org.deeplearning4j.nn.conf.layers.recurrent.Bidirectional 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.conf.layers.recurrent;

import lombok.*;
import org.deeplearning4j.nn.api.ParamInitializer;
import org.deeplearning4j.nn.conf.GradientNormalization;
import org.deeplearning4j.nn.conf.InputPreProcessor;
import org.deeplearning4j.nn.conf.NeuralNetConfiguration;
import org.deeplearning4j.nn.conf.RNNFormat;
import org.deeplearning4j.nn.conf.inputs.InputType;
import org.deeplearning4j.nn.conf.layers.BaseRecurrentLayer;
import org.deeplearning4j.nn.conf.layers.FeedForwardLayer;
import org.deeplearning4j.nn.conf.layers.Layer;
import org.deeplearning4j.nn.conf.layers.wrapper.BaseWrapperLayer;
import org.deeplearning4j.nn.conf.memory.LayerMemoryReport;
import org.deeplearning4j.nn.layers.recurrent.BidirectionalLayer;
import org.deeplearning4j.nn.params.BidirectionalParamInitializer;
import org.deeplearning4j.optimize.api.TrainingListener;
import org.deeplearning4j.util.TimeSeriesUtils;
import org.nd4j.linalg.api.buffer.DataType;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.learning.config.IUpdater;
import org.nd4j.linalg.learning.regularization.Regularization;
import org.nd4j.shade.jackson.annotation.JsonIgnoreProperties;

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

import static org.nd4j.linalg.indexing.NDArrayIndex.interval;
import static org.nd4j.linalg.indexing.NDArrayIndex.point;

@NoArgsConstructor
@Data
@EqualsAndHashCode(callSuper = true, exclude = {"initializer"})
@JsonIgnoreProperties({"initializer"})
public class Bidirectional extends Layer {

    /**
     * This Mode enumeration defines how the activations for the forward and backward networks should be combined.

     * ADD: out = forward + backward (elementwise addition)
 MUL: out = forward * backward (elementwise
     * multiplication)
 AVERAGE: out = 0.5 * (forward + backward)
 CONCAT: Concatenate the activations.
 Where
     * 'forward' is the activations for the forward RNN, and 'backward' is the activations for the backward RNN. In all
     * cases except CONCAT, the output activations size is the same size as the standard RNN that is being wrapped by
     * this layer. In the CONCAT case, the output activations size (dimension 1) is 2x larger than the standard RNN's
     * activations array.
     */
    public enum Mode {
        ADD, MUL, AVERAGE, CONCAT
    }

    private Layer fwd;
    private Layer bwd;
    private Mode mode;
    private transient BidirectionalParamInitializer initializer;

    private Bidirectional(Builder builder) {
        super(builder);
    }

    /**
     * Create a Bidirectional wrapper, with the default Mode (CONCAT) for the specified layer
     *
     * @param layer layer to wrap
     */
    public Bidirectional(@NonNull Layer layer) {
        this(Mode.CONCAT, layer);
    }

    /**
     * Create a Bidirectional wrapper for the specified layer
     *
     * @param mode Mode to use to combine activations. See {@link Mode} for details
     * @param layer layer to wrap
     */
    public Bidirectional(@NonNull Mode mode, @NonNull Layer layer) {
        if (!(layer instanceof BaseRecurrentLayer || layer instanceof LastTimeStep
                        || layer instanceof BaseWrapperLayer)) {
            throw new IllegalArgumentException("Cannot wrap a non-recurrent layer: "
                            + "config must extend BaseRecurrentLayer or LastTimeStep " + "Got class: "
                            + layer.getClass());
        }
        this.fwd = layer;
        this.bwd = layer.clone();
        this.mode = mode;
    }

    public long getNOut() {
        if (this.fwd instanceof LastTimeStep) {
            return ((FeedForwardLayer) ((LastTimeStep) this.fwd).getUnderlying()).getNOut();
        } else if(this.mode == Mode.CONCAT) {
            return ((FeedForwardLayer) this.fwd).getNOut() * 2;
        }
        else {
            return ((FeedForwardLayer) this.fwd).getNOut();
        }
    }

    public long getNIn() {
        if (this.fwd instanceof LastTimeStep) {
            return ((FeedForwardLayer) ((LastTimeStep) this.fwd).getUnderlying()).getNIn();
        } else {
            return ((FeedForwardLayer) this.fwd).getNIn();
        }
    }

    public RNNFormat getRNNDataFormat(){
        return TimeSeriesUtils.getFormatFromRnnLayer(fwd);
    }

    @Override
    public org.deeplearning4j.nn.api.Layer instantiate(NeuralNetConfiguration conf,
                                                       Collection trainingListeners, int layerIndex, INDArray layerParamsView,
                                                       boolean initializeParams, DataType networkDataType) {
        NeuralNetConfiguration c1 = conf.clone();
        NeuralNetConfiguration c2 = conf.clone();
        c1.setLayer(fwd);
        c2.setLayer(bwd);

        INDArray layerParamsReshape = layerParamsView.reshape(layerParamsView.length());
        long n = layerParamsView.length() / 2;
        INDArray fp = layerParamsReshape.get(interval(0, n));
        INDArray bp = layerParamsReshape.get(interval(n, 2 * n));
        org.deeplearning4j.nn.api.Layer f = fwd.instantiate(c1, trainingListeners, layerIndex, fp, initializeParams, networkDataType);

        org.deeplearning4j.nn.api.Layer b = bwd.instantiate(c2, trainingListeners, layerIndex, bp, initializeParams, networkDataType);

        BidirectionalLayer ret = new BidirectionalLayer(conf, f, b, layerParamsView);
        Map paramTable = initializer().init(conf, layerParamsView, initializeParams);
        ret.setParamTable(paramTable);
        ret.setConf(conf);

        return ret;
    }

    @Override
    public ParamInitializer initializer() {
        if (initializer == null) {
            initializer = new BidirectionalParamInitializer(this);
        }
        return initializer;
    }

    @Override
    public InputType getOutputType(int layerIndex, InputType inputType) {
        InputType outOrig = fwd.getOutputType(layerIndex, inputType);

        if (fwd instanceof LastTimeStep) {
            InputType.InputTypeFeedForward ff = (InputType.InputTypeFeedForward) outOrig;
            if (mode == Mode.CONCAT) {
                return InputType.feedForward(2 * ff.getSize());
            } else {
                return ff;
            }
        } else {
            InputType.InputTypeRecurrent r = (InputType.InputTypeRecurrent) outOrig;
            if (mode == Mode.CONCAT) {
                return InputType.recurrent(2 * r.getSize(), getRNNDataFormat());
            } else {
                return r;
            }
        }
    }

    @Override
    public void setNIn(InputType inputType, boolean override) {
        fwd.setNIn(inputType, override);
        bwd.setNIn(inputType, override);
    }

    @Override
    public InputPreProcessor getPreProcessorForInputType(InputType inputType) {
        return fwd.getPreProcessorForInputType(inputType);
    }

    @Override
    public List getRegularizationByParam(String paramName){
        //Strip forward/backward prefix from param name
        return fwd.getRegularizationByParam(paramName.substring(1));
    }

    @Override
    public boolean isPretrainParam(String paramName) {
        return fwd.isPretrainParam(paramName.substring(1));
    }

    /**
     * Get the updater for the given parameter. Typically the same updater will be used for all updaters, but this is
     * not necessarily the case
     *
     * @param paramName Parameter name
     * @return IUpdater for the parameter
     */
    public IUpdater getUpdaterByParam(String paramName) {
        String sub = paramName.substring(1);
        return fwd.getUpdaterByParam(sub);
    }

    @Override
    public GradientNormalization getGradientNormalization() {
        return fwd.getGradientNormalization();
    }

    @Override
    public double getGradientNormalizationThreshold() {
        return fwd.getGradientNormalizationThreshold();
    }

    @Override
    public void setLayerName(String layerName) {
        this.layerName = layerName;
        fwd.setLayerName(layerName);
        bwd.setLayerName(layerName);
    }

    @Override
    public LayerMemoryReport getMemoryReport(InputType inputType) {
        LayerMemoryReport lmr = fwd.getMemoryReport(inputType);
        lmr.scale(2); //Double all memory use
        return lmr;
    }

    @AllArgsConstructor
    @Getter
    @Setter
    public static class Builder extends Layer.Builder {

        private Mode mode;
        private Layer layer;

        public void setLayer(Layer layer) {
            rnnLayer(layer);
        }

        public Builder mode(Mode mode) {
            this.setMode(mode);
            return this;
        }

        public Builder rnnLayer(Layer layer) {
            if (!(layer instanceof BaseRecurrentLayer || layer instanceof LastTimeStep
                            || layer instanceof BaseWrapperLayer)) {
                throw new IllegalArgumentException("Cannot wrap a non-recurrent layer: "
                                + "config must extend BaseRecurrentLayer or LastTimeStep " + "Got class: "
                                + layer.getClass());
            }
            this.setLayer(layer);
            return this;
        }

        @SuppressWarnings("unchecked")
        public Bidirectional build() {
            return new Bidirectional(this);
        }
    }
}