Examples with CategoricalFeature edu.neu.ccs.pyramid.feature.CategoricalFeature used on opensource projects

Example 1 with CategoricalFeature

use of edu.neu.ccs.pyramid.feature.CategoricalFeature in project pyramid by cheng-li.

the class BRCalibration method calibrate.

private static void calibrate(Config config, Logger logger) throws Exception {
    logger.info("start training calibrators");
    DataSetType dataSetType;
    switch(config.getString("dataSetType")) {
        case "sparse_random":
            dataSetType = DataSetType.ML_CLF_SPARSE;
            break;
        case "sparse_sequential":
            dataSetType = DataSetType.ML_CLF_SEQ_SPARSE;
            break;
        default:
            throw new IllegalArgumentException("unknown dataSetType");
    }
    MultiLabelClfDataSet train = TRECFormat.loadMultiLabelClfDataSet(config.getString("input.trainData"), dataSetType, true);
    MultiLabelClfDataSet cal = TRECFormat.loadMultiLabelClfDataSet(config.getString("input.calibrationData"), dataSetType, true);
    MultiLabelClfDataSet valid = TRECFormat.loadMultiLabelClfDataSet(config.getString("input.validData"), dataSetType, true);
    List<MultiLabel> support = (List<MultiLabel>) Serialization.deserialize(Paths.get(config.getString("output.dir"), "model_predictions", config.getString("output.modelFolder"), "models", "support").toFile());
    MultiLabelClassifier.ClassProbEstimator classProbEstimator = (MultiLabelClassifier.ClassProbEstimator) Serialization.deserialize(Paths.get(config.getString("output.dir"), "model_predictions", config.getString("output.modelFolder"), "models", "classifier"));
    List<Integer> labelCalIndices = IntStream.range(0, cal.getNumDataPoints()).filter(i -> i % 2 == 0).boxed().collect(Collectors.toList());
    List<Integer> setCalIndices = IntStream.range(0, cal.getNumDataPoints()).filter(i -> i % 2 == 1).boxed().collect(Collectors.toList());
    MultiLabelClfDataSet labelCalData = DataSetUtil.sampleData(cal, labelCalIndices);
    MultiLabelClfDataSet setCalData = DataSetUtil.sampleData(cal, setCalIndices);
    logger.info("start training label calibrator");
    LabelCalibrator labelCalibrator = null;
    switch(config.getString("labelCalibrator")) {
        case "isotonic":
            labelCalibrator = new IsoLabelCalibrator(classProbEstimator, labelCalData, false);
            break;
        case "identity":
            labelCalibrator = new IdentityLabelCalibrator();
            break;
    }
    logger.info("finish training label calibrator");
    logger.info("start training set calibrator");
    List<PredictionFeatureExtractor> extractors = new ArrayList<>();
    if (config.getBoolean("brProb")) {
        // todo order matters; the first one will be used by iso, card iso
        extractors.add(new BRProbFeatureExtractor());
    }
    if (config.getBoolean("expectedF1")) {
        extractors.add(new ExpectedF1FeatureExtractor());
    }
    if (config.getBoolean("expectedPrecision")) {
        extractors.add(new ExpectedPrecisionFeatureExtractor());
    }
    if (config.getBoolean("expectedRecall")) {
        extractors.add(new ExpectedRecallFeatureExtractor());
    }
    if (config.getBoolean("setPrior")) {
        extractors.add(new PriorFeatureExtractor(train));
    }
    if (config.getBoolean("card")) {
        extractors.add(new CardFeatureExtractor());
    }
    if (config.getBoolean("encodeLabel")) {
        extractors.add(new LabelBinaryFeatureExtractor(classProbEstimator.getNumClasses(), train.getLabelTranslator()));
    }
    if (config.getBoolean("useInitialFeatures")) {
        Set<String> prefixes = new HashSet<>(config.getStrings("featureFieldPrefix"));
        FeatureList featureList = train.getFeatureList();
        List<Integer> featureIds = new ArrayList<>();
        for (int j = 0; j < featureList.size(); j++) {
            Feature feature = featureList.get(j);
            if (feature instanceof CategoricalFeature) {
                if (matchPrefixes(((CategoricalFeature) feature).getVariableName(), prefixes)) {
                    featureIds.add(j);
                }
            } else {
                if (!(feature instanceof Ngram)) {
                    if (matchPrefixes(feature.getName(), prefixes)) {
                        featureIds.add(j);
                    }
                }
            }
        }
        extractors.add(new InstanceFeatureExtractor(featureIds, train.getFeatureList()));
    }
    PredictionFeatureExtractor predictionFeatureExtractor = new CombinedPredictionFeatureExtractor(extractors);
    CalibrationDataGenerator calibrationDataGenerator = new CalibrationDataGenerator(labelCalibrator, predictionFeatureExtractor);
    CalibrationDataGenerator.TrainData caliTrainingData;
    CalibrationDataGenerator.TrainData caliValidData;
    caliTrainingData = calibrationDataGenerator.createCaliTrainingData(setCalData, classProbEstimator, config.getInt("numCandidates"), config.getString("calibrate.target"), support, 10);
    caliValidData = calibrationDataGenerator.createCaliTrainingData(valid, classProbEstimator, config.getInt("numCandidates"), config.getString("calibrate.target"), support, 10);
    RegDataSet calibratorTrainData = caliTrainingData.regDataSet;
    double[] weights = caliTrainingData.instanceWeights;
    VectorCalibrator setCalibrator = null;
    switch(config.getString("setCalibrator")) {
        case "cardinality_isotonic":
            setCalibrator = new VectorCardIsoSetCalibrator(calibratorTrainData, 0, 2, false);
            break;
        case "reranker":
            RerankerTrainer rerankerTrainer = RerankerTrainer.newBuilder().numCandidates(config.getInt("numCandidates")).numLeaves(config.getInt("numLeaves")).monotonicityType("weak").build();
            setCalibrator = rerankerTrainer.trainWithSigmoid(calibratorTrainData, weights, classProbEstimator, predictionFeatureExtractor, labelCalibrator, caliValidData.regDataSet);
            break;
        case "isotonic":
            setCalibrator = new VectorIsoSetCalibrator(calibratorTrainData, 0, false);
            break;
        case "identity":
            setCalibrator = new VectorIdentityCalibrator(0);
            break;
        case "zero":
            setCalibrator = new ZeroCalibrator();
            break;
        default:
            throw new IllegalArgumentException("illegal setCalibrator");
    }
    logger.info("finish training set calibrator");
    Serialization.serialize(labelCalibrator, Paths.get(config.getString("output.dir"), "model_predictions", config.getString("output.modelFolder"), "models", "calibrators", config.getString("output.calibratorFolder"), "label_calibrator").toFile());
    Serialization.serialize(setCalibrator, Paths.get(config.getString("output.dir"), "model_predictions", config.getString("output.modelFolder"), "models", "calibrators", config.getString("output.calibratorFolder"), "set_calibrator").toFile());
    Serialization.serialize(predictionFeatureExtractor, Paths.get(config.getString("output.dir"), "model_predictions", config.getString("output.modelFolder"), "models", "calibrators", config.getString("output.calibratorFolder"), "prediction_feature_extractor").toFile());
    logger.info("finish training calibrators");
    MultiLabelClassifier classifier = null;
    switch(config.getString("predict.mode")) {
        case "independent":
            classifier = new IndependentPredictor(classProbEstimator, labelCalibrator);
            break;
        case "support":
            classifier = new edu.neu.ccs.pyramid.multilabel_classification.predictor.SupportPredictor(classProbEstimator, labelCalibrator, setCalibrator, predictionFeatureExtractor, support);
            break;
        case "reranker":
            Reranker reranker = (Reranker) setCalibrator;
            reranker.setMinPredictionSize(config.getInt("predict.minSize"));
            reranker.setMaxPredictionSize(config.getInt("predict.maxSize"));
            classifier = reranker;
            break;
        default:
            throw new IllegalArgumentException("illegal predict.mode");
    }
    MultiLabel[] predictions = classifier.predict(cal);
    MultiLabel[] predictions_valid = classifier.predict(valid);
    if (true) {
        logger.info("calibration performance on " + config.getString("input.calibrationFolder") + " set");
        List<CalibrationDataGenerator.CalibrationInstance> instances = IntStream.range(0, cal.getNumDataPoints()).parallel().boxed().map(i -> calibrationDataGenerator.createInstance(classProbEstimator, cal.getRow(i), predictions[i], cal.getMultiLabels()[i], config.getString("calibrate.target"))).collect(Collectors.toList());
        eval(instances, setCalibrator, logger, config.getString("calibrate.target"));
    }
    logger.info("classification performance on " + config.getString("input.validFolder") + " set");
    logger.info(new MLMeasures(valid.getNumClasses(), valid.getMultiLabels(), predictions_valid).toString());
    if (true) {
        logger.info("calibration performance on " + config.getString("input.validFolder") + " set");
        List<CalibrationDataGenerator.CalibrationInstance> instances = IntStream.range(0, valid.getNumDataPoints()).parallel().boxed().map(i -> calibrationDataGenerator.createInstance(classProbEstimator, valid.getRow(i), predictions_valid[i], valid.getMultiLabels()[i], config.getString("calibrate.target"))).collect(Collectors.toList());
        eval(instances, setCalibrator, logger, config.getString("calibrate.target"));
    }
}