org.nd4j.linalg.api.ops.impl.transforms.custom.BatchToSpace Maven / Gradle / Ivy

Show more of this group Show more artifacts with this name
Show all versions of nd4j-api Show documentation
There is a newer version: 1.0.0-M2.1
/*******************************************************************************
 * Copyright (c) 2015-2018 Skymind, Inc.
 *
 * This program and the accompanying materials are made available under the
 * terms of the Apache License, Version 2.0 which is available at
 * https://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.
 *
 * SPDX-License-Identifier: Apache-2.0
 ******************************************************************************/

package org.nd4j.linalg.api.ops.impl.transforms.custom;


import lombok.val;
import org.nd4j.autodiff.samediff.SDVariable;
import org.nd4j.autodiff.samediff.SameDiff;
import org.nd4j.base.Preconditions;
import org.nd4j.linalg.api.buffer.DataType;
import org.nd4j.linalg.api.ops.DynamicCustomOp;

import java.util.Arrays;
import java.util.Collections;
import java.util.List;

/**
 * N-dimensional batch to space operation. Transforms data from a tensor from batch dimension into M spatial dimensions
 * according to the "blocks" specified (a vector of length M). Afterwards the spatial dimensions are optionally cropped,
 * as specified in "crops", a tensor of dim (M, 2), denoting the crop range.
 * 
 * Example:
 * input:        [[[[1]]], [[[2]]], [[[3]]], [[[4]]]]
 * input shape:  [4, 1, 1, 1]
 * blocks:       [2, 2]
 * crops:        [[0, 0], [0, 0]]
 * 
 * output:       [[[[1], [2]], [[3], [4]]]]
 * output shape: [1, 2, 2, 1]
 *
 * @author Max Pumperla
 */
public class BatchToSpace extends DynamicCustomOp {

    private int[] blocks;
    private int[][] crops;

    public BatchToSpace() {
    }

    public BatchToSpace(SameDiff sameDiff, SDVariable[] args, int[] blocks, int[][] crops, boolean inPlace) {
        super(null, sameDiff, args, inPlace);

        this.blocks = blocks;
        this.crops = crops;

        for (val b : blocks)
            addIArgument(b);

        for (int e = 0; e < crops.length; e++)
            addIArgument(crops[e][0], crops[e][1]);
    }

    @Override
    public String opName() {
        return "batch_to_space";
    }

    @Override
    public String onnxName() {
        return "batch_to_space";
    }

    @Override
    public String tensorflowName() {
        return "BatchToSpace";
    }

    @Override
    public List doDiff(List i_v) {
        // Inverse of batch to space is space to batch with same blocks and padding as crops
        SDVariable gradient = sameDiff.setupFunction(i_v.get(0));
        return Arrays.asList(sameDiff.cnn().spaceToBatch(gradient, blocks, crops));
    }

    @Override
    public List calculateOutputDataTypes(List dataTypes){
        return Collections.singletonList(dataTypes.get(0));
    }
}