Package cc.mallet.fst

Source Code of cc.mallet.fst.SumLatticeDefault

package cc.mallet.fst;

import java.io.IOException;
import java.io.ObjectInputStream;
import java.io.ObjectOutputStream;
import java.io.Serializable;
import java.util.logging.Level;
import java.util.logging.Logger;

import cc.mallet.fst.Transducer.State;
import cc.mallet.fst.Transducer.TransitionIterator;
import cc.mallet.types.DenseVector;
import cc.mallet.types.LabelAlphabet;
import cc.mallet.types.LabelVector;
import cc.mallet.types.MatrixOps;
import cc.mallet.types.Sequence;
import cc.mallet.util.MalletLogger;

/** Default, full dynamic programming implementation of the Forward-Backward "Sum-(Product)-Lattice" algorithm */
public class SumLatticeDefault implements SumLattice
{
  private static Logger logger = MalletLogger.getLogger(SumLatticeDefault.class.getName());
  //{logger.setLevel(Level.FINE);}
 
  // Static variables acting as default values for the correspondingly-named instance variables.
  // Can be overridden sort of like named parameters, like this:
  // SumLattice lattice = new SumLatticeDefault(transducer, input) {{ saveXis=true; }}
  protected static boolean saveXis = false;

  // "ip" == "input position", "op" == "output position", "i" == "state index"
  Transducer t;
  double totalWeight;
  Sequence input, output;
  LatticeNode[][] nodes;       // indexed by ip,i
  int latticeLength;
  double[][] gammas;           // indexed by ip,i
  double[][][] xis;            // indexed by ip,i,j; saved only if saveXis is true;

  LabelVector labelings[];       // indexed by op, created only if "outputAlphabet" is non-null in constructor

 
  // Ensure that instances cannot easily be created by a zero arg constructor.
  protected SumLatticeDefault() {  }

  protected LatticeNode getLatticeNode (int ip, int stateIndex)
  {
    if (nodes[ip][stateIndex] == null)
      nodes[ip][stateIndex] = new LatticeNode (ip, t.getState (stateIndex));
    return nodes[ip][stateIndex];
  }

  public SumLatticeDefault (Transducer trans, Sequence input)
  {
    this (trans, input, null, (Transducer.Incrementor)null, saveXis, null);
  }

  public SumLatticeDefault (Transducer trans, Sequence input, boolean saveXis)
  {
    this (trans, input, null, (Transducer.Incrementor)null, saveXis, null);
  }

  public SumLatticeDefault (Transducer trans, Sequence input, Transducer.Incrementor incrementor)
  {
    this (trans, input, null, incrementor, saveXis, null);
  }

  public SumLatticeDefault (Transducer trans, Sequence input, Sequence output)
  {
    this (trans, input, output, (Transducer.Incrementor)null, saveXis, null);
  }

  // You may pass null for output, meaning that the lattice
  // is not constrained to match the output
  public SumLatticeDefault (Transducer trans, Sequence input, Sequence output, Transducer.Incrementor incrementor)
  {
    this (trans, input, output, incrementor, saveXis, null);
  }
 
  public SumLatticeDefault (Transducer trans, Sequence input, Sequence output, Transducer.Incrementor incrementor, LabelAlphabet outputAlphabet)
  {
    this (trans, input, output, incrementor, saveXis, outputAlphabet);
  }

  // You may pass null for output, meaning that the lattice
  // is not constrained to match the output
  public SumLatticeDefault (Transducer trans, Sequence input, Sequence output, Transducer.Incrementor incrementor, boolean saveXis)
  {
    this (trans, input, output, incrementor, saveXis, null);
  }

  // If outputAlphabet is non-null, this will create a LabelVector
  // for each position in the output sequence indicating the
  // probability distribution over possible outputs at that time
  // index
  public SumLatticeDefault (Transducer trans, Sequence input, Sequence output, Transducer.Incrementor incrementor, boolean saveXis, LabelAlphabet outputAlphabet)
  {
    assert (output == null || input.size() == output.size());
    if (false && logger.isLoggable (Level.FINE)) {
      logger.fine ("Starting Lattice");
      logger.fine ("Input: ");
      for (int ip = 0; ip < input.size(); ip++)
        logger.fine (" " + input.get(ip));
      logger.fine ("\nOutput: ");
      if (output == null)
        logger.fine ("null");
      else
        for (int op = 0; op < output.size(); op++)
          logger.fine (" " + output.get(op));
      logger.fine ("\n");
    }

    // Initialize some structures
    this.t = trans;
    this.input = input;
    this.output = output;
    // xxx Not very efficient when the lattice is actually sparse,
    // especially when the number of states is large and the
    // sequence is long.
    latticeLength = input.size()+1;
    int numStates = t.numStates();
    nodes = new LatticeNode[latticeLength][numStates];
    // xxx Yipes, this could get big; something sparse might be better?
    gammas = new double[latticeLength][numStates];
    if (saveXis) xis = new double[latticeLength][numStates][numStates];

    double outputCounts[][] = null;
    if (outputAlphabet != null)
      outputCounts = new double[latticeLength][outputAlphabet.size()];

    for (int i = 0; i < numStates; i++) {
      for (int ip = 0; ip < latticeLength; ip++)
        gammas[ip][i] = Transducer.IMPOSSIBLE_WEIGHT;
      if (saveXis)
        for (int j = 0; j < numStates; j++)
          for (int ip = 0; ip < latticeLength; ip++)
            xis[ip][i][j] = Transducer.IMPOSSIBLE_WEIGHT;
    }

    // Forward pass
    logger.fine ("Starting Foward pass");
    boolean atLeastOneInitialState = false;
    for (int i = 0; i < numStates; i++) {
      double initialWeight = t.getState(i).getInitialWeight();
      //System.out.println ("Forward pass initialCost = "+initialCost);
      if (initialWeight > Transducer.IMPOSSIBLE_WEIGHT) {
        getLatticeNode(0, i).alpha = initialWeight;
        //System.out.println ("nodes[0][i].alpha="+nodes[0][i].alpha);
        atLeastOneInitialState = true;
      }
    }
    if (atLeastOneInitialState == false)
      logger.warning ("There are no starting states!");

    for (int ip = 0; ip < latticeLength-1; ip++)
      for (int i = 0; i < numStates; i++) {
        if (nodes[ip][i] == null || nodes[ip][i].alpha == Transducer.IMPOSSIBLE_WEIGHT)
          // xxx if we end up doing this a lot,
          // we could save a list of the non-null ones
          continue;
        State s = t.getState(i);
        TransitionIterator iter = s.transitionIterator (input, ip, output, ip);
        if (logger.isLoggable (Level.FINE))
          logger.fine (" Starting Foward transition iteration from state "
              + s.getName() + " on input " + input.get(ip).toString()
              + " and output "
              + (output==null ? "(null)" : output.get(ip).toString()));
        while (iter.hasNext()) {
          State destination = iter.nextState();
          if (logger.isLoggable (Level.FINE))
            logger.fine ("Forward Lattice[inputPos="+ip+"][source="+s.getName()+"][dest="+destination.getName()+"]");
          LatticeNode destinationNode = getLatticeNode (ip+1, destination.getIndex());
          destinationNode.output = iter.getOutput();
          double transitionWeight = iter.getWeight();
          if (logger.isLoggable (Level.FINE))
            logger.fine ("BEFORE update: destinationNode.alpha="+destinationNode.alpha);
          destinationNode.alpha = Transducer.sumLogProb (destinationNode.alpha,  nodes[ip][i].alpha + transitionWeight);
          if (logger.isLoggable (Level.FINE))
            logger.fine ("transitionWeight="+transitionWeight+" nodes["+ip+"]["+i+"].alpha="+nodes[ip][i].alpha
                +" destinationNode.alpha="+destinationNode.alpha);
          //System.out.println ("destinationNode.alpha <- "+destinationNode.alpha);
        }
      }
   
    if (logger.isLoggable (Level.FINE)) {
      logger.fine("Forward Lattice:");
      for (int ip = 0; ip < latticeLength; ip++) {
        StringBuffer sb = new StringBuffer();
        for (int i = 0; i < numStates; i++)
          sb.append (" "+(nodes[ip][i] == null ? "<null>" : nodes[ip][i].alpha));
        logger.fine(sb.toString());
      }
    }

   
    // Calculate total weight of Lattice.  This is the normalizer
    totalWeight = Transducer.IMPOSSIBLE_WEIGHT;
    for (int i = 0; i < numStates; i++)
      if (nodes[latticeLength-1][i] != null) {
        //System.out.println ("Ending alpha, state["+i+"] = "+nodes[latticeLength-1][i].alpha);
        //System.out.println ("Ending beta,  state["+i+"] = "+t.getState(i).getFinalWeight());
        totalWeight = Transducer.sumLogProb (totalWeight,  (nodes[latticeLength-1][i].alpha + t.getState(i).getFinalWeight()));
      }
    logger.fine ("totalWeight="+totalWeight);
    // totalWeight is now an "unnormalized weight" of the entire Lattice

    // If the sequence has -infinite weight, just return.
    // Usefully this avoids calling any incrementX methods.
    // It also relies on the fact that the gammas[][] and .alpha (but not .beta) values
    // are already initialized to values that reflect -infinite weight
    // TODO Is it important to fill in the betas before we return?
    if (totalWeight == Transducer.IMPOSSIBLE_WEIGHT)
      return;

    // Backward pass
    for (int i = 0; i < numStates; i++)
      if (nodes[latticeLength-1][i] != null) {
        State s = t.getState(i);
        nodes[latticeLength-1][i].beta = s.getFinalWeight();
        gammas[latticeLength-1][i] = nodes[latticeLength-1][i].alpha + nodes[latticeLength-1][i].beta - totalWeight;
        if (incrementor != null) {
          double p = Math.exp(gammas[latticeLength-1][i]);
          // gsc: reducing from 1e-10 to 1e-6
          // gsc: removing the isNaN check, range check will catch the NaN error as well
          // assert (p >= 0.0 && p <= 1.0+1e-10 && !Double.isNaN(p)) : "p="+p+" gamma="+gammas[latticeLength-1][i];
          assert (p >= 0.0 && p <= 1.0+1e-6) : "p="+p+", gamma="+gammas[latticeLength-1][i];
          incrementor.incrementFinalState (s, p);
        }
      }

    for (int ip = latticeLength-2; ip >= 0; ip--) {
      for (int i = 0; i < numStates; i++) {
        if (nodes[ip][i] == null || nodes[ip][i].alpha == Transducer.IMPOSSIBLE_WEIGHT)
          // Note that skipping here based on alpha means that beta values won't
          // be correct, but since alpha is infinite anyway, it shouldn't matter.
          continue;
        State s = t.getState(i);
        TransitionIterator iter = s.transitionIterator (input, ip, output, ip);
        while (iter.hasNext()) {
          State destination = iter.nextState();
          if (logger.isLoggable (Level.FINE))
            logger.fine ("Backward Lattice[inputPos="+ip+"][source="+s.getName()+"][dest="+destination.getName()+"]");
          int j = destination.getIndex();
          LatticeNode destinationNode = nodes[ip+1][j];
          if (destinationNode != null) {
            double transitionWeight = iter.getWeight();
            assert (!Double.isNaN(transitionWeight));
            double oldBeta = nodes[ip][i].beta;
            assert (!Double.isNaN(nodes[ip][i].beta));
            nodes[ip][i].beta = Transducer.sumLogProb (nodes[ip][i].beta,  destinationNode.beta + transitionWeight);
            assert (!Double.isNaN(nodes[ip][i].beta))
            : "dest.beta="+destinationNode.beta+" trans="+transitionWeight+" sum="+(destinationNode.beta+transitionWeight+ " oldBeta="+oldBeta;
            double xi = nodes[ip][i].alpha + transitionWeight + nodes[ip+1][j].beta - totalWeight;
            if (saveXis) xis[ip][i][j] = xi;
            assert (!Double.isNaN(nodes[ip][i].alpha));
            assert (!Double.isNaN(transitionWeight));
            assert (!Double.isNaN(nodes[ip+1][j].beta));
            assert (!Double.isNaN(totalWeight));
            if (incrementor != null || outputAlphabet != null) {
              double p = Math.exp(xi);
              // gsc: reducing from 1e-10 to 1e-6
              // gsc: removing the isNaN check, range check will catch the NaN error as well
              // assert (p >= 0.0 && p <= 1.0+1e-10 && !Double.isNaN(p)) : "xis["+ip+"]["+i+"]["+j+"]="+xi;
              assert (p >= 0.0 && p <= 1.0+1e-6) : "p="+p+", xis["+ip+"]["+i+"]["+j+"]="+xi;
              if (incrementor != null)
                incrementor.incrementTransition(iter, p);
              if (outputAlphabet != null) {
                int outputIndex = outputAlphabet.lookupIndex (iter.getOutput(), false);
                assert (outputIndex >= 0);
                // xxx This assumes that "ip" == "op"!
                outputCounts[ip][outputIndex] += p;
                //System.out.println ("CRF Lattice outputCounts["+ip+"]["+outputIndex+"]+="+p);
              }
            }
          }
        }
        gammas[ip][i] = nodes[ip][i].alpha + nodes[ip][i].beta - totalWeight;
      }
    }
    if (incrementor != null)
      for (int i = 0; i < numStates; i++) {
        double p = Math.exp(gammas[0][i]);
        // gsc: reducing from 1e-10 to 1e-6
        // gsc: removing the isNaN check, range check will catch the NaN error as well
        // assert (p >= 0.0 && p <= 1.0+1e-10 && !Double.isNaN(p)) : "p="+p;
        assert (p >= 0.0 && p <= 1.0+1e-6) : "p="+p;
        incrementor.incrementInitialState(t.getState(i), p);
      }
    if (outputAlphabet != null) {
      labelings = new LabelVector[latticeLength];
      for (int ip = latticeLength-2; ip >= 0; ip--) {
        assert (Math.abs(1.0-MatrixOps.sum (outputCounts[ip])) < 0.000001);;
        labelings[ip] = new LabelVector (outputAlphabet, outputCounts[ip]);
      }
    }
   
    if (logger.isLoggable (Level.FINE)) {
      logger.fine("Lattice:");
      for (int ip = 0; ip < latticeLength; ip++) {
        StringBuffer sb = new StringBuffer();
        for (int i = 0; i < numStates; i++)
          sb.append (" "+gammas[ip][i]);
        logger.fine(sb.toString());
      }
    }
  }



  public double[][][] getXis(){
    return xis;
  }

  public double[][] getGammas(){
    return gammas;
  }

  public double getTotalWeight () {
    assert (!Double.isNaN(totalWeight));
    return totalWeight; }

  public double getGammaWeight(int inputPosition, State s) {
    return gammas[inputPosition][s.getIndex()]; }

  public double getGammaWeight(int inputPosition, int stateIndex) {
    return gammas[inputPosition][stateIndex]; }

  public double getGammaProbability (int inputPosition, State s) {
    return Math.exp (gammas[inputPosition][s.getIndex()]); }

  public double getGammaProbability (int inputPosition, int stateIndex) {
    return Math.exp (gammas[inputPosition][stateIndex]); }

  public double getXiProbability (int ip, State s1, State s2) {
    if (xis == null)
      throw new IllegalStateException ("xis were not saved.");
    int i = s1.getIndex ();
    int j = s2.getIndex ();
    return Math.exp (xis[ip][i][j]);
  }

  public double getXiWeight(int ip, State s1, State s2)
  {
    if (xis == null)
      throw new IllegalStateException ("xis were not saved.");

    int i = s1.getIndex ();
    int j = s2.getIndex ();
    return xis[ip][i][j];
  }

  public int length () { return latticeLength; }

  public Sequence getInput() {
    return input;
  }
 
  public double getAlpha (int ip, State s) {
    LatticeNode node = getLatticeNode (ip, s.getIndex ());
    return node.alpha;
  }

  public double getBeta (int ip, State s) {
    LatticeNode node = getLatticeNode (ip, s.getIndex ());
    return node.beta;
  }

  public LabelVector getLabelingAtPosition (int outputPosition)  {
    if (labelings != null)
      return labelings[outputPosition];
    return null;
  }

  public Transducer getTransducer ()
  {
    return t;
  }


  // A container for some information about a particular input position and state
  protected class LatticeNode
  {
    int inputPosition;
    // outputPosition not really needed until we deal with asymmetric epsilon.
    State state;
    Object output;
    double alpha = Transducer.IMPOSSIBLE_WEIGHT;
    double beta = Transducer.IMPOSSIBLE_WEIGHT;
    LatticeNode (int inputPosition, State state)  {
      this.inputPosition = inputPosition;
      this.state = state;
      assert (this.alpha == Transducer.IMPOSSIBLE_WEIGHT)// xxx Remove this check
    }
  }
 
  public static class Factory extends SumLatticeFactory implements Serializable
  {
    public SumLattice newSumLattice (Transducer trans, Sequence input, Sequence output,
        Transducer.Incrementor incrementor, boolean saveXis, LabelAlphabet outputAlphabet)
    {
      return new SumLatticeDefault (trans, input, output, incrementor, saveXis, outputAlphabet);
    }

    private static final long serialVersionUID = 1;
    private static final int CURRENT_SERIAL_VERSION = 1;

    private void writeObject(ObjectOutputStream out) throws IOException {
      out.writeInt(CURRENT_SERIAL_VERSION);
    }
    private void readObject(ObjectInputStream in) throws IOException, ClassNotFoundException {
      int version = in.readInt();
    }

  }
 

}
TOP

Related Classes of cc.mallet.fst.SumLatticeDefault

TOP
Copyright © 2018 www.massapi.com. All rights reserved.
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.