Flatten in Eclipse Deeplearning4j, For example to flatten Rank 4 tensor to Rank 2 (batch size, nIn) input to the Dense Layer.

Flatten.java

// Take note this config is just an example
// Take note that my input is a Rank 4 tensor, a DataSetIterator that uses ImageRecordReader 
// Method 1: using .setInputTypes(InputType.convolutional(height,width,channels))
ComputationGraphConfiguration config = new NeuralNetConfiguration.Builder()
        .seed(seed)
        .updater(new RmsProp(1e-3))
        .weightInit(WeightInit.XAVIER)
        .l2(1e-4)
        .graphBuilder()
        .addInputs("input")
        .layer("vae",new VariationalAutoencoder.Builder()
                .activation(Activation.LEAKYRELU)
                .encoderLayerSizes(20,10)                    //1 encoder layer with 20 nodes
                .decoderLayerSizes(5,15)                    //2 decoder layers with 15 and 5 nodes respectively
                .pzxActivationFunction(Activation.IDENTITY)  //p(z|data) activation function
                .reconstructionDistribution(new BernoulliReconstructionDistribution(Activation.SIGMOID.getActivationFunction()))     //Bernoulli distribution for p(data|z) (binary or 0 to 1 data only)
                .nIn(width*height*channels)
                .nOut(2)                            //Size of the latent variable space: p(z|x). 2 dimensions here for plotting, use more in general
                .build(),"input")
        .setInputTypes(InputType.convolutional(height,width,channels))
        .setOutputs("vae")
        .build();

// Method 2: add a CnnToFeedForwardPreprocessor as a vertex, your network must be a ComputationGraph
// .addVertex("flatten",new PreprocessorVertex(new CnnToFeedForwardPreProcessor(height,width,channels)),"input")

ComputationGraphConfiguration config = new NeuralNetConfiguration.Builder()
        .seed(seed)
        .updater(new RmsProp(1e-3))
        .weightInit(WeightInit.XAVIER)
        .l2(1e-4)
        .graphBuilder()
        .addInputs("input")
        .addVertex("flatten",new PreprocessorVertex(new CnnToFeedForwardPreProcessor(height,width,channels)),"input")
        .layer("vae",new VariationalAutoencoder.Builder()
                .activation(Activation.LEAKYRELU)
                .encoderLayerSizes(20,10)                    //1 encoder layer with 20 nodes
                .decoderLayerSizes(5,15)                    //2 decoder layers with 15 and 5 nodes respectively
                .pzxActivationFunction(Activation.IDENTITY)  //p(z|data) activation function
                .reconstructionDistribution(new BernoulliReconstructionDistribution(Activation.SIGMOID.getActivationFunction()))     //Bernoulli distribution for p(data|z) (binary or 0 to 1 data only)
                .nIn(width*height*channels)
                .nOut(2)                            //Size of the latent variable space: p(z|x). 2 dimensions here for plotting, use more in general
                .build(),"flatten")
        .setInputTypes(InputType.convolutional(height,width,channels))
        .setOutputs("vae")
        .build();

// Reference: https://deeplearning4j.konduit.ai/getting-started/cheat-sheet#input-pre-processors
// I have tried ReshapeVertex but it wont work if batch size is more than 1 because the new shape requires a dynamic size.