home | career | drupal | java | mac | mysql | perl | scala | uml | unix  

Java example source code file (GravesLSTMParamInitializer.java)

This example Java source code file (GravesLSTMParamInitializer.java) is included in the alvinalexander.com "Java Source Code Warehouse" project. The intent of this project is to help you "Learn Java by Example" TM.

Learn more about this Java project at its project page.

Java - Java tags/keywords

bias_key, distribution, expected, illegalstateexception, indarray, indarrayindex, input_weight_key, map, neuralnetconfiguration, order, override, paraminitializer, recurrent_weight_key, string, util

The GravesLSTMParamInitializer.java Java example source code

 *  * Copyright 2015 Skymind,Inc.
 *  *
 *  *    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.

package org.deeplearning4j.nn.params;

import org.canova.api.conf.Configuration;
import org.deeplearning4j.nn.api.ParamInitializer;
import org.deeplearning4j.nn.conf.NeuralNetConfiguration;
import org.deeplearning4j.nn.conf.distribution.Distributions;
import org.deeplearning4j.nn.weights.WeightInitUtil;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.api.rng.distribution.Distribution;
import org.nd4j.linalg.factory.Nd4j;
import org.nd4j.linalg.indexing.INDArrayIndex;
import org.nd4j.linalg.indexing.NDArrayIndex;

import java.util.LinkedHashMap;
import java.util.Map;

/**LSTM Parameter initializer, for LSTM based on
 * Graves: Supervised Sequence Labelling with Recurrent Neural Networks
 * http://www.cs.toronto.edu/~graves/phd.pdf
public class GravesLSTMParamInitializer implements ParamInitializer {
	/** Weights for previous time step -> current time step connections */
    public final static String RECURRENT_WEIGHT_KEY = "RW";
    public final static String BIAS_KEY = DefaultParamInitializer.BIAS_KEY;
    public final static String INPUT_WEIGHT_KEY = DefaultParamInitializer.WEIGHT_KEY;

    public int numParams(NeuralNetConfiguration conf, boolean backprop) {
        org.deeplearning4j.nn.conf.layers.GravesLSTM layerConf =
                (org.deeplearning4j.nn.conf.layers.GravesLSTM) conf.getLayer();

        int nL = layerConf.getNOut();	//i.e., n neurons in this layer
        int nLast = layerConf.getNIn();	//i.e., n neurons in previous layer

        int nParams = nLast * (4*nL)   //"input" weights
                    + nL * (4 * nL + 3) //recurrent weights
                    + 4*nL;             //bias

        return nParams;

    public void init(Map<String, INDArray> params, NeuralNetConfiguration conf, INDArray paramsView, boolean initializeParams) {
        org.deeplearning4j.nn.conf.layers.GravesLSTM layerConf =
                (org.deeplearning4j.nn.conf.layers.GravesLSTM) conf.getLayer();
        double forgetGateInit = layerConf.getForgetGateBiasInit();

        Distribution dist = Distributions.createDistribution(layerConf.getDist());

        int nL = layerConf.getNOut();	//i.e., n neurons in this layer
        int nLast = layerConf.getNIn();	//i.e., n neurons in previous layer

        int length = numParams(conf,true);
        if(paramsView.length() != length) throw new IllegalStateException("Expected params view of length " + length + ", got length " + paramsView.length());

        int nParamsIn = nLast * (4*nL);
        int nParamsRecurrent = nL * (4*nL+3);
        int nBias = 4*nL;
        INDArray inputWeightView = paramsView.get(NDArrayIndex.point(0), NDArrayIndex.interval(0, nParamsIn));
        INDArray recurrentWeightView = paramsView.get(NDArrayIndex.point(0), NDArrayIndex.interval(nParamsIn, nParamsIn + nParamsRecurrent));
        INDArray biasView = paramsView.get(NDArrayIndex.point(0), NDArrayIndex.interval(nParamsIn+nParamsRecurrent, nParamsIn+nParamsRecurrent+nBias));

        if(initializeParams) {
            params.put(INPUT_WEIGHT_KEY, WeightInitUtil.initWeights(nLast, 4 * nL, layerConf.getWeightInit(), dist, inputWeightView));
            params.put(RECURRENT_WEIGHT_KEY, WeightInitUtil.initWeights(nL, 4 * nL + 3, layerConf.getWeightInit(), dist, recurrentWeightView));
            biasView.put(new INDArrayIndex[]{NDArrayIndex.point(0), NDArrayIndex.interval(nL, 2 * nL)}, Nd4j.ones(1, nL).muli(forgetGateInit));   //Order: input, forget, output, input modulation, i.e., IFOG}
            /*The above line initializes the forget gate biases to specified value.
             * See Sutskever PhD thesis, pg19:
             * "it is important for [the forget gate activations] to be approximately 1 at the early stages of learning,
             *  which is accomplished by initializing [the forget gate biases] to a large value (such as 5). If it is
             *  not done, it will be harder to learn long range dependencies because the smaller values of the forget
             *  gates will create a vanishing gradients problem."
             *  http://www.cs.utoronto.ca/~ilya/pubs/ilya_sutskever_phd_thesis.pdf
            params.put(BIAS_KEY, biasView);
        } else {
            params.put(INPUT_WEIGHT_KEY, WeightInitUtil.reshapeWeights(new int[]{nLast, 4 * nL}, inputWeightView));
            params.put(RECURRENT_WEIGHT_KEY, WeightInitUtil.reshapeWeights(new int[]{nL, 4 * nL + 3}, recurrentWeightView));
            params.put(BIAS_KEY, biasView);


    public Map<String, INDArray> getGradientsFromFlattened(NeuralNetConfiguration conf, INDArray gradientView) {
        org.deeplearning4j.nn.conf.layers.GravesLSTM layerConf = (org.deeplearning4j.nn.conf.layers.GravesLSTM) conf.getLayer();

        int nL = layerConf.getNOut();	//i.e., n neurons in this layer
        int nLast = layerConf.getNIn();	//i.e., n neurons in previous layer

        int length = numParams(conf,true);
        if(gradientView.length() != length) throw new IllegalStateException("Expected gradient view of length " + length + ", got length " + gradientView.length());

        int nParamsIn = nLast * (4*nL);
        int nParamsRecurrent = nL * (4*nL+3);
        int nBias = 4*nL;
        INDArray inputWeightGradView = gradientView.get(NDArrayIndex.point(0), NDArrayIndex.interval(0, nParamsIn)).reshape('f',nLast, 4*nL);
        INDArray recurrentWeightGradView = gradientView.get(NDArrayIndex.point(0), NDArrayIndex.interval(nParamsIn, nParamsIn + nParamsRecurrent)).reshape('f', nL, 4*nL + 3);
        INDArray biasGradView = gradientView.get(NDArrayIndex.point(0), NDArrayIndex.interval(nParamsIn+nParamsRecurrent, nParamsIn+nParamsRecurrent+nBias));   //already a row vector

        Map<String,INDArray> out = new LinkedHashMap<>();
        out.put(RECURRENT_WEIGHT_KEY, recurrentWeightGradView);
        out.put(BIAS_KEY, biasGradView);

        return out;

Other Java examples (source code examples)

Here is a short list of links related to this Java GravesLSTMParamInitializer.java source code file:

my book on functional programming


new blog posts


Copyright 1998-2019 Alvin Alexander, alvinalexander.com
All Rights Reserved.

A percentage of advertising revenue from
pages under the /java/jwarehouse URI on this website is
paid back to open source projects.