Started gibbs sampling for pr(Z|Yi)

ajaynagesh · Nov 19, 2013 · 49a3701 · 49a3701
1 parent aa0a28d
commit 49a3701
Showing 1 changed file with 40 additions and 13 deletions.
diff --git a/src/edu/stanford/nlp/kbp/slotfilling/classify/SelPrefORExtractor.java b/src/edu/stanford/nlp/kbp/slotfilling/classify/SelPrefORExtractor.java
@@ -1,13 +1,10 @@
 package edu.stanford.nlp.kbp.slotfilling.classify;
 
-import edu.stanford.nlp.classify.Dataset;
 import edu.stanford.nlp.ie.machinereading.structure.RelationMention;
 import edu.stanford.nlp.kbp.slotfilling.KBPTrainer;
-import edu.stanford.nlp.kbp.slotfilling.classify.HoffmannExtractor.LabelWeights;
 import edu.stanford.nlp.kbp.slotfilling.common.Constants;
 import edu.stanford.nlp.kbp.slotfilling.common.Log;
 import edu.stanford.nlp.kbp.slotfilling.common.Props;
-import edu.stanford.nlp.sequences.SequenceGibbsSampler;
 import edu.stanford.nlp.stats.ClassicCounter;
 import edu.stanford.nlp.stats.Counter;
 import edu.stanford.nlp.util.ErasureUtils;
@@ -487,6 +484,18 @@ void generateYZTPredicted(List<Counter<Integer>> ys, List<Counter<Integer>> zs,
   List<Counter<Integer>> ComputePrZ(int [][] crtGroup) {
 	  List<Counter<Integer>> prZs = estimateZ(crtGroup);
 
+	  for(Counter<Integer> pr_z : prZs){
+		  double scoreTotal = 0.0;
+		  for(double score : pr_z.values())
+			  scoreTotal += Math.exp(score);
+
+		  for(int z : pr_z.keySet()){
+			  double score = Math.exp(pr_z.getCount(z));
+			  pr_z.setCount(z, score/scoreTotal);
+		  }
+
+	  }
+
 	  return prZs;
   }
 
@@ -507,20 +516,18 @@ private void trainJointly(
 		  // 1. estimate Pr(Z) .. for now estimating \hat{Z}
 		  List<Counter<Integer>> pr_z = ComputePrZ(crtGroup);
 		  zPredicted = generateZPredicted(pr_z);
+
 		  // 2. estimate Pr(Y|Z,T)
 		  Counter<Integer> pr_y = ComputePrY_ZiTi(zPredicted, arg1Type, arg2Type, goldPos, yLabels);
-
-		  Counter<Integer> yPredicted = null;
-
-		  yPredicted = generateYPredicted(pr_y, yLabels, 0.01); // TODO: temporary hack .. need to parameterize this
+		  Counter<Integer> yPredicted = generateYPredicted(pr_y, yLabels, 0.01); // TODO: temporary hack .. need to parameterize this
 
 		  Set<Integer> [] zUpdate;
 
 		  if(updateCondition(yPredicted.keySet(), goldPos)){
 
 			  zUpdate = generateZUpdate(goldPos, pr_z);
-			  updateZModel(zUpdate, zPredicted, crtGroup, posUpdateStats, negUpdateStats);
-		  }  
+			  //updateZModel(zUpdate, zPredicted, crtGroup, posUpdateStats, negUpdateStats);
+		  }
 	  }
 
 	  else if(ALGO_TYPE == 2){
@@ -627,11 +634,31 @@ private static boolean updateCondition(Set<Integer> y, Set<Integer> yPos) {
    */
   private Set<Integer> [] generateZUpdate(
           Set<Integer> goldPos,
-          List<Counter<Integer>> zs) {
-    Set<Integer> [] zUpdate = ErasureUtils.uncheckedCast(new Set[zs.size()]);
-
+          List<Counter<Integer>> pr_z) {
+
+	  /**
+	   * while (not converged) {
+	   * 	choose a random permutation for Z variable Pi
+	   * 	for (j = 1 to |Z|) {
+	   * 		update Z_{Pi_j} given the rest is fixed
+	   * 	}
+	   * }
+	   */
+
+	  Set<Integer> [] zUpdate = null; //ErasureUtils.uncheckedCast(new Set[zs.size()]);
+
+	  Counter<Integer> yLabels = new ClassicCounter<Integer>();
+	  for(int y : goldPos){
+		  yLabels.incrementCount(y);
+	  }
+
+
+
+	  for(int i = 0; i < epochsInf; i++){
+
+	  }
 
-    return zUpdate;
+	  return zUpdate;
   }
 
   /**