org.deeplearning4j.nn.conf.layers.AbstractLSTM Maven / Gradle / Ivy

Go to download
Show more of this group Show more artifacts with this name
Show all versions of deeplearning4j-nn Show documentation
There is a newer version: 1.0.0-M2.1
/*-
 *
 *  * 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.conf.layers;

import lombok.*;
import org.deeplearning4j.nn.api.Layer;
import org.deeplearning4j.nn.api.ParamInitializer;
import org.deeplearning4j.nn.conf.NeuralNetConfiguration;
import org.deeplearning4j.nn.params.GravesLSTMParamInitializer;
import org.deeplearning4j.nn.params.LSTMParamInitializer;
import org.deeplearning4j.optimize.api.IterationListener;
import org.deeplearning4j.util.LayerValidation;
import org.nd4j.linalg.activations.Activation;
import org.nd4j.linalg.activations.IActivation;
import org.nd4j.linalg.activations.impl.ActivationSigmoid;
import org.nd4j.linalg.api.ndarray.INDArray;

import java.util.Collection;
import java.util.Map;

/**
 * LSTM recurrent net, based on Graves: Supervised Sequence Labelling with Recurrent Neural Networks
 * http://www.cs.toronto.edu/~graves/phd.pdf
 */
@Data
@NoArgsConstructor
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
public abstract class AbstractLSTM extends BaseRecurrentLayer {

    protected double forgetGateBiasInit;
    protected IActivation gateActivationFn = new ActivationSigmoid();

    protected AbstractLSTM(Builder builder) {
        super(builder);
        this.forgetGateBiasInit = builder.forgetGateBiasInit;
        this.gateActivationFn = builder.gateActivationFn;
    }

    @Override
    public double getL1ByParam(String paramName) {
        switch (paramName) {
            case LSTMParamInitializer.INPUT_WEIGHT_KEY:
            case LSTMParamInitializer.RECURRENT_WEIGHT_KEY:
                return l1;
            case LSTMParamInitializer.BIAS_KEY:
                return l1Bias;
            default:
                throw new IllegalArgumentException("Unknown parameter name: \"" + paramName + "\"");
        }
    }

    @Override
    public double getL2ByParam(String paramName) {
        switch (paramName) {
            case LSTMParamInitializer.INPUT_WEIGHT_KEY:
            case LSTMParamInitializer.RECURRENT_WEIGHT_KEY:
                return l2;
            case LSTMParamInitializer.BIAS_KEY:
                return l2Bias;
            default:
                throw new IllegalArgumentException("Unknown parameter name: \"" + paramName + "\"");
        }
    }

    @Override
    public double getLearningRateByParam(String paramName) {
        switch (paramName) {
            case LSTMParamInitializer.INPUT_WEIGHT_KEY:
            case LSTMParamInitializer.RECURRENT_WEIGHT_KEY:
                return learningRate;
            case LSTMParamInitializer.BIAS_KEY:
                if (!Double.isNaN(biasLearningRate)) {
                    //Bias learning rate has been explicitly set
                    return biasLearningRate;
                } else {
                    return learningRate;
                }
            default:
                throw new IllegalArgumentException("Unknown parameter name: \"" + paramName + "\"");
        }
    }

    @AllArgsConstructor
    @NoArgsConstructor
    public static abstract class Builder> extends BaseRecurrentLayer.Builder {

        protected double forgetGateBiasInit = 1.0;
        protected IActivation gateActivationFn = new ActivationSigmoid();

        /** Set forget gate bias initalizations. Values in range 1-5 can potentially
         * help with learning or longer-term dependencies.
         */
        public T forgetGateBiasInit(double biasInit) {
            this.forgetGateBiasInit = biasInit;
            return (T) this;
        }

        /**
         * Activation function for the LSTM gates.
         * Note: This should be bounded to range 0-1: sigmoid or hard sigmoid, for example
         *
         * @param gateActivationFn Activation function for the LSTM gates
         */
        public T gateActivationFunction(String gateActivationFn) {
            return (T) gateActivationFunction(Activation.fromString(gateActivationFn));
        }

        /**
         * Activation function for the LSTM gates.
         * Note: This should be bounded to range 0-1: sigmoid or hard sigmoid, for example
         *
         * @param gateActivationFn Activation function for the LSTM gates
         */
        public T gateActivationFunction(Activation gateActivationFn) {
            return (T) gateActivationFunction(gateActivationFn.getActivationFunction());
        }

        /**
         * Activation function for the LSTM gates.
         * Note: This should be bounded to range 0-1: sigmoid or hard sigmoid, for example
         *
         * @param gateActivationFn Activation function for the LSTM gates
         */
        public T gateActivationFunction(IActivation gateActivationFn) {
            this.gateActivationFn = gateActivationFn;
            return (T) this;
        }

    }

}