org.deeplearning4j.nn.params.DepthwiseConvolutionParamInitializer Maven / Gradle / Ivy

Go to download
/*
 *  ******************************************************************************
 *  *
 *  *
 *  * 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.params;


import lombok.val;
import org.deeplearning4j.nn.api.ParamInitializer;
import org.deeplearning4j.nn.conf.NeuralNetConfiguration;
import org.deeplearning4j.nn.conf.layers.DepthwiseConvolution2D;
import org.deeplearning4j.nn.conf.layers.Layer;
import org.deeplearning4j.nn.weights.WeightInitUtil;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.indexing.NDArrayIndex;

import java.util.*;

public class DepthwiseConvolutionParamInitializer implements ParamInitializer {

    private static final DepthwiseConvolutionParamInitializer INSTANCE = new DepthwiseConvolutionParamInitializer();

    public static DepthwiseConvolutionParamInitializer getInstance() {
        return INSTANCE;
    }

    public final static String WEIGHT_KEY = DefaultParamInitializer.WEIGHT_KEY;
    public final static String BIAS_KEY = DefaultParamInitializer.BIAS_KEY;

    @Override
    public long numParams(NeuralNetConfiguration conf) {
        return numParams(conf.getLayer());
    }

    @Override
    public long numParams(Layer l) {
        DepthwiseConvolution2D layerConf = (DepthwiseConvolution2D) l;

        val depthWiseParams = numDepthWiseParams(layerConf);
        val biasParams = numBiasParams(layerConf);

        return depthWiseParams + biasParams;
    }

    private long numBiasParams(DepthwiseConvolution2D layerConf) {
        val nOut = layerConf.getNOut();
        return (layerConf.hasBias() ? nOut : 0);
    }

    /**
     * For each input feature we separately compute depthMultiplier many
     * output maps for the given kernel size
     *
     * @param layerConf layer configuration of the separable conv2d layer
     * @return number of parameters of the channels-wise convolution operation
     */
    private long numDepthWiseParams(DepthwiseConvolution2D layerConf) {
        int[] kernel = layerConf.getKernelSize();
        val nIn = layerConf.getNIn();
        val depthMultiplier = layerConf.getDepthMultiplier();

        return nIn * depthMultiplier * kernel[0] * kernel[1];
    }

    @Override
    public List paramKeys(Layer layer) {
        DepthwiseConvolution2D layerConf =
                (DepthwiseConvolution2D) layer;
        if(layerConf.hasBias()){
            return Arrays.asList(WEIGHT_KEY, BIAS_KEY);
        } else {
            return weightKeys(layer);
        }
    }

    @Override
    public List weightKeys(Layer layer) {
        return Arrays.asList(WEIGHT_KEY);
    }

    @Override
    public List biasKeys(Layer layer) {
        DepthwiseConvolution2D layerConf =
                (DepthwiseConvolution2D) layer;
        if(layerConf.hasBias()){
            return Collections.singletonList(BIAS_KEY);
        } else {
            return Collections.emptyList();
        }
    }

    @Override
    public boolean isWeightParam(Layer layer, String key) {
        return WEIGHT_KEY.equals(key);
    }

    @Override
    public boolean isBiasParam(Layer layer, String key) {
        return BIAS_KEY.equals(key);
    }


    @Override
    public Map init(NeuralNetConfiguration conf, INDArray paramsView, boolean initializeParams) {
        DepthwiseConvolution2D layer = (DepthwiseConvolution2D) conf.getLayer();
        if (layer.getKernelSize().length != 2) throw new IllegalArgumentException("Filter size must be == 2");

        Map params = Collections.synchronizedMap(new LinkedHashMap());
        DepthwiseConvolution2D layerConf = (DepthwiseConvolution2D) conf.getLayer();

        val depthWiseParams = numDepthWiseParams(layerConf);
        val biasParams = numBiasParams(layerConf);

        INDArray paramsViewReshape = paramsView.reshape(paramsView.length());
        INDArray depthWiseWeightView = paramsViewReshape.get(
                NDArrayIndex.interval(biasParams, biasParams + depthWiseParams));

        params.put(WEIGHT_KEY, createDepthWiseWeightMatrix(conf, depthWiseWeightView, initializeParams));
        conf.addVariable(WEIGHT_KEY);

        if(layer.hasBias()){
            INDArray biasView = paramsViewReshape.get(NDArrayIndex.interval(0, biasParams));
            params.put(BIAS_KEY, createBias(conf, biasView, initializeParams));
            conf.addVariable(BIAS_KEY);
        }

        return params;
    }

    @Override
    public Map getGradientsFromFlattened(NeuralNetConfiguration conf, INDArray gradientView) {

        DepthwiseConvolution2D layerConf = (DepthwiseConvolution2D) conf.getLayer();

        int[] kernel = layerConf.getKernelSize();
        val nIn = layerConf.getNIn();
        val depthMultiplier = layerConf.getDepthMultiplier();
        val nOut = layerConf.getNOut();

        Map out = new LinkedHashMap<>();

        val depthWiseParams = numDepthWiseParams(layerConf);
        val biasParams = numBiasParams(layerConf);
        INDArray gradientViewReshape = gradientView.reshape(gradientView.length());
        INDArray depthWiseWeightGradientView = gradientViewReshape.get(
                        NDArrayIndex.interval(biasParams, biasParams + depthWiseParams))
                .reshape('c', kernel[0], kernel[1], nIn, depthMultiplier);
        out.put(WEIGHT_KEY, depthWiseWeightGradientView);

        if(layerConf.hasBias()) {
            INDArray biasGradientView = gradientViewReshape.get(NDArrayIndex.interval(0, nOut));
            out.put(BIAS_KEY, biasGradientView);
        }
        return out;
    }

    protected INDArray createBias(NeuralNetConfiguration conf, INDArray biasView, boolean initializeParams) {
        DepthwiseConvolution2D layerConf = (DepthwiseConvolution2D) conf.getLayer();
        if (initializeParams)
            biasView.assign(layerConf.getBiasInit());
        return biasView;
    }


    protected INDArray createDepthWiseWeightMatrix(NeuralNetConfiguration conf, INDArray weightView, boolean initializeParams) {
        /*
         Create a 4d weight matrix of: (channels multiplier, num input channels, kernel height, kernel width)
         Inputs to the convolution layer are: (batch size, num input feature maps, image height, image width)
         */
        DepthwiseConvolution2D layerConf =
                (DepthwiseConvolution2D) conf.getLayer();
        int depthMultiplier = layerConf.getDepthMultiplier();

        if (initializeParams) {
            int[] kernel = layerConf.getKernelSize();
            int[] stride = layerConf.getStride();

            val inputDepth = layerConf.getNIn();

            double fanIn = inputDepth * kernel[0] * kernel[1];
            double fanOut = depthMultiplier * kernel[0] * kernel[1] / ((double) stride[0] * stride[1]);

            val weightsShape = new long[] {kernel[0], kernel[1], inputDepth, depthMultiplier};

            return layerConf.getWeightInitFn().init(fanIn, fanOut, weightsShape, 'c',
                    weightView);
        } else {
            int[] kernel = layerConf.getKernelSize();
            return WeightInitUtil.reshapeWeights(
                    new long[] {kernel[0], kernel[1], layerConf.getNIn(), depthMultiplier}, weightView, 'c');
        }
    }
}