opennlp-ml/src/main/java/org/apache/opennlp/ml/perceptron/SimplePerceptronSequenceTrainer.java

/*
 * Licensed to the Apache Software Foundation (ASF) under one
 * or more contributor license agreements.  See the NOTICE file
 * distributed with this work for additional information
 * regarding copyright ownership.  The ASF licenses this file
 * to you under the Apache License, Version 2.0 (the
 * "License"); you may not use this file except in compliance
 * with the License.  You may obtain a copy of the License at
 * 
 *   http://www.apache.org/licenses/LICENSE-2.0
 * 
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
 * KIND, either express or implied.  See the License for the
 * specific language governing permissions and limitations
 * under the License.
 */

package org.apache.opennlp.ml.perceptron;

import java.io.IOException;
import java.util.HashMap;
import java.util.Map;

import org.apache.opennlp.ml.model.AbstractModel;
import org.apache.opennlp.ml.model.DataIndexer;
import org.apache.opennlp.ml.model.Event;
import org.apache.opennlp.ml.model.IndexHashTable;
import org.apache.opennlp.ml.model.MutableContext;
import org.apache.opennlp.ml.model.OnePassDataIndexer;
import org.apache.opennlp.ml.model.Sequence;
import org.apache.opennlp.ml.model.SequenceStream;
import org.apache.opennlp.ml.model.SequenceStreamEventStream;
import org.apache.opennlp.ml.model.TwoPassDataIndexer;

/**
 * Trains models for sequences using the perceptron algorithm.  Each outcome is represented as
 * a binary perceptron classifier.  This supports standard (integer) weighting as well
 * average weighting.  Sequence information is used in a simplified was to that described in:
 * Discriminative Training Methods for Hidden Markov Models: Theory and Experiments
 * with the Perceptron Algorithm. Michael Collins, EMNLP 2002.
 * Specifically only updates are applied to tokens which were incorrectly tagged by a sequence tagger
 * rather than to all feature across the sequence which differ from the training sequence.
 */
public class SimplePerceptronSequenceTrainer {

  private boolean printMessages = true;
  private int iterations;
  private SequenceStream sequenceStream;
  /** Number of events in the event set. */
  private int numEvents;

  /** Number of predicates. */
  private int numPreds; 
  private int numOutcomes;

  /** List of outcomes for each event i, in context[i]. */
  private int[] outcomeList;
  
  private String[] outcomeLabels;

  double[] modelDistribution;

  /** Stores the average parameter values of each predicate during iteration. */
  private MutableContext[] averageParams;
  
  /** Mapping between context and an integer */ 
  private IndexHashTable<String> pmap;

  private Map<String,Integer> omap;
  
  /** Stores the estimated parameter value of each predicate during iteration. */
  private MutableContext[] params;
  private boolean useAverage;
  private int[][][] updates;
  private int VALUE = 0;
  private int ITER = 1;
  private int EVENT = 2;
  
  private int[] allOutcomesPattern;
  private String[] predLabels;
  int numSequences;

  public AbstractModel trainModel(int iterations, SequenceStream sequenceStream, int cutoff, boolean useAverage) throws IOException {
    this.iterations = iterations;
    this.sequenceStream = sequenceStream;
    DataIndexer di = new OnePassDataIndexer(new SequenceStreamEventStream(sequenceStream),cutoff,false);
    numSequences = 0;
    for (Sequence s : sequenceStream) {
      numSequences++;
    }
    outcomeList  = di.getOutcomeList();
    predLabels = di.getPredLabels();
    pmap = new IndexHashTable<String>(predLabels, 0.7d);
      
    display("Incorporating indexed data for training...  \n");
    this.useAverage = useAverage;
    numEvents = di.getNumEvents();

    this.iterations = iterations;
    outcomeLabels = di.getOutcomeLabels();
    omap = new HashMap<String,Integer>();
    for (int oli=0;oli<outcomeLabels.length;oli++) {
      omap.put(outcomeLabels[oli], oli);
    }
    outcomeList = di.getOutcomeList();

    numPreds = predLabels.length;
    numOutcomes = outcomeLabels.length;
    if (useAverage) {
      updates = new int[numPreds][numOutcomes][3];
    }
    
    display("done.\n");
    
    display("\tNumber of Event Tokens: " + numEvents + "\n");
    display("\t    Number of Outcomes: " + numOutcomes + "\n");
    display("\t  Number of Predicates: " + numPreds + "\n");
    

    params = new MutableContext[numPreds];
    if (useAverage) averageParams = new MutableContext[numPreds];
    
    allOutcomesPattern= new int[numOutcomes];
    for (int oi = 0; oi < numOutcomes; oi++) {
      allOutcomesPattern[oi] = oi;
    }
    
    for (int pi = 0; pi < numPreds; pi++) {
      params[pi]=new MutableContext(allOutcomesPattern,new double[numOutcomes]);
      if (useAverage) averageParams[pi] = new MutableContext(allOutcomesPattern,new double[numOutcomes]);
      for (int aoi=0;aoi<numOutcomes;aoi++) {
        params[pi].setParameter(aoi, 0.0);
        if (useAverage) averageParams[pi].setParameter(aoi, 0.0);
      }
    }
    modelDistribution = new double[numOutcomes];

    display("Computing model parameters...\n");
    findParameters(iterations);
    display("...done.\n");

    /*************** Create and return the model ******************/
    String[] updatedPredLabels = predLabels;
    /*
    String[] updatedPredLabels = new String[pmap.size()];
    for (String pred : pmap.keySet()) {
      updatedPredLabels[pmap.get(pred)]=pred;
    }
    */
    if (useAverage) {
      return new PerceptronModel(averageParams, updatedPredLabels, outcomeLabels);
    }
    else {
      return new PerceptronModel(params, updatedPredLabels, outcomeLabels);
    }
  }

  private void findParameters(int iterations) {
    display("Performing " + iterations + " iterations.\n");
    for (int i = 1; i <= iterations; i++) {
      if (i < 10)
        display("  " + i + ":  ");
      else if (i < 100)
        display(" " + i + ":  ");
      else
        display(i + ":  ");
      nextIteration(i);
    }
    if (useAverage) {
      trainingStats(averageParams);
    }
    else {
      trainingStats(params);
    }
  }

  private void display(String s) {
    if (printMessages)
      System.out.print(s);
  }

  public void nextIteration(int iteration) {
    iteration--; //move to 0-based index
    int numCorrect = 0;
    int oei=0;
    int si=0;
    Map<String,Float>[] featureCounts = new Map[numOutcomes];
    for (int oi=0;oi<numOutcomes;oi++) {
      featureCounts[oi] = new HashMap<String,Float>();
    }
    PerceptronModel model = new PerceptronModel(params,predLabels,pmap,outcomeLabels);
    for (Sequence sequence : sequenceStream) {
      Event[] taggerEvents = sequenceStream.updateContext(sequence, model);
      Event[] events = sequence.getEvents();
      boolean update = false;
      for (int ei=0;ei<events.length;ei++,oei++) {
        if (!taggerEvents[ei].getOutcome().equals(events[ei].getOutcome())) {
          update = true;
          //break;
        }
        else {
          numCorrect++;
        }
      }
      if (update) {
        for (int oi=0;oi<numOutcomes;oi++) {
          featureCounts[oi].clear();
        }
        //System.err.print("train:");for (int ei=0;ei<events.length;ei++) {System.err.print(" "+events[ei].getOutcome());} System.err.println();
        //training feature count computation
        for (int ei=0;ei<events.length;ei++,oei++) {
          String[] contextStrings = events[ei].getContext();
          float values[] = events[ei].getValues();
          int oi = omap.get(events[ei].getOutcome());
          for (int ci=0;ci<contextStrings.length;ci++) {
            float value = 1;
            if (values != null) {
              value = values[ci];
            }
            Float c = featureCounts[oi].get(contextStrings[ci]);
            if (c == null) {
              c = value;
            }
            else {
              c+=value;
            }
            featureCounts[oi].put(contextStrings[ci], c);
          }
        }
        //evaluation feature count computation
        //System.err.print("test: ");for (int ei=0;ei<taggerEvents.length;ei++) {System.err.print(" "+taggerEvents[ei].getOutcome());} System.err.println();
        for (Event taggerEvent : taggerEvents) {
          String[] contextStrings = taggerEvent.getContext();
          float values[] = taggerEvent.getValues();
          int oi = omap.get(taggerEvent.getOutcome());
          for (int ci = 0; ci < contextStrings.length; ci++) {
            float value = 1;
            if (values != null) {
              value = values[ci];
            }
            Float c = featureCounts[oi].get(contextStrings[ci]);
            if (c == null) {
              c = -1*value;
            }
            else {
              c-=value;
            }
            if (c == 0f) {
              featureCounts[oi].remove(contextStrings[ci]);
            }
            else {
              featureCounts[oi].put(contextStrings[ci], c);
            }
          }
        }
        for (int oi=0;oi<numOutcomes;oi++) {
          for (String feature : featureCounts[oi].keySet()) {
            int pi = pmap.get(feature);
            if (pi != -1) {
              //System.err.println(si+" "+outcomeLabels[oi]+" "+feature+" "+featureCounts[oi].get(feature));
              params[pi].updateParameter(oi, featureCounts[oi].get(feature));
              if (useAverage) {
                if (updates[pi][oi][VALUE] != 0) {
                  averageParams[pi].updateParameter(oi,updates[pi][oi][VALUE]*(numSequences*(iteration-updates[pi][oi][ITER])+(si-updates[pi][oi][EVENT])));
                  //System.err.println("p avp["+pi+"]."+oi+"="+averageParams[pi].getParameters()[oi]);
                }
                //System.err.println("p updates["+pi+"]["+oi+"]=("+updates[pi][oi][ITER]+","+updates[pi][oi][EVENT]+","+updates[pi][oi][VALUE]+") + ("+iteration+","+oei+","+params[pi].getParameters()[oi]+") -> "+averageParams[pi].getParameters()[oi]);
                updates[pi][oi][VALUE] = (int) params[pi].getParameters()[oi];
                updates[pi][oi][ITER] = iteration;
                updates[pi][oi][EVENT] = si;
              }
            }
          }
        }
        model = new PerceptronModel(params,predLabels,pmap,outcomeLabels);
      }
      si++;
    }
    //finish average computation
    double totIterations = (double) iterations*si;
    if (useAverage && iteration == iterations-1) {
      for (int pi = 0; pi < numPreds; pi++) {
        double[] predParams = averageParams[pi].getParameters();
        for (int oi = 0;oi<numOutcomes;oi++) {
          if (updates[pi][oi][VALUE] != 0) {
            predParams[oi] +=  updates[pi][oi][VALUE]*(numSequences*(iterations-updates[pi][oi][ITER])-updates[pi][oi][EVENT]);
          }
          if (predParams[oi] != 0) {
            predParams[oi] /=totIterations;  
            averageParams[pi].setParameter(oi, predParams[oi]);
            //System.err.println("updates["+pi+"]["+oi+"]=("+updates[pi][oi][ITER]+","+updates[pi][oi][EVENT]+","+updates[pi][oi][VALUE]+") + ("+iterations+","+0+","+params[pi].getParameters()[oi]+") -> "+averageParams[pi].getParameters()[oi]);
          }
        }
      }
    }
    display(". ("+numCorrect+"/"+numEvents+") "+((double) numCorrect / numEvents) + "\n");
  }
  
  private void trainingStats(MutableContext[] params) {
    int numCorrect = 0;
    int oei=0;
    for (Sequence sequence : sequenceStream) {
      Event[] taggerEvents = sequenceStream.updateContext(sequence, new PerceptronModel(params,predLabels,pmap,outcomeLabels));
      for (int ei=0;ei<taggerEvents.length;ei++,oei++) {
        int max = omap.get(taggerEvents[ei].getOutcome());
        if (max == outcomeList[oei]) {
          numCorrect ++;
        }
      }
    }
    display(". ("+numCorrect+"/"+numEvents+") "+((double) numCorrect / numEvents) + "\n");
  }
}