Examples with Duration de.lmu.ifi.dbs.elki.logging.statistics.Duration used on opensource projects

Example 1 with Duration

use of de.lmu.ifi.dbs.elki.logging.statistics.Duration in project elki by elki-project.

the class PerplexityAffinityMatrixBuilder method computePij.

/**
 * Compute the pij from the distance matrix.
 *
 * @param dist Distance matrix.
 * @param perplexity Desired perplexity
 * @param initialScale Initial scale
 * @return Affinity matrix pij
 */
protected static double[][] computePij(double[][] dist, double perplexity, double initialScale) {
    final int size = dist.length;
    final double logPerp = FastMath.log(perplexity);
    double[][] pij = new double[size][size];
    FiniteProgress prog = LOG.isVerbose() ? new FiniteProgress("Optimizing perplexities", size, LOG) : null;
    Duration timer = LOG.isStatistics() ? LOG.newDuration(PerplexityAffinityMatrixBuilder.class.getName() + ".runtime.pijmatrix").begin() : null;
    MeanVariance mv = LOG.isStatistics() ? new MeanVariance() : null;
    for (int i = 0; i < size; i++) {
        double beta = computePi(i, dist[i], pij[i], perplexity, logPerp);
        if (mv != null) {
            // Sigma
            mv.put(beta > 0 ? FastMath.sqrt(.5 / beta) : 0.);
        }
        LOG.incrementProcessed(prog);
    }
    LOG.ensureCompleted(prog);
    if (LOG.isStatistics()) {
        // timer != null, mv != null
        LOG.statistics(timer.end());
        LOG.statistics(new DoubleStatistic(PerplexityAffinityMatrixBuilder.class.getName() + ".sigma.average", mv.getMean()));
        LOG.statistics(new DoubleStatistic(PerplexityAffinityMatrixBuilder.class.getName() + ".sigma.stddev", mv.getSampleStddev()));
    }
    // Scale pij to have the desired sum EARLY_EXAGGERATION
    double sum = 0.;
    for (int i = 1; i < size; i++) {
        final double[] pij_i = pij[i];
        for (int j = 0; j < i; j++) {
            // Nur über halbe Matrix!
            // Symmetrie herstellen
            sum += (pij_i[j] += pij[j][i]);
        }
    }
    // Scaling taken from original tSNE code:
    final double scale = initialScale / (2. * sum);
    for (int i = 1; i < size; i++) {
        final double[] pij_i = pij[i];
        for (int j = 0; j < i; j++) {
            pij_i[j] = pij[j][i] = MathUtil.max(pij_i[j] * scale, MIN_PIJ);
        }
    }
    return pij;
}

Example 2 with Duration

use of de.lmu.ifi.dbs.elki.logging.statistics.Duration in project elki by elki-project.

the class SNE method optimizeSNE.

/**
 * Perform the actual tSNE optimization.
 *
 * @param pij Initial affinity matrix
 * @param sol Solution output array (preinitialized)
 */
protected void optimizeSNE(AffinityMatrix pij, double[][] sol) {
    final int size = pij.size();
    if (size * 3L * dim > 0x7FFF_FFFAL) {
        throw new AbortException("Memory exceeds Java array size limit.");
    }
    // Meta information on each point; joined for memory locality.
    // Gradient, Momentum, and learning rate
    // For performance, we use a flat memory layout!
    double[] meta = new double[size * 3 * dim];
    final int dim3 = dim * 3;
    for (int off = 2 * dim; off < meta.length; off += dim3) {
        // Initial learning rate
        Arrays.fill(meta, off, off + dim, 1.);
    }
    // Affinity matrix in projected space
    double[][] qij = new double[size][size];
    FiniteProgress prog = LOG.isVerbose() ? new FiniteProgress("Iterative Optimization", iterations, LOG) : null;
    Duration timer = LOG.isStatistics() ? LOG.newDuration(this.getClass().getName() + ".runtime.optimization").begin() : null;
    // Optimize
    for (int it = 0; it < iterations; it++) {
        double qij_sum = computeQij(qij, sol);
        computeGradient(pij, qij, qij_sum, sol, meta);
        updateSolution(sol, meta, it);
        LOG.incrementProcessed(prog);
    }
    LOG.ensureCompleted(prog);
    if (timer != null) {
        LOG.statistics(timer.end());
    }
}

Example 3 with Duration

use of de.lmu.ifi.dbs.elki.logging.statistics.Duration in project elki by elki-project.

the class TSNE method optimizetSNE.

/**
 * Perform the actual tSNE optimization.
 *
 * @param pij Initial affinity matrix
 * @param sol Solution output array (preinitialized)
 */
protected void optimizetSNE(AffinityMatrix pij, double[][] sol) {
    final int size = pij.size();
    if (size * 3L * dim > 0x7FFF_FFFAL) {
        throw new AbortException("Memory exceeds Java array size limit.");
    }
    // Meta information on each point; joined for memory locality.
    // Gradient, Momentum, and learning rate
    // For performance, we use a flat memory layout!
    double[] meta = new double[size * 3 * dim];
    final int dim3 = dim * 3;
    for (int off = 2 * dim; off < meta.length; off += dim3) {
        // Initial learning rate
        Arrays.fill(meta, off, off + dim, 1.);
    }
    // Affinity matrix in projected space
    double[][] qij = new double[size][size];
    FiniteProgress prog = LOG.isVerbose() ? new FiniteProgress("Iterative Optimization", iterations, LOG) : null;
    Duration timer = LOG.isStatistics() ? LOG.newDuration(this.getClass().getName() + ".runtime.optimization").begin() : null;
    // Optimize
    for (int it = 0; it < iterations; it++) {
        double qij_sum = computeQij(qij, sol);
        computeGradient(pij, qij, qij_sum, sol, meta);
        updateSolution(sol, meta, it);
        // Undo early exaggeration
        if (it == EARLY_EXAGGERATION_ITERATIONS) {
            pij.scale(1. / EARLY_EXAGGERATION);
        }
        LOG.incrementProcessed(prog);
    }
    LOG.ensureCompleted(prog);
    if (timer != null) {
        LOG.statistics(timer.end());
    }
}

Example 4 with Duration

use of de.lmu.ifi.dbs.elki.logging.statistics.Duration in project elki by elki-project.

the class BarnesHutTSNE method optimizetSNE.

/**
 * Perform the actual tSNE optimization.
 *
 * @param pij Sparse initial affinity matrix
 * @param sol Solution output array (preinitialized)
 */
protected void optimizetSNE(AffinityMatrix pij, double[][] sol) {
    final int size = pij.size();
    if (size * 3L * dim > 0x7FFF_FFFAL) {
        throw new AbortException("Memory exceeds Java array size limit.");
    }
    // Meta information on each point; joined for memory locality.
    // Gradient, Momentum, and learning rate
    // For performance, we use a flat memory layout!
    double[] meta = new double[size * 3 * dim];
    final int dim3 = dim * 3;
    for (int off = 2 * dim; off < meta.length; off += dim3) {
        // Initial learning rate
        Arrays.fill(meta, off, off + dim, 1.);
    }
    FiniteProgress prog = LOG.isVerbose() ? new FiniteProgress("Iterative Optimization", iterations, LOG) : null;
    Duration timer = LOG.isStatistics() ? LOG.newDuration(this.getClass().getName() + ".runtime.optimization").begin() : null;
    // Optimize
    for (int i = 0; i < iterations; i++) {
        computeGradient(pij, sol, meta);
        updateSolution(sol, meta, i);
        // Undo early exaggeration
        if (i == EARLY_EXAGGERATION_ITERATIONS) {
            pij.scale(1. / EARLY_EXAGGERATION);
        }
        LOG.incrementProcessed(prog);
    }
    LOG.ensureCompleted(prog);
    if (timer != null) {
        LOG.statistics(timer.end());
    }
}

Example 5 with Duration

use of de.lmu.ifi.dbs.elki.logging.statistics.Duration in project elki by elki-project.

the class GaussianAffinityMatrixBuilder method computePij.

/**
 * Compute the pij from the distance matrix.
 *
 * @param dist Distance matrix.
 * @param sigma Kernel bandwidth sigma
 * @param initialScale Initial scale
 * @return Affinity matrix pij
 */
protected static double[][] computePij(double[][] dist, double sigma, double initialScale) {
    final int size = dist.length;
    final double msigmasq = -.5 / (sigma * sigma);
    double[][] pij = new double[size][size];
    FiniteProgress prog = LOG.isVerbose() ? new FiniteProgress("Computing affinities", size, LOG) : null;
    Duration timer = LOG.isStatistics() ? LOG.newDuration(GaussianAffinityMatrixBuilder.class.getName() + ".runtime.pijmatrix").begin() : null;
    MeanVariance mv = LOG.isStatistics() ? new MeanVariance() : null;
    for (int i = 0; i < size; i++) {
        double logP = computeH(i, dist[i], pij[i], msigmasq);
        if (mv != null) {
            mv.put(FastMath.exp(logP));
        }
        LOG.incrementProcessed(prog);
    }
    LOG.ensureCompleted(prog);
    if (LOG.isStatistics()) {
        // timer != null, mv != null
        LOG.statistics(timer.end());
        LOG.statistics(new DoubleStatistic(GaussianAffinityMatrixBuilder.class.getName() + ".perplexity.average", mv.getMean()));
        LOG.statistics(new DoubleStatistic(GaussianAffinityMatrixBuilder.class.getName() + ".perplexity.stddev", mv.getSampleStddev()));
    }
    // Scale pij to have the desired sum EARLY_EXAGGERATION
    double sum = 0.;
    for (int i = 1; i < size; i++) {
        final double[] pij_i = pij[i];
        for (int j = 0; j < i; j++) {
            // Nur über halbe Matrix!
            // Symmetrie herstellen
            sum += (pij_i[j] += pij[j][i]);
        }
    }
    // Scaling taken from original tSNE code:
    final double scale = initialScale / (2. * sum);
    for (int i = 1; i < size; i++) {
        final double[] pij_i = pij[i];
        for (int j = 0; j < i; j++) {
            pij_i[j] = pij[j][i] = MathUtil.max(pij_i[j] * scale, MIN_PIJ);
        }
    }
    return pij;
}