@tensorflow/tfjs-core/src/ops/space_to_batch_nd.ts

Hardware-accelerated JavaScript library for machine intelligence

/**
 * @license
 * Copyright 2020 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 {ENGINE} from '../engine';
import {SpaceToBatchND, SpaceToBatchNDAttrs, SpaceToBatchNDInputs} from '../kernel_names';
import {NamedAttrMap} from '../kernel_registry';
import {Tensor} from '../tensor';
import {NamedTensorMap} from '../tensor_types';
import {convertToTensor} from '../tensor_util_env';
import {TensorLike} from '../types';
import * as util from '../util';

import {op} from './operation';

/**
 * This operation divides "spatial" dimensions `[1, ..., M]` of the input into
 * a grid of blocks of shape `blockShape`, and interleaves these blocks with
 * the "batch" dimension (0) such that in the output, the spatial
 * dimensions `[1, ..., M]` correspond to the position within the grid,
 * and the batch dimension combines both the position within a spatial block
 * and the original batch position. Prior to division into blocks,
 * the spatial dimensions of the input are optionally zero padded
 * according to `paddings`. See below for a precise description.
 *
 * ```js
 * const x = tf.tensor4d([1, 2, 3, 4], [1, 2, 2, 1]);
 * const blockShape = [2, 2];
 * const paddings = [[0, 0], [0, 0]];
 *
 * x.spaceToBatchND(blockShape, paddings).print();
 * ```
 *
 * @param x A `tf.Tensor`. N-D with `x.shape` = `[batch] + spatialShape +
 * remainingShape`, where spatialShape has `M` dimensions.
 * @param blockShape A 1-D array. Must have shape `[M]`, all values must
 * be >= 1.
 * @param paddings A 2-D array. Must have shape `[M, 2]`, all values must be >=
 *     0. `paddings[i] = [padStart, padEnd]` specifies the amount to zero-pad
 * from input dimension `i + 1`, which corresponds to spatial dimension `i`. It
 * is required that
 * `(inputShape[i + 1] + padStart + padEnd) % blockShape[i] === 0`
 *
 * This operation is equivalent to the following steps:
 *
 * 1. Zero-pad the start and end of dimensions `[1, ..., M]` of the input
 * according to `paddings` to produce `padded` of shape paddedShape.
 *
 * 2. Reshape `padded` to `reshapedPadded` of shape:
 * `[batch] + [paddedShape[1] / blockShape[0], blockShape[0], ...,
 * paddedShape[M] / blockShape[M-1], blockShape[M-1]] + remainingShape`
 *
 * 3. Permute dimensions of `reshapedPadded` to produce `permutedReshapedPadded`
 * of shape: `blockShape + [batch] + [paddedShape[1] / blockShape[0], ...,
 * paddedShape[M] / blockShape[M-1]] + remainingShape`
 *
 * 4. Reshape `permutedReshapedPadded` to flatten `blockShape` into the
 * batch dimension, producing an output tensor of shape:
 * `[batch * prod(blockShape)] + [paddedShape[1] / blockShape[0], ...,
 * paddedShape[M] / blockShape[M-1]] + remainingShape`
 *
 * @doc {heading: 'Tensors', subheading: 'Transformations'}
 */
function spaceToBatchND_<T extends Tensor>(
    x: T|TensorLike, blockShape: number[], paddings: number[][]): T {
  const $x = convertToTensor(x, 'x', 'spaceToBatchND');

  util.assert(
      $x.rank >= 1 + blockShape.length,
      () => `input rank ${$x.rank} should be > than [blockShape] ${
          blockShape.length}`);

  util.assert(
      paddings.length === blockShape.length,
      () => `paddings.shape[0] ${
          paddings.length} must be equal to [blockShape] ${blockShape.length}`);

  util.assert(
      $x.shape.reduce(
          (a, b, i) => {
            if (i > 0 && i <= blockShape.length) {
              return a &&
                  ((b + paddings[i - 1][0] + paddings[i - 1][1]) %
                       blockShape[i - 1] ===
                   0);
            }
            return a;
          },
          true),
      () => `input spatial dimensions ${$x.shape.slice(1)} with paddings ${
          paddings.toString()} must be divisible by blockShapes ${
          blockShape.toString()}`);

  const inputs: SpaceToBatchNDInputs = {x: $x};
  const attrs: SpaceToBatchNDAttrs = {blockShape, paddings};

  return ENGINE.runKernel(
      SpaceToBatchND, inputs as {} as NamedTensorMap,
      attrs as {} as NamedAttrMap);
}

export const spaceToBatchND = op({spaceToBatchND_});