Examples with DenseVector org.apache.mahout.math.DenseVector used on opensource projects

Example 26 with DenseVector

use of org.apache.mahout.math.DenseVector in project pyramid by cheng-li.

the class DBR method predict.

@Override
public MultiLabel predict(Vector vector) {
    Vector stage1Predictions = stage1BR.predict(vector).toVectorRandomSparse(stage1BR.getNumClasses());
    Vector stage2input;
    if (useXStage2) {
        stage2input = Vectors.concatenate(stage1Predictions, vector);
    } else {
        stage2input = new DenseVector(stage1Predictions);
    }
    return stage2BR.predict(stage2input);
}

Example 27 with DenseVector

use of org.apache.mahout.math.DenseVector in project pyramid by cheng-li.

the class BackTrackingLineSearcher method moveAlongDirection.

/**
 * move to a new position along the direction
 */
public MoveInfo moveAlongDirection(Vector searchDirection) {
    Vector localSearchDir;
    if (logger.isDebugEnabled()) {
        logger.debug("start line search");
        // don't want to show too much; only show it on small problems
        if (searchDirection.size() < 100) {
            logger.debug("direction=" + searchDirection);
        }
    }
    MoveInfo moveInfo = new MoveInfo();
    double stepLength = initialStepLength;
    double value = function.getValue();
    moveInfo.setOldValue(value);
    Vector gradient = function.getGradient();
    double product = gradient.dot(searchDirection);
    if (product < 0) {
        localSearchDir = searchDirection;
    } else {
        if (logger.isWarnEnabled()) {
            logger.warn("Bad search direction! Use negative gradient instead. Product of gradient and search direction = " + product);
        }
        localSearchDir = gradient.times(-1);
    }
    Vector initialPosition;
    // keep a copy of initial parameters
    if (function.getParameters().isDense()) {
        initialPosition = new DenseVector(function.getParameters());
    } else {
        initialPosition = new RandomAccessSparseVector(function.getParameters());
    }
    while (true) {
        Vector step = localSearchDir.times(stepLength);
        Vector target = initialPosition.plus(step);
        function.setParameters(target);
        double targetValue = function.getValue();
        if (logger.isDebugEnabled()) {
            logger.debug("step length = " + stepLength + ", target value = " + targetValue);
        // logger.debug("requirement = "+(value + c*stepLength*product));
        }
        // todo: if equal ok?
        if ((targetValue <= value + c * stepLength * product && value < Double.POSITIVE_INFINITY) || stepLength == 0) {
            moveInfo.setStep(step);
            moveInfo.setStepLength(stepLength);
            moveInfo.setNewValue(targetValue);
            break;
        }
        stepLength *= shrinkage;
    }
    if (logger.isDebugEnabled()) {
        logger.debug("line search done. " + moveInfo);
    }
    return moveInfo;
}

Example 28 with DenseVector

use of org.apache.mahout.math.DenseVector in project pyramid by cheng-li.

the class SupervisedEmbeddingLoss method getGradient.

public Vector getGradient() {
    int numData = this.updatedEmbeddingMatrix.getNumDataPoints();
    int numFeatures = this.updatedEmbeddingMatrix.getNumFeatures();
    int vecSize = numData * numFeatures;
    Vector finalGradient = new DenseVector(vecSize);
    for (int i = 0; i < numData; i++) {
        Vector gradient = new DenseVector(numFeatures);
        Vector q_i = this.updatedEmbeddingMatrix.getRow(i);
        Vector q_i_orig = this.embeddingMatrix.getRow(i);
        gradient = gradient.plus(q_i.minus(q_i_orig).times(2.0 * this.alpha));
        for (int j = 0; j < numData; j++) {
            Vector q_j = this.updatedEmbeddingMatrix.getRow(j);
            double pi_x = this.projMatrix.getColumn(0).dot(q_i);
            double pi_y = this.projMatrix.getColumn(1).dot(q_i);
            double pj_x = this.projMatrix.getColumn(0).dot(q_j);
            double pj_y = this.projMatrix.getColumn(1).dot(q_j);
            double p_sq = (pi_x - pj_x) * (pi_x - pj_x) + (pi_y - pj_y) * (pi_y - pj_y);
            double d_sq = this.distMatrix.getRow(i).get(j) * this.distMatrix.getRow(i).get(j);
            Vector p_dist_vec = new DenseVector(2);
            p_dist_vec.set(0, pi_x - pj_x);
            p_dist_vec.set(1, pi_y - pj_y);
            Vector tmp = new DenseVector(this.projMatrix.getNumDataPoints());
            for (int k = 0; k < this.projMatrix.getNumDataPoints(); k++) {
                tmp.set(k, this.projMatrix.getRow(k).dot(p_dist_vec));
            }
            gradient = gradient.plus(tmp.times(4.0 * this.beta * (p_sq - d_sq)));
        }
        for (int j = 0; j < numFeatures; j++) {
            finalGradient.set(i * numFeatures + j, gradient.get(j));
        }
    }
    return finalGradient;
}

Example 29 with DenseVector

use of org.apache.mahout.math.DenseVector in project pyramid by cheng-li.

the class PlattScaling method transform.

public double transform(double uncalibrated) {
    Vector vector = new DenseVector(1);
    vector.set(0, uncalibrated);
    return logisticRegression.predictClassProb(vector, 1);
}

Example 30 with DenseVector

use of org.apache.mahout.math.DenseVector in project pyramid by cheng-li.

the class KMeans method updateCenters.

private void updateCenters(int k) {
    Vector center = new DenseVector(dataSet.getNumFeatures());
    double count = 0;
    for (int i = 0; i < dataSet.getNumDataPoints(); i++) {
        if (assignments[i] == k) {
            Vector instance = dataSet.getRow(i);
            for (int j = 0; j < instance.size(); j++) {
                center.set(j, center.get(j) + instance.get(j));
            }
            count += 1;
        }
    }
    center = center.divide(count);
    centers[k] = center;
    System.out.println("update the centroid of cluster " + (k + 1) + " based on " + (int) count + " instances in the cluster");
}