Java example source code file (RnnToFeedForwardPreProcessor.java)

This example Java source code file (RnnToFeedForwardPreProcessor.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

clonenotsupportedexception, data, edge, illegalargumentexception, indarray, inputpreprocessor, invalid, like, override, permute, rnntofeedforwardpreprocessor, runtimeexception

The RnnToFeedForwardPreProcessor.java Java example source code

package org.deeplearning4j.nn.conf.preprocessor;

import lombok.Data;
import org.deeplearning4j.nn.conf.InputPreProcessor;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.api.shape.Shape;

/**
 * A preprocessor to allow RNN and feed-forward network layers to be used together.<br>
 * For example, GravesLSTM -> OutputLayer or GravesLSTM -> DenseLayer<br>
 * This does two things:<br>
 * (a) Reshapes activations out of RNN layer (which is 3D with shape
 * [miniBatchSize,layerSize,timeSeriesLength]) into 2d activations (with shape
 * [miniBatchSize*timeSeriesLength,layerSize]) suitable for use in feed-forward layers.<br>
 * (b) Reshapes 2d epsilons (weights*deltas from feed forward layer, with shape
 * [miniBatchSize*timeSeriesLength,layerSize]) into 3d epsilons (with shape
 * [miniBatchSize,layerSize,timeSeriesLength]) for use in RNN layer
 *
 * @author Alex Black
 * @see FeedForwardToRnnPreProcessor for opposite case (i.e., DenseLayer -> GravesLSTM etc)
 */
@Data
public class RnnToFeedForwardPreProcessor implements InputPreProcessor {

    @Override
    public INDArray preProcess(INDArray input, int miniBatchSize) {
        //Need to reshape RNN activations (3d) activations to 2d (for input into feed forward layer)
        if (input.rank() != 3)
            throw new IllegalArgumentException("Invalid input: expect NDArray with rank 3 (i.e., activations for RNN layer)");

        if (input.ordering() != 'f') input = input.dup('f');

        int[] shape = input.shape();
        if (shape[0] == 1) return input.tensorAlongDimension(0, 1, 2).permutei(1, 0);    //Edge case: miniBatchSize==1
        if (shape[2] == 1) return input.tensorAlongDimension(0, 1, 0);    //Edge case: timeSeriesLength=1
        INDArray permuted = input.permute(0, 2, 1);    //Permute, so we get correct order after reshaping
        return permuted.reshape('f', shape[0] * shape[2], shape[1]);
    }

    @Override
    public INDArray backprop(INDArray output, int miniBatchSize) {
        if (output == null)
            return null;    //In a few cases: output may be null, and this is valid. Like time series data -> embedding layer
        //Need to reshape FeedForward layer epsilons (2d) to 3d (for use in RNN layer backprop calculations)
        if (output.rank() != 2)
            throw new IllegalArgumentException("Invalid input: expect NDArray with rank 2 (i.e., epsilons from feed forward layer)");
        if (output.ordering() == 'c') output = Shape.toOffsetZeroCopy(output, 'f');

        int[] shape = output.shape();
        INDArray reshaped = output.reshape('f', miniBatchSize, shape[0] / miniBatchSize, shape[1]);
        return reshaped.permute(0, 2, 1);
    }

    @Override
    public RnnToFeedForwardPreProcessor clone() {
        try {
            RnnToFeedForwardPreProcessor clone = (RnnToFeedForwardPreProcessor) super.clone();
            return clone;
        } catch (CloneNotSupportedException e) {
            throw new RuntimeException(e);
        }
    }
}