UNPKG

@tensorflow/tfjs-core

Version:

Hardware-accelerated JavaScript library for machine intelligence

270 lines 13.2 kB
"use strict"; /** * @license * Copyright 2019 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. * ============================================================================= */ Object.defineProperty(exports, "__esModule", { value: true }); var engine_1 = require("../engine"); var conv_1 = require("../ops/conv"); var conv_util = require("../ops/conv_util"); var operation_1 = require("../ops/operation"); var tensor_util_1 = require("../tensor_util"); var tensor_util_env_1 = require("../tensor_util_env"); var util = require("../util"); var broadcast_util = require("./broadcast_util"); /** * Computes the dot product of two matrices with optional activation and bias. * * ```js * const a = tf.tensor2d([-1, -2], [1, 2]); * const b = tf.tensor2d([1, 2, 3, 4], [2, 2]); * const bias = tf.tensor2d([1, 2], [1, 2]); * * tf.fused.matMul(a, b, false, false, bias, 'relu').print(); * ``` * * @param a First matrix in dot product operation. * @param b Second matrix in dot product operation. * @param transposeA If true, `a` is transposed before multiplication. * @param transposeB If true, `b` is transposed before multiplication. * @param bias Matrix to be added to the result. * @param activation Name of activation kernel (defaults to `linear`). */ /** @doc {heading: 'Operations', subheading: 'Matrices', namespace: 'fused'} */ function matMul_(a, b, transposeA, transposeB, bias, activation) { if (transposeA === void 0) { transposeA = false; } if (transposeB === void 0) { transposeB = false; } if (activation === void 0) { activation = 'linear'; } var _a; var $a = tensor_util_env_1.convertToTensor(a, 'a', 'fused matMul'); var $b = tensor_util_env_1.convertToTensor(b, 'b', 'fused matMul'); _a = tensor_util_1.makeTypesMatch($a, $b), $a = _a[0], $b = _a[1]; var innerShapeA = transposeA ? $a.shape[$a.rank - 2] : $a.shape[$a.rank - 1]; var innerShapeB = transposeB ? $b.shape[$b.rank - 1] : $b.shape[$b.rank - 2]; var outerShapeA = transposeA ? $a.shape[$a.rank - 1] : $a.shape[$a.rank - 2]; var outerShapeB = transposeB ? $b.shape[$b.rank - 2] : $b.shape[$b.rank - 1]; var outerDimsA = $a.shape.slice(0, -2); var outerDimsB = $b.shape.slice(0, -2); var batchDimA = util.sizeFromShape(outerDimsA); var batchDimB = util.sizeFromShape(outerDimsB); util.assert($a.rank >= 2 && $b.rank >= 2 && $a.rank === $b.rank, function () { return "Error in fused matMul: inputs must have the same rank of at least " + ("2, got ranks " + $a.rank + " and " + $b.rank + "."); }); util.assert(util.arraysEqual(outerDimsA, outerDimsB), function () { return "Error in fused matMul: outer dimensions (" + outerDimsA + ") and (" + (outerDimsB + ") of Tensors with shapes " + $a.shape + " and ") + ($b.shape + " must match."); }); util.assert(innerShapeA === innerShapeB, function () { return "Error in fused matMul: inner shapes (" + innerShapeA + ") and (" + (innerShapeB + ") of Tensors with shapes " + $a.shape + " and ") + ($b.shape + " and transposeA=" + transposeA) + (" and transposeB=" + transposeB + " must match."); }); var outShape = $a.shape.slice(0, -2).concat([outerShapeA, outerShapeB]); var a3D = transposeA ? $a.as3D(batchDimA, innerShapeA, outerShapeA) : $a.as3D(batchDimA, outerShapeA, innerShapeA); var b3D = transposeB ? $b.as3D(batchDimB, outerShapeB, innerShapeB) : $b.as3D(batchDimB, innerShapeB, outerShapeB); var $bias; if (bias != null) { $bias = tensor_util_env_1.convertToTensor(bias, 'bias', 'fused matMul'); $bias = tensor_util_1.makeTypesMatch($bias, $a)[0]; broadcast_util.assertAndGetBroadcastShape(outShape, $bias.shape); } var grad = function (dy, saved) { var a3D = saved[0], b3D = saved[1], y = saved[2]; var dyActivation; if (activation == null || activation === 'linear') { dyActivation = dy; } else if (activation === 'relu') { dyActivation = dy.mul(y.step()); } else { throw new Error("Gradient for activation " + activation + " has not been " + "implemented yet."); } var biasGradient = {}; if (bias != null) { biasGradient = { $bias: function () { var res = dyActivation; // Using dyActivation as reference shape because outputShape does not // account for the fact that we temporarily reshape inputs to 3D as // part of batched matMul. var reduceAxes = broadcast_util.getReductionAxes($bias.shape, dyActivation.shape); if (reduceAxes.length > 0) { res = res.sum(reduceAxes); } return res.reshape($bias.shape); } }; } if (!transposeA && !transposeB) { return Object.assign({ $a: function () { return dyActivation.matMul(b3D, false, true); }, $b: function () { return a3D.matMul(dyActivation, true, false); } }, biasGradient); } else if (!transposeA && transposeB) { return Object.assign({ $a: function () { return dyActivation.matMul(b3D, false, false); }, $b: function () { return dyActivation.matMul(a3D, true, false); } }, biasGradient); } else if (transposeA && !transposeB) { return Object.assign({ $a: function () { return b3D.matMul(dyActivation, false, true); }, $b: function () { return a3D.matMul(dyActivation, false, false); } }, biasGradient); } else { return Object.assign({ $a: function () { return b3D.matMul(dyActivation, true, true); }, $b: function () { return dyActivation.matMul(a3D, true, true); } }, biasGradient); } }; var inputs = { $a: a3D, $b: b3D }; if (bias != null) { inputs.$bias = $bias; } var res = engine_1.ENGINE.runKernel(function (backend, save) { var y = backend.fusedBatchMatMul(a3D, b3D, transposeA, transposeB, $bias, activation); save([a3D, b3D, y]); return y; }, inputs, grad); return res.reshape(outShape); } /** * Computes a 2D convolution over the input x, optionally fused with adding a * bias and applying an activation. * * @param x The input tensor, of rank 4 or rank 3, of shape * `[batch, height, width, inChannels]`. If rank 3, batch of 1 is * assumed. * @param filter The filter, rank 4, of shape * `[filterHeight, filterWidth, inDepth, outDepth]`. * @param strides The strides of the convolution: `[strideHeight, * strideWidth]`. * @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 dataFormat: An optional string from: "NHWC", "NCHW". Defaults to * "NHWC". Specify the data format of the input and output data. With the * default format "NHWC", the data is stored in the order of: [batch, * height, width, channels]. Only "NHWC" is currently supported. * @param dilations The dilation rates: `[dilationHeight, dilationWidth]` * in which we sample input values across the height and width dimensions * in atrous convolution. Defaults to `[1, 1]`. If `dilations` is a single * number, then `dilationHeight == dilationWidth`. If it is greater than * 1, then all values of `strides` must be 1. * @param dimRoundingMode The rounding mode used when computing output * dimensions if pad is a number. If none is provided, it will not round * and error if the output is of fractional size. * @param bias Tensor to be added to the result. * @param activation Name of activation kernel (defaults to `linear`). */ /** @doc {heading: 'Operations', subheading: 'Convolution'} */ function conv2d_(x, filter, strides, pad, dataFormat, dilations, dimRoundingMode, bias, activation) { if (dataFormat === void 0) { dataFormat = 'NHWC'; } if (dilations === void 0) { dilations = [1, 1]; } if (activation === void 0) { activation = 'linear'; } var $x = tensor_util_env_1.convertToTensor(x, 'x', 'conv2d'); var $filter = tensor_util_env_1.convertToTensor(filter, 'filter', 'conv2d'); var x4D = $x; var reshapedTo4D = false; if ($x.rank === 3) { reshapedTo4D = true; x4D = $x.as4D(1, $x.shape[0], $x.shape[1], $x.shape[2]); } util.assert(x4D.rank === 4, function () { return "Error in fused conv2d: input must be rank 4, but got rank " + (x4D.rank + "."); }); util.assert($filter.rank === 4, function () { return "Error in fused conv2d: filter must be rank 4, but got rank " + ($filter.rank + "."); }); if (dimRoundingMode != null) { util.assert(util.isInt(pad), function () { return "Error in fused conv2d: pad must be an integer when using, " + ("dimRoundingMode " + dimRoundingMode + " but got pad " + pad + "."); }); } util.assert(x4D.shape[3] === $filter.shape[2], function () { return "Error in conv2d: depth of input (" + x4D.shape[3] + ") must match " + ("input depth for filter " + $filter.shape[2] + "."); }); util.assert(conv_util.eitherStridesOrDilationsAreOne(strides, dilations), function () { return 'Error in conv2D: Either strides or dilations must be 1. ' + ("Got strides " + strides + " and dilations '" + dilations + "'"); }); util.assert(dataFormat === 'NHWC', function () { return "Error in conv2d: got dataFormat of " + dataFormat + " but only NHWC is currently supported."; }); var convInfo = conv_util.computeConv2DInfo(x4D.shape, $filter.shape, strides, dilations, pad, dimRoundingMode); var $bias; if (bias != null) { $bias = tensor_util_env_1.convertToTensor(bias, 'bias', 'fused conv2d'); $bias = tensor_util_1.makeTypesMatch($bias, $x)[0]; broadcast_util.assertAndGetBroadcastShape(convInfo.outShape, $bias.shape); } var grad = function (dy, saved) { var _a = saved, $filter = _a[0], x4D = _a[1], y = _a[2]; var dyActivation; if (activation == null || activation === 'linear') { dyActivation = dy; } else if (activation === 'relu') { dyActivation = dy.mul(y.step()); } else { throw new Error("Gradient for activation " + activation + " has not been " + "implemented yet."); } util.assert(conv_util.tupleValuesAreOne(dilations), function () { return 'Error in gradient of fused conv2D: ' + "dilation rates greater than 1 " + ("are not yet supported in gradients. Got dilations '" + dilations + "'"); }); var biasGradient = {}; if (bias != null) { biasGradient = { $bias: function () { var res = dyActivation; var reduceAxes = broadcast_util.getReductionAxes($bias.shape, dyActivation.shape); if (reduceAxes.length > 0) { res = res.sum(reduceAxes); } return res.reshape($bias.shape); } }; } return Object.assign({ x: function () { return conv_1.conv2dDerInput(x4D.shape, dyActivation, $filter, strides, pad); }, $filter: function () { return conv_1.conv2dDerFilter(x4D, dyActivation, $filter.shape, strides, pad); } }, biasGradient); }; var inputs = { x: x4D, $filter: $filter }; if (bias != null) { inputs.$bias = $bias; } var res = engine_1.ENGINE.runKernel(function (backend, save) { var res = backend.fusedConv2d(x4D, $filter, convInfo, $bias, activation); save([$filter, x4D, res]); return res; }, inputs, grad); if (reshapedTo4D) { return res.as3D(res.shape[1], res.shape[2], res.shape[3]); } return res; } exports.matMul = operation_1.op({ matMul_: matMul_ }); exports.conv2d = operation_1.op({ conv2d_: conv2d_ }); //# sourceMappingURL=fused_ops.js.map