NeuralNetConfiguration.java

package org.deeplearning4j.nn.conf;

import org.apache.commons.math3.random.RandomGenerator;
import org.deeplearning4j.nn.WeightInit;
import org.deeplearning4j.nn.api.NeuralNetwork;

import java.io.Serializable;
import java.util.HashMap;
import java.util.Map;

/**
 * A Serializable configuration
 */
public class NeuralNetConfiguration implements Serializable {

    private float sparsity;
    private boolean useAdaGrad = true;
    private float lr = 1e-1f;
    /* momentum for learning */
    protected float momentum = 0.5f;
    /* L2 Regularization constant */
    protected float l2 = 0.1f;
    protected boolean useRegularization = false;
    //momentum after n iterations
    protected Map<Integer,Float> momentumAfter = new HashMap<>();
    //reset adagrad historical gradient after n iterations
    protected int resetAdaGradIterations = -1;
    protected float dropOut = 0;
    //use only when binary hidden layers are active
    protected boolean applySparsity = false;
    //weight init scheme, this can either be a distribution or a applyTransformToDestination scheme
    protected WeightInit weightInit;
    protected NeuralNetwork.OptimizationAlgorithm optimizationAlgo;
    protected int renderWeightsEveryNumEpochs = -1;
    //whether to concat hidden bias or add it
    protected  boolean concatBiases = false;
    //whether to constrain the gradient to unit norm or not
    protected boolean constrainGradientToUnitNorm = false;
    /* RNG for sampling. */
    protected RandomGenerator rng;
    protected long seed = 123;

    public NeuralNetConfiguration(float sparsity, boolean useAdaGrad, float lr, float momentum, float l2, boolean useRegularization, Map<Integer, Float> momentumAfter, int resetAdaGradIterations, float dropOut, boolean applySparsity, WeightInit weightInit, NeuralNetwork.OptimizationAlgorithm optimizationAlgo, int renderWeightsEveryNumEpochs, boolean concatBiases, boolean constrainGradientToUnitNorm, RandomGenerator rng, long seed) {
        this.sparsity = sparsity;
        this.useAdaGrad = useAdaGrad;
        this.lr = lr;
        this.momentum = momentum;
        this.l2 = l2;
        this.useRegularization = useRegularization;
        this.momentumAfter = momentumAfter;
        this.resetAdaGradIterations = resetAdaGradIterations;
        this.dropOut = dropOut;
        this.applySparsity = applySparsity;
        this.weightInit = weightInit;
        this.optimizationAlgo = optimizationAlgo;
        this.renderWeightsEveryNumEpochs = renderWeightsEveryNumEpochs;
        this.concatBiases = concatBiases;
        this.constrainGradientToUnitNorm = constrainGradientToUnitNorm;
        this.rng = rng;
        this.seed = seed;
    }
}