Merge branch 'master' of https://github.com/automl/fanova.git

automl · Jun 17, 2015 · cd933ee · cd933ee
2 parents 861d3d4 + ca7b249
commit cd933ee
Show file tree

Hide file tree

Showing 4 changed files with 41 additions and 8 deletions.
diff --git a/fanova/src/net/aeatk/fanova/model/FunctionalANOVARunner.java b/fanova/src/net/aeatk/fanova/model/FunctionalANOVARunner.java
@@ -67,6 +67,17 @@ public static void decomposeVariance(RandomForest existingForest, List<Algorithm
 		double[][] allObservations = new double[numDim][];
 		double[][] allIntervalSizes = new double[numDim][];
 
+		//=== Loop over trees to get number of trees with positive variance.
+		double[] treeTotalVariance = new double[forest.Trees.length];
+		int numTreesWithPositiveVariance = 0;
+		for(int numTree=0; numTree<forest.Trees.length; numTree++){
+			double thisTreeTotalVariance = RandomForestPreprocessor.computeTotalVarianceOfRegressionTree(forest.Trees[numTree], configSpace);
+			treeTotalVariance[numTree] = thisTreeTotalVariance;
+			if (thisTreeTotalVariance > 0.0){
+				numTreesWithPositiveVariance++;
+			}
+		}
+
 		//=== Loop over trees.
 		for(int numTree=0; numTree<forest.Trees.length; numTree++){
 			HashSet<Integer> allVariableIndices = new HashSet<Integer>();
@@ -75,7 +86,7 @@ public static void decomposeVariance(RandomForest existingForest, List<Algorithm
 			}
 
 			//=== Get the tree's total variance (only works for marginal trees, i.e., in the absence of instance features).
-			double thisTreeTotalVariance = RandomForestPreprocessor.computeTotalVarianceOfRegressionTree(forest.Trees[numTree], configSpace);
+			double thisTreeTotalVariance = treeTotalVariance[numTree]; 
 			if (thisTreeTotalVariance == 0.0){
 				s = "Tree " + numTree + " has no variance -> skipping.";
 				log.info(s);
@@ -271,7 +282,7 @@ public static void decomposeVariance(RandomForest existingForest, List<Algorithm
 				log.debug("ThisTreeVarianceContributions of index" + indexSet.toString() + " for Tree" + numTree + " : " + thisTreeVarianceContributions.get(indexSet));
 				log.debug("ThisTreeTotalVariance for Tree" + numTree + " : " + thisTreeTotalVariance);
 				tmpThisTreeFractionExplained += thisFractionExplained;
-				totalFractionsExplained.put(indexSet, previousFractionExplained + 1.0/forest.Trees.length * thisFractionExplained);
+				totalFractionsExplained.put(indexSet, previousFractionExplained + 1.0/numTreesWithPositiveVariance * thisFractionExplained);
 				log.debug("TotalFractionExplained for Tree" + numTree + " : " + totalFractionsExplained.get(indexSet));
 
 			}	

diff --git a/fanova/src/net/aeatk/fanova/model/FunctionalANOVAVarianceDecompose.java b/fanova/src/net/aeatk/fanova/model/FunctionalANOVAVarianceDecompose.java
@@ -28,7 +28,8 @@ public class FunctionalANOVAVarianceDecompose {
 	private Vector <HashMap<Integer,Double>> singleVarianceContributions;
 	private HashMap<HashSet<Integer>,Double> thisTreeVarianceContributions = new HashMap<HashSet<Integer>,Double>(); 
 	private HashMap<HashSet<Integer>,Double> totalFractionsExplained = new HashMap<HashSet<Integer>,Double>();
-
+	private int numTreesWithPositiveVariance;
+
 	public FunctionalANOVAVarianceDecompose(RandomForest existingForest, List<AlgorithmRunResult> testRuns,
 			ParameterConfigurationSpace configSpace, Random rand, 
 			boolean compareToDef, double quantileToCompareTo, boolean logModel) throws IOException, InterruptedException 
@@ -42,11 +43,13 @@ public FunctionalANOVAVarianceDecompose(RandomForest existingForest, List<Algori
 
 		//=== Initialize variables to be incrementally updated.
 		String s;
-
+
+
 		int numDim = configSpace.getCategoricalSize().length;
 		allObservations = new double[forest.Trees.length][numDim][];
 		allIntervalSizes = new double[forest.Trees.length][numDim][];
 		thisTreeTotalVariance = new double[forest.Trees.length];
+		this.numTreesWithPositiveVariance = 0;
 		//=== Loop over trees.
 		for(int numTree=0; numTree<forest.Trees.length; numTree++){
 			HashSet<Integer> allVariableIndices = new HashSet<Integer>();
@@ -61,6 +64,9 @@ public FunctionalANOVAVarianceDecompose(RandomForest existingForest, List<Algori
 				log.info(s);
 				continue;
 			}
+			else{
+				this.numTreesWithPositiveVariance++;
+			}
 			s = "Tree " + numTree + ": Total variance of predictor: " + thisTreeTotalVariance[numTree];
 			log.info(s);
 
@@ -130,6 +136,10 @@ public double getMarginal(int dim)
 
 		for(int numTree=0; numTree<forest.Trees.length; numTree++){
 
+			if (thisTreeTotalVariance[numTree] == 0.0){
+				continue;
+			}
+
 			int[] indicesOfObservations = new int[1];
 			//=== Compute marginal predictions for each instantiation of this categorical parameter.
 			indicesOfObservations[0] = dim;
@@ -170,7 +180,7 @@ public double getMarginal(int dim)
 				previousFractionExplained = totalFractionsExplained.get(set);
 			}
 			double thisFractionExplained = thisTreeVarianceContributions.get(set)/thisTreeTotalVariance[numTree]*100;
-			totalFractionsExplained.put(set, previousFractionExplained + 1.0/forest.Trees.length * thisFractionExplained);
+			totalFractionsExplained.put(set, previousFractionExplained + 1.0/this.numTreesWithPositiveVariance* thisFractionExplained);
 		}
 		return totalFractionsExplained.get(set);
 	}
@@ -193,6 +203,10 @@ public double getPairwiseMarginal(int dim1, int dim2)
 
 		for(int numTree=0; numTree<forest.Trees.length; numTree++){
 
+			if (thisTreeTotalVariance[numTree] == 0.0){
+				continue;
+			}
+
 			int[] indicesOfObservations = new int[2];
 			indicesOfObservations[0] = dim1;
 			ArrayList<Double> as = new ArrayList<Double>();
@@ -234,7 +248,7 @@ public double getPairwiseMarginal(int dim1, int dim2)
 				previousFractionExplained = totalFractionsExplained.get(set);
 			}
 			double thisFractionExplained = thisTreeVarianceContributions.get(set)/thisTreeTotalVariance[numTree]*100;
-			totalFractionsExplained.put(set, previousFractionExplained + 1.0/forest.Trees.length * thisFractionExplained);
+			totalFractionsExplained.put(set, previousFractionExplained + 1.0/this.numTreesWithPositiveVariance * thisFractionExplained);
 
 		}
 		return totalFractionsExplained.get(set);

diff --git a/pyfanova/fanova/fanova.jar b/pyfanova/fanova/fanova.jar
diff --git a/setup.py b/setup.py
@@ -34,7 +34,7 @@ def check_java_exists():
         error_msg = """
         Java not found!
 
-        pysmac needs java in order to work. You can download java from:
+        Fanova needs java in order to work. You can download java from:
         http://java.com/getjava
         """
         raise RuntimeError(error_msg)
@@ -46,7 +46,15 @@ def check_java_exists():
     name = "pyfanova",
     version = "1.0",
     packages = find_packages(),
-    install_requires = ['numpy', 'docutils>=0.3', 'setuptools', 'matplotlib>=1.4.2'],
+    install_requires = [
+                        'numpy',
+                        'docutils>=0.3',
+                        'setuptools',
+                        'matplotlib>=1.4.2',
+                        'ParameterConfigSpace'],
+
+    dependency_links=['https://github.com/automl/ParameterConfigSpace/archive/master.zip'],
+
     author = "Tobias Domhan, Aaron Klein (python wrapper). Frank Hutter (FANOVA)",
     author_email = "kleinaa@cs.uni-freiburg.de",
     description = "Functional ANOVA: an implementation of the ICML 2014 paper 'An Efficient Approach for Assessing Hyperparameter Importance' by Frank Hutter, Holger Hoos and Kevin Leyton-Brown.",