add support for adaboost bdt classifiers, and add workaround for inco…

…nsistent behaviour of TMVA decision tree evaluation in different root versions

CondFormats/EgammaObjects/interface/GBRForest.h

@@ -36,7 +36,11 @@
        virtual ~GBRForest();
 
        double GetResponse(const float* vector) const;
-       double GetClassifier(const float* vector) const;
+       double GetGradBoostClassifier(const float* vector) const;
+       double GetAdaBoostClassifier(const float* vector) const { return GetResponse(vector); }
+
+       //for backwards-compatibility
+       double GetClassifier(const float* vector) const { return GetGradBoostClassifier(vector); }
 
        void SetInitialResponse(double response) { fInitialResponse = response; }
 
@@ -61,7 +65,7 @@ inline double GBRForest::GetResponse(const float* vector) const {
 }
 
 //_______________________________________________________________________
-inline double GBRForest::GetClassifier(const float* vector) const {
+inline double GBRForest::GetGradBoostClassifier(const float* vector) const {
   double response = GetResponse(vector);
   return 2.0/(1.0+exp(-2.0*response))-1; //MVA output between -1 and 1
 }

CondFormats/EgammaObjects/interface/GBRTree.h

@@ -38,7 +38,7 @@
     public:
 
        GBRTree();
-       explicit GBRTree(const TMVA::DecisionTree *tree);
+       explicit GBRTree(const TMVA::DecisionTree *tree, double scale, bool useyesnoleaf, bool adjustboundary);
        virtual ~GBRTree();
 
        double GetResponse(const float* vector) const;
@@ -65,7 +65,7 @@
         unsigned int CountIntermediateNodes(const TMVA::DecisionTreeNode *node);
         unsigned int CountTerminalNodes(const TMVA::DecisionTreeNode *node);
 
-        void AddNode(const TMVA::DecisionTreeNode *node);
+        void AddNode(const TMVA::DecisionTreeNode *node, double scale, bool isregression, bool useyesnoleaf, bool adjustboundary);
 
 	std::vector<unsigned char> fCutIndices;
 	std::vector<float> fCutVals;

CondFormats/EgammaObjects/src/GBRForest.cxx

@@ -21,17 +21,34 @@ GBRForest::~GBRForest()
 GBRForest::GBRForest(const TMVA::MethodBDT *bdt)
 {
 
-  if (bdt->DoRegression()) {
+  //special handling for non-gradient-boosted (ie ADABoost) classifiers, where tree responses
+  //need to be renormalized after the training for evaluation purposes
+  bool isadaclassifier = !bdt->DoRegression() && !bdt->GetOptions().Contains("~BoostType=Grad");  
+  bool useyesnoleaf = isadaclassifier && bdt->GetOptions().Contains("~UseYesNoLeaf=True");
+  bool isregression = bdt->DoRegression();
+  //newer tmva versions use >= instead of > in decision tree splits, so adjust cut value
+  //to reproduce the correct behaviour  
+  bool adjustboundaries = (bdt->GetTrainingROOTVersionCode()>=ROOT_VERSION(5,34,20) && bdt->GetTrainingROOTVersionCode()<ROOT_VERSION(6,0,0)) || bdt->GetTrainingROOTVersionCode()>=ROOT_VERSION(6,2,0);
+
+  if (isregression) {
     fInitialResponse = bdt->GetBoostWeights().front();
   }
   else {
     fInitialResponse = 0.;
   }
 
+  double norm = 0;
+  if (isadaclassifier) {
+    for (std::vector<double>::const_iterator it=bdt->GetBoostWeights().begin(); it!=bdt->GetBoostWeights().end(); ++it) {
+      norm += *it;  
+    }
+  }
+
   const std::vector<TMVA::DecisionTree*> &forest = bdt->GetForest();
   fTrees.reserve(forest.size());
-  for (std::vector<TMVA::DecisionTree*>::const_iterator it=forest.begin(); it!=forest.end(); ++it) {
-    fTrees.push_back(GBRTree(*it));
+  for (unsigned int itree=0; itree<forest.size(); ++itree) {
+    double scale = isadaclassifier ? bdt->GetBoostWeights()[itree]/norm : 1.0;
+    fTrees.push_back(GBRTree(forest[itree],scale,useyesnoleaf,adjustboundaries));
   }
 
 }

CondFormats/EgammaObjects/src/GBRTree.cxx

@@ -13,7 +13,7 @@ GBRTree::GBRTree()
 }
 
 //_______________________________________________________________________
-GBRTree::GBRTree(const TMVA::DecisionTree *tree)
+GBRTree::GBRTree(const TMVA::DecisionTree *tree, double scale, bool useyesnoleaf, bool adjustboundary)
 {
 
   //printf("boostweights size = %i, forest size = %i\n",bdt->GetBoostWeights().size(),bdt->GetForest().size());
@@ -29,7 +29,7 @@ GBRTree::GBRTree(const TMVA::DecisionTree *tree)
   fRightIndices.reserve(nIntermediate);
   fResponses.reserve(nTerminal);
 
-  AddNode((TMVA::DecisionTreeNode*)tree->GetRoot());
+  AddNode((TMVA::DecisionTreeNode*)tree->GetRoot(), scale, tree->DoRegression(), useyesnoleaf, adjustboundary);
 
   //special case, root node is terminal, create fake intermediate node at root
   if (fCutIndices.size()==0) {
@@ -73,17 +73,36 @@ unsigned int GBRTree::CountTerminalNodes(const TMVA::DecisionTreeNode *node) {
 
 
 //_______________________________________________________________________
-void GBRTree::AddNode(const TMVA::DecisionTreeNode *node) {
+void GBRTree::AddNode(const TMVA::DecisionTreeNode *node, double scale, bool isregression, bool useyesnoleaf, bool adjustboundary) {
 
   if (!node->GetLeft() || !node->GetRight() || node->IsTerminal()) {
-    fResponses.push_back(node->GetResponse());
+    double response = 0.;
+    if (isregression) {
+      response = node->GetResponse();
+    }
+    else {
+      if (useyesnoleaf) {
+        response = double(node->GetNodeType());
+      }
+      else {
+        response  = node->GetPurity();
+      }
+    }
+    response *= scale;
+    fResponses.push_back(response);
     return;
   }
   else {    
     int thisidx = fCutIndices.size();
 
     fCutIndices.push_back(node->GetSelector());
-    fCutVals.push_back(node->GetCutValue());
+    float cutval = node->GetCutValue();
+    //newer tmva versions use >= instead of > in decision tree splits, so adjust cut value
+    //to reproduce the correct behaviour
+    if (adjustboundary) {
+      cutval = std::nextafter(cutval,std::numeric_limits<float>::lowest());
+    }
+    fCutVals.push_back(cutval);
     fLeftIndices.push_back(0);   
     fRightIndices.push_back(0);
 
@@ -105,15 +124,15 @@ void GBRTree::AddNode(const TMVA::DecisionTreeNode *node) {
     else {
       fLeftIndices[thisidx] = fCutIndices.size();
     }
-    AddNode(left);
+    AddNode(left, scale, isregression, useyesnoleaf, adjustboundary);
 
     if (!right->GetLeft() || !right->GetRight() || right->IsTerminal()) {
       fRightIndices[thisidx] = -fResponses.size();
     }
     else {
       fRightIndices[thisidx] = fCutIndices.size();
     }
-    AddNode(right);    
+    AddNode(right, scale, isregression, useyesnoleaf, adjustboundary);
 
   }