UNPKG

@tensorflow/tfjs-core

Version:

Hardware-accelerated JavaScript library for machine intelligence

342 lines 15 kB
/** * @license * Copyright 2017 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('conv2dTranspose', ALL_ENVS, () => { it('input=2x2x1,d2=1,f=2,s=1,p=0', async () => { const origInputDepth = 1; const origOutputDepth = 1; const inputShape = [1, 1, origOutputDepth]; const fSize = 2; const origPad = 0; const origStride = 1; const x = tf.tensor3d([2], inputShape); const w = tf.tensor4d([3, 1, 5, 0], [fSize, fSize, origInputDepth, origOutputDepth]); const result = tf.conv2dTranspose(x, w, [2, 2, 1], origStride, origPad); const expected = [6, 2, 10, 0]; expect(result.shape).toEqual([2, 2, 1]); expectArraysClose(await result.data(), expected); }); it('input=2x2x1,d2=1,f=2,s=1,p=0, batch=2', async () => { const origInputDepth = 1; const origOutputDepth = 1; const inputShape = [2, 1, 1, origOutputDepth]; const fSize = 2; const origPad = 0; const origStride = 1; const x = tf.tensor4d([2, 3], inputShape); const w = tf.tensor4d([3, 1, 5, 0], [fSize, fSize, origInputDepth, origOutputDepth]); const result = tf.conv2dTranspose(x, w, [2, 2, 2, 1], origStride, origPad); const expected = [6, 2, 10, 0, 9, 3, 15, 0]; expect(result.shape).toEqual([2, 2, 2, 1]); expectArraysClose(await result.data(), expected); }); // Reference (Python) TensorFlow code: // // ```py // import numpy as np // import tensorflow as tf // // tf.enable_eager_execution() // // x = tf.constant(np.array([[ // [[-0.14656299], [0.32942239], [-1.90302866]], // [[-0.06487813], [-2.02637842], [-1.83669377]], // [[0.82650784], [-0.89249092], [0.01207666]] // ]]).astype(np.float32)) // filt = tf.constant(np.array([ // [[[-0.48280062], [1.26770487]], [[-0.83083738], [0.54341856]]], // [[[-0.274904], [0.73111374]], [[2.01885189], [-2.68975237]]] // ]).astype(np.float32)) // // with tf.GradientTape() as g: // g.watch(x) // g.watch(filt) // y = tf.keras.backend.conv2d_transpose(x, filt, [1, 4, 4, 2]) // print(y) // (x_grad, filt_grad) = g.gradient(y, [x, filt]) // // print("x_grad = %s" % x_grad) // print("filt_grad = %s" % filt_grad) // ``` it('gradient with clones input=[1,3,3,1] f=[2,2,2,1] s=1 padding=valid', async () => { const inputDepth = 1; const outputDepth = 2; const inputShape = [1, 3, 3, inputDepth]; const filterSize = 2; const stride = 1; const pad = 'valid'; const filterShape = [filterSize, filterSize, outputDepth, inputDepth]; const x = tf.tensor4d([[ [[-0.14656299], [0.32942239], [-1.90302866]], [[-0.06487813], [-2.02637842], [-1.83669377]], [[0.82650784], [-0.89249092], [0.01207666]] ]], inputShape); const filt = tf.tensor4d([ [[[-0.48280062], [1.26770487]], [[-0.83083738], [0.54341856]]], [[[-0.274904], [0.73111374]], [[2.01885189], [-2.68975237]]] ], filterShape); const grads = tf.grads((x, filter) => tf.conv2dTranspose(x.clone(), filter.clone(), [1, 4, 4, outputDepth], stride, pad) .clone()); const dy = tf.ones([1, 4, 4, outputDepth]); const [xGrad, filtGrad] = grads([x, filt], dy); const expectedXGrad = tf.ones([1, 3, 3, 1]).mul(tf.scalar(0.2827947)); expectArraysClose(await xGrad.data(), await expectedXGrad.data()); const expectedFiltGrad = tf.ones([2, 2, 2, 1]).mul(tf.scalar(-5.70202599)); expectArraysClose(await filtGrad.data(), await expectedFiltGrad.data()); }); // Reference (Python) TensorFlow code: // // ```py // import numpy as np // import tensorflow as tf // // tf.enable_eager_execution() // // x = tf.constant(np.array([ // [[[-0.36541713], [-0.53973116]], [[0.01731674], [0.90227772]]] // ]).astype(np.float32)) // filt = tf.constant(np.array([ // [[[-0.01423461], [-1.00267384]], [[1.61163029], [0.66302646]]], // [[[-0.46900087], [-0.78649444]], [[0.87780536], [-0.84551637]]] // ]).astype(np.float32)) // // with tf.GradientTape() as g: // g.watch(x) // g.watch(filt) // y = tf.keras.backend.conv2d_transpose(x, filt, [1, 4, 4, 2], strides=(2, // 2)) print(y) // (x_grad, filt_grad) = g.gradient(y, [x, filt]) // // print("x_grad = %s" % -x_grad) // print("filt_grad = %s" % -filt_grad) // ``` it('gradient input=[1,2,2,1] f=[2,2,2,1] s=[2,2] padding=valid', async () => { const inputDepth = 1; const outputDepth = 2; const inputShape = [1, 2, 2, inputDepth]; const filterSize = 2; const stride = [2, 2]; const pad = 'valid'; const filterShape = [filterSize, filterSize, outputDepth, inputDepth]; const x = tf.tensor4d([[[[-0.36541713], [-0.53973116]], [[0.01731674], [0.90227772]]]], inputShape); const filt = tf.tensor4d([ [[[-0.01423461], [-1.00267384]], [[1.61163029], [0.66302646]]], [[[-0.46900087], [-0.78649444]], [[0.87780536], [-0.84551637]]] ], filterShape); const grads = tf.grads((x, filter) => tf.conv2dTranspose(x, filter, [1, 4, 4, outputDepth], stride, pad)); const dy = tf.ones([1, 4, 4, outputDepth]).mul(tf.scalar(-1)); const [xGrad, filtGrad] = grads([x, filt], dy); const expectedXGrad = tf.ones([1, 2, 2, 1]).mul(tf.scalar(-0.03454196)); expectArraysClose(await xGrad.data(), await expectedXGrad.data()); expect(xGrad.shape).toEqual([1, 2, 2, 1]); const expectedFiltGrad = tf.ones([2, 2, 2, 1]).mul(tf.scalar(-0.01444618)); expectArraysClose(await filtGrad.data(), await expectedFiltGrad.data()); expect(filtGrad.shape).toEqual([2, 2, 2, 1]); }); // Reference (Python) TensorFlow code: // // ```py // import numpy as np // import tensorflow as tf // // tf.enable_eager_execution() // // x = tf.constant(np.array([[ // [[1.52433065], [-0.77053435], [-0.64562341]], // [[0.77962889], [1.58413887], [-0.25581856]], // [[-0.58966221], [0.05411662], [0.70749138]] // ]]).astype(np.float32)) // filt = tf.constant(np.array([ // [[[0.11178388], [-0.96654977]], [[1.21021296], [0.84121729]]], // [[[0.34968338], [-0.42306114]], [[1.27395733], [-1.09014535]]] // ]).astype(np.float32)) // // with tf.GradientTape() as g: // g.watch(x) // g.watch(filt) // y = tf.keras.backend.conv2d_transpose( // x, filt, [1, 3, 3, 2], strides=(1, 1), padding='same') // (x_grad, filt_grad) = g.gradient(y, [x, filt]) // // print("x_grad = %s" % x_grad) // print("filt_grad = %s" % filt_grad) // ``` it('gradient input=[1,3,3,1] f=[2,2,2,1] s=[1,1] padding=same', async () => { const inputDepth = 1; const outputDepth = 2; const inputShape = [1, 3, 3, inputDepth]; const filterSize = 2; const stride = [1, 1]; const pad = 'same'; const filterShape = [filterSize, filterSize, outputDepth, inputDepth]; const x = tf.tensor4d([[ [[1.52433065], [-0.77053435], [-0.64562341]], [[0.77962889], [1.58413887], [-0.25581856]], [[-0.58966221], [0.05411662], [0.70749138]] ]], inputShape); const filt = tf.tensor4d([ [[[0.11178388], [-0.96654977]], [[1.21021296], [0.84121729]]], [[[0.34968338], [-0.42306114]], [[1.27395733], [-1.09014535]]] ], filterShape); const grads = tf.grads((x, filter) => tf.conv2dTranspose(x, filter, [1, 3, 3, outputDepth], stride, pad)); const dy = tf.ones([1, 3, 3, outputDepth]); const [xGrad, filtGrad] = grads([x, filt], dy); expectArraysClose(await xGrad.array(), [[ [[1.30709858], [1.30709858], [-0.92814366]], [[1.30709858], [1.30709858], [-0.92814366]], [[1.19666437], [1.19666437], [-0.85476589]] ]]); expectArraysClose(await filtGrad.array(), [ [[[2.38806788], [2.38806788]], [[2.58201847], [2.58201847]]], [[[2.2161221], [2.2161221]], [[3.11756406], [3.11756406]]] ]); }); // Reference (Python) TensorFlow code: // // ```py // import numpy as np // import tensorflow as tf // // tf.enable_eager_execution() // // x = tf.constant(np.array([[ // [[1.52433065], [-0.77053435]], [[0.77962889], [1.58413887]], // ]]).astype(np.float32)) // filt = tf.constant(np.array([ // [[[0.11178388], [-0.96654977]], [[1.21021296], [0.84121729]]], // [[[0.34968338], [-0.42306114]], [[1.27395733], [-1.09014535]]] // ]).astype(np.float32)) // // with tf.GradientTape() as g: // g.watch(x) // g.watch(filt) // y = tf.keras.backend.conv2d_transpose( // x, filt, [1, 3, 3, 2], strides=(2, 2), padding='same') // print(y.shape) // (x_grad, filt_grad) = g.gradient(y, [x, filt]) // // print("x_grad = %s" % x_grad) // print("filt_grad = %s" % filt_grad) // ``` it('gradient input=[1,2,2,2] f=[2,2,2,1] s=[2,2] padding=same', async () => { const inputDepth = 2; const outputDepth = 2; const inputShape = [1, 2, 2, inputDepth]; const filterSize = 2; const stride = [2, 2]; const pad = 'same'; const filterShape = [filterSize, filterSize, outputDepth, inputDepth]; const x = tf.tensor4d([[ [[-1.81506593, 1.00900095], [-0.05199118, 0.26311377]], [[-1.18469792, -0.34780521], [2.04971242, -0.65154692]] ]], inputShape); const filt = tf.tensor4d([ [ [[0.19529686, -0.79594708], [0.70314057, -0.06081263]], [[0.28724744, 0.88522715], [-0.51824096, -0.97120989]] ], [ [[0.51872197, -1.17569193], [1.28316791, -0.81225092]], [[-0.44221532, 0.70058174], [-0.4849217, 0.03806348]] ] ], filterShape); const grads = tf.grads((x, filter) => tf.conv2dTranspose(x, filter, [1, 3, 3, outputDepth], stride, pad)); const dy = tf.ones([1, 3, 3, outputDepth]); const [xGrad, filtGrad] = grads([x, filt], dy); expectArraysClose(await xGrad.data(), [ 1.54219678, -2.19204008, 2.70032732, -2.84470257, 0.66744391, -0.94274245, 0.89843743, -0.85675972 ]); expect(xGrad.shape).toEqual([1, 2, 2, 2]); expectArraysClose(await filtGrad.data(), [ -1.00204261, 0.27276259, -1.00204261, 0.27276259, -2.99976385, 0.66119574, -2.99976385, 0.66119574, -1.86705711, 1.27211472, -1.86705711, 1.27211472, -1.81506593, 1.00900095, -1.81506593, 1.00900095 ]); expect(filtGrad.shape).toEqual([2, 2, 2, 2]); }); it('throws when x is not rank 3', () => { const origInputDepth = 1; const origOutputDepth = 1; const fSize = 2; const origPad = 0; const origStride = 1; // tslint:disable-next-line:no-any const x = tf.tensor2d([2, 2], [2, 1]); const w = tf.tensor4d([3, 1, 5, 0], [fSize, fSize, origInputDepth, origOutputDepth]); expect(() => tf.conv2dTranspose(x, w, [2, 2, 1], origStride, origPad)) .toThrowError(); }); it('throws when weights is not rank 4', () => { const origInputDepth = 1; const origOutputDepth = 1; const inputShape = [1, 1, origOutputDepth]; const fSize = 2; const origPad = 0; const origStride = 1; const x = tf.tensor3d([2], inputShape); // tslint:disable-next-line:no-any const w = tf.tensor3d([3, 1, 5, 0], [fSize, fSize, origInputDepth]); expect(() => tf.conv2dTranspose(x, w, [2, 2, 1], origStride, origPad)) .toThrowError(); }); it('throws when x depth does not match weights original output depth', () => { const origInputDepth = 1; const origOutputDepth = 2; const wrongOrigOutputDepth = 3; const inputShape = [1, 1, origOutputDepth]; const fSize = 2; const origPad = 0; const origStride = 1; const x = tf.tensor3d([2, 2], inputShape); const w = tf.randomNormal([fSize, fSize, origInputDepth, wrongOrigOutputDepth]); expect(() => tf.conv2dTranspose(x, w, [2, 2, 2], origStride, origPad)) .toThrowError(); }); it('throws when passed x as a non-tensor', () => { const origInputDepth = 1; const origOutputDepth = 1; const fSize = 2; const origPad = 0; const origStride = 1; const w = tf.tensor4d([3, 1, 5, 0], [fSize, fSize, origInputDepth, origOutputDepth]); expect(() => tf.conv2dTranspose({}, w, [2, 2, 1], origStride, origPad)) .toThrowError(/Argument 'x' passed to 'conv2dTranspose' must be a Tensor/); }); it('throws when passed filter as a non-tensor', () => { const origOutputDepth = 1; const inputShape = [1, 1, origOutputDepth]; const origPad = 0; const origStride = 1; const x = tf.tensor3d([2], inputShape); expect(() => tf.conv2dTranspose(x, {}, [2, 2, 1], origStride, origPad)) .toThrowError(/Argument 'filter' passed to 'conv2dTranspose' must be a Tensor/); }); it('accepts a tensor-like object', async () => { const origPad = 0; const origStride = 1; const x = [[[2]]]; // 1x1x1 const w = [[[[3]], [[1]]], [[[5]], [[0]]]]; // 2x2x1x1 const result = tf.conv2dTranspose(x, w, [2, 2, 1], origStride, origPad); const expected = [6, 2, 10, 0]; expect(result.shape).toEqual([2, 2, 1]); expectArraysClose(await result.data(), expected); }); }); //# sourceMappingURL=conv2d_transpose_test.js.map