Examples with ClassifierTrainingData qupath.process.gui.commands.ml.PixelClassifierTraining.ClassifierTrainingData used on opensource projects

Example 1 with ClassifierTrainingData

use of qupath.process.gui.commands.ml.PixelClassifierTraining.ClassifierTrainingData in project qupath by qupath.

the class PixelClassifierPane method doClassification.

private void doClassification() {
    // if (helper == null || helper.getFeatureServer() == null) {
    // //			updateFeatureCalculator();
    // //			updateClassifier();
    // if (helper == null) {
    // logger.error("No pixel classifier helper available!");
    // return;
    // }
    // }
    var imageData = qupath.getImageData();
    if (imageData == null) {
        if (!qupath.getViewers().stream().anyMatch(v -> v.getImageData() != null)) {
            logger.debug("doClassification() called, but no images are open");
            return;
        }
    }
    var model = selectedClassifier.get();
    if (model == null) {
        Dialogs.showErrorNotification("Pixel classifier", "No classifier selected!");
        return;
    }
    ClassifierTrainingData trainingData;
    try {
        var trainingImages = getTrainingImageData();
        if (trainingImages.size() > 1)
            logger.info("Creating training data from {} images", trainingImages.size());
        trainingData = helper.createTrainingData(trainingImages);
    } catch (Exception e) {
        logger.error("Error when updating training data", e);
        return;
    }
    if (trainingData == null) {
        resetPieChart();
        return;
    }
    // TODO: Optionally limit the number of training samples we use
    // var trainData = classifier.createTrainData(matFeatures, matTargets);
    // Ensure we seed the RNG for reproducibility
    opencv_core.setRNGSeed(rngSeed);
    // TODO: Prevent training K nearest neighbor with a huge number of samples (very slow!)
    var actualMaxSamples = this.maxSamples;
    var trainData = trainingData.getTrainData();
    if (actualMaxSamples > 0 && trainData.getNTrainSamples() > actualMaxSamples)
        trainData.setTrainTestSplit(actualMaxSamples, true);
    else
        trainData.shuffleTrainTest();
    // System.err.println("Train: " + trainData.getTrainResponses());
    // System.err.println("Test: " + trainData.getTestResponses());
    // Apply normalization, if we need to
    FeaturePreprocessor preprocessor = normalization.build(trainData.getTrainSamples(), false);
    if (preprocessor.doesSomething()) {
        preprocessingOp = ImageOps.ML.preprocessor(preprocessor);
    } else
        preprocessingOp = null;
    var labels = trainingData.getLabelMap();
    // Using getTrainNormCatResponses() causes confusion if classes are not represented
    // var targets = trainData.getTrainNormCatResponses();
    var targets = trainData.getTrainResponses();
    IntBuffer buffer = targets.createBuffer();
    int n = (int) targets.total();
    var rawCounts = new int[labels.size()];
    for (int i = 0; i < n; i++) {
        rawCounts[buffer.get(i)] += 1;
    }
    Map<PathClass, Integer> counts = new LinkedHashMap<>();
    for (var entry : labels.entrySet()) {
        counts.put(entry.getKey(), rawCounts[entry.getValue()]);
    }
    updatePieChart(counts);
    Mat weights = null;
    if (reweightSamples) {
        weights = new Mat(n, 1, opencv_core.CV_32FC1);
        FloatIndexer bufferWeights = weights.createIndexer();
        float[] weightArray = new float[rawCounts.length];
        for (int i = 0; i < weightArray.length; i++) {
            int c = rawCounts[i];
            // weightArray[i] = c == 0 ? 1 : (float)1.f/c;
            weightArray[i] = c == 0 ? 1 : (float) n / c;
        }
        for (int i = 0; i < n; i++) {
            int label = buffer.get(i);
            bufferWeights.put(i, weightArray[label]);
        }
        bufferWeights.release();
    }
    // Create TrainData in an appropriate format (e.g. labels or one-hot encoding)
    var trainSamples = trainData.getTrainSamples();
    var trainResponses = trainData.getTrainResponses();
    preprocessor.apply(trainSamples, false);
    trainData = model.createTrainData(trainSamples, trainResponses, weights, false);
    logger.info("Training data: {} x {}, Target data: {} x {}", trainSamples.rows(), trainSamples.cols(), trainResponses.rows(), trainResponses.cols());
    model.train(trainData);
    // Calculate accuracy using whatever we can, as a rough guide to progress
    var test = trainData.getTestSamples();
    String testSet = "HELD-OUT TRAINING SET";
    if (test.empty()) {
        test = trainSamples;
        testSet = "TRAINING SET";
    } else {
        preprocessor.apply(test, false);
        buffer = trainData.getTestNormCatResponses().createBuffer();
    }
    var testResults = new Mat();
    model.predict(test, testResults, null);
    IntBuffer bufferResults = testResults.createBuffer();
    int nTest = (int) testResults.rows();
    int nCorrect = 0;
    for (int i = 0; i < nTest; i++) {
        if (bufferResults.get(i) == buffer.get(i))
            nCorrect++;
    }
    logger.info("Current accuracy on the {}: {} %", testSet, GeneralTools.formatNumber(nCorrect * 100.0 / n, 1));
    if (model instanceof RTreesClassifier) {
        var trees = (RTreesClassifier) model;
        if (trees.hasFeatureImportance() && imageData != null)
            logVariableImportance(trees, helper.getFeatureOp().getChannels(imageData).stream().map(c -> c.getName()).collect(Collectors.toList()));
    }
    trainData.close();
    var featureCalculator = helper.getFeatureOp();
    if (preprocessingOp != null)
        featureCalculator = featureCalculator.appendOps(preprocessingOp);
    // TODO: CHECK IF INPUT SIZE SHOULD BE DEFINED
    int inputWidth = 512;
    int inputHeight = 512;
    // int inputWidth = featureCalculator.getInputSize().getWidth();
    // int inputHeight = featureCalculator.getInputSize().getHeight();
    var cal = helper.getResolution();
    var channelType = ImageServerMetadata.ChannelType.CLASSIFICATION;
    if (model.supportsProbabilities()) {
        channelType = selectedOutputType.get();
    }
    // Channels are needed for probability output (and work for classification as well)
    var labels2 = new TreeMap<Integer, PathClass>();
    for (var entry : labels.entrySet()) {
        var previous = labels2.put(entry.getValue(), entry.getKey());
        if (previous != null)
            logger.warn("Duplicate label found! {} matches with {} and {}, only the latter be used", entry.getValue(), previous, entry.getKey());
    }
    var channels = PathClassifierTools.classificationLabelsToChannels(labels2, true);
    PixelClassifierMetadata metadata = new PixelClassifierMetadata.Builder().inputResolution(cal).inputShape(inputWidth, inputHeight).setChannelType(channelType).outputChannels(channels).build();
    currentClassifier.set(PixelClassifiers.createClassifier(model, featureCalculator, metadata, true));
    var overlay = PixelClassificationOverlay.create(qupath.getOverlayOptions(), currentClassifier.get(), getLivePredictionThreads());
    replaceOverlay(overlay);
}