Example usage for org.deeplearning4j.nn.conf.distribution GaussianDistribution GaussianDistribution

Introduction

In this page you can find the example usage for org.deeplearning4j.nn.conf.distribution GaussianDistribution GaussianDistribution.

Prototype

@JsonCreator
public GaussianDistribution(@JsonProperty("mean") double mean, @JsonProperty("std") double std) 

Source Link

Document

Create a gaussian distribution (equivalent to normal) with the given mean and std

Usage

From source file:org.eigengo.rsa.identity.v100.AlexNet.java

License:Open Source License

public MultiLayerConfiguration conf() {
    double nonZeroBias = 1;
    double dropOut = 0.5;
    SubsamplingLayer.PoolingType poolingType = SubsamplingLayer.PoolingType.MAX;

    // TODO split and link kernel maps on GPUs - 2nd, 4th, 5th convolution should only connect maps on the same gpu, 3rd connects to all in 2nd
    MultiLayerConfiguration.Builder conf = new NeuralNetConfiguration.Builder().seed(seed)
            .weightInit(WeightInit.DISTRIBUTION).dist(new NormalDistribution(0.0, 0.01)).activation("relu")
            .updater(Updater.NESTEROVS).iterations(iterations)
            .gradientNormalization(GradientNormalization.RenormalizeL2PerLayer) // normalize to prevent vanishing or exploding gradients
            .optimizationAlgo(OptimizationAlgorithm.STOCHASTIC_GRADIENT_DESCENT).learningRate(1e-2)
            .biasLearningRate(1e-2 * 2).learningRateDecayPolicy(LearningRatePolicy.Step).lrPolicyDecayRate(0.1)
            .lrPolicySteps(100000).regularization(true).l2(5 * 1e-4).momentum(0.9).miniBatch(false).list()
            .layer(0,/*from w  ww  . j a va2 s  .co m*/
                    new ConvolutionLayer.Builder(new int[] { 11, 11 }, new int[] { 4, 4 }, new int[] { 3, 3 })
                            .name("cnn1").nIn(channels).nOut(96).build())
            .layer(1, new LocalResponseNormalization.Builder().name("lrn1").build())
            .layer(2,
                    new SubsamplingLayer.Builder(poolingType, new int[] { 3, 3 }, new int[] { 2, 2 })
                            .name("maxpool1").build())
            .layer(3,
                    new ConvolutionLayer.Builder(new int[] { 5, 5 }, new int[] { 1, 1 }, new int[] { 2, 2 })
                            .name("cnn2").nOut(256).biasInit(nonZeroBias).build())
            .layer(4,
                    new LocalResponseNormalization.Builder().name("lrn2").k(2).n(5).alpha(1e-4).beta(0.75)
                            .build())
            .layer(5,
                    new SubsamplingLayer.Builder(poolingType, new int[] { 3, 3 }, new int[] { 2, 2 })
                            .name("maxpool2").build())
            .layer(6,
                    new ConvolutionLayer.Builder(new int[] { 3, 3 }, new int[] { 1, 1 }, new int[] { 1, 1 })
                            .name("cnn3").nOut(384).build())
            .layer(7,
                    new ConvolutionLayer.Builder(new int[] { 3, 3 }, new int[] { 1, 1 }, new int[] { 1, 1 })
                            .name("cnn4").nOut(384).biasInit(nonZeroBias).build())
            .layer(8,
                    new ConvolutionLayer.Builder(new int[] { 3, 3 }, new int[] { 1, 1 }, new int[] { 1, 1 })
                            .name("cnn5").nOut(256).biasInit(nonZeroBias).build())
            .layer(9,
                    new SubsamplingLayer.Builder(poolingType, new int[] { 3, 3 }, new int[] { 2, 2 })
                            .name("maxpool3").build())
            .layer(10,
                    new DenseLayer.Builder().name("ffn1").nOut(4096).dist(new GaussianDistribution(0, 0.005))
                            .biasInit(nonZeroBias).dropOut(dropOut).build())
            .layer(11,
                    new DenseLayer.Builder().name("ffn2").nOut(4096).dist(new GaussianDistribution(0, 0.005))
                            .biasInit(nonZeroBias).dropOut(dropOut).build())
            .layer(12,
                    new OutputLayer.Builder(LossFunctions.LossFunction.NEGATIVELOGLIKELIHOOD).name("output")
                            .nOut(numLabels).activation("softmax").build())
            .backprop(true).pretrain(false).cnnInputSize(height, width, channels);

    return conf.build();
}

---