Examples with Tensor org.tribuo.math.la.Tensor used on opensource projects

Example 1 with Tensor

use of org.tribuo.math.la.Tensor in project tribuo by oracle.

the class CRFParameters method getEmptyCopy.

/**
 * Returns a 3 element {@link Tensor} array.
 *
 * The first element is a {@link DenseVector} of label biases.
 * The second element is a {@link DenseMatrix} of feature-label weights.
 * The third element is a {@link DenseMatrix} of label-label transition weights.
 * @return A {@link Tensor} array.
 */
@Override
public Tensor[] getEmptyCopy() {
    Tensor[] output = new Tensor[3];
    output[0] = new DenseVector(biases.size());
    output[1] = new DenseMatrix(featureLabelWeights.getDimension1Size(), featureLabelWeights.getDimension2Size());
    output[2] = new DenseMatrix(labelLabelWeights.getDimension1Size(), labelLabelWeights.getDimension2Size());
    return output;
}

Example 2 with Tensor

use of org.tribuo.math.la.Tensor in project tribuo by oracle.

the class CRFParameters method merge.

@Override
public Tensor[] merge(Tensor[][] gradients, int size) {
    DenseVector biasUpdate = new DenseVector(biases.size());
    List<DenseSparseMatrix> updates = new ArrayList<>(size);
    DenseMatrix denseUpdates = null;
    DenseMatrix labelLabelUpdate = new DenseMatrix(labelLabelWeights.getDimension1Size(), labelLabelWeights.getDimension2Size());
    for (int j = 0; j < gradients.length; j++) {
        biasUpdate.intersectAndAddInPlace(gradients[j][0]);
        Matrix tmpUpdate = (Matrix) gradients[j][1];
        if (tmpUpdate instanceof DenseSparseMatrix) {
            updates.add((DenseSparseMatrix) tmpUpdate);
        } else {
            // is dense
            if (denseUpdates == null) {
                denseUpdates = (DenseMatrix) tmpUpdate;
            } else {
                denseUpdates.intersectAndAddInPlace(tmpUpdate);
            }
        }
        labelLabelUpdate.intersectAndAddInPlace(gradients[j][2]);
    }
    // Merge the combination of any dense and sparse updates
    Matrix featureLabelUpdate;
    if (updates.size() > 0) {
        featureLabelUpdate = merger.merge(updates.toArray(new DenseSparseMatrix[0]));
        if (denseUpdates != null) {
            denseUpdates.intersectAndAddInPlace(featureLabelUpdate);
            featureLabelUpdate = denseUpdates;
        }
    } else {
        featureLabelUpdate = denseUpdates;
    }
    return new Tensor[] { biasUpdate, featureLabelUpdate, labelLabelUpdate };
}

Example 3 with Tensor

use of org.tribuo.math.la.Tensor in project tribuo by oracle.

the class CRFTrainer method train.

@Override
public CRFModel train(SequenceDataset<Label> sequenceExamples, Map<String, Provenance> runProvenance) {
    if (sequenceExamples.getOutputInfo().getUnknownCount() > 0) {
        throw new IllegalArgumentException("The supplied Dataset contained unknown Outputs, and this Trainer is supervised.");
    }
    // Creates a new RNG, adds one to the invocation count, generates a local optimiser.
    SplittableRandom localRNG;
    TrainerProvenance trainerProvenance;
    StochasticGradientOptimiser localOptimiser;
    synchronized (this) {
        localRNG = rng.split();
        localOptimiser = optimiser.copy();
        trainerProvenance = getProvenance();
        trainInvocationCounter++;
    }
    ImmutableOutputInfo<Label> labelIDMap = sequenceExamples.getOutputIDInfo();
    ImmutableFeatureMap featureIDMap = sequenceExamples.getFeatureIDMap();
    SGDVector[][] sgdFeatures = new SGDVector[sequenceExamples.size()][];
    int[][] sgdLabels = new int[sequenceExamples.size()][];
    double[] weights = new double[sequenceExamples.size()];
    int n = 0;
    for (SequenceExample<Label> example : sequenceExamples) {
        weights[n] = example.getWeight();
        Pair<int[], SGDVector[]> pair = CRFModel.convertToVector(example, featureIDMap, labelIDMap);
        sgdFeatures[n] = pair.getB();
        sgdLabels[n] = pair.getA();
        n++;
    }
    logger.info(String.format("Training SGD CRF with %d examples", n));
    CRFParameters crfParameters = new CRFParameters(featureIDMap.size(), labelIDMap.size());
    localOptimiser.initialise(crfParameters);
    double loss = 0.0;
    int iteration = 0;
    for (int i = 0; i < epochs; i++) {
        if (shuffle) {
            Util.shuffleInPlace(sgdFeatures, sgdLabels, weights, localRNG);
        }
        if (minibatchSize == 1) {
            /*
                 * Special case a minibatch of size 1. Directly updates the parameters after each
                 * example rather than aggregating.
                 */
            for (int j = 0; j < sgdFeatures.length; j++) {
                Pair<Double, Tensor[]> output = crfParameters.valueAndGradient(sgdFeatures[j], sgdLabels[j]);
                loss += output.getA() * weights[j];
                // Update the gradient with the current learning rates
                Tensor[] updates = localOptimiser.step(output.getB(), weights[j]);
                // Apply the update to the current parameters.
                crfParameters.update(updates);
                iteration++;
                if ((iteration % loggingInterval == 0) && (loggingInterval != -1)) {
                    logger.info("At iteration " + iteration + ", average loss = " + loss / loggingInterval);
                    loss = 0.0;
                }
            }
        } else {
            Tensor[][] gradients = new Tensor[minibatchSize][];
            for (int j = 0; j < sgdFeatures.length; j += minibatchSize) {
                double tempWeight = 0.0;
                int curSize = 0;
                // Aggregate the gradient updates for each example in the minibatch
                for (int k = j; k < j + minibatchSize && k < sgdFeatures.length; k++) {
                    Pair<Double, Tensor[]> output = crfParameters.valueAndGradient(sgdFeatures[j], sgdLabels[j]);
                    loss += output.getA() * weights[k];
                    tempWeight += weights[k];
                    gradients[k - j] = output.getB();
                    curSize++;
                }
                // Merge the values into a single gradient update
                Tensor[] updates = crfParameters.merge(gradients, curSize);
                for (Tensor update : updates) {
                    update.scaleInPlace(minibatchSize);
                }
                tempWeight /= minibatchSize;
                // Update the gradient with the current learning rates
                updates = localOptimiser.step(updates, tempWeight);
                // Apply the gradient.
                crfParameters.update(updates);
                iteration++;
                if ((loggingInterval != -1) && (iteration % loggingInterval == 0)) {
                    logger.info("At iteration " + iteration + ", average loss = " + loss / loggingInterval);
                    loss = 0.0;
                }
            }
        }
    }
    localOptimiser.finalise();
    // public CRFModel(String name, String description, ImmutableInfoMap featureIDMap, ImmutableInfoMap outputIDInfo, CRFParameters parameters) {
    ModelProvenance provenance = new ModelProvenance(CRFModel.class.getName(), OffsetDateTime.now(), sequenceExamples.getProvenance(), trainerProvenance, runProvenance);
    CRFModel model = new CRFModel("crf-sgd-model", provenance, featureIDMap, labelIDMap, crfParameters);
    localOptimiser.reset();
    return model;
}

Example 4 with Tensor

use of org.tribuo.math.la.Tensor in project tribuo by oracle.

the class LinearParameters method getEmptyCopy.

/**
 * This returns a {@link DenseMatrix} the same size as the Parameters.
 * @return A {@link Tensor} array containing a single {@link DenseMatrix}.
 */
@Override
public Tensor[] getEmptyCopy() {
    DenseMatrix matrix = new DenseMatrix(weightMatrix.getDimension1Size(), weightMatrix.getDimension2Size());
    Tensor[] output = new Tensor[1];
    output[0] = matrix;
    return output;
}

Example 5 with Tensor

use of org.tribuo.math.la.Tensor in project tribuo by oracle.

the class Pegasos method step.

@Override
public Tensor[] step(Tensor[] updates, double weight) {
    double eta_t = baseRate / (lambda * iteration);
    for (Tensor t : updates) {
        t.scaleInPlace(eta_t * weight);
    }
    iteration++;
    return updates;
}