GBRTreeD.h

Go to the documentation of this file.

#ifndef EGAMMAOBJECTS_GBRTreeD
#define EGAMMAOBJECTS_GBRTreeD

//                                                                      //
// GBRForest                                                            //
//                                                                      //
// A fast minimal implementation of Gradient-Boosted Regression Trees   //
// which has been especially optimized for size on disk and in memory.  //                                                                  
//                                                                      //
// Designed to be built from TMVA-trained trees, but could also be      //
// generalized to otherwise-trained trees, classification,              //
//  or other boosting methods in the future                             //
//                                                                      //
//  Josh Bendavid - CERN                                                 //

// The decision tree is implemented here as a set of two arrays, one for
// intermediate nodes, containing the variable index and cut value, as well
// as the indices of the 'left' and 'right' daughter nodes.  Positive indices
// indicate further intermediate nodes, whereas negative indices indicate
// terminal nodes, which are stored simply as a vector of regression responses

#include "CondFormats/Serialization/interface/Serializable.h"

#include <vector>
#include <map>
#include <stdio.h>
#include <cmath>
#include "Rtypes.h"

class GBRTreeD {

  public:

    GBRTreeD() {}
    template<typename InputTreeT> GBRTreeD(const InputTreeT &tree);
    virtual ~GBRTreeD();
          
    //double GetResponse(const float* vector) const;
    double GetResponse(int termidx) const { return fResponses[termidx]; }
    int TerminalIndex(const float *vector) const;
          
    std::vector<double> &Responses() { return fResponses; }       
    const std::vector<double> &Responses() const { return fResponses; }
    
    std::vector<unsigned short> &CutIndices() { return fCutIndices; }
    const std::vector<unsigned short> &CutIndices() const { return fCutIndices; }
    
    std::vector<float> &CutVals() { return fCutVals; }
    const std::vector<float> &CutVals() const { return fCutVals; }
    
    std::vector<int> &LeftIndices() { return fLeftIndices; }
    const std::vector<int> &LeftIndices() const { return fLeftIndices; } 
    
    std::vector<int> &RightIndices() { return fRightIndices; }
    const std::vector<int> &RightIndices() const { return fRightIndices; }
    
  protected:              
    std::vector<unsigned short> fCutIndices;
    std::vector<float> fCutVals;
    std::vector<int> fLeftIndices;
    std::vector<int> fRightIndices;
    std::vector<double> fResponses;

  COND_SERIALIZABLE;    
};


//_______________________________________________________________________
inline int GBRTreeD::TerminalIndex(const float* vector) const {
  
  int index = 0;
  
  unsigned short cutindex = fCutIndices[0];
  float cutval = fCutVals[0];
  
  while (true) {
    if (vector[cutindex] > cutval) {
      index = fRightIndices[index];
    }
    else {
      index = fLeftIndices[index];
    }
    
    if (index>0) {
      cutindex = fCutIndices[index];
      cutval = fCutVals[index];
    }
    else {
      return (-index);
    }
    
  }
  
}

//_______________________________________________________________________
template<typename InputTreeT> GBRTreeD::GBRTreeD(const InputTreeT &tree) :
 fCutIndices(tree.CutIndices()),
 fCutVals(tree.CutVals()),
 fLeftIndices(tree.LeftIndices()),
 fRightIndices(tree.RightIndices()),
 fResponses(tree.Responses()) {
  
}

#endif