL=conv3d_test.js.map