Method train

JSAT/src/jsat/classifiers/linear/NewGLMNET.java:308–825

(ClassificationDataSet dataSet, Vec w_init, double b_init, boolean useInit)

    }
    
    private void train(ClassificationDataSet dataSet, Vec w_init, double b_init, boolean useInit)
    {
        /*
         * The original NewGLMNET paper describes the algorithm as minimizing 
         * f(w) = ||w||_1 + L(w), where L(w) is the logistic loss summed over 
         * all the variables. To make adapation to elastic net easier, we define
         * f(w) = alpha ||w||_1 + L(w), where L(w) = (1-alpha) ||w||_2 + loss sum. 
         * This way we keep all the framework for L_1 regularization and 
         * shrinking, and just update the appropriate terms where necessary. 
         */
        
        //paper uses n= #features so we will follow their lead
        final int n = dataSet.getNumNumericalVars();
        //l = # data points
        final int l = dataSet.getSampleSize();
        
        if(useInit)
        {
            w = new DenseVector(w_init);
            b = useBias ? b_init : 0;
        }
        else
        {
            w = new DenseVector(n);
            b = 0;
        }
        List<Vec> X = dataSet.getDataVectors();
        
        double first_M_bar = 0;
        double e_in = 1.0;//set later when first_M_bar is set
        
        double[] w_dot_x = new double[l];
        double[] exp_w_dot_x = new double[l];
        double[] exp_w_dot_x_plus_dx = new double[l];
        /**
         * Used in the linear search step at the end
         */
        double[] d_dot_x = new double[l];
        /**
         * Contains the value 1/(1+e^(w^T x)). This is used in computing D and the partial derivatives. 
         */
        double[] D_part = new double[l];
        double[] D = new double[l];
        
        /**
         * Stores the value H<sup>k</sup><sub>j,j</sub> computer at the start of each iteration
         */
        double[] H = new double[n];
        /**
         * Stores the value H<sup>k</sup><sub>j,j</sub> computer at the start of
         * each iteration for the bias term
         */
        double H_bias = 0;
        /**
         * Stores the value &nambla; L<sub>j</sub>
         */
        double[] delta_L = new double[n];
        /**