NICE_VisLearning/classifier/kernelclassifier/KCGPApproxOneClass.cpp

/** 
* @file KCGPApproxOneClass.cpp
* @brief One-Class Gaussian Process Regression for Classification: we approximate the inverse of the regularized kernel matrix using a diagonal matrix
* @author Alexander Lütz
* @date 22-05-2012 (dd-mm-yyyy)

*/
#include <iostream>
#include <typeinfo>
#include <cstring>

#include "core/vector/Algorithms.h"
#include "core/vector/VVector.h"

#include "KCGPApproxOneClass.h"


using namespace std;
using namespace NICE;
using namespace OBJREC;


KCGPApproxOneClass::KCGPApproxOneClass( const Config *conf, Kernel *kernel, const string & section ) : KernelClassifier ( conf, kernel )
{
//   this->kernelFunction = kernel;  
  //overwrite the default optimization options, since we don't want to perform standard loo or marginal likelihood stuff
  Config config(*conf);
  string modestr = config.gS(section,"detection_mode"); 

  if(strcmp("mean",modestr.c_str())==0){
    this->mode=MEAN_DETECTION_MODE;cerr << "One-class classification via GP predictive _mean_ !!!"<<endl;
  }
  if(strcmp("variance",modestr.c_str())==0){
    mode=VARIANCE_DETECTION_MODE;cerr << "One-class classification via GP predictive _variance_ !!!"<<endl;
  }

  this->staticNoise = conf->gD(section, "static_noise", 0.0);
}



KCGPApproxOneClass::KCGPApproxOneClass( const KCGPApproxOneClass & src ) : KernelClassifier ( src )
{
  this->matrixDInv = src.matrixDInv;
  this->InvDY = src.InvDY;
  this->mode = src.mode;
  this->staticNoise = src.staticNoise;
}

KCGPApproxOneClass::~KCGPApproxOneClass()
{
}


void KCGPApproxOneClass::teach ( KernelData *kernelData, const NICE::Vector & y )
{
    fthrow( Exception, "KCGPApproxOneClass::teach: this method is not implemented for this specific type of classifier. Please use the second teach-method." );  
}

void KCGPApproxOneClass::teach (const LabeledSetVector &teachSet)
{
  if ( this->kernelFunction == NULL )
    fthrow( Exception, "KernelClassifier::teach: To use this function, you have to specify a kernel function using the constructor" );  
  
  //we do not have to allocate new storage here since these variables come from the interface KernelClassifier
//   NICE::VVector vecSet;

  teachSet.getFlatRepresentation (this->vecSet, this->vecSetLabels);
    
  if ( (this->vecSetLabels.Min() != 1) || (this->vecSetLabels.Max() != 1) ) {
    fthrow(Exception, "This classifier is suitable only for one-class classification problems, i.e. max(y) = min(y) = 1");
  }  

  this->matrixDInv.resize(this->vecSetLabels.size());
 
  //compute D 
  //start with adding some noise, if necessary
  if (this->staticNoise != 0.0)
    this->matrixDInv.set(this->staticNoise);
  else
    this->matrixDInv.set(0.0);
  
  //now sum up all entries of each row in the original kernel matrix
  double kernelScore(0.0);
  for (int i = 0; i < (int)this->vecSetLabels.size(); i++)
  {
    for (int j = i; j < (int)this->vecSetLabels.size(); j++)
    {
      kernelScore = this->kernelFunction->K(vecSet[i],vecSet[j]);
      this->matrixDInv[i] += kernelScore;
      if (i != j)
        this->matrixDInv[j] += kernelScore; 
    }
  }  
  
  //compute its inverse
  for (int i = 0; i < (int)this->vecSetLabels.size(); i++)
  {
    this->matrixDInv[i] = 1.0 / this->matrixDInv[i];
  }
  
  //and multiply it from right with the label vector (precalculation for mean computation)
  if(this->mode==MEAN_DETECTION_MODE)
  {
    this->InvDY.resize ( this->vecSetLabels.size() );
    for (int i = 0; i < (int)this->vecSetLabels.size(); i++)
    {
      this->InvDY[i] = this->vecSetLabels[i] * this->matrixDInv[i];
    }
  }  
}

ClassificationResult KCGPApproxOneClass::classifyKernel ( const NICE::Vector & kernelVector, double kernelSelf ) const
{
  FullVector scores ( 2 );
  scores[0] = 0.0;

  if(this->mode==MEAN_DETECTION_MODE)
  {
    // kernelSelf is not needed for the regression type of GP

    if ( kernelVector.size() != this->vecSetLabels.size() ) 
      fthrow(Exception, "KCGPApproxOneClass::classifyKernel: size of kernel value vector " << 
        kernelVector.size() << " does not match number of training points " << this->vecSetLabels.size() );
      
    double yEstimate = kernelVector.scalarProduct ( InvDY );    
    scores[1] = yEstimate;
  }
  if(this->mode==VARIANCE_DETECTION_MODE)
  {
    if ( kernelVector.size() != this->vecSetLabels.size() ) 
      fthrow(Exception, "KCGPApproxOneClass::classifyKernel: size of kernel value vector " << 
        kernelVector.size() << " does not match number of training points " << this->vecSetLabels.size() );
      
    NICE::Vector rightPart (this->vecSetLabels.size());
    for (int i = 0; i < (int)this->vecSetLabels.size(); i++)
    {
      rightPart[i] = kernelVector[i] * this->matrixDInv[i];
    }

    double uncertainty = kernelSelf - kernelVector.scalarProduct ( rightPart );
    scores[1] = 1.0 - uncertainty;

  }
  ClassificationResult r ( scores[1]<0.5 ? 0 : 1, scores );

  return r;
}

KCGPApproxOneClass *KCGPApproxOneClass::clone() const
{
  return new KCGPApproxOneClass ( *this );
}

void KCGPApproxOneClass::store(std::ostream& ofs, int type) const
{
  ofs << this->matrixDInv << std::endl;
  ofs << this->InvDY << std::endl;
  ofs << this->mode << std::endl;
  ofs << this->staticNoise << std::endl;
}

void KCGPApproxOneClass::restore(std::istream& ifs, int type)
{
  ifs >> this->matrixDInv;
  ifs >> this->InvDY;
  ifs >> this->mode;
  ifs >> this->staticNoise;
} 

void KCGPApproxOneClass::clear()
{
}