NICE_VisLearning/regression/npregression/RegKNN.cpp

/**
* @file RegKNN.cpp
* @brief Implementation of k-Nearest-Neighbor algorithm for regression purposes
* @author Frank Prüfer
* @date 08/29/2013

*/

#ifdef NICE_USELIB_OPENMP
#include <omp.h>
#endif

#include <iostream>

#include "vislearning/regression/npregression/RegKNN.h"

#include "vislearning/math/mathbase/FullVector.h"

using namespace OBJREC;

using namespace NICE;


RegKNN::RegKNN ( const Config *_conf, NICE::VectorDistance<double> *_distancefunc ) : distancefunc (_distancefunc)
{
  K = _conf->gI("RegKNN", "K", 1 );
  if ( _distancefunc == NULL )
  distancefunc = new EuclidianDistance<double>();
}

RegKNN::RegKNN ( const RegKNN & src ) : RegressionAlgorithm ( src )
{
  dataSet = src.dataSet;
  labelSet = src.labelSet;
  distancefunc = src.distancefunc;
  K = src.K;
}

RegKNN::~RegKNN ()
{
}

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


void RegKNN::teach ( const NICE::VVector & _dataSet, const NICE::Vector & _labelSet)
{
  fprintf (stderr, "teach using all !\n");
  //NOTE this is crucial if we clear _teachSet afterwards!
  //therefore, take care NOT to call _techSet.clear() somewhere out of this method
  this->dataSet = _dataSet;
  this->labelSet = _labelSet.std_vector();
  
  std::cerr << "number of known training samples: " << this->dataSet.size() << std::endl;   
    
}

void RegKNN::teach ( const NICE::Vector & x, const double & y )
{
  std::cerr << "RegKNN::teach one new example" << std::endl;

  for ( size_t i = 0 ; i < x.size() ; i++ )
    if ( isnan(x[i]) ) 
    {
        fprintf (stderr, "There is a NAN value within this vector: x[%d] = %f\n", (int)i, x[i]);
        std::cerr << x << std::endl;
        exit(-1);
    }

  dataSet.push_back ( x );

  labelSet.push_back ( y );

  std::cerr << "number of known training samples: " << dataSet.size()<< std::endl;
}

double RegKNN::predict ( const NICE::Vector & x )
{
  FullVector distances(dataSet.size());

  if ( dataSet.size() <= 0 )
  {
    fprintf (stderr, "RegKNN: please use the teach method first\n");
    exit(-1);
  }

#pragma omp parallel for
  for(uint i = 0; i < dataSet.size(); i++)
  {
    double distance = distancefunc->calculate (x,dataSet[i]);

    if ( isnan(distance) )
    {
      fprintf (stderr, "RegKNN::predict: NAN value found !!\n");
      std::cerr << x << std::endl;
      std::cerr << dataSet[i] << std::endl;
    }
// #pragma omp critical      
    distances[i] = distance;     
  }
    
  std::vector<int> ind;
  distances.getSortedIndices(ind);

  double response = 0.0;  
    
  if ( dataSet.size() < K )
  {
    std::cerr << K << std::endl;
    K = dataSet.size();
    std::cerr<<"RegKNN: Not enough datapoints! Setting K to: "<< K << std::endl;
  }

  if ( distances[ind[0]] == 0.0 ) {
    std::cerr<<"RegKNN: Warning: datapoint was already seen during training... using its label as prediction."<< std::endl;
    return labelSet[ind[0]];  
  }

  double maxElement = distances.max();	//normalize distances
  distances.multiply(1.0/maxElement);

  double weightSum = 0.0;

  for(uint i = 0; i < K; i++)
  {
    response += 1.0/distances[ind[i]] * labelSet[ind[i]];
    weightSum += 1.0/distances[ind[i]];
  }

  return ( response / weightSum );
}