Examples with EarlyStopper edu.neu.ccs.pyramid.optimization.EarlyStopper used on opensource projects

Example 11 with EarlyStopper

use of edu.neu.ccs.pyramid.optimization.EarlyStopper in project pyramid by cheng-li.

the class RerankerTrainer method train.

public Reranker train(RegDataSet regDataSet, double[] instanceWeights, MultiLabelClassifier.ClassProbEstimator classProbEstimator, PredictionFeatureExtractor predictionFeatureExtractor, LabelCalibrator labelCalibrator, RegDataSet validation) {
    LSBoost lsBoost = new LSBoost();
    RegTreeConfig regTreeConfig = new RegTreeConfig().setMaxNumLeaves(numLeaves).setMinDataPerLeaf(minDataPerLeaf).setMonotonicityType(monotonicityType);
    RegTreeFactory regTreeFactory = new RegTreeFactory(regTreeConfig);
    LSBoostOptimizer optimizer = new LSBoostOptimizer(lsBoost, regDataSet, regTreeFactory, instanceWeights, regDataSet.getLabels());
    if (!monotonicityType.equals("none")) {
        int[][] mono = new int[1][regDataSet.getNumFeatures()];
        mono[0] = predictionFeatureExtractor.featureMonotonicity();
        optimizer.setMonotonicity(mono);
    }
    optimizer.setShrinkage(shrinkage);
    optimizer.initialize();
    EarlyStopper earlyStopper = new EarlyStopper(EarlyStopper.Goal.MINIMIZE, 5);
    LSBoost bestModel = null;
    for (int i = 1; i <= maxIter; i++) {
        optimizer.iterate();
        if (i % 10 == 0 || i == maxIter) {
            double mse = MSE.mse(lsBoost, validation);
            earlyStopper.add(i, mse);
            if (earlyStopper.getBestIteration() == i) {
                try {
                    bestModel = (LSBoost) Serialization.deepCopy(lsBoost);
                } catch (IOException e) {
                    e.printStackTrace();
                } catch (ClassNotFoundException e) {
                    e.printStackTrace();
                }
            }
            if (earlyStopper.shouldStop()) {
                break;
            }
        }
    }
    // System.out.println("best iteration = "+earlyStopper.getBestIteration());
    return new Reranker(bestModel, classProbEstimator, numCandidates, predictionFeatureExtractor, labelCalibrator);
}

Example 12 with EarlyStopper

use of edu.neu.ccs.pyramid.optimization.EarlyStopper in project pyramid by cheng-li.

the class BRGB method train.

static void train(Config config, Logger logger) throws Exception {
    String output = config.getString("output.folder");
    int numIterations = config.getInt("train.numIterations");
    int numLeaves = config.getInt("train.numLeaves");
    double learningRate = config.getDouble("train.learningRate");
    int minDataPerLeaf = config.getInt("train.minDataPerLeaf");
    int randomSeed = config.getInt("train.randomSeed");
    StopWatch stopWatch = new StopWatch();
    stopWatch.start();
    MultiLabelClfDataSet allTrainData = loadData(config, config.getString("input.trainData"));
    double[] instanceWeights = new double[allTrainData.getNumDataPoints()];
    Arrays.fill(instanceWeights, 1.0);
    if (config.getBoolean("train.useInstanceWeights")) {
        instanceWeights = loadInstanceWeights(config);
    }
    MultiLabelClfDataSet trainSetForEval = minibatch(allTrainData, instanceWeights, config.getInt("train.showProgress.sampleSize"), 0 + randomSeed).getFirst();
    MultiLabelClfDataSet validSet = loadData(config, config.getString("input.validData"));
    List<MultiLabel> support = DataSetUtil.gatherMultiLabels(allTrainData);
    Serialization.serialize(support, Paths.get(output, "model_predictions", config.getString("output.modelFolder"), "models", "support"));
    int numClasses = allTrainData.getNumClasses();
    logger.info("number of class = " + numClasses);
    IMLGradientBoosting boosting;
    List<EarlyStopper> earlyStoppers;
    List<Terminator> terminators;
    boolean[] shouldStop;
    int numLabelsLeftToTrain;
    int startIter;
    List<Pair<Integer, Double>> trainingTime;
    List<Pair<Integer, Double>> accuracy;
    double startTime = 0;
    boolean earlyStop = config.getBoolean("train.earlyStop");
    CheckPoint checkPoint;
    if (config.getBoolean("train.warmStart")) {
        checkPoint = (CheckPoint) Serialization.deserialize(Paths.get(output, "model_predictions", config.getString("output.modelFolder"), "models", "checkpoint"));
        boosting = checkPoint.boosting;
        earlyStoppers = checkPoint.earlyStoppers;
        terminators = checkPoint.terminators;
        shouldStop = checkPoint.shouldStop;
        numLabelsLeftToTrain = checkPoint.numLabelsLeftToTrain;
        startIter = checkPoint.lastIter + 1;
        trainingTime = checkPoint.trainingTime;
        accuracy = checkPoint.accuracy;
        startTime = checkPoint.trainingTime.get(trainingTime.size() - 1).getSecond();
    } else {
        boosting = new IMLGradientBoosting(numClasses);
        earlyStoppers = new ArrayList<>();
        terminators = new ArrayList<>();
        trainingTime = new ArrayList<>();
        accuracy = new ArrayList<>();
        if (earlyStop) {
            for (int l = 0; l < numClasses; l++) {
                EarlyStopper earlyStopper = new EarlyStopper(EarlyStopper.Goal.MINIMIZE, config.getInt("train.earlyStop.patience"));
                earlyStopper.setMinimumIterations(config.getInt("train.earlyStop.minIterations"));
                earlyStoppers.add(earlyStopper);
            }
            for (int l = 0; l < numClasses; l++) {
                Terminator terminator = new Terminator();
                terminator.setMaxStableIterations(config.getInt("train.earlyStop.patience")).setMinIterations(config.getInt("train.earlyStop.minIterations") / config.getInt("train.showProgress.interval")).setAbsoluteEpsilon(config.getDouble("train.earlyStop.absoluteChange")).setRelativeEpsilon(config.getDouble("train.earlyStop.relativeChange")).setOperation(Terminator.Operation.OR);
                terminators.add(terminator);
            }
        }
        shouldStop = new boolean[allTrainData.getNumClasses()];
        numLabelsLeftToTrain = numClasses;
        checkPoint = new CheckPoint();
        checkPoint.boosting = boosting;
        checkPoint.earlyStoppers = earlyStoppers;
        checkPoint.terminators = terminators;
        checkPoint.shouldStop = shouldStop;
        // this is not a pointer, has to be updated
        checkPoint.numLabelsLeftToTrain = numLabelsLeftToTrain;
        checkPoint.lastIter = 0;
        checkPoint.trainingTime = trainingTime;
        checkPoint.accuracy = accuracy;
        startIter = 1;
    }
    logger.info("During training, the performance is reported using Hamming loss optimal predictor. The performance is computed approximately with " + config.getInt("train.showProgress.sampleSize") + " instances.");
    int progressInterval = config.getInt("train.showProgress.interval");
    int interval = config.getInt("train.fullScanInterval");
    int minibatchLifeSpan = config.getInt("train.minibatchLifeSpan");
    int numActiveFeatures = config.getInt("train.numActiveFeatures");
    int numofLabels = allTrainData.getNumClasses();
    List<Integer>[] activeFeaturesLists = new ArrayList[numofLabels];
    for (int labelnum = 0; labelnum < numofLabels; labelnum++) {
        activeFeaturesLists[labelnum] = new ArrayList<>();
    }
    MultiLabelClfDataSet trainBatch = null;
    IMLGBTrainer trainer = null;
    StopWatch timeWatch = new StopWatch();
    timeWatch.start();
    for (int i = startIter; i <= numIterations; i++) {
        logger.info("iteration " + i);
        if (i % minibatchLifeSpan == 1 || i == startIter) {
            Pair<MultiLabelClfDataSet, double[]> sampled = minibatch(allTrainData, instanceWeights, config.getInt("train.batchSize"), i + randomSeed);
            trainBatch = sampled.getFirst();
            IMLGBConfig imlgbConfig = new IMLGBConfig.Builder(trainBatch).learningRate(learningRate).minDataPerLeaf(minDataPerLeaf).numLeaves(numLeaves).numSplitIntervals(config.getInt("train.numSplitIntervals")).usePrior(config.getBoolean("train.usePrior")).numActiveFeatures(numActiveFeatures).build();
            trainer = new IMLGBTrainer(imlgbConfig, boosting, shouldStop);
            trainer.setInstanceWeights(sampled.getSecond());
        }
        if (i % interval == 1) {
            trainer.iterate(activeFeaturesLists, true);
        } else {
            trainer.iterate(activeFeaturesLists, false);
        }
        checkPoint.lastIter += 1;
        if (earlyStop && (i % progressInterval == 0 || i == numIterations)) {
            for (int l = 0; l < numClasses; l++) {
                EarlyStopper earlyStopper = earlyStoppers.get(l);
                Terminator terminator = terminators.get(l);
                if (!shouldStop[l]) {
                    double kl = KL(boosting, validSet, l);
                    earlyStopper.add(i, kl);
                    terminator.add(kl);
                    if (earlyStopper.shouldStop() || terminator.shouldTerminate()) {
                        logger.info("training for label " + l + " (" + allTrainData.getLabelTranslator().toExtLabel(l) + ") should stop now");
                        logger.info("the best number of training iterations for the label is " + earlyStopper.getBestIteration());
                        if (i != earlyStopper.getBestIteration()) {
                            boosting.cutTail(l, earlyStopper.getBestIteration());
                            logger.info("roll back the model for this label to iteration " + earlyStopper.getBestIteration());
                        }
                        shouldStop[l] = true;
                        numLabelsLeftToTrain -= 1;
                        checkPoint.numLabelsLeftToTrain = numLabelsLeftToTrain;
                        logger.info("the number of labels left to be trained on = " + numLabelsLeftToTrain);
                    }
                }
            }
        }
        if (config.getBoolean("train.showTrainProgress") && (i % progressInterval == 0 || i == numIterations)) {
            logger.info("training set performance (computed approximately with Hamming loss predictor on " + config.getInt("train.showProgress.sampleSize") + " instances).");
            logger.info(new MLMeasures(boosting, trainSetForEval).toString());
        }
        if (config.getBoolean("train.showValidProgress") && (i % progressInterval == 0 || i == numIterations)) {
            logger.info("validation set performance (computed approximately with Hamming loss predictor)");
            MLMeasures validPerformance = new MLMeasures(boosting, validSet);
            logger.info(validPerformance.toString());
            accuracy.add(new Pair<>(i, validPerformance.getInstanceAverage().getF1()));
        }
        trainingTime.add(new Pair<>(i, startTime + timeWatch.getTime() / 1000.0));
        Serialization.serialize(checkPoint, Paths.get(output, "model_predictions", config.getString("output.modelFolder"), "models", "checkpoint"));
        Serialization.serialize(boosting, Paths.get(output, "model_predictions", config.getString("output.modelFolder"), "models", "classifier"));
        if (numLabelsLeftToTrain == 0) {
            logger.info("all label training finished");
            break;
        }
    }
    logger.info("training done");
    logger.info(stopWatch.toString());
    File analysisFolder = Paths.get(output, "model_predictions", config.getString("output.modelFolder"), "analysis").toFile();
    if (true) {
        ObjectMapper objectMapper = new ObjectMapper();
        List<LabelModel> labelModels = IMLGBInspector.getAllRules(boosting);
        new File(analysisFolder, "decision_rules").mkdirs();
        for (int l = 0; l < boosting.getNumClasses(); l++) {
            objectMapper.writeValue(Paths.get(analysisFolder.toString(), "decision_rules", l + ".json").toFile(), labelModels.get(l));
        }
    }
    boolean topFeaturesToFile = true;
    if (topFeaturesToFile) {
        logger.info("start writing top features");
        List<TopFeatures> topFeaturesList = IntStream.range(0, boosting.getNumClasses()).mapToObj(k -> IMLGBInspector.topFeatures(boosting, k, Integer.MAX_VALUE)).collect(Collectors.toList());
        ObjectMapper mapper = new ObjectMapper();
        String file = "top_features.json";
        mapper.writeValue(new File(analysisFolder, file), topFeaturesList);
        StringBuilder sb = new StringBuilder();
        for (int l = 0; l < boosting.getNumClasses(); l++) {
            sb.append("-------------------------").append("\n");
            sb.append(allTrainData.getLabelTranslator().toExtLabel(l)).append(":").append("\n");
            for (Feature feature : topFeaturesList.get(l).getTopFeatures()) {
                sb.append(feature.simpleString()).append(", ");
            }
            sb.append("\n");
        }
        FileUtils.writeStringToFile(new File(analysisFolder, "top_features.txt"), sb.toString());
        logger.info("finish writing top features");
    }
}

Example 13 with EarlyStopper

use of edu.neu.ccs.pyramid.optimization.EarlyStopper in project pyramid by cheng-li.

the class CBMEN method tune.

private static TuneResult tune(Config config, HyperParameters hyperParameters, MultiLabelClfDataSet trainSet, MultiLabelClfDataSet validSet) throws Exception {
    CBM cbm = newCBM(config, trainSet, hyperParameters);
    EarlyStopper earlyStopper = loadNewEarlyStopper(config);
    ENCBMOptimizer optimizer = getOptimizer(config, hyperParameters, cbm, trainSet);
    if (config.getBoolean("train.randomInitialize")) {
        optimizer.randInitialize();
    } else {
        optimizer.initialize();
    }
    MultiLabelClassifier classifier;
    String predictTarget = config.getString("tune.targetMetric");
    switch(predictTarget) {
        case "instance_set_accuracy":
            AccPredictor accPredictor = new AccPredictor(cbm);
            accPredictor.setComponentContributionThreshold(config.getDouble("predict.piThreshold"));
            classifier = accPredictor;
            break;
        case "instance_f1":
            PluginF1 pluginF1 = new PluginF1(cbm);
            List<MultiLabel> support = DataSetUtil.gatherMultiLabels(trainSet);
            pluginF1.setSupport(support);
            pluginF1.setPiThreshold(config.getDouble("predict.piThreshold"));
            classifier = pluginF1;
            break;
        case "instance_hamming_loss":
            MarginalPredictor marginalPredictor = new MarginalPredictor(cbm);
            marginalPredictor.setPiThreshold(config.getDouble("predict.piThreshold"));
            classifier = marginalPredictor;
            break;
        case "label_map":
            AccPredictor accPredictor2 = new AccPredictor(cbm);
            accPredictor2.setComponentContributionThreshold(config.getDouble("predict.piThreshold"));
            classifier = accPredictor2;
            break;
        default:
            throw new IllegalArgumentException("predictTarget should be instance_set_accuracy, instance_f1 or instance_hamming_loss");
    }
    int interval = config.getInt("tune.monitorInterval");
    for (int iter = 1; true; iter++) {
        if (VERBOSE) {
            System.out.println("iteration " + iter);
        }
        optimizer.iterate();
        if (iter % interval == 0) {
            MLMeasures validMeasures = new MLMeasures(classifier, validSet);
            if (VERBOSE) {
                System.out.println("validation performance with " + predictTarget + " optimal predictor:");
                System.out.println(validMeasures);
            }
            switch(predictTarget) {
                case "instance_set_accuracy":
                    earlyStopper.add(iter, validMeasures.getInstanceAverage().getAccuracy());
                    break;
                case "instance_f1":
                    earlyStopper.add(iter, validMeasures.getInstanceAverage().getF1());
                    break;
                case "instance_hamming_loss":
                    earlyStopper.add(iter, validMeasures.getInstanceAverage().getHammingLoss());
                    break;
                case "label_map":
                    List<MultiLabel> support = DataSetUtil.gatherMultiLabels(trainSet);
                    double map = MAP.mapBySupport(cbm, validSet, support);
                    earlyStopper.add(iter, map);
                    break;
                default:
                    throw new IllegalArgumentException("predictTarget should be instance_set_accuracy or instance_f1");
            }
            if (earlyStopper.shouldStop()) {
                if (VERBOSE) {
                    System.out.println("Early Stopper: the training should stop now!");
                }
                break;
            }
        }
    }
    if (VERBOSE) {
        System.out.println("done!");
    }
    hyperParameters.iterations = earlyStopper.getBestIteration();
    TuneResult tuneResult = new TuneResult();
    tuneResult.hyperParameters = hyperParameters;
    tuneResult.performance = earlyStopper.getBestValue();
    return tuneResult;
}