nn.java

        double learningRate = 0.01;
        int numHiddenNodes = 64*2;
        int numInnerNodes = 32;
        int numOutputs = 2;

        double regularization = 0.1;

		MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder()
                .seed(seed)
                .optimizationAlgo(OptimizationAlgorithm.STOCHASTIC_GRADIENT_DESCENT)
                .learningRate(learningRate)
                .updater(Updater.NESTEROVS).momentum(0.9)
                .regularization(true).l2(regularization)
                .list()
                .layer(0, new DenseLayer.Builder().nIn(numInputs).nOut(numHiddenNodes)
                        .weightInit(WeightInit.RELU_UNIFORM)
                        .activation(Activation.LEAKYRELU)
                        //.dropOut(0.9)
                        .build())
                .layer(1, new DenseLayer.Builder().nIn(numHiddenNodes).nOut(numHiddenNodes)
                        .weightInit(WeightInit.RELU_UNIFORM)
                        .activation(Activation.LEAKYRELU)
                        //.dropOut(0.5)
                        .build())
                .layer(2, new DenseLayer.Builder().nIn(numHiddenNodes).nOut(numInnerNodes)
                        .weightInit(WeightInit.RELU_UNIFORM)
                        .activation(Activation.LEAKYRELU)
                        //.dropOut(0.5)
                        .build())
                .layer(3, new DenseLayer.Builder().nIn(numInnerNodes).nOut(numInnerNodes)
                        .weightInit(WeightInit.XAVIER)
                        .activation(Activation.LEAKYRELU)
                        .build())
                .layer(4, new DenseLayer.Builder().nIn(numInnerNodes).nOut(numInnerNodes)
                        .weightInit(WeightInit.XAVIER)
                        .activation(Activation.TANH)
                        .build())
                .layer(5, new OutputLayer.Builder(LossFunction.MSE)
                        .weightInit(WeightInit.XAVIER)
                        .activation(Activation.SOFTMAX)
                        .nIn(numInnerNodes).nOut(numOutputs).build())
                .pretrain(false).backprop(true).build();