Package cc.mallet.fst

Class CRFTrainerByLabelLikelihood

  • All Implemented Interfaces:
    TransducerTrainer.ByOptimization
    Direct Known Subclasses:
    CRFTrainerByL1LabelLikelihood
    public class CRFTrainerByLabelLikelihood
extends TransducerTrainer
implements TransducerTrainer.ByOptimization
    Unlike ClassifierTrainer, TransducerTrainer is not "stateless" between calls to train. A TransducerTrainer is constructed paired with a specific Transducer, and can only train that Transducer. CRF stores and has methods for FeatureSelection and weight freezing. CRFTrainer stores and has methods for determining the contents/dimensions/sparsity/FeatureInduction of the CRF's weights as determined by training data.

    Note: In the future this class may go away in favor of some default version of CRFTrainerByValueGradients.

    • Field Detail

      • printGradient

        public boolean printGradient

    • Constructor Detail

      • CRFTrainerByLabelLikelihood

        public CRFTrainerByLabelLikelihood​(CRF crf)

    • Method Detail

      • getCRF

        public CRF getCRF()

      • isConverged

        public boolean isConverged()

      • setAddNoFactors

        public void setAddNoFactors​(boolean flag)
        Use this method to specify whether or not factors are added to the CRF by this trainer. If you have already setup the factors in your CRF, you may not want the trainer to add additional factors.
        Parameters:
        flag - If true, this trainer adds no factors to the CRF.

      • trainIncremental

        public boolean trainIncremental​(InstanceList training)

      • train

        public boolean train​(InstanceList trainingSet,
                     int numIterations)
        Description copied from class: TransducerTrainer
        Train the transducer associated with this TransducerTrainer. You should be able to call this method with different trainingSet objects. Whether this causes the TransducerTrainer to combine both trainingSets or to view the second as a new alternative is at the discretion of the particular TransducerTrainer subclass involved.
        Specified by:
        train in class TransducerTrainer

      • train

        public boolean train​(InstanceList training,
                     int numIterationsPerProportion,
                     double[] trainingProportions)
        Train a CRF on various-sized subsets of the data. This method is typically used to accelerate training by quickly getting to reasonable parameters on only a subset of the parameters first, then on progressively more data.
        Parameters:
        training - The training Instances.
        numIterationsPerProportion - Maximum number of Maximizer iterations per training proportion.
        trainingProportions - Train on increasingly larger portions of the data, e.g. new double[] {0.2, 0.5, 1.0}. This can sometimes speed up convergence, similar to SGD. Be sure to end in 1.0 if you want to train on all the data in the end.
        Returns:
        True if training has converged.

      • trainWithFeatureInduction

        public boolean trainWithFeatureInduction​(InstanceList trainingData,
                                         InstanceList validationData,
                                         InstanceList testingData,
                                         TransducerEvaluator eval,
                                         int numIterations,
                                         int numIterationsBetweenFeatureInductions,
                                         int numFeatureInductions,
                                         int numFeaturesPerFeatureInduction,
                                         double trueLabelProbThreshold,
                                         boolean clusteredFeatureInduction,
                                         double[] trainingProportions)

      • trainWithFeatureInduction

        public boolean trainWithFeatureInduction​(InstanceList trainingData,
                                         InstanceList validationData,
                                         InstanceList testingData,
                                         TransducerEvaluator eval,
                                         int numIterations,
                                         int numIterationsBetweenFeatureInductions,
                                         int numFeatureInductions,
                                         int numFeaturesPerFeatureInduction,
                                         double trueLabelProbThreshold,
                                         boolean clusteredFeatureInduction,
                                         double[] trainingProportions,
                                         java.lang.String gainName)
        Train a CRF using feature induction to generate conjunctions of features. Feature induction is run periodically during training. The features are added to improve performance on the mislabeled instances, with the specific scoring criterion given by the FeatureInducer specified by gainName
        Parameters:
        training - The training Instances.
        validation - The validation Instances.
        testing - The testing instances.
        eval - For evaluation during training.
        numIterations - Maximum number of Maximizer iterations.
        numIterationsBetweenFeatureInductions - Number of maximizer iterations between each call to the Feature Inducer.
        numFeatureInductions - Maximum number of rounds of feature induction.
        numFeaturesPerFeatureInduction - Maximum number of features to induce at each round of induction.
        trueLabelProbThreshold - If the model's probability of the true Label of an Instance is less than this value, it is added as an error instance to the FeatureInducer.
        clusteredFeatureInduction - If true, a separate FeatureInducer is constructed for each label pair. This can avoid inducing a disproportionate number of features for a single label.
        trainingProportions - If non-null, train on increasingly larger portions of the data (e.g. [0.2, 0.5, 1.0]. This can sometimes speedup convergence.
        gainName - The type of FeatureInducer to use. One of "exp", "grad", or "info" for ExpGain, GradientGain, or InfoGain.
        Returns:
        True if training has converged.

      • setUseHyperbolicPrior

        public void setUseHyperbolicPrior​(boolean f)

      • setHyperbolicPriorSlope

        public void setHyperbolicPriorSlope​(double p)

      • setHyperbolicPriorSharpness

        public void setHyperbolicPriorSharpness​(double p)

      • getUseHyperbolicPriorSlope

        public double getUseHyperbolicPriorSlope()

      • getUseHyperbolicPriorSharpness

        public double getUseHyperbolicPriorSharpness()

      • setGaussianPriorVariance

        public void setGaussianPriorVariance​(double p)

      • getGaussianPriorVariance

        public double getGaussianPriorVariance()

      • setUseSparseWeights

        public void setUseSparseWeights​(boolean b)

      • getUseSparseWeights

        public boolean getUseSparseWeights()

      • setUseSomeUnsupportedTrick

        public void setUseSomeUnsupportedTrick​(boolean b)
        Sets whether to use the 'some unsupported trick.' This trick is, if training a CRF where some training has been done and sparse weights are used, to add a few weights for feaures that do not occur in the tainig data.

        This generally leads to better accuracy at only a small memory cost.

        Parameters:
        b - Whether to use the trick