How to configure DL4j for local images

Question

I'm trying to use DeepLearning4j to categorize 32x32 images in numbers from 0-9. I've looked up a number of examples and tutorials, but always run into some exception when fitting the dataset to a network.

Im currently trying to use a ImageRecordReader with ParentPathLabelGenerator and RecordReaderDataSetIterator.

The images seem to load fine but i always run into a DL4JInvalidInputException when fitting.

        File parentDir = new File(dataPath);
        FileSplit filesInDir = new FileSplit(parentDir, NativeImageLoader.ALLOWED_FORMATS);
        ParentPathLabelGenerator labelMaker = new ParentPathLabelGenerator();

        BalancedPathFilter pathFilter = new BalancedPathFilter(new Random(), labelMaker, 100);
        InputSplit[] filesInDirSplit = filesInDir.sample(pathFilter, 80, 20);
        InputSplit trainData = filesInDirSplit[0];
        InputSplit testData = filesInDirSplit[1];

        ImageRecordReader recordReader = new ImageRecordReader(numRows, numColumns, 3, labelMaker);
        recordReader.initialize(trainData);

        DataSetIterator dataIter = new RecordReaderDataSetIterator(recordReader, 1, 1, outputNum);

When using DenseLayer:

Exception in thread "main" org.deeplearning4j.exception.DL4JInvalidInputException: Input that is not a matrix; expected matrix (rank 2), got rank 4 array with shape [1, 3, 32, 32]. Missing preprocessor or wrong input type? (layer name: layer0, layer index: 0, layer type: DenseLayer)

When using ConvolutionLayer the error occures at the OutputLayer:

Exception in thread "main" org.deeplearning4j.exception.DL4JInvalidInputException: Input that is not a matrix; expected matrix (rank 2), got rank 4 array with shape [1, 1000, 28, 28]. Missing preprocessor or wrong input type? (layer name: layer1, layer index: 1, layer type: OutputLayer)

Is my attempt at loading the images incorrect or is my network misconfigured?

Configuration:

MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder()
                .list()
                .layer(0, new ConvolutionLayer.Builder()
                        .nIn(3) // Number of input datapoints.
                        .nOut(1000) // Number of output datapoints.
                        .activation(Activation.RELU) // Activation function.
                        .weightInit(WeightInit.XAVIER) // Weight initialization.
                        .build())
                .layer(1, new OutputLayer.Builder(LossFunctions.LossFunction.NEGATIVELOGLIKELIHOOD)
                        .nIn(1000)
                        .nOut(outputNum)
                        .activation(Activation.SOFTMAX)
                        .weightInit(WeightInit.XAVIER)
                        .build())
                .build();

Answer 1

The easiest way is to use the .setInputType configuration option when defining the network. It will set up all the necessary pre-processors for you, and it will calculate all the correct .nIn values too.

Take another look at this example https://github.com/eclipse/deeplearning4j-examples/blob/master/dl4j-examples/src/main/java/org/deeplearning4j/examples/convolution/mnist/MnistClassifier.java#L156

When you use the .setInputType way of setting up your network, you don't have to set any .nIn values at all - you still can, as is evident in the example I've linked, but usually there is no good reason to do so.