Examples with DynamicProgramming edu.neu.ccs.pyramid.multilabel_classification.DynamicProgramming used on opensource projects

Example 1 with DynamicProgramming

use of edu.neu.ccs.pyramid.multilabel_classification.DynamicProgramming in project pyramid by cheng-li.

the class CBMPredictor method predictByDynamic.

public MultiLabel predictByDynamic() {
    // initialization
    DynamicProgramming[] DPs = new DynamicProgramming[numClusters];
    double[] minClusterProb = new double[numClusters];
    for (int k = 0; k < numClusters; k++) {
        DPs[k] = new DynamicProgramming(probs[k], logProbs[k]);
        minClusterProb[k] = Double.POSITIVE_INFINITY;
    }
    double maxLogProb = Double.NEGATIVE_INFINITY;
    MultiLabel bestMultiLabel = new MultiLabel();
    int maxIter = 10;
    for (int iter = 0; iter < maxIter; iter++) {
        for (int k = 0; k < numClusters; k++) {
            DynamicProgramming dp = DPs[k];
            double prob = dp.nextHighestProb();
            MultiLabel multiLabel = dp.nextHighestVector();
            minClusterProb[k] = prob;
            double threshold = computeThreshold(minClusterProb);
            boolean isCandidateValid = true;
            if ((multiLabel.getNumMatchedLabels() == 0) && !allowEmpty) {
                isCandidateValid = false;
            }
            if (isCandidateValid) {
                double logProb = logProbYnGivenXnLogisticProb(multiLabel);
                if (logProb >= maxLogProb) {
                    bestMultiLabel = multiLabel;
                    maxLogProb = logProb;
                }
            }
            if (Math.exp(maxLogProb) >= threshold) {
                return bestMultiLabel;
            }
        }
    }
    return bestMultiLabel;
}

Example 2 with DynamicProgramming

use of edu.neu.ccs.pyramid.multilabel_classification.DynamicProgramming in project pyramid by cheng-li.

the class SupportPredictor method predict.

@Override
public MultiLabel predict(Vector vector) {
    double[] uncali = classifier.predictClassProbs(vector);
    double[] marginals = labelCalibrator.calibratedClassProbs(uncali);
    List<Pair<MultiLabel, Double>> candidates = new ArrayList<>();
    DynamicProgramming dynamicProgramming = new DynamicProgramming(marginals);
    List<Pair<MultiLabel, Double>> sparseJoint = dynamicProgramming.topK(50);
    for (MultiLabel candidate : support) {
        PredictionCandidate predictionCandidate = new PredictionCandidate();
        predictionCandidate.x = vector;
        predictionCandidate.labelProbs = marginals;
        predictionCandidate.multiLabel = candidate;
        predictionCandidate.sparseJoint = sparseJoint;
        Vector feature = predictionFeatureExtractor.extractFeatures(predictionCandidate);
        double score = setCalibrator.calibrate(feature);
        candidates.add(new Pair<>(candidate, score));
    }
    Comparator<Pair<MultiLabel, Double>> comparator = Comparator.comparing(Pair::getSecond);
    return candidates.stream().max(comparator).map(Pair::getFirst).get();
}

Example 3 with DynamicProgramming

use of edu.neu.ccs.pyramid.multilabel_classification.DynamicProgramming in project pyramid by cheng-li.

the class Reranker method prob.

public double prob(Vector vector, MultiLabel multiLabel) {
    double[] marginals = labelCalibrator.calibratedClassProbs(classProbEstimator.predictClassProbs(vector));
    DynamicProgramming dynamicProgramming = new DynamicProgramming(marginals);
    List<Pair<MultiLabel, Double>> topK = dynamicProgramming.topK(numCandidate);
    PredictionCandidate predictionCandidate = new PredictionCandidate();
    predictionCandidate.x = vector;
    predictionCandidate.labelProbs = marginals;
    predictionCandidate.multiLabel = multiLabel;
    predictionCandidate.sparseJoint = topK;
    Vector feature = predictionFeatureExtractor.extractFeatures(predictionCandidate);
    double score = regressor.predict(feature);
    if (score > 1) {
        score = 1;
    }
    if (score < 0) {
        score = 0;
    }
    return score;
}

Example 4 with DynamicProgramming

use of edu.neu.ccs.pyramid.multilabel_classification.DynamicProgramming in project pyramid by cheng-li.

the class TopKFinder method topK.

public static List<Pair<MultiLabel, Double>> topK(Vector x, MultiLabelClassifier.ClassProbEstimator classProbEstimator, LabelCalibrator labelCalibrator, VectorCalibrator vectorCalibrator, PredictionFeatureExtractor predictionFeatureExtractor, int minSetSize, int maxSetSize, int top) {
    double[] uncalibratedMarginals = classProbEstimator.predictClassProbs(x);
    double[] marginals = labelCalibrator.calibratedClassProbs(uncalibratedMarginals);
    DynamicProgramming dynamicProgramming = new DynamicProgramming(marginals);
    // todo better
    List<Pair<MultiLabel, Double>> sparseJoint = dynamicProgramming.topK(50);
    List<MultiLabel> multiLabels = sparseJoint.stream().map(pair -> pair.getFirst()).filter(candidate -> candidate.getNumMatchedLabels() >= minSetSize && candidate.getNumMatchedLabels() <= maxSetSize).collect(Collectors.toList());
    List<Pair<MultiLabel, Double>> candidates = new ArrayList<>();
    if (multiLabels.isEmpty()) {
        int[] sorted = ArgSort.argSortDescending(marginals);
        MultiLabel multiLabel = new MultiLabel();
        for (int i = 0; i < minSetSize; i++) {
            multiLabel.addLabel(sorted[i]);
        }
        multiLabels.add(multiLabel);
    }
    for (MultiLabel candidate : multiLabels) {
        PredictionCandidate predictionCandidate = new PredictionCandidate();
        predictionCandidate.x = x;
        predictionCandidate.labelProbs = marginals;
        predictionCandidate.multiLabel = candidate;
        predictionCandidate.sparseJoint = sparseJoint;
        Vector feature = predictionFeatureExtractor.extractFeatures(predictionCandidate);
        double score = vectorCalibrator.calibrate(feature);
        candidates.add(new Pair<>(candidate, score));
    }
    Comparator<Pair<MultiLabel, Double>> comparator = Comparator.comparing(pair -> pair.getSecond());
    return candidates.stream().sorted(comparator.reversed()).limit(top).collect(Collectors.toList());
}

Example 5 with DynamicProgramming

use of edu.neu.ccs.pyramid.multilabel_classification.DynamicProgramming in project pyramid by cheng-li.

the class TopKFinder method topKinSupport.

public static List<Pair<MultiLabel, Double>> topKinSupport(Vector x, MultiLabelClassifier.ClassProbEstimator classProbEstimator, LabelCalibrator labelCalibrator, VectorCalibrator vectorCalibrator, PredictionFeatureExtractor predictionFeatureExtractor, List<MultiLabel> support, int top) {
    double[] marginals = labelCalibrator.calibratedClassProbs(classProbEstimator.predictClassProbs(x));
    DynamicProgramming dynamicProgramming = new DynamicProgramming(marginals);
    // todo better
    List<Pair<MultiLabel, Double>> sparseJoint = dynamicProgramming.topK(50);
    List<Pair<MultiLabel, Double>> list = new ArrayList<>();
    for (MultiLabel candidate : support) {
        PredictionCandidate predictionCandidate = new PredictionCandidate();
        predictionCandidate.x = x;
        predictionCandidate.labelProbs = marginals;
        predictionCandidate.multiLabel = candidate;
        predictionCandidate.sparseJoint = sparseJoint;
        Vector feature = predictionFeatureExtractor.extractFeatures(predictionCandidate);
        double pro = vectorCalibrator.calibrate(feature);
        list.add(new Pair<>(candidate, pro));
    }
    Comparator<Pair<MultiLabel, Double>> comparator = Comparator.comparing(pair -> pair.getSecond());
    return list.stream().sorted(comparator.reversed()).limit(top).collect(Collectors.toList());
}