Examples with PriorProbClassifier edu.neu.ccs.pyramid.classification.PriorProbClassifier used on opensource projects

Example 6 with PriorProbClassifier

use of edu.neu.ccs.pyramid.classification.PriorProbClassifier in project pyramid by cheng-li.

the class AbstractRecoverCBMOptimizer method skipOrUpdateBinaryClassifier.

protected void skipOrUpdateBinaryClassifier(int component, int label, MultiLabelClfDataSet activeDataSet, double[] activeGammas, double totalWeight) {
    StopWatch stopWatch = new StopWatch();
    stopWatch.start();
    double effectivePositives = effectivePositives(component, label);
    double nonSmoothedPositiveProb = effectivePositives / totalWeight;
    // smooth the component-wise label fraction with global label fraction
    double positiveCount = labelMatrix.getColumn(label).getNumNonZeroElements();
    double smoothedPositiveProb = (effectivePositives + smoothingStrength * positiveCount) / (totalWeight + smoothingStrength * groundTruth.getNumDataPoints());
    StringBuilder sb = new StringBuilder();
    sb.append("for component ").append(component).append(", label ").append(label);
    sb.append(", weighted positives = ").append(effectivePositives);
    sb.append(", non-smoothed positive fraction = " + (effectivePositives / totalWeight));
    sb.append(", global positive fraction = " + (positiveCount / groundTruth.getNumDataPoints()));
    sb.append(", smoothed positive fraction = " + smoothedPositiveProb);
    // it be happen that p >1 for numerical reasons
    if (smoothedPositiveProb >= 1) {
        smoothedPositiveProb = 1;
    }
    if (nonSmoothedPositiveProb < skipLabelThreshold || nonSmoothedPositiveProb > 1 - skipLabelThreshold) {
        double[] probs = { 1 - smoothedPositiveProb, smoothedPositiveProb };
        cbm.binaryClassifiers[component][label] = new PriorProbClassifier(probs);
        sb.append(", skip, use prior = ").append(smoothedPositiveProb);
        sb.append(", time spent = ").append(stopWatch.toString());
        if (logger.isDebugEnabled()) {
            logger.debug(sb.toString());
        }
        return;
    }
    if (logger.isDebugEnabled()) {
        logger.debug(sb.toString());
    }
    updateBinaryClassifier(component, label, activeDataSet, activeGammas);
}

Example 7 with PriorProbClassifier

use of edu.neu.ccs.pyramid.classification.PriorProbClassifier in project pyramid by cheng-li.

the class LKBoostOptimizer method addPriors.

@Override
protected void addPriors() {
    PriorProbClassifier priorProbClassifier = new PriorProbClassifier(numClasses);
    priorProbClassifier.fit(dataSet, targetDistribution, weights);
    double[] probs = priorProbClassifier.getClassProbs();
    double[] scores = MathUtil.inverseSoftMax(probs);
    // weaken the priors
    for (int i = 0; i < scores.length; i++) {
        if (scores[i] > 5) {
            scores[i] = 5;
        }
        if (scores[i] < -5) {
            scores[i] = -5;
        }
    }
    for (int k = 0; k < numClasses; k++) {
        Regressor constant = new ConstantRegressor(scores[k]);
        boosting.getEnsemble(k).add(constant);
    }
}

Example 8 with PriorProbClassifier

use of edu.neu.ccs.pyramid.classification.PriorProbClassifier in project pyramid by cheng-li.

the class GBCBMOptimizer method updateBinaryClassifier.

@Override
protected void updateBinaryClassifier(int component, int label, MultiLabelClfDataSet activeDataset, double[] activeGammas) {
    StopWatch stopWatch = new StopWatch();
    stopWatch.start();
    if (cbm.binaryClassifiers[component][label] == null || cbm.binaryClassifiers[component][label] instanceof PriorProbClassifier) {
        cbm.binaryClassifiers[component][label] = new LKBoost(2);
    }
    int[] binaryLabels = DataSetUtil.toBinaryLabels(activeDataset.getMultiLabels(), label);
    double[][] targetsDistributions = DataSetUtil.labelsToDistributions(binaryLabels, 2);
    LKBoost boost = (LKBoost) this.cbm.binaryClassifiers[component][label];
    RegTreeConfig regTreeConfig = new RegTreeConfig().setMaxNumLeaves(numLeaves);
    RegTreeFactory regTreeFactory = new RegTreeFactory(regTreeConfig);
    regTreeFactory.setLeafOutputCalculator(new LKBOutputCalculator(2));
    LKBoostOptimizer optimizer = new LKBoostOptimizer(boost, activeDataset, regTreeFactory, activeGammas, targetsDistributions);
    optimizer.setShrinkage(shrinkage);
    optimizer.initialize();
    optimizer.iterate(binaryUpdatesPerIter);
    if (logger.isDebugEnabled()) {
        logger.debug("time spent on updating component " + component + " label " + label + " = " + stopWatch);
    }
}

Example 9 with PriorProbClassifier

use of edu.neu.ccs.pyramid.classification.PriorProbClassifier in project pyramid by cheng-li.

the class SparseCBMOptimzer method updateBinaryLogisticRegression.

private void updateBinaryLogisticRegression(int componentIndex, int labelIndex) {
    StopWatch stopWatch = new StopWatch();
    stopWatch.start();
    double effectivePositives = effectivePositives(componentIndex, labelIndex);
    StringBuilder sb = new StringBuilder();
    sb.append("for component ").append(componentIndex).append(", label ").append(labelIndex);
    sb.append(", effective positives = ").append(effectivePositives);
    if (effectivePositives <= 1) {
        double positiveProb = prior(componentIndex, labelIndex);
        double[] probs = { 1 - positiveProb, positiveProb };
        cbm.binaryClassifiers[componentIndex][labelIndex] = new PriorProbClassifier(probs);
        sb.append(", skip, use prior = ").append(positiveProb);
        sb.append(", time spent = " + stopWatch.toString());
        System.out.println(sb.toString());
        return;
    }
    if (cbm.binaryClassifiers[componentIndex][labelIndex] == null || cbm.binaryClassifiers[componentIndex][labelIndex] instanceof PriorProbClassifier) {
        cbm.binaryClassifiers[componentIndex][labelIndex] = new LogisticRegression(2, dataSet.getNumFeatures());
    }
    RidgeLogisticOptimizer ridgeLogisticOptimizer;
    int[] binaryLabels = DataSetUtil.toBinaryLabels(dataSet.getMultiLabels(), labelIndex);
    // no parallelism
    ridgeLogisticOptimizer = new RidgeLogisticOptimizer((LogisticRegression) cbm.binaryClassifiers[componentIndex][labelIndex], dataSet, binaryLabels, activeGammas, priorVarianceBinary, false);
    // TODO maximum iterations
    ridgeLogisticOptimizer.getOptimizer().getTerminator().setMaxIteration(numBinaryUpdates);
    ridgeLogisticOptimizer.optimize();
    sb.append(", time spent = " + stopWatch.toString());
    System.out.println(sb.toString());
}

Example 10 with PriorProbClassifier

use of edu.neu.ccs.pyramid.classification.PriorProbClassifier in project pyramid by cheng-li.

the class RobustLRCBMOptimizer method updateBinaryClassifier.

@Override
protected void updateBinaryClassifier(int component, int label, MultiLabelClfDataSet activeDataset, double[] activeInstanceWeights) {
    StopWatch stopWatch = new StopWatch();
    stopWatch.start();
    if (cbm.binaryClassifiers[component][label] == null || cbm.binaryClassifiers[component][label] instanceof PriorProbClassifier) {
        cbm.binaryClassifiers[component][label] = new LogisticRegression(2, activeDataset.getNumFeatures());
    }
    RidgeLogisticOptimizer ridgeLogisticOptimizer;
    int[] binaryLabels = DataSetUtil.toBinaryLabels(activeDataset.getMultiLabels(), label);
    // no parallelism
    ridgeLogisticOptimizer = new RidgeLogisticOptimizer((LogisticRegression) cbm.binaryClassifiers[component][label], activeDataset, binaryLabels, activeInstanceWeights, priorVarianceBinary, false);
    ridgeLogisticOptimizer.getOptimizer().getTerminator().setMaxIteration(binaryUpdatesPerIter);
    ridgeLogisticOptimizer.optimize();
    if (logger.isDebugEnabled()) {
        logger.debug("time spent on updating component " + component + " label " + label + " = " + stopWatch);
    }
}