Examples of cc.mallet.fst.CRFTrainerByLabelLikelihood

cc.mallet.fst.CRFTrainerByLabelLikelihood
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.

    InstanceList instances = new InstanceList(p);
    instances.addThruPipe(new ArrayIterator(data));


    CRF crf1 = new CRF(p.getDataAlphabet(), p.getTargetAlphabet());
    crf1.addFullyConnectedStatesForLabels();
    CRFTrainerByLabelLikelihood crft1 = new CRFTrainerByLabelLikelihood(
        crf1);
    crft1.trainIncremental(instances);


    CRF crf2 = new CRF(p.getDataAlphabet(), p.getTargetAlphabet());
    crf2.addFullyConnectedStatesForLabels();
    // Freeze some weights, before training
    for (int i = 0; i < crf2.getWeights().length; i += 2)
      crf2.freezeWeights(i);
    CRFTrainerByLabelLikelihood crft2 = new CRFTrainerByLabelLikelihood(
        crf2);
    crft2.trainIncremental(instances);


    SparseVector[] w = crf2.getWeights();
    double[] b = crf2.getDefaultWeights();
    for (int i = 0; i < w.length; i += 2) {
      assertEquals(0.0, b[i], 1e-10);
      for (int loc = 0; loc < w[i].numLocations(); loc++) {
        assertEquals(0.0, w[i].valueAtLocation(loc), 1e-10);
      }
    }


    // Check that the frozen weights has worse likelihood
    Optimizable.ByGradientValue optable1 = crft1
        .getOptimizableCRF(instances);
    Optimizable.ByGradientValue optable2 = crft2
        .getOptimizableCRF(instances);
    double val1 = optable1.getValue();
    double val2 = optable2.getValue();
    assertTrue(
        "Error: Freezing weights does not harm log-likelihood!  Full "

View Full Code Here

    InstanceList[] lists = instances.split(new Random(777), new double[] {
        .5, .5 });


    CRF crf = new CRF(p.getDataAlphabet(), p.getTargetAlphabet());
    crf.addFullyConnectedStatesForLabels();
    CRFTrainerByLabelLikelihood crft = new CRFTrainerByLabelLikelihood(crf);
    crft.setUseSparseWeights(true);


    crft.trainIncremental(lists[0]);


    TokenAccuracyEvaluator eval = new TokenAccuracyEvaluator(lists,
        new String[] { "Train", "Test" });
    eval.evaluateInstanceList(crft, lists[1], "Test");

View Full Code Here

    InstanceList one = new InstanceList(p);
    String[] data = new String[] { "ABCDE", };
    one.addThruPipe(new ArrayIterator(data));
    CRF crf = new CRF(p, null);
    crf.addFullyConnectedStatesForThreeQuarterLabels(one);
    CRFTrainerByLabelLikelihood crft = new CRFTrainerByLabelLikelihood(crf);
    crf.setWeightsDimensionAsIn(one, false);
    Optimizable mcrf = crft.getOptimizableCRF(one);
    double[] params = new double[mcrf.getNumParameters()];
    for (int i = 0; i < params.length; i++) {
      params[i] = i;
    }
    mcrf.setParameters(params);

View Full Code Here

    InstanceList one = new InstanceList(p);
    String[] data = new String[] { "ABCDE", };
    one.addThruPipe(new ArrayIterator(data));
    CRF crf = new CRF(p, null);
    crf.addFullyConnectedStatesForLabels();
    CRFTrainerByLabelLikelihood crft = new CRFTrainerByLabelLikelihood(crf);
    crf.setWeightsDimensionAsIn(one, false);
    Optimizable.ByGradientValue mcrf = crft.getOptimizableCRF(one);
    double[] params = new double[mcrf.getNumParameters()];
    for (int i = 0; i < params.length; i++) {
      params[i] = i;
    }
    mcrf.setParameters(params);


    StringWriter out = new StringWriter();
    crf.print(new PrintWriter(out, true));
    System.out.println("------------- CRF1 -------------");
    crf.print();


    // Make a copy of this CRF
    CRF crf2 = new CRF(crf);


    StringWriter out2 = new StringWriter();
    crf2.print(new PrintWriter(out2, true));
    System.out.println("------------- CRF2 -------------");
    crf2.print();


    assertEquals(out.toString(), out2.toString());


    double val1 = mcrf.getValue();
    CRFTrainerByLabelLikelihood crft2 = new CRFTrainerByLabelLikelihood(
        crf2);
    double val2 = crft2.getOptimizableCRF(one).getValue();
    assertEquals(val1, val2, 1e-5);
  }

View Full Code Here


    CRF crf = new CRF(p, null);
    crf.print();
    crf.addStatesForLabelsConnectedAsIn(data);
    crf.addStartState();
    CRFTrainerByLabelLikelihood crft = new CRFTrainerByLabelLikelihood(crf);


    Optimizable.ByGradientValue maxable = crft.getOptimizableCRF(data);
    assertEquals(-1.3862, maxable.getValue(), 1e-4);


    crf = new CRF(p, null);
    crf
        .addOrderNStates(data, new int[] { 1 }, null, "A", null, null,
            false);
    crf.print();
    crft = new CRFTrainerByLabelLikelihood(crf);


    maxable = crft.getOptimizableCRF(data);
    assertEquals(-3.09104245335831, maxable.getValue(), 1e-4);
  }

View Full Code Here

    // Test that dense observations wights aren't added for
    // "default-feature" edges.
    CRF crf1 = new CRF(p, null);
    crf1.addOrderNStates(instances, new int[] { 0 }, null, "start", null,
        null, true);
    CRFTrainerByLabelLikelihood crft1 = new CRFTrainerByLabelLikelihood(
        crf1);
    crft1.setUseSparseWeights(false);
    crft1.train(instances, 1); // Set weights dimension
    int nParams1 = crft1.getOptimizableCRF(instances).getNumParameters();


    CRF crf2 = new CRF(p, null);
    crf2.addOrderNStates(instances, new int[] { 0, 1 }, new boolean[] {
        false, true }, "start", null, null, true);
    CRFTrainerByLabelLikelihood crft2 = new CRFTrainerByLabelLikelihood(
        crf2);
    crft2.setUseSparseWeights(false);
    crft2.train(instances, 1); // Set weights dimension
    int nParams2 = crft2.getOptimizableCRF(instances).getNumParameters();


    assertEquals(nParams2, nParams1 + 4);


  }

View Full Code Here

    InstanceList instances = new InstanceList(p);
    instances.addThruPipe(new ArrayIterator(data));


    CRF crf1 = new CRF(p, null);
    crf1.addFullyConnectedStatesForLabels();
    CRFTrainerByLabelLikelihood crft1 = new CRFTrainerByLabelLikelihood(
        crf1);
    crft1.train(instances, 10); // Let's get some parameters


    Instance inst = instances.get(0);
    Sequence input = (Sequence) inst.getData();
    SumLatticeDefault lattice = new SumLatticeDefault(crf1, input,
        (Sequence) inst.getTarget(), null, true);

View Full Code Here

    // don't know why -akm
    // 12/2007


    CRF crf = new CRF(p, null);
    crf.addFullyConnectedStatesForLabels();
    CRFTrainerByLabelLikelihood crft = new CRFTrainerByLabelLikelihood(crf);
    crft.trainIncremental(training);


    // Check that the notstart state is used at test time
    Sequence input = (Sequence) training.get(0).getData();
    Sequence output = new MaxLatticeDefault(crf, input)
        .bestOutputSequence();

View Full Code Here

0 1

TOP

Related Classes of cc.mallet.fst.CRFTrainerByLabelLikelihood

cc.mallet.extract.test.TestDocumentViewer

cc.mallet.extract.test.TestLatticeViewer

cc.mallet.fst.semi_supervised.CRFTrainerByLikelihoodAndGE

cc.mallet.fst.tests.TestCRF

cc.mallet.optimize.LimitedMemoryBFGS

cc.mallet.types.FeatureInducer

cc.mallet.types.FeatureSelection

cc.mallet.types.Instance

cc.mallet.types.InstanceList

cc.mallet.types.Label

All source code are property of their respective owners. Java is a trademark of Sun Microsystems, Inc and owned by ORACLE Inc. Contact coftware#gmail.com.