/**
|
* @license
|
* Copyright 2018 Google Inc. 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 {ENGINE} from '../engine';
|
import {Tensor, Tensor2D, Tensor3D, Tensor4D, Tensor5D} from '../tensor';
|
import {convertToTensor} from '../tensor_util_env';
|
import {TensorLike} from '../types';
|
import * as util from '../util';
|
import * as conv_util from './conv_util';
|
import {op} from './operation';
|
|
/**
|
* Computes a 1D convolution over the input x.
|
*
|
* @param x The input tensor, of rank 3 or rank 2, of shape
|
* `[batch, width, inChannels]`. If rank 2, batch of 1 is assumed.
|
* @param filter The filter, rank 3, of shape
|
* `[filterWidth, inDepth, outDepth]`.
|
* @param stride The number of entries by which the filter is moved right at
|
* each step.
|
* @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 "NWC", "NCW". Defaults to "NWC",
|
* the data is stored in the order of [batch, in_width, in_channels]. Only
|
* "NWC" is currently supported.
|
* @param dilation The dilation rate in which we sample input values in
|
* atrous convolution. Defaults to `1`. If it is greater than 1, then
|
* stride 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.
|
*/
|
/** @doc {heading: 'Operations', subheading: 'Convolution'} */
|
function conv1d_<T extends Tensor2D|Tensor3D>(
|
x: T|TensorLike, filter: Tensor3D|TensorLike, stride: number,
|
pad: 'valid'|'same'|number, dataFormat: 'NWC'|'NCW' = 'NWC', dilation = 1,
|
dimRoundingMode?: 'floor'|'round'|'ceil'): T {
|
const $x = convertToTensor(x, 'x', 'conv1d');
|
const $filter = convertToTensor(filter, 'filter', 'conv1d');
|
|
let x3D = $x as Tensor3D;
|
let reshapedTo3D = false;
|
if ($x.rank === 2) {
|
reshapedTo3D = true;
|
x3D = $x.as3D(1, $x.shape[0], $x.shape[1]);
|
}
|
|
util.assert(
|
x3D.rank === 3,
|
() => `Error in conv1d: input must be rank 3, but got rank ${x3D.rank}.`);
|
util.assert(
|
$filter.rank === 3,
|
() => `Error in conv1d: filter must be rank 3, but got rank ` +
|
`${$filter.rank}.`);
|
if (dimRoundingMode != null) {
|
util.assert(
|
util.isInt(pad as number),
|
() => `Error in conv1d: pad must be an integer when using, ` +
|
`dimRoundingMode ${dimRoundingMode} but got pad ${pad}.`);
|
}
|
|
util.assert(
|
x3D.shape[2] === $filter.shape[1],
|
() => `Error in conv1d: depth of input (${x3D.shape[2]}) must match ` +
|
`input depth for filter ${$filter.shape[1]}.`);
|
util.assert(
|
conv_util.eitherStridesOrDilationsAreOne(stride, dilation),
|
() => 'Error in conv1D: Either stride or dilation must be 1. ' +
|
`Got stride ${stride} and dilation '${dilation}'`);
|
util.assert(
|
dataFormat === 'NWC',
|
() => `Error in conv1d: got dataFormat of ${
|
dataFormat} but only NWC is currently supported.`);
|
|
const filter4D =
|
$filter.as4D(1, $filter.shape[0], $filter.shape[1], $filter.shape[2]);
|
const input4D = x3D.as4D(x3D.shape[0], 1, x3D.shape[1], x3D.shape[2]);
|
const strides: [number, number] = [1, stride];
|
const dilations: [number, number] = [1, dilation];
|
|
const conv2dDataFormat = 'NHWC';
|
|
const res = conv2d(
|
input4D, filter4D, strides, pad, conv2dDataFormat, dilations,
|
dimRoundingMode);
|
|
if (reshapedTo3D) {
|
return res.as2D(res.shape[2], res.shape[3]) as T;
|
}
|
return res.as3D(res.shape[0], res.shape[2], res.shape[3]) as T;
|
}
|
|
/**
|
* Computes a 2D convolution over the input x.
|
*
|
* @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].
|
* @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.
|
*/
|
/** @doc {heading: 'Operations', subheading: 'Convolution'} */
|
function conv2d_<T extends Tensor3D|Tensor4D>(
|
x: T|TensorLike, filter: Tensor4D|TensorLike,
|
strides: [number, number]|number, pad: 'valid'|'same'|number,
|
dataFormat: 'NHWC'|'NCHW' = 'NHWC',
|
dilations: [number, number]|number = [1, 1],
|
dimRoundingMode?: 'floor'|'round'|'ceil'): T {
|
const $x = convertToTensor(x, 'x', 'conv2d');
|
const $filter = convertToTensor(filter, 'filter', 'conv2d');
|
|
let x4D = $x as Tensor4D;
|
let reshapedTo4D = false;
|
|
if ($x.rank === 3) {
|
reshapedTo4D = true;
|
x4D = $x.as4D(1, $x.shape[0], $x.shape[1], $x.shape[2]);
|
}
|
util.assert(
|
x4D.rank === 4,
|
() => `Error in conv2d: input must be rank 4, but got rank ${x4D.rank}.`);
|
util.assert(
|
$filter.rank === 4,
|
() => `Error in conv2d: filter must be rank 4, but got rank ` +
|
`${$filter.rank}.`);
|
if (dimRoundingMode != null) {
|
util.assert(
|
util.isInt(pad as number),
|
() => `Error in conv2d: pad must be an integer when using, ` +
|
`dimRoundingMode ${dimRoundingMode} but got pad ${pad}.`);
|
}
|
const inDepth = dataFormat === 'NHWC' ? x4D.shape[3] : x4D.shape[1];
|
util.assert(
|
inDepth === $filter.shape[2],
|
() => `Error in conv2d: depth of input (${inDepth}) must match ` +
|
`input depth for filter ${$filter.shape[2]}.`);
|
util.assert(
|
conv_util.eitherStridesOrDilationsAreOne(strides, dilations),
|
() => 'Error in conv2D: Either strides or dilations must be 1. ' +
|
`Got strides ${strides} and dilations '${dilations}'`);
|
|
const $dataFormat = conv_util.convertConv2DDataFormat(dataFormat);
|
const convInfo = conv_util.computeConv2DInfo(
|
x4D.shape, $filter.shape, strides, dilations, pad, dimRoundingMode, false,
|
$dataFormat);
|
|
const grad = (dy: Tensor4D, saved: Tensor[]) => {
|
const [$filter, x4D] = saved as [Tensor4D, Tensor4D];
|
util.assert(
|
conv_util.tupleValuesAreOne(dilations),
|
() => 'Error in gradient of conv2D: dilation rates greater than 1 ' +
|
`are not yet supported in gradients. Got dilations '${dilations}'`);
|
|
return {
|
x: () => conv2dDerInput(x4D.shape, dy, $filter, strides, pad, dataFormat),
|
filter: () =>
|
conv2dDerFilter(x4D, dy, $filter.shape, strides, pad, dataFormat)
|
};
|
};
|
|
const inputsToSave = [$filter, x4D];
|
const res = ENGINE.runKernelFunc((backend, save) => {
|
const res = backend.conv2d(x4D, $filter, convInfo);
|
save([$filter, x4D]);
|
|
return res;
|
}, {x: x4D, filter: $filter}, grad, 'Conv2D', convInfo, inputsToSave);
|
|
if (reshapedTo4D) {
|
return res.as3D(res.shape[1], res.shape[2], res.shape[3]) as T;
|
}
|
return res as T;
|
}
|
|
/**
|
* Computes the derivative of the input of a 2D convolution.
|
*
|
* @param xShape The shape of the input: [batch, height, width, inDepth].
|
* If length of 3, batch of 1 is assumed.
|
* @param dy The derivative of the output, of rank 4 or rank 3 of shape
|
* `[batch, outHeight, outWidth, outDepth]`. If rank 3, batch of 1 is
|
* assumed.
|
* @param filter The filter, rank 4, of shape
|
* `[filterHeight, filterWidth, inDepth, outDepth]`.
|
* @param strides The strides of the convolution: `[strideHeight,
|
* strideWidth]`.
|
* @param pad The type of padding algorithm used:
|
* - `same` and stride 1: output will be of same size as input,
|
* regardless of filter size.
|
* - `valid`: output will be smaller than input if filter is larger
|
* than 1x1.
|
* @param dataFormat: An optional string from: "NHWC", "NCHW". Defaults to
|
* "NHWC". Specify the data format of the input and output data. With the
|
* default format "NHWC", the data is stored in the order of: [batch,
|
* height, width, channels].
|
* @param dimRoundingMode The rounding mode used when computing output
|
* dimensions if pad is a number. If none is provided, it will not round
|
* and error if the output is of fractional size.
|
*/
|
function conv2dDerInput_<T extends Tensor3D|Tensor4D>(
|
xShape: [number, number, number, number]|[number, number, number], dy: T,
|
filter: Tensor4D, strides: [number, number]|number,
|
pad: 'valid'|'same'|number, dataFormat: 'NHWC'|'NCHW' = 'NHWC',
|
dimRoundingMode?: 'floor'|'round'|'ceil'): T {
|
util.assert(
|
xShape.length === dy.rank,
|
() => `Length of inShape ` +
|
`(${xShape.length}) and rank of dy (${dy.rank}) must match`);
|
|
let xShape4D = xShape as [number, number, number, number];
|
let dy4D = dy as Tensor4D;
|
let reshapedTo4D = false;
|
if (dy.rank === 3) {
|
reshapedTo4D = true;
|
dy4D = dy.as4D(1, dy.shape[0], dy.shape[1], dy.shape[2]);
|
xShape4D = [1, xShape[0], xShape[1], xShape[2]];
|
}
|
|
util.assert(
|
xShape4D.length === 4,
|
() =>
|
`Error in conv2dDerInput: inShape must be length 4, but got length ` +
|
`${xShape4D.length}.`);
|
util.assert(
|
dy4D.rank === 4,
|
() => `Error in conv2dDerInput: dy must be rank 4, but got ` +
|
`rank ${dy4D.rank}`);
|
util.assert(
|
filter.rank === 4,
|
() => `Error in conv2dDerInput: filter must be rank 4, but got ` +
|
`rank ${filter.rank}`);
|
const inDepth = dataFormat === 'NHWC' ? xShape4D[3] : xShape4D[1];
|
const outDepth = dataFormat === 'NHWC' ? dy4D.shape[3] : dy4D.shape[1];
|
util.assert(
|
inDepth === filter.shape[2],
|
() => `Error in conv2dDerInput: depth of input (${inDepth}) must ` +
|
`match input depth for filter ${filter.shape[2]}.`);
|
util.assert(
|
outDepth === filter.shape[3],
|
() => `Error in conv2dDerInput: depth of output (${outDepth}) must ` +
|
`match output depth for filter ${filter.shape[3]}.`);
|
if (dimRoundingMode != null) {
|
util.assert(
|
util.isInt(pad as number),
|
() => `Error in conv2dDerInput: pad must be an integer when using, ` +
|
`dimRoundingMode ${dimRoundingMode} but got pad ${pad}.`);
|
}
|
|
const dilations = 1;
|
|
const grad = (ddx: Tensor4D, saved: Tensor[]) => {
|
const [filter, dy4D] = saved;
|
return {
|
dy4D: () => conv2d(
|
ddx, filter as Tensor4D, strides, pad, dataFormat, dilations,
|
dimRoundingMode),
|
filter: () => conv2dDerFilter(
|
ddx, dy4D as Tensor4D, (filter as Tensor4D).shape, strides, pad,
|
dataFormat, dimRoundingMode)
|
};
|
};
|
|
const $dataFormat = conv_util.convertConv2DDataFormat(dataFormat);
|
const convInfo = conv_util.computeConv2DInfo(
|
xShape4D, filter.shape, strides, dilations, pad, dimRoundingMode, false,
|
$dataFormat);
|
const res = ENGINE.runKernelFunc((backend, save) => {
|
const res = backend.conv2dDerInput(dy4D, filter, convInfo);
|
save([filter, dy4D]);
|
return res;
|
}, {dy4D, filter}, grad);
|
if (reshapedTo4D) {
|
return res.as3D(res.shape[1], res.shape[2], res.shape[3]) as T;
|
}
|
return res as T;
|
}
|
|
/**
|
* Computes the derivative of the filter of a 2D convolution.
|
*
|
* @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 dy The dy image, of rank 4 or rank 3, of shape
|
* [batch, height, width, outDepth]. If rank 3, batch of 1 is assumed.
|
* @param filterShape The shape of the filter, length 4,
|
* [filterHeight, filterWidth, inDepth, outDepth].
|
* @param strides The strides of the convolution: [strideHeight,
|
* strideWidth].
|
* @param pad A string from: 'same', 'valid'. The type of padding algorithm
|
* used in the forward prop of the op.
|
* @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].
|
* @param dimRoundingMode A string from: 'ceil', 'round', 'floor'. 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.
|
*/
|
function conv2dDerFilter_<T extends Tensor3D|Tensor4D>(
|
x: T, dy: T, filterShape: [number, number, number, number],
|
strides: [number, number]|number, pad: 'valid'|'same'|number,
|
dataFormat: 'NHWC'|'NCHW' = 'NHWC',
|
dimRoundingMode?: 'floor'|'round'|'ceil'): Tensor4D {
|
let x4D = x as Tensor4D;
|
if (x.rank === 3) {
|
x4D = x.as4D(1, x.shape[0], x.shape[1], x.shape[2]);
|
}
|
let dy4D = dy as Tensor4D;
|
if (dy4D.rank === 3) {
|
dy4D = dy.as4D(1, dy.shape[0], dy.shape[1], dy.shape[2]);
|
}
|
util.assert(
|
x4D.rank === 4,
|
() => `Error in conv2dDerFilter: input must be rank 4, but got shape ` +
|
`${x4D.shape}.`);
|
util.assert(
|
dy4D.rank === 4,
|
() => `Error in conv2dDerFilter: dy must be rank 4, but got shape ` +
|
`${dy4D.shape}.`);
|
util.assert(
|
filterShape.length === 4,
|
() => `Error in conv2dDerFilter: filterShape must be length 4, but got ` +
|
`${filterShape}.`);
|
const inDepth = dataFormat === 'NHWC' ? x4D.shape[3] : x4D.shape[1];
|
const outDepth = dataFormat === 'NHWC' ? dy4D.shape[3] : dy4D.shape[1];
|
util.assert(
|
inDepth === filterShape[2],
|
() => `Error in conv2dDerFilter: depth of input ${inDepth}) must ` +
|
`match input depth in filter (${filterShape[2]}.`);
|
util.assert(
|
outDepth === filterShape[3],
|
() => `Error in conv2dDerFilter: depth of dy (${outDepth}) must ` +
|
`match output depth for filter (${filterShape[3]}).`);
|
if (dimRoundingMode != null) {
|
util.assert(
|
util.isInt(pad as number),
|
() => `Error in conv2dDerFilter: pad must be an integer when using, ` +
|
`dimRoundingMode ${dimRoundingMode} but got pad ${pad}.`);
|
}
|
|
const dilations = 1;
|
const $dataFormat = conv_util.convertConv2DDataFormat(dataFormat);
|
const convInfo = conv_util.computeConv2DInfo(
|
x4D.shape, filterShape, strides, dilations, pad, dimRoundingMode, false,
|
$dataFormat);
|
return ENGINE.runKernelFunc(
|
backend => backend.conv2dDerFilter(x4D, dy4D, convInfo), {x4D, dy4D});
|
}
|
|
/**
|
* Computes the transposed 2D convolution of an image, also known as a
|
* deconvolution.
|
*
|
* @param x The input image, of rank 4 or rank 3, of shape
|
* `[batch, height, width, inDepth]`. If rank 3, batch of 1 is assumed.
|
* @param filter The filter, rank 4, of shape
|
* `[filterHeight, filterWidth, outDepth, inDepth]`.
|
* `inDepth` must match `inDepth` in `x`.
|
* @param outputShape Output shape, of rank 4 or rank 3:
|
* `[batch, height, width, outDepth]`. If rank 3, batch of 1 is assumed.
|
* @param strides The strides of the original convolution:
|
* `[strideHeight, strideWidth]`.
|
* @param pad The type of padding algorithm used in the non-transpose version
|
* of the op.
|
* @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.
|
*/
|
/** @doc {heading: 'Operations', subheading: 'Convolution'} */
|
function conv2dTranspose_<T extends Tensor3D|Tensor4D>(
|
x: T|TensorLike, filter: Tensor4D|TensorLike,
|
outputShape: [number, number, number, number]|[number, number, number],
|
strides: [number, number]|number, pad: 'valid'|'same'|number,
|
dimRoundingMode?: 'floor'|'round'|'ceil'): T {
|
const $x = convertToTensor(x, 'x', 'conv2dTranspose');
|
const $filter = convertToTensor(filter, 'filter', 'conv2dTranspose');
|
|
return conv2dDerInput_(
|
outputShape, $x, $filter, strides, pad, 'NHWC', dimRoundingMode);
|
}
|
|
/**
|
* Depthwise 2D convolution.
|
*
|
* Given a 4D `input` array and a `filter` array of shape
|
* `[filterHeight, filterWidth, inChannels, channelMultiplier]` containing
|
* `inChannels` convolutional filters of depth 1, this op applies a
|
* different filter to each input channel (expanding from 1 channel to
|
* `channelMultiplier` channels for each), then concatenates the results
|
* together. The output has `inChannels * channelMultiplier` channels.
|
*
|
* See
|
* [https://www.tensorflow.org/api_docs/python/tf/nn/depthwise_conv2d](
|
* https://www.tensorflow.org/api_docs/python/tf/nn/depthwise_conv2d)
|
* for more details.
|
*
|
* @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 tensor, rank 4, of shape
|
* `[filterHeight, filterWidth, inChannels, channelMultiplier]`.
|
* @param strides The strides of the convolution: `[strideHeight,
|
* strideWidth]`. If strides is a single number, then `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 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 `rate` is a single
|
* number, then `dilationHeight == dilationWidth`. If it is greater than
|
* 1, then all values of `strides` must be 1.
|
* @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 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.
|
*/
|
/** @doc {heading: 'Operations', subheading: 'Convolution'} */
|
function depthwiseConv2d_<T extends Tensor3D|Tensor4D>(
|
x: T|TensorLike, filter: Tensor4D|TensorLike,
|
strides: [number, number]|number, pad: 'valid'|'same'|number,
|
dataFormat: 'NHWC'|'NCHW' = 'NHWC',
|
dilations: [number, number]|number = [1, 1],
|
dimRoundingMode?: 'floor'|'round'|'ceil'): T {
|
const $x = convertToTensor(x, 'x', 'depthwiseConv2d');
|
const $filter = convertToTensor(filter, 'filter', 'depthwiseConv2d');
|
|
let x4D = $x as Tensor4D;
|
let reshapedTo4D = false;
|
if ($x.rank === 3) {
|
reshapedTo4D = true;
|
x4D = $x.as4D(1, $x.shape[0], $x.shape[1], $x.shape[2]);
|
}
|
util.assert(
|
x4D.rank === 4,
|
() => `Error in depthwiseConv2d: input must be rank 4, but got ` +
|
`rank ${x4D.rank}.`);
|
util.assert(
|
$filter.rank === 4,
|
() => `Error in depthwiseConv2d: filter must be rank 4, but got rank ` +
|
`${$filter.rank}.`);
|
util.assert(
|
x4D.shape[3] === $filter.shape[2],
|
() => `Error in depthwiseConv2d: number of input channels ` +
|
`(${x4D.shape[3]}) must match the inChannels dimension in ` +
|
`filter ${$filter.shape[2]}.`);
|
if (dilations == null) {
|
dilations = [1, 1];
|
}
|
util.assert(
|
conv_util.eitherStridesOrDilationsAreOne(strides, dilations),
|
() =>
|
'Error in depthwiseConv2d: Either strides or dilations must be 1. ' +
|
`Got strides ${strides} and dilations '${dilations}'`);
|
|
if (dimRoundingMode != null) {
|
util.assert(
|
util.isInt(pad as number),
|
() => `Error in depthwiseConv2d: pad must be an integer when using, ` +
|
`dimRoundingMode ${dimRoundingMode} but got pad ${pad}.`);
|
}
|
|
const convInfo = conv_util.computeConv2DInfo(
|
x4D.shape, $filter.shape, strides, dilations, pad, dimRoundingMode,
|
true /* depthwise */);
|
|
const grad = (dy: Tensor4D, saved: Tensor[]) => {
|
util.assert(
|
conv_util.tupleValuesAreOne(dilations),
|
() => 'Error in gradient of depthwiseConv2d: dilation rates ' +
|
`greater than 1 are not yet supported. Got dilations ` +
|
`'${dilations}'`);
|
const [x4D, $filter] = saved;
|
return {
|
x: () => depthwiseConv2dDerInput(
|
(x4D as Tensor4D).shape, dy, $filter as Tensor4D, convInfo),
|
filter: () => depthwiseConv2dDerFilter(
|
x4D as Tensor4D, dy, ($filter as Tensor4D).shape, convInfo),
|
};
|
};
|
|
const inputsToSave = [x4D, $filter];
|
const res = ENGINE.runKernelFunc(
|
(backend, save) => {
|
const res = backend.depthwiseConv2D(x4D, $filter, convInfo);
|
save([x4D, $filter]);
|
return res;
|
},
|
{x: x4D, filter: $filter}, grad, 'DepthwiseConv2dNative', convInfo,
|
inputsToSave);
|
if (reshapedTo4D) {
|
return res.as3D(res.shape[1], res.shape[2], res.shape[3]) as T;
|
}
|
return res as T;
|
}
|
|
/**
|
* 2-D convolution with separable filters.
|
*
|
* Performs a depthwise convolution that acts separately on channels followed
|
* by a pointwise convolution that mixes channels. Note that this is
|
* separability between dimensions [1, 2] and 3, not spatial separability
|
* between dimensions 1 and 2.
|
*
|
* See
|
* [https://www.tensorflow.org/api_docs/python/tf/nn/separable_conv2d](
|
* https://www.tensorflow.org/api_docs/python/tf/nn/separable_conv2d)
|
* for more details.
|
*
|
* @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 depthwiseFilter The depthwise filter tensor, rank 4, of shape
|
* `[filterHeight, filterWidth, inChannels, channelMultiplier]`. This is
|
* the filter used in the first step.
|
* @param pointwiseFilter The pointwise filter tensor, rank 4, of shape
|
* `[1, 1, inChannels * channelMultiplier, outChannels]`. This is
|
* the filter used in the second step.
|
* @param strides The strides of the convolution: `[strideHeight,
|
* strideWidth]`. If strides is a single number, then `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 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 `rate` is a single
|
* number, then `dilationHeight == dilationWidth`. If it is greater than
|
* 1, then all values of `strides` must be 1.
|
* @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.
|
*/
|
/** @doc {heading: 'Operations', subheading: 'Convolution'} */
|
function separableConv2d_<T extends Tensor3D|Tensor4D>(
|
x: T|TensorLike, depthwiseFilter: Tensor4D|TensorLike,
|
pointwiseFilter: Tensor4D|TensorLike, strides: [number, number]|number,
|
pad: 'valid'|'same', dilation: [number, number]|number = [1, 1],
|
dataFormat: 'NHWC'|'NCHW' = 'NHWC'): T {
|
const $x = convertToTensor(x, 'x', 'separableConv2d');
|
const $depthwiseFilter =
|
convertToTensor(depthwiseFilter, 'depthwiseFilter', 'separableConv2d');
|
const $pointwiseFilter =
|
convertToTensor(pointwiseFilter, 'pointwiseFilter', 'separableConv2d');
|
|
let x4D = $x as Tensor4D;
|
let reshapedTo4D = false;
|
if ($x.rank === 3) {
|
reshapedTo4D = true;
|
x4D = $x.as4D(1, $x.shape[0], $x.shape[1], $x.shape[2]);
|
}
|
|
if (dataFormat === 'NCHW') {
|
throw new Error(
|
'separableConv2d currently does not support dataFormat NCHW; only ' +
|
'NHWC is supported');
|
}
|
|
util.assert(
|
x4D.rank === 4,
|
() => `Error in separableConv2d: input must be rank 4, but got ` +
|
`rank ${x4D.rank}.`);
|
util.assert(
|
$depthwiseFilter.rank === 4,
|
() => `Error in separableConv2d: depthwise filter must be rank 4, but ` +
|
`got rank ${$depthwiseFilter.rank}.`);
|
util.assert(
|
$pointwiseFilter.rank === 4,
|
() => `Error in separableConv2d: pointwise filter must be rank 4, but ` +
|
`got rank ${$depthwiseFilter.rank}.`);
|
util.assert(
|
$pointwiseFilter.shape[0] === 1,
|
() =>
|
`Error in separableConv2d: the first dimension of pointwise filter ` +
|
` must be 1, but got ${$pointwiseFilter.shape[0]}.`);
|
util.assert(
|
$pointwiseFilter.shape[1] === 1,
|
() => `Error in separableConv2d: the second dimension of pointwise ` +
|
`filter must be 1, but got ${$pointwiseFilter.shape[1]}.`);
|
|
const inChannels = $depthwiseFilter.shape[2];
|
const channelMultiplier = $depthwiseFilter.shape[3];
|
util.assert(
|
$pointwiseFilter.shape[2] === inChannels * channelMultiplier,
|
() =>
|
`Error in separableConv2d: the third dimension of pointwise filter ` +
|
`must be ${inChannels * channelMultiplier}, ` +
|
`but got ${$pointwiseFilter.shape[2]}.`);
|
|
const depthwise = depthwiseConv2d(
|
x4D, $depthwiseFilter, strides, pad, dataFormat, dilation);
|
const pointwiseStride = 1;
|
const res =
|
conv2d(depthwise, $pointwiseFilter, pointwiseStride, 'valid', dataFormat);
|
if (reshapedTo4D) {
|
return res.as3D(res.shape[1], res.shape[2], res.shape[3]) as T;
|
}
|
return res as T;
|
}
|
|
function parseTupleParam(
|
param: number|[number, number]|[number, number, number]):
|
[number, number, number] {
|
if (typeof param === 'number') {
|
return [param, param, param];
|
}
|
if (param.length === 2) {
|
return [param[0], param[1], 1];
|
}
|
return param;
|
}
|
|
function tupleValuesAreOne(
|
param: number|[number, number]|[number, number, number]): boolean {
|
const [dimA, dimB, dimC] = parseTupleParam(param);
|
return dimA === 1 && dimB === 1 && dimC === 1;
|
}
|
|
function eitherStridesOrDilationsAreOne(
|
strides: number|[number, number]|[number, number, number],
|
dilations: number|[number, number]|[number, number, number]): boolean {
|
return tupleValuesAreOne(strides) || tupleValuesAreOne(dilations);
|
}
|
|
function depthwiseConv2dDerInput_<T extends Tensor3D|Tensor4D>(
|
xShape: [number, number, number, number]|[number, number, number], dy: T,
|
filter: Tensor4D, convInfo: conv_util.Conv2DInfo): T {
|
let dy4D = dy as Tensor4D;
|
let reshapedTo4D = false;
|
if (dy.rank === 3) {
|
reshapedTo4D = true;
|
dy4D = dy.as4D(1, dy.shape[0], dy.shape[1], dy.shape[2]);
|
}
|
const res = ENGINE.runKernelFunc(
|
backend => backend.depthwiseConv2DDerInput(dy4D, filter, convInfo),
|
{dy4D});
|
if (reshapedTo4D) {
|
return res.as3D(res.shape[1], res.shape[2], res.shape[3]) as T;
|
}
|
return res as T;
|
}
|
|
function depthwiseConv2dDerFilter_<T extends Tensor3D|Tensor4D>(
|
x: T, dy: T, filterShape: [number, number, number, number],
|
convInfo: conv_util.Conv2DInfo): Tensor4D {
|
let x4D = x as Tensor4D;
|
if (x.rank === 3) {
|
x4D = x.as4D(1, x.shape[0], x.shape[1], x.shape[2]);
|
}
|
let dy4D = dy as Tensor4D;
|
if (dy4D.rank === 3) {
|
dy4D = dy.as4D(1, dy.shape[0], dy.shape[1], dy.shape[2]);
|
}
|
return ENGINE.runKernelFunc(
|
backend => backend.depthwiseConv2DDerFilter(x4D, dy4D, convInfo),
|
{x4D, dy4D});
|
}
|
|
/**
|
* Computes a 3D convolution over the input x.
|
*
|
* @param x The input tensor, of rank 5 or rank 4, of shape
|
* `[batch, depth, height, width, channels]`. If rank 4,
|
* batch of 1 is assumed.
|
* @param filter The filter, rank 5, of shape
|
* `[filterDepth, filterHeight, filterWidth, inChannels, outChannels]`.
|
* inChannels must match between input and filter.
|
* @param strides The strides of the convolution: `[strideDepth, 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: "NDHWC", "NCDHW". Defaults to
|
* "NDHWC". Specify the data format of the input and output data. With the
|
* default format "NDHWC", the data is stored in the order of: [batch,
|
* depth, height, width, channels]. Only "NDHWC" is currently supported.
|
* @param dilations The dilation rates: `[dilationDepth, dilationHeight,
|
* dilationWidth]` in which we sample input values across the height
|
* and width dimensions in atrous convolution. Defaults to `[1, 1, 1]`.
|
* If `dilations` is a single number, then
|
* `dilationDepth == dilationHeight == dilationWidth`. If it is greater
|
* than 1, then all values of `strides` must be 1.
|
*/
|
|
/** @doc {heading: 'Operations', subheading: 'Convolution'} */
|
function conv3d_<T extends Tensor4D|Tensor5D>(
|
x: T|TensorLike, filter: Tensor5D|TensorLike,
|
strides: [number, number, number]|number, pad: 'valid'|'same',
|
dataFormat: 'NDHWC'|'NCDHW' = 'NDHWC',
|
dilations: [number, number, number]|number = [1, 1, 1]): T {
|
const $x = convertToTensor(x, 'x', 'conv3d');
|
const $filter = convertToTensor(filter, 'filter', 'conv3d');
|
|
let x5D = $x as Tensor5D;
|
let reshapedTo5D = false;
|
|
if ($x.rank === 4) {
|
reshapedTo5D = true;
|
x5D = $x.as5D(1, $x.shape[0], $x.shape[1], $x.shape[2], $x.shape[3]);
|
}
|
util.assert(
|
x5D.rank === 5,
|
() => `Error in conv3d: input must be rank 5, but got rank ${x5D.rank}.`);
|
util.assert(
|
$filter.rank === 5,
|
() => `Error in conv3d: filter must be rank 5, but got rank ` +
|
`${$filter.rank}.`);
|
util.assert(
|
x5D.shape[4] === $filter.shape[3],
|
() => `Error in conv3d: depth of input (${x5D.shape[4]}) must match ` +
|
`input depth for filter ${$filter.shape[3]}.`);
|
util.assert(
|
eitherStridesOrDilationsAreOne(strides, dilations),
|
() => 'Error in conv3D: Either strides or dilations must be 1. ' +
|
`Got strides ${strides} and dilations '${dilations}'`);
|
util.assert(
|
dataFormat === 'NDHWC',
|
() => `Error in conv3d: got dataFormat of ${
|
dataFormat} but only NDHWC is currently supported.`);
|
|
const convInfo = conv_util.computeConv3DInfo(
|
x5D.shape, $filter.shape, strides, dilations, pad);
|
|
const grad = (dy: Tensor5D, saved: Tensor[]) => {
|
util.assert(
|
tupleValuesAreOne(dilations),
|
() =>
|
'Error in gradient of conv3D: dilation rates greater than 1 are ' +
|
`not yet supported in gradients. Got dilations '${dilations}'`);
|
const [x5D, $filter] = saved;
|
return {
|
x: () => conv3dDerInput_(
|
(x5D as Tensor5D).shape, dy, $filter as Tensor5D, strides, pad),
|
$filter: () => conv3dDerFilter_(
|
x5D as Tensor5D, dy, ($filter as Tensor5D).shape, strides, pad)
|
};
|
};
|
|
const res = ENGINE.runKernelFunc((backend, save) => {
|
const res = backend.conv3d(x5D, $filter, convInfo);
|
save([x5D, $filter]);
|
return res;
|
}, {x: x5D, $filter}, grad);
|
if (reshapedTo5D) {
|
return res.as4D(res.shape[1], res.shape[2], res.shape[3], res.shape[4]) as
|
T;
|
}
|
return res as T;
|
}
|
|
/**
|
* Computes the derivative of the input of a 3D convolution.
|
*
|
* @param xShape The shape of the input: [batch, depth, height, width,
|
* in_channels]. If length of 4, batch of 1 is assumed.
|
* @param dy The derivative of the output, of rank 5 or rank 4 of shape
|
* `[batch, outDepth, outHeight, outWidth, in_channels]`.
|
* If rank 4, batch of 1 is assumed.
|
* @param filter The filter, rank 5, of shape
|
* `[filterDepth, filterHeight, filterWidth, inDepth, outDepth]`.
|
* @param strides The strides of the convolution: `[strideDepth, strideHeight,
|
* strideWidth]`.
|
* @param pad The type of padding algorithm used:
|
* - `same` and stride 1: output will be of same size as input,
|
* regardless of filter size.
|
* - `valid`: output will be smaller than input if filter is larger
|
* than 1x1.
|
*/
|
function conv3dDerInput_<T extends Tensor4D|Tensor5D>(
|
xShape:
|
[number, number, number, number,
|
number]|[number, number, number, number],
|
dy: T, filter: Tensor5D, strides: [number, number, number]|number,
|
pad: 'valid'|'same'): T {
|
util.assert(
|
xShape.length === dy.rank,
|
() => `Length of inShape ` +
|
`(${xShape.length}) and rank of dy (${dy.rank}) must match`);
|
|
let xShape5D = xShape as [number, number, number, number, number];
|
let dy5D = dy as Tensor5D;
|
let reshapedTo5D = false;
|
if (dy.rank === 4) {
|
reshapedTo5D = true;
|
dy5D = dy.as5D(1, dy.shape[0], dy.shape[1], dy.shape[2], dy.shape[3]);
|
xShape5D = [1, xShape[0], xShape[1], xShape[2], xShape[3]];
|
}
|
|
const inDepth = xShape5D[4];
|
const outDepth = dy5D.shape[4];
|
util.assert(
|
xShape5D.length === 5,
|
() =>
|
`Error in conv3dDerInput: inShape must be length 5, but got length ` +
|
`${xShape5D.length}.`);
|
util.assert(
|
dy5D.rank === 5,
|
() => `Error in conv3dDerInput: dy must be rank 5, but got ` +
|
`rank ${dy5D.rank}`);
|
util.assert(
|
filter.rank === 5,
|
() => `Error in conv3dDerInput: filter must be rank 5, but got ` +
|
`rank ${filter.rank}`);
|
util.assert(
|
inDepth === filter.shape[3],
|
() => `Error in conv3dDerInput: depth of input (${inDepth}) must ` +
|
`match input depth for filter ${filter.shape[3]}.`);
|
util.assert(
|
outDepth === filter.shape[4],
|
() => `Error in conv3dDerInput: depth of output (${outDepth}) must ` +
|
`match output depth for filter ${filter.shape[4]}.`);
|
|
const dilations = 1;
|
|
const convInfo = conv_util.computeConv3DInfo(
|
xShape5D, filter.shape, strides, dilations, pad);
|
const res = ENGINE.runKernelFunc(
|
backend => backend.conv3dDerInput(dy5D, filter, convInfo), {dy5D});
|
if (reshapedTo5D) {
|
return res.as4D(res.shape[1], res.shape[2], res.shape[3], res.shape[4]) as
|
T;
|
}
|
return res as T;
|
}
|
|
/**
|
* Computes the derivative of the filter of a 3D convolution.
|
*
|
* @param x The input tensor, of rank 5 or rank 4 of shape
|
* [batch, depth, height, width, inChannels]. If rank 4, batch of 1 is
|
* assumed.
|
* @param dy The dy image, of rank 5 or rank 4, of shape
|
* [batch, depth, height, width, outDepth]. If rank 4, batch of 1 is
|
* assumed.
|
* @param filterShape The shape of the filter, length 5,
|
* [filterDepth, filterHeight, filterWidth, inDepth, outDepth].
|
* @param strides The strides of the convolution: [strideDepth, strideHeight,
|
* strideWidth].
|
* @param pad A string from: 'same', 'valid'. The type of padding algorithm
|
* used in the forward prop of the op.
|
*/
|
function conv3dDerFilter_<T extends Tensor4D|Tensor5D>(
|
x: T, dy: T, filterShape: [number, number, number, number, number],
|
strides: [number, number, number]|number, pad: 'valid'|'same'): Tensor5D {
|
let x5D = x as Tensor5D;
|
if (x.rank === 4) {
|
x5D = x.as5D(1, x.shape[0], x.shape[1], x.shape[2], x.shape[3]);
|
}
|
let dy5D = dy as Tensor5D;
|
if (dy5D.rank === 4) {
|
dy5D = dy.as5D(1, dy.shape[0], dy.shape[1], dy.shape[2], dy.shape[3]);
|
}
|
util.assert(
|
x5D.rank === 5,
|
() => `Error in conv3dDerFilter: input must be rank 5, but got shape ` +
|
`${x5D.shape}.`);
|
util.assert(
|
dy5D.rank === 5,
|
() => `Error in conv3dDerFilter: dy must be rank 5, but got shape ` +
|
`${dy5D.shape}.`);
|
util.assert(
|
filterShape.length === 5,
|
() => `Error in conv3dDerFilter: filterShape must be length 5, but got ` +
|
`${filterShape}.`);
|
util.assert(
|
x5D.shape[4] === filterShape[3],
|
() => `Error in conv3dDerFilter: depth of input ${x5D.shape[4]}) must ` +
|
`match input depth in filter (${filterShape[3]}.`);
|
util.assert(
|
dy5D.shape[4] === filterShape[4],
|
() => `Error in conv3dDerFilter: depth of dy (${dy5D.shape[4]}) must ` +
|
`match output depth for filter (${filterShape[4]}).`);
|
|
const dilations = 1;
|
|
const convInfo = conv_util.computeConv3DInfo(
|
x5D.shape, filterShape, strides, dilations, pad);
|
return ENGINE.runKernelFunc(
|
backend => backend.conv3dDerFilter(x5D, dy5D, convInfo), {x5D, dy5D});
|
}
|
|
/**
|
* Computes the transposed 3D convolution of a volume, also known as a
|
* deconvolution.
|
*
|
* @param x The input image, of rank 5 or rank 4, of shape
|
* `[batch, depth, height, width, inDepth]`. If rank 4, batch of 1 is assumed.
|
* @param filter The filter, rank 4, of shape
|
* `[depth, filterHeight, filterWidth, outDepth, inDepth]`.
|
* `inDepth` must match `inDepth` in `x`.
|
* @param outputShape Output shape, of rank 5 or rank 4:
|
* `[batch, depth, height, width, outDepth]`. If rank 3, batch of 1 is
|
* assumed.
|
* @param strides The strides of the original convolution:
|
* `[strideDepth, strideHeight, strideWidth]`.
|
* @param pad The type of padding algorithm used in the non-transpose version
|
* of the op.
|
*/
|
/** @doc {heading: 'Operations', subheading: 'Convolution'} */
|
function conv3dTranspose_<T extends Tensor4D|Tensor5D>(
|
x: T|TensorLike, filter: Tensor5D|TensorLike,
|
outputShape:
|
[number, number, number, number,
|
number]|[number, number, number, number],
|
strides: [number, number, number]|number, pad: 'valid'|'same'): T {
|
const $x = convertToTensor(x, 'x', 'conv3dTranspose');
|
const $filter = convertToTensor(filter, 'filter', 'conv3dTranspose');
|
|
return conv3dDerInput_(outputShape, $x, $filter, strides, pad);
|
}
|
|
export const conv1d = op({conv1d_});
|
export const conv2d = op({conv2d_});
|
export const conv3d = op({conv3d_});
|
export const conv2dDerFilter = op({conv2dDerFilter_});
|
export const conv2dDerInput = op({conv2dDerInput_});
|
export const depthwiseConv2d = op({depthwiseConv2d_});
|
export const depthwiseConv2dDerInput = op({depthwiseConv2dDerInput_});
|
export const depthwiseConv2dDerFilter = op({depthwiseConv2dDerFilter_});
|
export const separableConv2d = op({separableConv2d_});
|
export const conv2dTranspose = op({conv2dTranspose_});
|
export const conv3dTranspose = op({conv3dTranspose_});
|
