My goal was to build a CNN autoencoder as I have never tried to build a CNN yet. I must be getting used to the api because it only took a single day. I used this dataset http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html
It works basically as I expected but surprisingly well for my first attempt. I’m posting in case you have any corrections I could implement or If your looking for an example as I couldn’t find one myself of a CNNAE.
edit: I think I quoted it wrong but this is a single class file I pasted even though it looks all broken up. You should be able to do a copy and paste. Instantiate and call run with any image dataset.
package ML;
import nu.pattern.OpenCV;
import org.datavec.api.split.FileSplit;
import org.datavec.image.loader.Java2DNativeImageLoader;
import org.datavec.image.recordreader.ImageRecordReader;
import org.deeplearning4j.datasets.datavec.RecordReaderDataSetIterator;
import org.deeplearning4j.datasets.fetchers.DataSetType;
import org.deeplearning4j.datasets.iterator.impl.Cifar10DataSetIterator;
import org.deeplearning4j.datasets.iterator.impl.MnistDataSetIterator;
import org.deeplearning4j.nn.api.Model;
import org.deeplearning4j.nn.api.OptimizationAlgorithm;
import org.deeplearning4j.nn.conf.MultiLayerConfiguration;
import org.deeplearning4j.nn.conf.NeuralNetConfiguration;
import org.deeplearning4j.nn.conf.inputs.InputType;
import org.deeplearning4j.nn.conf.layers.*;
import org.deeplearning4j.nn.conf.layers.variational.BernoulliReconstructionDistribution;
import org.deeplearning4j.nn.conf.layers.variational.VariationalAutoencoder;
import org.deeplearning4j.nn.conf.ocnn.OCNNOutputLayer;
import org.deeplearning4j.nn.conf.preprocessor.CnnToFeedForwardPreProcessor;
import org.deeplearning4j.nn.conf.preprocessor.FeedForwardToCnnPreProcessor;
import org.deeplearning4j.nn.multilayer.MultiLayerNetwork;
import org.deeplearning4j.nn.weights.WeightInit;
import org.deeplearning4j.nn.workspace.LayerWorkspaceMgr;
import org.deeplearning4j.optimize.api.BaseTrainingListener;
import org.deeplearning4j.optimize.listeners.ScoreIterationListener;
import org.nd4j.linalg.activations.Activation;
import org.nd4j.linalg.api.buffer.DataType;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.dataset.api.DataSet;
import org.nd4j.linalg.dataset.api.iterator.DataSetIterator;
import org.nd4j.linalg.factory.Nd4j;
import org.nd4j.linalg.indexing.NDArrayIndex;
import org.nd4j.linalg.learning.config.AdaDelta;
import org.nd4j.linalg.learning.config.RmsProp;
import org.nd4j.linalg.lossfunctions.LossFunctions;
import org.opencv.core.Mat;
import org.opencv.imgcodecs.Imgcodecs;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;import javax.swing.;
import java.awt.;
import java.awt.image.BufferedImage;
import java.io.File;
import java.util.ArrayList;
import java.util.List;public class NewExp {
private static JFrame frame; private static JLabel label; JLabel originalLaabel; JLabel resultLabel; Java2DNativeImageLoader imageLoader; Java2DNativeImageLoader imageLoaderTwo; private static final Logger log = LoggerFactory.getLogger(NewExp.class); int height = 32; int width = 32;// String source = “D:\Downloads\img_align_celeba\output\”;
String source = “D:\Downloads\img_align_celeba\img_align_celeba\”;int channels = 3; public NewExp() { OpenCV.loadLocally(); String[] file = new File(source).list(); Mat src = Imgcodecs.imread(source + file[0]); System.out.println("FINAL \n Width: " + src.width() + "\n Height: " + src.height()); frame = new JFrame("Results"); frame.setDefaultCloseOperation(JFrame.EXIT_ON_CLOSE); imageLoader = new Java2DNativeImageLoader(); imageLoaderTwo = new Java2DNativeImageLoader(); originalLaabel = new JLabel(); originalLaabel.setBounds(0,0, width * 2, height * 2); resultLabel = new JLabel(); resultLabel.setBounds(width * 2,0, width * 2, height * 2); frame.add(originalLaabel); frame.add(resultLabel); frame.setSize(width * 3, height * 4 ); frame.setLayout(null); frame.setVisible(true); } public void run() { DataSetIterator iter = getDataSetIter(); MultiLayerNetwork net = getNet(); for (int i = 0; i < 100; i++) { int count = 0; while (iter.hasNext()){ DataSet d = iter.next(); net.fit(d.getFeatures(), d.getFeatures()); if ((count % 100) == 0) { BufferedImage bf = imageLoader.asBufferedImage(d.getFeatures()); List<INDArray> a = net.feedForwardToLayer(10, d.getFeatures()); BufferedImage bfO = imageLoaderTwo.asBufferedImage(a.get(a.size() - 1)); originalLaabel.setIcon(new ImageIcon(bf)); resultLabel.setIcon(new ImageIcon(bfO)); System.out.println(""); } count++; } } } public DataSetIterator getDataSetIter() { try{ ImageRecordReader recordReader = new ImageRecordReader(height,width, 3); recordReader.initialize(new FileSplit(new File(source))); DataSetIterator dataSetIterator = new RecordReaderDataSetIterator(recordReader, 1); return dataSetIterator; }catch (Exception e){ e.printStackTrace(); } return null; } private MultiLayerNetwork getNet() { int rngSeed = 123; int dimensions = 16; //Neural net configuration Nd4j.getRandom().setSeed(rngSeed); MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder() .seed(rngSeed) .updater(new RmsProp(1e-3)) .weightInit(WeightInit.XAVIER) .l2(1e-4) .list() //encode .layer(0, new ConvolutionLayer.Builder().kernelSize(3,3).stride(1,1).activation(Activation.RELU).nIn(channels).nOut(32).build()) //(32-3)/1 + 1 = 32x30x30 //(64-3)/1 + 1 = 32x62x62 .layer( new BatchNormalization()) .layer(1, new SubsamplingLayer.Builder().kernelSize(2,2).stride(2,2).poolingType(SubsamplingLayer.PoolingType.MAX).build()) //(30-2)/2 + 1 = 32x15x15 //(62-3)/2 + 1 = 32x30x30 .layer(2, new ConvolutionLayer.Builder().kernelSize(2,2).stride(1,1).activation(Activation.RELU).nIn(32).nOut(16).build()) //(15-2)/1 + 1 = 16x14x14 //(35-2)/1 + 1 = 16x29x29 .layer(new BatchNormalization()) .layer(3, new SubsamplingLayer.Builder().kernelSize(2,2).stride(2,2).poolingType(SubsamplingLayer.PoolingType.MAX).build()) //(14-2)/2 + 1 = 16x7x7 //(29-2)/2 + 1 = 16x15x15 //double stuff oreo center .layer(4, new DenseLayer.Builder().nIn(784).nOut(dimensions).weightInit(WeightInit.XAVIER).activation(Activation.IDENTITY).build()) .layer(5, new DenseLayer.Builder().nIn(dimensions).nOut(784).weightInit(WeightInit.XAVIER).activation(Activation.IDENTITY).build()) //decode .layer( 6, new Upsampling2D.Builder().size(2) .build()) .layer( new BatchNormalization()) .layer(7, new Deconvolution2D.Builder().kernelSize(2,2) .stride(1,1).nIn(16).nOut(32).activation(Activation.RELU).build()) .layer(8, new Upsampling2D.Builder().size(2).build()) .layer( new BatchNormalization()) .layer(9, new Deconvolution2D.Builder().kernelSize(3,3).stride(1,1).activation(Activation.RELU).nIn(32).nOut(channels).build()) .layer(10, new CnnLossLayer.Builder(LossFunctions.LossFunction.MSE).activation(Activation.IDENTITY).build()) //.layer(10, new OutputLayer.Builder(LossFunctions.LossFunction.MSE).nIn(3072).nOut(3072).activation(Activation.IDENTITY).build()) // say hail marys here .inputPreProcessor(4, new CnnToFeedForwardPreProcessor(7,7,16)) .inputPreProcessor(6, new FeedForwardToCnnPreProcessor(7,7,16)) //.inputPreProcessor(10, new CnnToFeedForwardPreProcessor(32,32,3)) .build(); MultiLayerNetwork net = new MultiLayerNetwork(conf); net.init(); net.setListeners(new ScoreIterationListener(100)); return net; }}