Java example source code file (RnnToCnnPreProcessor.java)

This example Java source code file (RnnToCnnPreProcessor.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

data, indarray, inputpreprocessor, jsonproperty, override, permute, rnntocnnpreprocessor

The RnnToCnnPreProcessor.java Java example source code

package org.deeplearning4j.nn.conf.preprocessor;

import com.fasterxml.jackson.annotation.JsonProperty;
import lombok.AccessLevel;
import lombok.Data;
import lombok.Getter;
import lombok.Setter;
import org.deeplearning4j.nn.conf.InputPreProcessor;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.api.shape.Shape;
import org.nd4j.linalg.util.ArrayUtil;

/**
 * A preprocessor to allow RNN and CNN layers to be used together<br>
 * For example, time series (video) input -> ConvolutionLayer, or conceivable GravesLSTM -> ConvolutionLayer<br>
 * Functionally equivalent to combining RnnToFeedForwardPreProcessor + FeedForwardToCnnPreProcessor<br>
 * Specifically, this does two things:<br>
 * (a) Reshape 3d activations out of RNN layer, with shape [miniBatchSize, numChannels*inputHeight*inputWidth, timeSeriesLength])
 * into 4d (CNN) activations (with shape [numExamples*timeSeriesLength, numChannels, inputWidth, inputHeight]) <br>
 * (b) Reshapes 4d epsilons (weights.*deltas) out of CNN layer (with shape
 * [numExamples*timeSeriesLength, numChannels, inputHeight, inputWidth]) into 3d epsilons with shape
 * [miniBatchSize, numChannels*inputHeight*inputWidth, timeSeriesLength] suitable to feed into CNN layers.
 * Note: numChannels is equivalent to depth or featureMaps referenced in different literature
 *
 * @author Alex Black
 */
@Data
public class RnnToCnnPreProcessor implements InputPreProcessor {

    private int inputHeight;
    private int inputWidth;
    private int numChannels;

    @Getter(AccessLevel.NONE)
    @Setter(AccessLevel.NONE)
    private int product;

    public RnnToCnnPreProcessor(@JsonProperty("inputHeight") int inputHeight,
                                @JsonProperty("inputWidth") int inputWidth,
                                @JsonProperty("numChannels") int numChannels) {
        this.inputHeight = inputHeight;
        this.inputWidth = inputWidth;
        this.numChannels = numChannels;
        this.product = inputHeight * inputWidth * numChannels;
    }


    @Override
    public INDArray preProcess(INDArray input, int miniBatchSize) {
        if(input.ordering() == 'c') input = input.dup('f');
        //Input: 3d activations (RNN)
        //Output: 4d activations (CNN)
        int[] shape = input.shape();
        INDArray in2d;
        if (shape[0] == 1) {
            //Edge case: miniBatchSize = 1
            in2d = input.tensorAlongDimension(0, 1, 2).permutei(1,0);
        } else if (shape[2] == 1) {
            //Edge case: time series length = 1
            in2d = input.tensorAlongDimension(0, 1, 0);
        } else {
            INDArray permuted = input.permute(0, 2, 1);    //Permute, so we get correct order after reshaping
            in2d = permuted.reshape('f',shape[0] * shape[2], shape[1]);
        }

        return in2d.dup('c').reshape('c',shape[0] * shape[2], numChannels, inputHeight, inputWidth);
    }

    @Override
    public INDArray backprop(INDArray output, int miniBatchSize) {
        //Input: 4d epsilons (CNN)
        //Output: 3d epsilons (RNN)
        if(output.ordering() == 'f') output = output.dup('c');
        int[] shape = output.shape();
        //First: reshape 4d to 2d
        INDArray twod = output.reshape('c',output.size(0), ArrayUtil.prod(output.shape())/output.size(0));
        //Second: reshape 2d to 4d
        INDArray reshaped = twod.dup('f').reshape('f',miniBatchSize,shape[0]/miniBatchSize,product);
        return reshaped.permute(0,2,1);
    }

    @Override
    public RnnToCnnPreProcessor clone() {
        return new RnnToCnnPreProcessor(inputHeight, inputWidth, numChannels);
    }
}