Package de.bwaldvogel.liblinear

Source Code of de.bwaldvogel.liblinear.Linear$GroupClassesReturn

package de.bwaldvogel.liblinear;

import java.io.BufferedReader;
import java.io.BufferedWriter;
import java.io.Closeable;
import java.io.EOFException;
import java.io.File;
import java.io.FileInputStream;
import java.io.FileOutputStream;
import java.io.IOException;
import java.io.InputStreamReader;
import java.io.OutputStreamWriter;
import java.io.PrintStream;
import java.io.Reader;
import java.io.Writer;
import java.nio.charset.Charset;
import java.util.Formatter;
import java.util.Locale;
import java.util.Random;
import java.util.regex.Pattern;

// LIBLINEAR is a library for Large Linear Classification
/**
* <h2>Java port of <a href="http://www.csie.ntu.edu.tw/~cjlin/liblinear/">liblinear</a></h2>
*
* <p>The usage should be pretty similar to the C version of <tt>liblinear</tt>.</p>
* <p>Please consider reading the <tt>README</tt> file of <tt>liblinear</tt>.</p>
*
* <p><em>The port was done by Benedikt Waldvogel (mail at bwaldvogel.de)</em></p>
*
* @version 1.92
*/
public class Linear {

    static final Charset       FILE_CHARSET        = Charset.forName("ISO-8859-1");

    static final Locale        DEFAULT_LOCALE      = Locale.ENGLISH;

    private static Object      OUTPUT_MUTEX        = new Object();
    private static PrintStream DEBUG_OUTPUT        = System.out;

    private static final long  DEFAULT_RANDOM_SEED = 0L;
    static Random              random              = new Random(DEFAULT_RANDOM_SEED);

    /**
     * @param target predicted classes
     */
    public static void crossValidation(Problem prob, Parameter param, int nr_fold, double[] target) {
        int i;
        int[] fold_start = new int[nr_fold + 1];
        int l = prob.l;
        int[] perm = new int[l];

        for (i = 0; i < l; i++)
            perm[i] = i;
        for (i = 0; i < l; i++) {
            int j = i + random.nextInt(l - i);
            swap(perm, i, j);
        }
        for (i = 0; i <= nr_fold; i++)
            fold_start[i] = i * l / nr_fold;

        for (i = 0; i < nr_fold; i++) {
            int begin = fold_start[i];
            int end = fold_start[i + 1];
            int j, k;
            Problem subprob = new Problem();

            subprob.bias = prob.bias;
            subprob.n = prob.n;
            subprob.l = l - (end - begin);
            subprob.x = new Feature[subprob.l][];
            subprob.y = new double[subprob.l];

            k = 0;
            for (j = 0; j < begin; j++) {
                subprob.x[k] = prob.x[perm[j]];
                subprob.y[k] = prob.y[perm[j]];
                ++k;
            }
            for (j = end; j < l; j++) {
                subprob.x[k] = prob.x[perm[j]];
                subprob.y[k] = prob.y[perm[j]];
                ++k;
            }
            Model submodel = train(subprob, param);
            for (j = begin; j < end; j++)
                target[perm[j]] = predict(submodel, prob.x[perm[j]]);
        }
    }

    /** used as complex return type */
    private static class GroupClassesReturn {

        final int[] count;
        final int[] label;
        final int   nr_class;
        final int[] start;

        GroupClassesReturn( int nr_class, int[] label, int[] start, int[] count ) {
            this.nr_class = nr_class;
            this.label = label;
            this.start = start;
            this.count = count;
        }
    }

    private static GroupClassesReturn groupClasses(Problem prob, int[] perm) {
        int l = prob.l;
        int max_nr_class = 16;
        int nr_class = 0;

        int[] label = new int[max_nr_class];
        int[] count = new int[max_nr_class];
        int[] data_label = new int[l];
        int i;

        for (i = 0; i < l; i++) {
            int this_label = (int)prob.y[i];
            int j;
            for (j = 0; j < nr_class; j++) {
                if (this_label == label[j]) {
                    ++count[j];
                    break;
                }
            }
            data_label[i] = j;
            if (j == nr_class) {
                if (nr_class == max_nr_class) {
                    max_nr_class *= 2;
                    label = copyOf(label, max_nr_class);
                    count = copyOf(count, max_nr_class);
                }
                label[nr_class] = this_label;
                count[nr_class] = 1;
                ++nr_class;
            }
        }

        int[] start = new int[nr_class];
        start[0] = 0;
        for (i = 1; i < nr_class; i++)
            start[i] = start[i - 1] + count[i - 1];
        for (i = 0; i < l; i++) {
            perm[start[data_label[i]]] = i;
            ++start[data_label[i]];
        }
        start[0] = 0;
        for (i = 1; i < nr_class; i++)
            start[i] = start[i - 1] + count[i - 1];

        return new GroupClassesReturn(nr_class, label, start, count);
    }

    static void info(String message) {
        synchronized (OUTPUT_MUTEX) {
            if (DEBUG_OUTPUT == null) return;
            DEBUG_OUTPUT.printf(message);
            DEBUG_OUTPUT.flush();
        }
    }

    static void info(String format, Object... args) {
        synchronized (OUTPUT_MUTEX) {
            if (DEBUG_OUTPUT == null) return;
            DEBUG_OUTPUT.printf(format, args);
            DEBUG_OUTPUT.flush();
        }
    }

    /**
     * @param s the string to parse for the double value
     * @throws IllegalArgumentException if s is empty or represents NaN or Infinity
     * @throws NumberFormatException see {@link Double#parseDouble(String)}
     */
    static double atof(String s) {
        if (s == null || s.length() < 1) throw new IllegalArgumentException("Can't convert empty string to integer");
        double d = Double.parseDouble(s);
        if (Double.isNaN(d) || Double.isInfinite(d)) {
            throw new IllegalArgumentException("NaN or Infinity in input: " + s);
        }
        return (d);
    }

    /**
     * @param s the string to parse for the integer value
     * @throws IllegalArgumentException if s is empty
     * @throws NumberFormatException see {@link Integer#parseInt(String)}
     */
    static int atoi(String s) throws NumberFormatException {
        if (s == null || s.length() < 1) throw new IllegalArgumentException("Can't convert empty string to integer");
        // Integer.parseInt doesn't accept '+' prefixed strings
        if (s.charAt(0) == '+') s = s.substring(1);
        return Integer.parseInt(s);
    }

    /**
     * Java5 'backport' of Arrays.copyOf
     */
    public static double[] copyOf(double[] original, int newLength) {
        double[] copy = new double[newLength];
        System.arraycopy(original, 0, copy, 0, Math.min(original.length, newLength));
        return copy;
    }

    /**
     * Java5 'backport' of Arrays.copyOf
     */
    public static int[] copyOf(int[] original, int newLength) {
        int[] copy = new int[newLength];
        System.arraycopy(original, 0, copy, 0, Math.min(original.length, newLength));
        return copy;
    }

    /**
     * Loads the model from inputReader.
     * It uses {@link java.util.Locale#ENGLISH} for number formatting.
     *
     * <p>Note: The inputReader is <b>NOT closed</b> after reading or in case of an exception.</p>
     */
    public static Model loadModel(Reader inputReader) throws IOException {
        Model model = new Model();

        model.label = null;

        Pattern whitespace = Pattern.compile("\\s+");

        BufferedReader reader = null;
        if (inputReader instanceof BufferedReader) {
            reader = (BufferedReader)inputReader;
        } else {
            reader = new BufferedReader(inputReader);
        }

        String line = null;
        while ((line = reader.readLine()) != null) {
            String[] split = whitespace.split(line);
            if (split[0].equals("solver_type")) {
                SolverType solver = SolverType.valueOf(split[1]);
                if (solver == null) {
                    throw new RuntimeException("unknown solver type");
                }
                model.solverType = solver;
            } else if (split[0].equals("nr_class")) {
                model.nr_class = atoi(split[1]);
                Integer.parseInt(split[1]);
            } else if (split[0].equals("nr_feature")) {
                model.nr_feature = atoi(split[1]);
            } else if (split[0].equals("bias")) {
                model.bias = atof(split[1]);
            } else if (split[0].equals("w")) {
                break;
            } else if (split[0].equals("label")) {
                model.label = new int[model.nr_class];
                for (int i = 0; i < model.nr_class; i++) {
                    model.label[i] = atoi(split[i + 1]);
                }
            } else {
                throw new RuntimeException("unknown text in model file: [" + line + "]");
            }
        }

        int w_size = model.nr_feature;
        if (model.bias >= 0) w_size++;

        int nr_w = model.nr_class;
        if (model.nr_class == 2 && model.solverType != SolverType.MCSVM_CS) nr_w = 1;

        model.w = new double[w_size * nr_w];
        int[] buffer = new int[128];

        for (int i = 0; i < w_size; i++) {
            for (int j = 0; j < nr_w; j++) {
                int b = 0;
                while (true) {
                    int ch = reader.read();
                    if (ch == -1) {
                        throw new EOFException("unexpected EOF");
                    }
                    if (ch == ' ') {
                        model.w[i * nr_w + j] = atof(new String(buffer, 0, b));
                        break;
                    } else {
                        buffer[b++] = ch;
                    }
                }
            }
        }

        return model;
    }

    /**
     * Loads the model from the file with ISO-8859-1 charset.
     * It uses {@link java.util.Locale#ENGLISH} for number formatting.
     */
    public static Model loadModel(File modelFile) throws IOException {
        BufferedReader inputReader = new BufferedReader(new InputStreamReader(new FileInputStream(modelFile), FILE_CHARSET));
        try {
            return loadModel(inputReader);
        }
        finally {
            inputReader.close();
        }
    }

    static void closeQuietly(Closeable c) {
        if (c == null) return;
        try {
            c.close();
        } catch (Throwable t) {}
    }

    public static double predict(Model model, Feature[] x) {
        double[] dec_values = new double[model.nr_class];
        return predictValues(model, x, dec_values);
    }

    /**
     * @throws IllegalArgumentException if model is not probabilistic (see {@link Model#isProbabilityModel()})
     */
    public static double predictProbability(Model model, Feature[] x, double[] prob_estimates) throws IllegalArgumentException {
        if (!model.isProbabilityModel()) {
            throw new IllegalArgumentException("probability output is only supported for logistic regression");
        }
        int nr_class = model.nr_class;
        int nr_w;
        if (nr_class == 2)
            nr_w = 1;
        else
            nr_w = nr_class;

        double label = predictValues(model, x, prob_estimates);
        for (int i = 0; i < nr_w; i++)
            prob_estimates[i] = 1 / (1 + Math.exp(-prob_estimates[i]));

        if (nr_class == 2) // for binary classification
            prob_estimates[1] = 1. - prob_estimates[0];
        else {
            double sum = 0;
            for (int i = 0; i < nr_class; i++)
                sum += prob_estimates[i];

            for (int i = 0; i < nr_class; i++)
                prob_estimates[i] = prob_estimates[i] / sum;
        }

        return label;
    }

    public static double predictValues(Model model, Feature[] x, double[] dec_values) {
        int n;
        if (model.bias >= 0)
            n = model.nr_feature + 1;
        else
            n = model.nr_feature;

        double[] w = model.w;

        int nr_w;
        if (model.nr_class == 2 && model.solverType != SolverType.MCSVM_CS)
            nr_w = 1;
        else
            nr_w = model.nr_class;

        for (int i = 0; i < nr_w; i++)
            dec_values[i] = 0;

        for (Feature lx : x) {
            int idx = lx.getIndex();
            // the dimension of testing data may exceed that of training
            if (idx <= n) {
                for (int i = 0; i < nr_w; i++) {
                    dec_values[i] += w[(idx - 1) * nr_w + i] * lx.getValue();
                }
            }
        }

        if (model.nr_class == 2) {
            if (model.solverType.isSupportVectorRegression())
                return dec_values[0];
            else
                return (dec_values[0] > 0) ? model.label[0] : model.label[1];
        } else {
            int dec_max_idx = 0;
            for (int i = 1; i < model.nr_class; i++) {
                if (dec_values[i] > dec_values[dec_max_idx]) dec_max_idx = i;
            }
            return model.label[dec_max_idx];
        }
    }

    static void printf(Formatter formatter, String format, Object... args) throws IOException {
        formatter.format(format, args);
        IOException ioException = formatter.ioException();
        if (ioException != null) throw ioException;
    }

    /**
     * Writes the model to the modelOutput.
     * It uses {@link java.util.Locale#ENGLISH} for number formatting.
     *
     * <p><b>Note: The modelOutput is closed after reading or in case of an exception.</b></p>
     */
    public static void saveModel(Writer modelOutput, Model model) throws IOException {
        int nr_feature = model.nr_feature;
        int w_size = nr_feature;
        if (model.bias >= 0) w_size++;

        int nr_w = model.nr_class;
        if (model.nr_class == 2 && model.solverType != SolverType.MCSVM_CS) nr_w = 1;

        Formatter formatter = new Formatter(modelOutput, DEFAULT_LOCALE);
        try {
            printf(formatter, "solver_type %s\n", model.solverType.name());
            printf(formatter, "nr_class %d\n", model.nr_class);

            if (model.label != null) {
                printf(formatter, "label");
                for (int i = 0; i < model.nr_class; i++) {
                    printf(formatter, " %d", model.label[i]);
                }
                printf(formatter, "\n");
            }

            printf(formatter, "nr_feature %d\n", nr_feature);
            printf(formatter, "bias %.16g\n", model.bias);

            printf(formatter, "w\n");
            for (int i = 0; i < w_size; i++) {
                for (int j = 0; j < nr_w; j++) {
                    double value = model.w[i * nr_w + j];

                    /** this optimization is the reason for {@link Model#equals(double[], double[])} */
                    if (value == 0.0) {
                        printf(formatter, "%d ", 0);
                    } else {
                        printf(formatter, "%.16g ", value);
                    }
                }
                printf(formatter, "\n");
            }

            formatter.flush();
            IOException ioException = formatter.ioException();
            if (ioException != null) throw ioException;
        }
        finally {
            formatter.close();
        }
    }

    /**
     * Writes the model to the file with ISO-8859-1 charset.
     * It uses {@link java.util.Locale#ENGLISH} for number formatting.
     */
    public static void saveModel(File modelFile, Model model) throws IOException {
        BufferedWriter modelOutput = new BufferedWriter(new OutputStreamWriter(new FileOutputStream(modelFile), FILE_CHARSET));
        saveModel(modelOutput, model);
    }

    /*
     * this method corresponds to the following define in the C version:
     * #define GETI(i) (y[i]+1)
     */
    private static int GETI(byte[] y, int i) {
        return y[i] + 1;
    }

    /**
     * A coordinate descent algorithm for
     * L1-loss and L2-loss SVM dual problems
     *<pre>
     *  min_\alpha  0.5(\alpha^T (Q + D)\alpha) - e^T \alpha,
     *    s.t.      0 <= \alpha_i <= upper_bound_i,
     *
     *  where Qij = yi yj xi^T xj and
     *  D is a diagonal matrix
     *
     * In L1-SVM case:
     *     upper_bound_i = Cp if y_i = 1
     *      upper_bound_i = Cn if y_i = -1
     *      D_ii = 0
     * In L2-SVM case:
     *      upper_bound_i = INF
     *      D_ii = 1/(2*Cp) if y_i = 1
     *      D_ii = 1/(2*Cn) if y_i = -1
     *
     * Given:
     * x, y, Cp, Cn
     * eps is the stopping tolerance
     *
     * solution will be put in w
     *
     * See Algorithm 3 of Hsieh et al., ICML 2008
     *</pre>
     */
    private static void solve_l2r_l1l2_svc(Problem prob, double[] w, double eps, double Cp, double Cn, SolverType solver_type) {
        int l = prob.l;
        int w_size = prob.n;
        int i, s, iter = 0;
        double C, d, G;
        double[] QD = new double[l];
        int max_iter = 1000;
        int[] index = new int[l];
        double[] alpha = new double[l];
        byte[] y = new byte[l];
        int active_size = l;

        // PG: projected gradient, for shrinking and stopping
        double PG;
        double PGmax_old = Double.POSITIVE_INFINITY;
        double PGmin_old = Double.NEGATIVE_INFINITY;
        double PGmax_new, PGmin_new;

        // default solver_type: L2R_L2LOSS_SVC_DUAL
        double diag[] = new double[] {0.5 / Cn, 0, 0.5 / Cp};
        double upper_bound[] = new double[] {Double.POSITIVE_INFINITY, 0, Double.POSITIVE_INFINITY};
        if (solver_type == SolverType.L2R_L1LOSS_SVC_DUAL) {
            diag[0] = 0;
            diag[2] = 0;
            upper_bound[0] = Cn;
            upper_bound[2] = Cp;
        }

        for (i = 0; i < l; i++) {
            if (prob.y[i] > 0) {
                y[i] = +1;
            } else {
                y[i] = -1;
            }
        }

        // Initial alpha can be set here. Note that
        // 0 <= alpha[i] <= upper_bound[GETI(i)]
        for (i = 0; i < l; i++)
            alpha[i] = 0;

        for (i = 0; i < w_size; i++)
            w[i] = 0;
        for (i = 0; i < l; i++) {
            QD[i] = diag[GETI(y, i)];

            for (Feature xi : prob.x[i]) {
                double val = xi.getValue();
                QD[i] += val * val;
                w[xi.getIndex() - 1] += y[i] * alpha[i] * val;
            }
            index[i] = i;
        }

        while (iter < max_iter) {
            PGmax_new = Double.NEGATIVE_INFINITY;
            PGmin_new = Double.POSITIVE_INFINITY;

            for (i = 0; i < active_size; i++) {
                int j = i + random.nextInt(active_size - i);
                swap(index, i, j);
            }

            for (s = 0; s < active_size; s++) {
                i = index[s];
                G = 0;
                byte yi = y[i];

                for (Feature xi : prob.x[i]) {
                    G += w[xi.getIndex() - 1] * xi.getValue();
                }
                G = G * yi - 1;

                C = upper_bound[GETI(y, i)];
                G += alpha[i] * diag[GETI(y, i)];

                PG = 0;
                if (alpha[i] == 0) {
                    if (G > PGmax_old) {
                        active_size--;
                        swap(index, s, active_size);
                        s--;
                        continue;
                    } else if (G < 0) {
                        PG = G;
                    }
                } else if (alpha[i] == C) {
                    if (G < PGmin_old) {
                        active_size--;
                        swap(index, s, active_size);
                        s--;
                        continue;
                    } else if (G > 0) {
                        PG = G;
                    }
                } else {
                    PG = G;
                }

                PGmax_new = Math.max(PGmax_new, PG);
                PGmin_new = Math.min(PGmin_new, PG);

                if (Math.abs(PG) > 1.0e-12) {
                    double alpha_old = alpha[i];
                    alpha[i] = Math.min(Math.max(alpha[i] - G / QD[i], 0.0), C);
                    d = (alpha[i] - alpha_old) * yi;

                    for (Feature xi : prob.x[i]) {
                        w[xi.getIndex() - 1] += d * xi.getValue();
                    }
                }
            }

            iter++;
            if (iter % 10 == 0) info(".");

            if (PGmax_new - PGmin_new <= eps) {
                if (active_size == l)
                    break;
                else {
                    active_size = l;
                    info("*");
                    PGmax_old = Double.POSITIVE_INFINITY;
                    PGmin_old = Double.NEGATIVE_INFINITY;
                    continue;
                }
            }
            PGmax_old = PGmax_new;
            PGmin_old = PGmin_new;
            if (PGmax_old <= 0) PGmax_old = Double.POSITIVE_INFINITY;
            if (PGmin_old >= 0) PGmin_old = Double.NEGATIVE_INFINITY;
        }

        info("%noptimization finished, #iter = %d%n", iter);
        if (iter >= max_iter) info("%nWARNING: reaching max number of iterations%nUsing -s 2 may be faster (also see FAQ)%n%n");

        // calculate objective value

        double v = 0;
        int nSV = 0;
        for (i = 0; i < w_size; i++)
            v += w[i] * w[i];
        for (i = 0; i < l; i++) {
            v += alpha[i] * (alpha[i] * diag[GETI(y, i)] - 2);
            if (alpha[i] > 0) ++nSV;
        }
        info("Objective value = %g%n", v / 2);
        info("nSV = %d%n", nSV);
    }

    // To support weights for instances, use GETI(i) (i)
    private static int GETI_SVR(int i) {
        return 0;
    }

    /**
     * A coordinate descent algorithm for
     * L1-loss and L2-loss epsilon-SVR dual problem
     *
     *  min_\beta  0.5\beta^T (Q + diag(lambda)) \beta - p \sum_{i=1}^l|\beta_i| + \sum_{i=1}^l yi\beta_i,
     *    s.t.      -upper_bound_i <= \beta_i <= upper_bound_i,
     *
     *  where Qij = xi^T xj and
     *  D is a diagonal matrix
     *
     * In L1-SVM case:
     *         upper_bound_i = C
     *         lambda_i = 0
     * In L2-SVM case:
     *         upper_bound_i = INF
     *         lambda_i = 1/(2*C)
     *
     * Given:
     * x, y, p, C
     * eps is the stopping tolerance
     *
     * solution will be put in w
     *
     * See Algorithm 4 of Ho and Lin, 2012
     */
    private static void solve_l2r_l1l2_svr(Problem prob, double[] w, Parameter param) {
        int l = prob.l;
        double C = param.C;
        double p = param.p;
        int w_size = prob.n;
        double eps = param.eps;
        int i, s, iter = 0;
        int max_iter = 1000;
        int active_size = l;
        int[] index = new int[l];

        double d, G, H;
        double Gmax_old = Double.POSITIVE_INFINITY;
        double Gmax_new, Gnorm1_new;
        double Gnorm1_init = 0; // initialize to 0 to get rid of Eclipse warning/error
        double[] beta = new double[l];
        double[] QD = new double[l];
        double[] y = prob.y;

        // L2R_L2LOSS_SVR_DUAL
        double[] lambda = new double[] {0.5 / C};
        double[] upper_bound = new double[] {Double.POSITIVE_INFINITY};

        if (param.solverType == SolverType.L2R_L1LOSS_SVR_DUAL) {
            lambda[0] = 0;
            upper_bound[0] = C;
        }

        // Initial beta can be set here. Note that
        // -upper_bound <= beta[i] <= upper_bound
        for (i = 0; i < l; i++)
            beta[i] = 0;

        for (i = 0; i < w_size; i++)
            w[i] = 0;
        for (i = 0; i < l; i++) {
            QD[i] = 0;
            for (Feature xi : prob.x[i]) {
                double val = xi.getValue();
                QD[i] += val * val;
                w[xi.getIndex() - 1] += beta[i] * val;
            }

            index[i] = i;
        }


        while (iter < max_iter) {
            Gmax_new = 0;
            Gnorm1_new = 0;

            for (i = 0; i < active_size; i++) {
                int j = i + random.nextInt(active_size - i);
                swap(index, i, j);
            }

            for (s = 0; s < active_size; s++) {
                i = index[s];
                G = -y[i] + lambda[GETI_SVR(i)] * beta[i];
                H = QD[i] + lambda[GETI_SVR(i)];

                for (Feature xi : prob.x[i]) {
                    int ind = xi.getIndex() - 1;
                    double val = xi.getValue();
                    G += val * w[ind];
                }

                double Gp = G + p;
                double Gn = G - p;
                double violation = 0;
                if (beta[i] == 0) {
                    if (Gp < 0)
                        violation = -Gp;
                    else if (Gn > 0)
                        violation = Gn;
                    else if (Gp > Gmax_old && Gn < -Gmax_old) {
                        active_size--;
                        swap(index, s, active_size);
                        s--;
                        continue;
                    }
                } else if (beta[i] >= upper_bound[GETI_SVR(i)]) {
                    if (Gp > 0)
                        violation = Gp;
                    else if (Gp < -Gmax_old) {
                        active_size--;
                        swap(index, s, active_size);
                        s--;
                        continue;
                    }
                } else if (beta[i] <= -upper_bound[GETI_SVR(i)]) {
                    if (Gn < 0)
                        violation = -Gn;
                    else if (Gn > Gmax_old) {
                        active_size--;
                        swap(index, s, active_size);
                        s--;
                        continue;
                    }
                } else if (beta[i] > 0)
                    violation = Math.abs(Gp);
                else
                    violation = Math.abs(Gn);

                Gmax_new = Math.max(Gmax_new, violation);
                Gnorm1_new += violation;

                // obtain Newton direction d
                if (Gp < H * beta[i])
                    d = -Gp / H;
                else if (Gn > H * beta[i])
                    d = -Gn / H;
                else
                    d = -beta[i];

                if (Math.abs(d) < 1.0e-12) continue;

                double beta_old = beta[i];
                beta[i] = Math.min(Math.max(beta[i] + d, -upper_bound[GETI_SVR(i)]), upper_bound[GETI_SVR(i)]);
                d = beta[i] - beta_old;

                if (d != 0) {
                    for (Feature xi : prob.x[i]) {
                        w[xi.getIndex() - 1] += d * xi.getValue();
                    }
                }
            }

            if (iter == 0) Gnorm1_init = Gnorm1_new;
            iter++;
            if (iter % 10 == 0) info(".");

            if (Gnorm1_new <= eps * Gnorm1_init) {
                if (active_size == l)
                    break;
                else {
                    active_size = l;
                    info("*");
                    Gmax_old = Double.POSITIVE_INFINITY;
                    continue;
                }
            }

            Gmax_old = Gmax_new;
        }

        info("%noptimization finished, #iter = %d%n", iter);
        if (iter >= max_iter) info("%nWARNING: reaching max number of iterations%nUsing -s 11 may be faster%n%n");

        // calculate objective value
        double v = 0;
        int nSV = 0;
        for (i = 0; i < w_size; i++)
            v += w[i] * w[i];
        v = 0.5 * v;
        for (i = 0; i < l; i++) {
            v += p * Math.abs(beta[i]) - y[i] * beta[i] + 0.5 * lambda[GETI_SVR(i)] * beta[i] * beta[i];
            if (beta[i] != 0) nSV++;
        }

        info("Objective value = %g%n", v);
        info("nSV = %d%n", nSV);
    }

    /**
     * A coordinate descent algorithm for
     * the dual of L2-regularized logistic regression problems
     *<pre>
     *  min_\alpha  0.5(\alpha^T Q \alpha) + \sum \alpha_i log (\alpha_i) + (upper_bound_i - \alpha_i) log (upper_bound_i - \alpha_i) ,
     *     s.t.      0 <= \alpha_i <= upper_bound_i,
     *
     *  where Qij = yi yj xi^T xj and
     *  upper_bound_i = Cp if y_i = 1
     *  upper_bound_i = Cn if y_i = -1
     *
     * Given:
     * x, y, Cp, Cn
     * eps is the stopping tolerance
     *
     * solution will be put in w
     *
     * See Algorithm 5 of Yu et al., MLJ 2010
     *</pre>
     *
     * @since 1.7
     */
    private static void solve_l2r_lr_dual(Problem prob, double w[], double eps, double Cp, double Cn) {
        int l = prob.l;
        int w_size = prob.n;
        int i, s, iter = 0;
        double xTx[] = new double[l];
        int max_iter = 1000;
        int index[] = new int[l];
        double alpha[] = new double[2 * l]; // store alpha and C - alpha
        byte y[] = new byte[l];
        int max_inner_iter = 100; // for inner Newton
        double innereps = 1e-2;
        double innereps_min = Math.min(1e-8, eps);
        double upper_bound[] = new double[] {Cn, 0, Cp};

        for (i = 0; i < l; i++) {
            if (prob.y[i] > 0) {
                y[i] = +1;
            } else {
                y[i] = -1;
            }
        }

        // Initial alpha can be set here. Note that
        // 0 < alpha[i] < upper_bound[GETI(i)]
        // alpha[2*i] + alpha[2*i+1] = upper_bound[GETI(i)]
        for (i = 0; i < l; i++) {
            alpha[2 * i] = Math.min(0.001 * upper_bound[GETI(y, i)], 1e-8);
            alpha[2 * i + 1] = upper_bound[GETI(y, i)] - alpha[2 * i];
        }

        for (i = 0; i < w_size; i++)
            w[i] = 0;
        for (i = 0; i < l; i++) {
            xTx[i] = 0;
            for (Feature xi : prob.x[i]) {
                double val = xi.getValue();
                xTx[i] += val * val;
                w[xi.getIndex() - 1] += y[i] * alpha[2 * i] * val;
            }
            index[i] = i;
        }

        while (iter < max_iter) {
            for (i = 0; i < l; i++) {
                int j = i + random.nextInt(l - i);
                swap(index, i, j);
            }
            int newton_iter = 0;
            double Gmax = 0;
            for (s = 0; s < l; s++) {
                i = index[s];
                byte yi = y[i];
                double C = upper_bound[GETI(y, i)];
                double ywTx = 0, xisq = xTx[i];
                for (Feature xi : prob.x[i]) {
                    ywTx += w[xi.getIndex() - 1] * xi.getValue();
                }
                ywTx *= y[i];
                double a = xisq, b = ywTx;

                // Decide to minimize g_1(z) or g_2(z)
                int ind1 = 2 * i, ind2 = 2 * i + 1, sign = 1;
                if (0.5 * a * (alpha[ind2] - alpha[ind1]) + b < 0) {
                    ind1 = 2 * i + 1;
                    ind2 = 2 * i;
                    sign = -1;
                }

                //  g_t(z) = z*log(z) + (C-z)*log(C-z) + 0.5a(z-alpha_old)^2 + sign*b(z-alpha_old)
                double alpha_old = alpha[ind1];
                double z = alpha_old;
                if (C - z < 0.5 * C) z = 0.1 * z;
                double gp = a * (z - alpha_old) + sign * b + Math.log(z / (C - z));
                Gmax = Math.max(Gmax, Math.abs(gp));

                // Newton method on the sub-problem
                final double eta = 0.1; // xi in the paper
                int inner_iter = 0;
                while (inner_iter <= max_inner_iter) {
                    if (Math.abs(gp) < innereps) break;
                    double gpp = a + C / (C - z) / z;
                    double tmpz = z - gp / gpp;
                    if (tmpz <= 0)
                        z *= eta;
                    else
                        // tmpz in (0, C)
                        z = tmpz;
                    gp = a * (z - alpha_old) + sign * b + Math.log(z / (C - z));
                    newton_iter++;
                    inner_iter++;
                }

                if (inner_iter > 0) // update w
                {
                    alpha[ind1] = z;
                    alpha[ind2] = C - z;
                    for (Feature xi : prob.x[i]) {
                        w[xi.getIndex() - 1] += sign * (z - alpha_old) * yi * xi.getValue();
                    }
                }
            }

            iter++;
            if (iter % 10 == 0) info(".");

            if (Gmax < eps) break;

            if (newton_iter <= l / 10) {
                innereps = Math.max(innereps_min, 0.1 * innereps);
            }

        }

        info("%noptimization finished, #iter = %d%n", iter);
        if (iter >= max_iter) info("%nWARNING: reaching max number of iterations%nUsing -s 0 may be faster (also see FAQ)%n%n");

        // calculate objective value

        double v = 0;
        for (i = 0; i < w_size; i++)
            v += w[i] * w[i];
        v *= 0.5;
        for (i = 0; i < l; i++)
            v += alpha[2 * i] * Math.log(alpha[2 * i]) + alpha[2 * i + 1] * Math.log(alpha[2 * i + 1]) - upper_bound[GETI(y, i)]
                * Math.log(upper_bound[GETI(y, i)]);
        info("Objective value = %g%n", v);
    }

    /**
     * A coordinate descent algorithm for
     * L1-regularized L2-loss support vector classification
     *
     *<pre>
     *  min_w \sum |wj| + C \sum max(0, 1-yi w^T xi)^2,
     *
     * Given:
     * x, y, Cp, Cn
     * eps is the stopping tolerance
     *
     * solution will be put in w
     *
     * See Yuan et al. (2010) and appendix of LIBLINEAR paper, Fan et al. (2008)
     *</pre>
     *
     * @since 1.5
     */
    private static void solve_l1r_l2_svc(Problem prob_col, double[] w, double eps, double Cp, double Cn) {
        int l = prob_col.l;
        int w_size = prob_col.n;
        int j, s, iter = 0;
        int max_iter = 1000;
        int active_size = w_size;
        int max_num_linesearch = 20;

        double sigma = 0.01;
        double d, G_loss, G, H;
        double Gmax_old = Double.POSITIVE_INFINITY;
        double Gmax_new, Gnorm1_new;
        double Gnorm1_init = 0; // eclipse moans this variable might not be initialized
        double d_old, d_diff;
        double loss_old = 0; // eclipse moans this variable might not be initialized
        double loss_new;
        double appxcond, cond;

        int[] index = new int[w_size];
        byte[] y = new byte[l];
        double[] b = new double[l]; // b = 1-ywTx
        double[] xj_sq = new double[w_size];

        double[] C = new double[] {Cn, 0, Cp};

        // Initial w can be set here.
        for (j = 0; j < w_size; j++)
            w[j] = 0;

        for (j = 0; j < l; j++) {
            b[j] = 1;
            if (prob_col.y[j] > 0)
                y[j] = 1;
            else
                y[j] = -1;
        }
        for (j = 0; j < w_size; j++) {
            index[j] = j;
            xj_sq[j] = 0;
            for (Feature xi : prob_col.x[j]) {
                int ind = xi.getIndex() - 1;
                xi.setValue(xi.getValue() * y[ind]); // x->value stores yi*xij
                double val = xi.getValue();
                b[ind] -= w[j] * val;

                xj_sq[j] += C[GETI(y, ind)] * val * val;
            }
        }

        while (iter < max_iter) {
            Gmax_new = 0;
            Gnorm1_new = 0;

            for (j = 0; j < active_size; j++) {
                int i = j + random.nextInt(active_size - j);
                swap(index, i, j);
            }

            for (s = 0; s < active_size; s++) {
                j = index[s];
                G_loss = 0;
                H = 0;

                for (Feature xi : prob_col.x[j]) {
                    int ind = xi.getIndex() - 1;
                    if (b[ind] > 0) {
                        double val = xi.getValue();
                        double tmp = C[GETI(y, ind)] * val;
                        G_loss -= tmp * b[ind];
                        H += tmp * val;
                    }
                }
                G_loss *= 2;

                G = G_loss;
                H *= 2;
                H = Math.max(H, 1e-12);

                double Gp = G + 1;
                double Gn = G - 1;
                double violation = 0;
                if (w[j] == 0) {
                    if (Gp < 0)
                        violation = -Gp;
                    else if (Gn > 0)
                        violation = Gn;
                    else if (Gp > Gmax_old / l && Gn < -Gmax_old / l) {
                        active_size--;
                        swap(index, s, active_size);
                        s--;
                        continue;
                    }
                } else if (w[j] > 0)
                    violation = Math.abs(Gp);
                else
                    violation = Math.abs(Gn);

                Gmax_new = Math.max(Gmax_new, violation);
                Gnorm1_new += violation;

                // obtain Newton direction d
                if (Gp < H * w[j])
                    d = -Gp / H;
                else if (Gn > H * w[j])
                    d = -Gn / H;
                else
                    d = -w[j];

                if (Math.abs(d) < 1.0e-12) continue;

                double delta = Math.abs(w[j] + d) - Math.abs(w[j]) + G * d;
                d_old = 0;
                int num_linesearch;
                for (num_linesearch = 0; num_linesearch < max_num_linesearch; num_linesearch++) {
                    d_diff = d_old - d;
                    cond = Math.abs(w[j] + d) - Math.abs(w[j]) - sigma * delta;

                    appxcond = xj_sq[j] * d * d + G_loss * d + cond;
                    if (appxcond <= 0) {
                        for (Feature x : prob_col.x[j]) {
                            b[x.getIndex() - 1] += d_diff * x.getValue();
                        }
                        break;
                    }

                    if (num_linesearch == 0) {
                        loss_old = 0;
                        loss_new = 0;
                        for (Feature x : prob_col.x[j]) {
                            int ind = x.getIndex() - 1;
                            if (b[ind] > 0) {
                                loss_old += C[GETI(y, ind)] * b[ind] * b[ind];
                            }
                            double b_new = b[ind] + d_diff * x.getValue();
                            b[ind] = b_new;
                            if (b_new > 0) {
                                loss_new += C[GETI(y, ind)] * b_new * b_new;
                            }
                        }
                    } else {
                        loss_new = 0;
                        for (Feature x : prob_col.x[j]) {
                            int ind = x.getIndex() - 1;
                            double b_new = b[ind] + d_diff * x.getValue();
                            b[ind] = b_new;
                            if (b_new > 0) {
                                loss_new += C[GETI(y, ind)] * b_new * b_new;
                            }
                        }
                    }

                    cond = cond + loss_new - loss_old;
                    if (cond <= 0)
                        break;
                    else {
                        d_old = d;
                        d *= 0.5;
                        delta *= 0.5;
                    }
                }

                w[j] += d;

                // recompute b[] if line search takes too many steps
                if (num_linesearch >= max_num_linesearch) {
                    info("#");
                    for (int i = 0; i < l; i++)
                        b[i] = 1;

                    for (int i = 0; i < w_size; i++) {
                        if (w[i] == 0) continue;
                        for (Feature x : prob_col.x[i]) {
                            b[x.getIndex() - 1] -= w[i] * x.getValue();
                        }
                    }
                }
            }

            if (iter == 0) {
                Gnorm1_init = Gnorm1_new;
            }
            iter++;
            if (iter % 10 == 0) info(".");

            if (Gmax_new <= eps * Gnorm1_init) {
                if (active_size == w_size)
                    break;
                else {
                    active_size = w_size;
                    info("*");
                    Gmax_old = Double.POSITIVE_INFINITY;
                    continue;
                }
            }

            Gmax_old = Gmax_new;
        }

        info("%noptimization finished, #iter = %d%n", iter);
        if (iter >= max_iter) info("%nWARNING: reaching max number of iterations%n");

        // calculate objective value

        double v = 0;
        int nnz = 0;
        for (j = 0; j < w_size; j++) {
            for (Feature x : prob_col.x[j]) {
                x.setValue(x.getValue() * prob_col.y[x.getIndex() - 1]); // restore x->value
            }
            if (w[j] != 0) {
                v += Math.abs(w[j]);
                nnz++;
            }
        }
        for (j = 0; j < l; j++)
            if (b[j] > 0) v += C[GETI(y, j)] * b[j] * b[j];

        info("Objective value = %g%n", v);
        info("#nonzeros/#features = %d/%d%n", nnz, w_size);
    }

    /**
     * A coordinate descent algorithm for
     * L1-regularized logistic regression problems
     *
     *<pre>
     *  min_w \sum |wj| + C \sum log(1+exp(-yi w^T xi)),
     *
     * Given:
     * x, y, Cp, Cn
     * eps is the stopping tolerance
     *
     * solution will be put in w
     *
     * See Yuan et al. (2011) and appendix of LIBLINEAR paper, Fan et al. (2008)
     *</pre>
     *
     * @since 1.5
     */
    private static void solve_l1r_lr(Problem prob_col, double[] w, double eps, double Cp, double Cn) {
        int l = prob_col.l;
        int w_size = prob_col.n;
        int j, s, newton_iter = 0, iter = 0;
        int max_newton_iter = 100;
        int max_iter = 1000;
        int max_num_linesearch = 20;
        int active_size;
        int QP_active_size;

        double nu = 1e-12;
        double inner_eps = 1;
        double sigma = 0.01;
        double w_norm, w_norm_new;
        double z, G, H;
        double Gnorm1_init = 0; // eclipse moans this variable might not be initialized
        double Gmax_old = Double.POSITIVE_INFINITY;
        double Gmax_new, Gnorm1_new;
        double QP_Gmax_old = Double.POSITIVE_INFINITY;
        double QP_Gmax_new, QP_Gnorm1_new;
        double delta, negsum_xTd, cond;

        int[] index = new int[w_size];
        byte[] y = new byte[l];
        double[] Hdiag = new double[w_size];
        double[] Grad = new double[w_size];
        double[] wpd = new double[w_size];
        double[] xjneg_sum = new double[w_size];
        double[] xTd = new double[l];
        double[] exp_wTx = new double[l];
        double[] exp_wTx_new = new double[l];
        double[] tau = new double[l];
        double[] D = new double[l];

        double[] C = {Cn, 0, Cp};

        // Initial w can be set here.
        for (j = 0; j < w_size; j++)
            w[j] = 0;

        for (j = 0; j < l; j++) {
            if (prob_col.y[j] > 0)
                y[j] = 1;
            else
                y[j] = -1;

            exp_wTx[j] = 0;
        }

        w_norm = 0;
        for (j = 0; j < w_size; j++) {
            w_norm += Math.abs(w[j]);
            wpd[j] = w[j];
            index[j] = j;
            xjneg_sum[j] = 0;
            for (Feature x : prob_col.x[j]) {
                int ind = x.getIndex() - 1;
                double val = x.getValue();
                exp_wTx[ind] += w[j] * val;
                if (y[ind] == -1) {
                    xjneg_sum[j] += C[GETI(y, ind)] * val;
                }
            }
        }
        for (j = 0; j < l; j++) {
            exp_wTx[j] = Math.exp(exp_wTx[j]);
            double tau_tmp = 1 / (1 + exp_wTx[j]);
            tau[j] = C[GETI(y, j)] * tau_tmp;
            D[j] = C[GETI(y, j)] * exp_wTx[j] * tau_tmp * tau_tmp;
        }

        while (newton_iter < max_newton_iter) {
            Gmax_new = 0;
            Gnorm1_new = 0;
            active_size = w_size;

            for (s = 0; s < active_size; s++) {
                j = index[s];
                Hdiag[j] = nu;
                Grad[j] = 0;

                double tmp = 0;
                for (Feature x : prob_col.x[j]) {
                    int ind = x.getIndex() - 1;
                    Hdiag[j] += x.getValue() * x.getValue() * D[ind];
                    tmp += x.getValue() * tau[ind];
                }
                Grad[j] = -tmp + xjneg_sum[j];

                double Gp = Grad[j] + 1;
                double Gn = Grad[j] - 1;
                double violation = 0;
                if (w[j] == 0) {
                    if (Gp < 0)
                        violation = -Gp;
                    else if (Gn > 0)
                        violation = Gn;
                    //outer-level shrinking
                    else if (Gp > Gmax_old / l && Gn < -Gmax_old / l) {
                        active_size--;
                        swap(index, s, active_size);
                        s--;
                        continue;
                    }
                } else if (w[j] > 0)
                    violation = Math.abs(Gp);
                else
                    violation = Math.abs(Gn);

                Gmax_new = Math.max(Gmax_new, violation);
                Gnorm1_new += violation;
            }

            if (newton_iter == 0) Gnorm1_init = Gnorm1_new;

            if (Gnorm1_new <= eps * Gnorm1_init) break;

            iter = 0;
            QP_Gmax_old = Double.POSITIVE_INFINITY;
            QP_active_size = active_size;

            for (int i = 0; i < l; i++)
                xTd[i] = 0;

            // optimize QP over wpd
            while (iter < max_iter) {
                QP_Gmax_new = 0;
                QP_Gnorm1_new = 0;

                for (j = 0; j < QP_active_size; j++) {
                    int i = random.nextInt(QP_active_size - j);
                    swap(index, i, j);
                }

                for (s = 0; s < QP_active_size; s++) {
                    j = index[s];
                    H = Hdiag[j];

                    G = Grad[j] + (wpd[j] - w[j]) * nu;
                    for (Feature x : prob_col.x[j]) {
                        int ind = x.getIndex() - 1;
                        G += x.getValue() * D[ind] * xTd[ind];
                    }

                    double Gp = G + 1;
                    double Gn = G - 1;
                    double violation = 0;
                    if (wpd[j] == 0) {
                        if (Gp < 0)
                            violation = -Gp;
                        else if (Gn > 0)
                            violation = Gn;
                        //inner-level shrinking
                        else if (Gp > QP_Gmax_old / l && Gn < -QP_Gmax_old / l) {
                            QP_active_size--;
                            swap(index, s, QP_active_size);
                            s--;
                            continue;
                        }
                    } else if (wpd[j] > 0)
                        violation = Math.abs(Gp);
                    else
                        violation = Math.abs(Gn);

                    QP_Gmax_new = Math.max(QP_Gmax_new, violation);
                    QP_Gnorm1_new += violation;

                    // obtain solution of one-variable problem
                    if (Gp < H * wpd[j])
                        z = -Gp / H;
                    else if (Gn > H * wpd[j])
                        z = -Gn / H;
                    else
                        z = -wpd[j];

                    if (Math.abs(z) < 1.0e-12) continue;
                    z = Math.min(Math.max(z, -10.0), 10.0);

                    wpd[j] += z;

                    for (Feature x : prob_col.x[j]) {
                        int ind = x.getIndex() - 1;
                        xTd[ind] += x.getValue() * z;
                    }
                }

                iter++;

                if (QP_Gnorm1_new <= inner_eps * Gnorm1_init) {
                    //inner stopping
                    if (QP_active_size == active_size)
                        break;
                    //active set reactivation
                    else {
                        QP_active_size = active_size;
                        QP_Gmax_old = Double.POSITIVE_INFINITY;
                        continue;
                    }
                }

                QP_Gmax_old = QP_Gmax_new;
            }

            if (iter >= max_iter) info("WARNING: reaching max number of inner iterations%n");

            delta = 0;
            w_norm_new = 0;
            for (j = 0; j < w_size; j++) {
                delta += Grad[j] * (wpd[j] - w[j]);
                if (wpd[j] != 0) w_norm_new += Math.abs(wpd[j]);
            }
            delta += (w_norm_new - w_norm);

            negsum_xTd = 0;
            for (int i = 0; i < l; i++)
                if (y[i] == -1) negsum_xTd += C[GETI(y, i)] * xTd[i];

            int num_linesearch;
            for (num_linesearch = 0; num_linesearch < max_num_linesearch; num_linesearch++) {
                cond = w_norm_new - w_norm + negsum_xTd - sigma * delta;

                for (int i = 0; i < l; i++) {
                    double exp_xTd = Math.exp(xTd[i]);
                    exp_wTx_new[i] = exp_wTx[i] * exp_xTd;
                    cond += C[GETI(y, i)] * Math.log((1 + exp_wTx_new[i]) / (exp_xTd + exp_wTx_new[i]));
                }

                if (cond <= 0) {
                    w_norm = w_norm_new;
                    for (j = 0; j < w_size; j++)
                        w[j] = wpd[j];
                    for (int i = 0; i < l; i++) {
                        exp_wTx[i] = exp_wTx_new[i];
                        double tau_tmp = 1 / (1 + exp_wTx[i]);
                        tau[i] = C[GETI(y, i)] * tau_tmp;
                        D[i] = C[GETI(y, i)] * exp_wTx[i] * tau_tmp * tau_tmp;
                    }
                    break;
                } else {
                    w_norm_new = 0;
                    for (j = 0; j < w_size; j++) {
                        wpd[j] = (w[j] + wpd[j]) * 0.5;
                        if (wpd[j] != 0) w_norm_new += Math.abs(wpd[j]);
                    }
                    delta *= 0.5;
                    negsum_xTd *= 0.5;
                    for (int i = 0; i < l; i++)
                        xTd[i] *= 0.5;
                }
            }

            // Recompute some info due to too many line search steps
            if (num_linesearch >= max_num_linesearch) {
                for (int i = 0; i < l; i++)
                    exp_wTx[i] = 0;

                for (int i = 0; i < w_size; i++) {
                    if (w[i] == 0) continue;
                    for (Feature x : prob_col.x[i]) {
                        exp_wTx[x.getIndex() - 1] += w[i] * x.getValue();
                    }
                }

                for (int i = 0; i < l; i++)
                    exp_wTx[i] = Math.exp(exp_wTx[i]);
            }

            if (iter == 1) inner_eps *= 0.25;

            newton_iter++;
            Gmax_old = Gmax_new;

            info("iter %3d  #CD cycles %d%n", newton_iter, iter);
        }

        info("=========================%n");
        info("optimization finished, #iter = %d%n", newton_iter);
        if (newton_iter >= max_newton_iter) info("WARNING: reaching max number of iterations%n");

        // calculate objective value

        double v = 0;
        int nnz = 0;
        for (j = 0; j < w_size; j++)
            if (w[j] != 0) {
                v += Math.abs(w[j]);
                nnz++;
            }
        for (j = 0; j < l; j++)
            if (y[j] == 1)
                v += C[GETI(y, j)] * Math.log(1 + 1 / exp_wTx[j]);
            else
                v += C[GETI(y, j)] * Math.log(1 + exp_wTx[j]);

        info("Objective value = %g%n", v);
        info("#nonzeros/#features = %d/%d%n", nnz, w_size);
    }

    // transpose matrix X from row format to column format
    static Problem transpose(Problem prob) {
        int l = prob.l;
        int n = prob.n;
        int[] col_ptr = new int[n + 1];
        Problem prob_col = new Problem();
        prob_col.l = l;
        prob_col.n = n;
        prob_col.y = new double[l];
        prob_col.x = new Feature[n][];

        for (int i = 0; i < l; i++)
            prob_col.y[i] = prob.y[i];

        for (int i = 0; i < l; i++) {
            for (Feature x : prob.x[i]) {
                col_ptr[x.getIndex()]++;
            }
        }

        for (int i = 0; i < n; i++) {
            prob_col.x[i] = new Feature[col_ptr[i + 1]];
            col_ptr[i] = 0; // reuse the array to count the nr of elements
        }

        for (int i = 0; i < l; i++) {
            for (int j = 0; j < prob.x[i].length; j++) {
                Feature x = prob.x[i][j];
                int index = x.getIndex() - 1;
                prob_col.x[index][col_ptr[index]] = new FeatureNode(i + 1, x.getValue());
                col_ptr[index]++;
            }
        }

        return prob_col;
    }

    static void swap(double[] array, int idxA, int idxB) {
        double temp = array[idxA];
        array[idxA] = array[idxB];
        array[idxB] = temp;
    }

    static void swap(int[] array, int idxA, int idxB) {
        int temp = array[idxA];
        array[idxA] = array[idxB];
        array[idxB] = temp;
    }

    static void swap(IntArrayPointer array, int idxA, int idxB) {
        int temp = array.get(idxA);
        array.set(idxA, array.get(idxB));
        array.set(idxB, temp);
    }


    /**
     * @throws IllegalArgumentException if the feature nodes of prob are not sorted in ascending order
     */
    public static Model train(Problem prob, Parameter param) {

        if (prob == null) throw new IllegalArgumentException("problem must not be null");
        if (param == null) throw new IllegalArgumentException("parameter must not be null");

        if (prob.n == 0) throw new IllegalArgumentException("problem has zero features");
        if (prob.l == 0) throw new IllegalArgumentException("problem has zero instances");

        for (Feature[] nodes : prob.x) {
            int indexBefore = 0;
            for (Feature n : nodes) {
                if (n.getIndex() <= indexBefore) {
                    throw new IllegalArgumentException("feature nodes must be sorted by index in ascending order");
                }
                indexBefore = n.getIndex();
            }
        }

        int l = prob.l;
        int n = prob.n;
        int w_size = prob.n;
        Model model = new Model();

        if (prob.bias >= 0)
            model.nr_feature = n - 1;
        else
            model.nr_feature = n;

        model.solverType = param.solverType;
        model.bias = prob.bias;

        if (param.solverType == SolverType.L2R_L2LOSS_SVR || //
            param.solverType == SolverType.L2R_L1LOSS_SVR_DUAL || //
            param.solverType == SolverType.L2R_L2LOSS_SVR_DUAL) {
            model.w = new double[w_size];
            model.nr_class = 2;
            model.label = null;

            checkProblemSize(n, model.nr_class);

            train_one(prob, param, model.w, 0, 0);
        } else {
            int[] perm = new int[l];

            // group training data of the same class
            GroupClassesReturn rv = groupClasses(prob, perm);
            int nr_class = rv.nr_class;
            int[] label = rv.label;
            int[] start = rv.start;
            int[] count = rv.count;

            checkProblemSize(n, nr_class);

            model.nr_class = nr_class;
            model.label = new int[nr_class];
            for (int i = 0; i < nr_class; i++)
                model.label[i] = label[i];

            // calculate weighted C
            double[] weighted_C = new double[nr_class];
            for (int i = 0; i < nr_class; i++)
                weighted_C[i] = param.C;
            for (int i = 0; i < param.getNumWeights(); i++) {
                int j;
                for (j = 0; j < nr_class; j++)
                    if (param.weightLabel[i] == label[j]) break;

                if (j == nr_class) throw new IllegalArgumentException("class label " + param.weightLabel[i] + " specified in weight is not found");
                weighted_C[j] *= param.weight[i];
            }

            // constructing the subproblem
            Feature[][] x = new Feature[l][];
            for (int i = 0; i < l; i++)
                x[i] = prob.x[perm[i]];

            Problem sub_prob = new Problem();
            sub_prob.l = l;
            sub_prob.n = n;
            sub_prob.x = new Feature[sub_prob.l][];
            sub_prob.y = new double[sub_prob.l];

            for (int k = 0; k < sub_prob.l; k++)
                sub_prob.x[k] = x[k];

            // multi-class svm by Crammer and Singer
            if (param.solverType == SolverType.MCSVM_CS) {
                model.w = new double[n * nr_class];
                for (int i = 0; i < nr_class; i++) {
                    for (int j = start[i]; j < start[i] + count[i]; j++) {
                        sub_prob.y[j] = i;
                    }
                }

                SolverMCSVM_CS solver = new SolverMCSVM_CS(sub_prob, nr_class, weighted_C, param.eps);
                solver.solve(model.w);
            } else {
                if (nr_class == 2) {
                    model.w = new double[w_size];

                    int e0 = start[0] + count[0];
                    int k = 0;
                    for (; k < e0; k++)
                        sub_prob.y[k] = +1;
                    for (; k < sub_prob.l; k++)
                        sub_prob.y[k] = -1;

                    train_one(sub_prob, param, model.w, weighted_C[0], weighted_C[1]);
                } else {
                    model.w = new double[w_size * nr_class];
                    double[] w = new double[w_size];
                    for (int i = 0; i < nr_class; i++) {
                        int si = start[i];
                        int ei = si + count[i];

                        int k = 0;
                        for (; k < si; k++)
                            sub_prob.y[k] = -1;
                        for (; k < ei; k++)
                            sub_prob.y[k] = +1;
                        for (; k < sub_prob.l; k++)
                            sub_prob.y[k] = -1;

                        train_one(sub_prob, param, w, weighted_C[i], param.C);

                        for (int j = 0; j < n; j++)
                            model.w[j * nr_class + i] = w[j];
                    }
                }
            }
        }
        return model;
    }

    /**
     * verify the size and throw an exception early if the problem is too large
     */
    private static void checkProblemSize(int n, int nr_class) {
        if (n >= Integer.MAX_VALUE / nr_class || n * nr_class < 0) {
            throw new IllegalArgumentException("'number of classes' * 'number of instances' is too large: " + nr_class + "*" + n);
        }
    }

    private static void train_one(Problem prob, Parameter param, double[] w, double Cp, double Cn) {
        double eps = param.eps;
        int pos = 0;
        for (int i = 0; i < prob.l; i++)
            if (prob.y[i] > 0) {
                pos++;
            }
        int neg = prob.l - pos;

        double primal_solver_tol = eps * Math.max(Math.min(pos, neg), 1) / prob.l;

        Function fun_obj = null;
        switch (param.solverType) {
            case L2R_LR: {
                double[] C = new double[prob.l];
                for (int i = 0; i < prob.l; i++) {
                    if (prob.y[i] > 0)
                        C[i] = Cp;
                    else
                        C[i] = Cn;
                }
                fun_obj = new L2R_LrFunction(prob, C);
                Tron tron_obj = new Tron(fun_obj, primal_solver_tol);
                tron_obj.tron(w);
                break;
            }
            case L2R_L2LOSS_SVC: {
                double[] C = new double[prob.l];
                for (int i = 0; i < prob.l; i++) {
                    if (prob.y[i] > 0)
                        C[i] = Cp;
                    else
                        C[i] = Cn;
                }
                fun_obj = new L2R_L2_SvcFunction(prob, C);
                Tron tron_obj = new Tron(fun_obj, primal_solver_tol);
                tron_obj.tron(w);
                break;
            }
            case L2R_L2LOSS_SVC_DUAL:
                solve_l2r_l1l2_svc(prob, w, eps, Cp, Cn, SolverType.L2R_L2LOSS_SVC_DUAL);
                break;
            case L2R_L1LOSS_SVC_DUAL:
                solve_l2r_l1l2_svc(prob, w, eps, Cp, Cn, SolverType.L2R_L1LOSS_SVC_DUAL);
                break;
            case L1R_L2LOSS_SVC: {
                Problem prob_col = transpose(prob);
                solve_l1r_l2_svc(prob_col, w, primal_solver_tol, Cp, Cn);
                break;
            }
            case L1R_LR: {
                Problem prob_col = transpose(prob);
                solve_l1r_lr(prob_col, w, primal_solver_tol, Cp, Cn);
                break;
            }
            case L2R_LR_DUAL:
                solve_l2r_lr_dual(prob, w, eps, Cp, Cn);
                break;
            case L2R_L2LOSS_SVR: {
                double[] C = new double[prob.l];
                for (int i = 0; i < prob.l; i++)
                    C[i] = param.C;

                fun_obj = new L2R_L2_SvrFunction(prob, C, param.p);
                Tron tron_obj = new Tron(fun_obj, param.eps);
                tron_obj.tron(w);
                break;
            }
            case L2R_L1LOSS_SVR_DUAL:
            case L2R_L2LOSS_SVR_DUAL:
                solve_l2r_l1l2_svr(prob, w, param);
                break;

            default:
                throw new IllegalStateException("unknown solver type: " + param.solverType);
        }
    }

    public static void disableDebugOutput() {
        setDebugOutput(null);
    }

    public static void enableDebugOutput() {
        setDebugOutput(System.out);
    }

    public static void setDebugOutput(PrintStream debugOutput) {
        synchronized (OUTPUT_MUTEX) {
            DEBUG_OUTPUT = debugOutput;
        }
    }

    /**
     * resets the PRNG
     *
     * this is i.a. needed for regression testing (eg. the Weka wrapper)
     */
    public static void resetRandom() {
        random = new Random(DEFAULT_RANDOM_SEED);
    }
}
TOP

Related Classes of de.bwaldvogel.liblinear.Linear$GroupClassesReturn

TOP
Copyright © 2018 www.massapi.com. All rights reserved.
All source code are property of their respective owners. Java is a trademark of Sun Microsystems, Inc and owned by ORACLE Inc. Contact coftware#gmail.com.