|| || || || || || || ||

---

PrecomputeKernelMatrix.h

Go to the documentation of this file.

#ifndef Impala_Core_Training_PrecomputeKernelMatrix_h
#define Impala_Core_Training_PrecomputeKernelMatrix_h

#include "Util/PropertySet.h"
#include "Core/Vector/Apply.h"
#include "Core/Training/KernelFunctions.h"
#include "Core/Array/PixSum.h"
#include "Core/Array/Pattern/PatMPixOp.h"
#include "Core/Feature/FeatureTable.h"

namespace Impala
{
namespace Core
{
namespace Training
{

Matrix::Mat* 
PrecomputeKernelMatrix(std::vector<Feature::FeatureTable*> features,
                       std::vector<double> weights,
                       Util::PropertySet* properties)
{
    ILOG_VAR(Training.PrecomputeKernelMatrix);
    if(!(weights.size() == features.size()))
    {
        ILOG_ERROR("size of list does not match!");
        return 0;
    }

    std::vector<Matrix::Mat*> DistanceMatrices;
    std::vector<double> AList;
    // perhaps we should make this a pattern?
    std::vector<Feature::FeatureTable*>::iterator it;
    for(it=features.begin() ; it!=features.end() ; ++it)
    {
        Feature::FeatureTable* f = *it;

        // fill a matrix with chi2 distance
        Matrix::Mat* m = Apply(&Training::Chi2Distance, f->GetColumn2(), f->GetColumn2());
        DistanceMatrices.push_back(m);

        // get A (average of all weights)
        double a = Array::PixSum(m);
        a /= m->W() * m->H();
        AList.push_back(a);
    }

    if(!(weights.size() == AList.size() == DistanceMatrices.size()))
    {
        ILOG_ERROR("size of list does not match!");
        return 0;
    }
    // fill matrix with K(vi,vj)
    MpoChi2Kernel kernel(weights, AList);
    Array::Pattern::PatMPixOp(DistanceMatrices[0], DistanceMatrices, kernel);

    for(int i=1 ; i<DistanceMatrices.size() ; ++i)
        delete DistanceMatrices[i];
    return DistanceMatrices[0];
}

Matrix::Mat* 
PrecomputeKernelMatrix(Feature::FeatureTable* features,
                       Util::PropertySet* properties)
{
    std::vector<Feature::FeatureTable*> f;
    std::vector<double> w;
    f.push_back(features);
    w.push_back(1);
    return PrecomputeKernelMatrix(f,w,properties);
}

}//namespace Core
}//namespace Training
}//Impala


#endif

---