org.deeplearning4j.nn.graph.vertex.impl.PoolHelperVertex Maven / Gradle / Ivy

/*
 *  ******************************************************************************
 *  *
 *  *
 *  * 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.
 *  *
 *  *  See the NOTICE file distributed with this work for additional
 *  *  information regarding copyright ownership.
 *  * 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.deeplearning4j.nn.graph.vertex.impl;

import org.deeplearning4j.nn.api.Layer;
import org.deeplearning4j.nn.api.MaskState;
import org.deeplearning4j.nn.gradient.Gradient;
import org.deeplearning4j.nn.graph.ComputationGraph;
import org.deeplearning4j.nn.graph.vertex.BaseGraphVertex;
import org.deeplearning4j.nn.graph.vertex.VertexIndices;
import org.nd4j.linalg.api.buffer.DataType;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.api.ops.impl.transforms.pairwise.bool.Or;
import org.nd4j.linalg.factory.Nd4j;
import org.nd4j.linalg.indexing.NDArrayIndex;
import org.nd4j.common.primitives.Pair;
import org.deeplearning4j.nn.workspace.ArrayType;
import org.deeplearning4j.nn.workspace.LayerWorkspaceMgr;

public class PoolHelperVertex extends BaseGraphVertex {

    public PoolHelperVertex(ComputationGraph graph, String name, int vertexIndex, DataType dataType) {
        this(graph, name, vertexIndex, null, null, dataType);
    }

    public PoolHelperVertex(ComputationGraph graph, String name, int vertexIndex, VertexIndices[] inputVertices,
                    VertexIndices[] outputVertices, DataType dataType) {
        super(graph, name, vertexIndex, inputVertices, outputVertices, dataType);
    }

    @Override
    public boolean hasLayer() {
        return false;
    }

    @Override
    public Layer getLayer() {
        return null;
    }

    @Override
    public INDArray doForward(boolean training, LayerWorkspaceMgr workspaceMgr) {
        if (!canDoForward())
            throw new IllegalStateException("Cannot do forward pass: inputs not set");

        if (inputs.length > 1)
            throw new IllegalStateException("PoolHelper vertex requires a single input.");

        INDArray strippedInput = inputs[0].get(NDArrayIndex.all(), NDArrayIndex.all(),
                        NDArrayIndex.interval(1, inputs[0].size(2)), NDArrayIndex.interval(1, inputs[0].size(3)));
        return workspaceMgr.dup(ArrayType.ACTIVATIONS, strippedInput);
    }

    @Override
    public Pair doBackward(boolean tbptt, LayerWorkspaceMgr workspaceMgr) {
        if (!canDoBackward())
            throw new IllegalStateException("Cannot do backward pass: errors not set");

        INDArray out = workspaceMgr.create(ArrayType.ACTIVATION_GRAD, epsilon.dataType(), epsilon.size(0), epsilon.size(1), 1+epsilon.size(2), 1+epsilon.size(2));
        out.get(NDArrayIndex.all(), NDArrayIndex.all(),NDArrayIndex.interval(1, inputs[0].size(2)), NDArrayIndex.interval(1, inputs[0].size(3)))
                .assign(epsilon);

        return new Pair<>(null, new INDArray[] {out});
    }

    @Override
    public void setBackpropGradientsViewArray(INDArray backpropGradientsViewArray) {
        if (backpropGradientsViewArray != null)
            throw new RuntimeException("Vertex does not have gradients; gradients view array cannot be set here");
    }

    @Override
    public Pair feedForwardMaskArrays(INDArray[] maskArrays, MaskState currentMaskState,
                    int minibatchSize) {
        if (maskArrays == null) {
            return new Pair<>(null, currentMaskState);
        }

        //Most common case: all or none.
        //If there's only *some* mask arrays: assume the others (missing) are equivalent to all 1s
        //And for handling multiple masks: best strategy seems to be an OR operation
        //i.e., output is 1 if any of the input are 1s
        //Which means: if any masks are missing, output null (equivalent to no mask, or all steps present)
        //Otherwise do an element-wise OR operation

        for (INDArray arr : maskArrays) {
            if (arr == null) {
                return new Pair<>(null, currentMaskState);
            }
        }

        //At this point: all present. Do OR operation
        if (maskArrays.length == 1) {
            return new Pair<>(maskArrays[0], currentMaskState);
        } else {
            INDArray ret = maskArrays[0].dup(maskArrays[0].ordering());
            Nd4j.getExecutioner().exec(new Or(maskArrays[0], maskArrays[1], ret));
            for (int i = 2; i < maskArrays.length; i++) {
                Nd4j.getExecutioner().exec(new Or(maskArrays[i], ret, ret));
            }
            return new Pair<>(ret, currentMaskState);
        }
    }

    @Override
    public String toString() {
        return "PoolHelperVertex(id=" + this.getVertexIndex() + ",name=\"" + this.getVertexName() + "\")";
    }
}