Training evaluation break in gradient descent weight learning.

linqs · Sep 8, 2023 · 5c49c87 · 5c49c87
1 parent 37fa46d
commit 5c49c87
Show file tree

Hide file tree

Showing 3 changed files with 47 additions and 5 deletions.
diff --git a/...ore/src/main/java/org/linqs/psl/application/learning/weight/gradient/GradientDescent.java b/...ore/src/main/java/org/linqs/psl/application/learning/weight/gradient/GradientDescent.java
@@ -20,8 +20,6 @@
 import org.linqs.psl.application.inference.InferenceApplication;
 import org.linqs.psl.application.learning.weight.WeightLearningApplication;
 import org.linqs.psl.application.learning.weight.gradient.batchgenerator.BatchGenerator;
-import org.linqs.psl.application.learning.weight.gradient.batchgenerator.ConnectedComponentBatchGenerator;
-import org.linqs.psl.application.learning.weight.gradient.batchgenerator.FullBatchGenerator;
 import org.linqs.psl.config.Options;
 import org.linqs.psl.database.AtomStore;
 import org.linqs.psl.database.Database;
@@ -78,6 +76,11 @@ public static enum GDExtension {
     protected List<DeepModelPredicate> trainFullDeepModelPredicates;
     protected TermState[] trainFullMAPTermState;
     protected float[] trainFullMAPAtomValueState;
+    double currentFullMAPEvaluationMetric;
+    double bestFullMAPEvaluationMetric;
+    protected boolean fullMAPEvaluationBreak;
+    protected int fullMAPEvaluationPatience;
+    protected int lastFullMAPImprovementEpoch;
 
     protected TermState[] trainMAPTermState;
     protected float[] trainMAPAtomValueState;
@@ -137,6 +140,11 @@ public GradientDescent(List<Rule> rules, Database trainTargetDatabase, Database
         trainFullDeepModelPredicates = null;
         trainFullMAPTermState = null;
         trainFullMAPAtomValueState = null;
+        currentFullMAPEvaluationMetric = Double.NEGATIVE_INFINITY;
+        bestFullMAPEvaluationMetric = Double.NEGATIVE_INFINITY;
+        fullMAPEvaluationBreak = Options.WLA_GRADIENT_DESCENT_FULL_MAP_EVALUATION_BREAK.getBoolean();
+        fullMAPEvaluationPatience = Options.WLA_GRADIENT_DESCENT_FULL_MAP_EVALUATION_PATIENCE.getInt();
+        lastFullMAPImprovementEpoch = 0;
 
         trainMAPTermState = null;
         trainMAPAtomValueState = null;
@@ -282,12 +290,17 @@ protected void initForLearning() {
             deepPredicate.predictDeepModel(true);
         }
 
+        currentFullMAPEvaluationMetric = Double.NEGATIVE_INFINITY;
+        bestFullMAPEvaluationMetric = Double.NEGATIVE_INFINITY;
+        lastFullMAPImprovementEpoch = 0;
+
         bestValidationWeights = new float[mutableRules.size()];
         for (int i = 0; i < mutableRules.size(); i++) {
             bestValidationWeights[i] = mutableRules.get(i).getWeight();
         }
         currentValidationEvaluationMetric = Double.NEGATIVE_INFINITY;
         bestValidationEvaluationMetric = Double.NEGATIVE_INFINITY;
+        lastValidationImprovementEpoch = 0;
 
         trainMAPTermState = trainFullMAPTermState;
         trainMAPAtomValueState = trainFullMAPAtomValueState;
@@ -309,7 +322,7 @@ protected void doLearn() {
             }
 
             if (log.isTraceEnabled() && (evaluation != null) && (epoch % trainingEvaluationComputePeriod == 0)) {
-                runMAPEvaluation();
+                runMAPEvaluation(epoch);
                 log.trace("MAP State Training Evaluation Metric: {}", evaluation.getNormalizedRepMetric());
             }
 
@@ -385,7 +398,7 @@ protected void doLearn() {
             if (saveBestValidationWeights) {
                 finalMAPStateEvaluation = bestValidationEvaluationMetric;
             } else {
-                runMAPEvaluation();
+                runMAPEvaluation(epoch);
                 finalMAPStateEvaluation = evaluation.getNormalizedRepMetric();
             }
             log.info("Final MAP State Evaluation Metric: {}", finalMAPStateEvaluation);
@@ -502,14 +515,21 @@ protected void setValidationModel() {
         }
     }
 
-    protected void runMAPEvaluation() {
+    protected void runMAPEvaluation(int epoch) {
         setFullTrainModel();
 
         // Compute the MAP state before evaluating so variables have assigned values.
         log.trace("Running MAP Inference.");
         computeMAPStateWithWarmStart(trainInferenceApplication, trainMAPTermState, trainMAPAtomValueState);
 
         evaluation.compute(trainingMap);
+        currentFullMAPEvaluationMetric = evaluation.getNormalizedRepMetric();
+
+        if (currentFullMAPEvaluationMetric > bestFullMAPEvaluationMetric) {
+            lastFullMAPImprovementEpoch = epoch;
+
+            bestFullMAPEvaluationMetric = currentFullMAPEvaluationMetric;
+        }
 
         // Evaluate the deep predicates. This calls predict with learning set to false.
         for (DeepPredicate deepPredicate : deepPredicates) {
@@ -556,6 +576,11 @@ protected boolean breakOptimization(int epoch) {
             return false;
         }
 
+        if (fullMAPEvaluationBreak && (epoch - lastFullMAPImprovementEpoch) > fullMAPEvaluationPatience) {
+            log.trace("Breaking Weight Learning. No improvement in training evaluation metric for {} epochs.", (epoch - lastFullMAPImprovementEpoch));
+            return true;
+        }
+
         if (validationBreak && (epoch - lastValidationImprovementEpoch) > validationPatience) {
             log.trace("Breaking Weight Learning. No improvement in validation evaluation metric for {} epochs.", (epoch - lastValidationImprovementEpoch));
             return true;

diff --git a/...src/main/java/org/linqs/psl/application/learning/weight/gradient/minimizer/Minimizer.java b/...src/main/java/org/linqs/psl/application/learning/weight/gradient/minimizer/Minimizer.java
@@ -320,6 +320,11 @@ protected boolean breakOptimization(int epoch) {
             return false;
         }
 
+        if (fullMAPEvaluationBreak && (epoch - lastFullMAPImprovementEpoch) > fullMAPEvaluationPatience) {
+            log.trace("Breaking Weight Learning. No improvement in training evaluation metric for {} epochs.", (epoch - lastFullMAPImprovementEpoch));
+            return true;
+        }
+
         if (validationBreak && (epoch - lastValidationImprovementEpoch) > validationPatience) {
             log.trace("Breaking Weight Learning. No improvement in validation evaluation metric for {} epochs.", (epoch - lastValidationImprovementEpoch));
             return true;

diff --git a/psl-core/src/main/java/org/linqs/psl/config/Options.java b/psl-core/src/main/java/org/linqs/psl/config/Options.java
@@ -419,6 +419,18 @@ public class Options {
         "Compute training evaluation every this many iterations of gradient descent weight learning."
     );
 
+    public static final Option WLA_GRADIENT_DESCENT_FULL_MAP_EVALUATION_BREAK = new Option(
+        "gradientdescent.trainingevaluationbreak",
+        false,
+        "Break gradient descent weight learning when the training evaluation stops improving."
+    );
+
+    public static final Option WLA_GRADIENT_DESCENT_FULL_MAP_EVALUATION_PATIENCE = new Option(
+        "gradientdescent.trainingevaluationpatience",
+        25,
+        "Break gradient descent weight learning when the training evaluation stops improving after this many epochs."
+    );
+
     public static final Option WLA_GRADIENT_DESCENT_VALIDATION_BREAK = new Option(
         "gradientdescent.validationbreak",
         false,