KernelSelection.cpp

Go to the documentation of this file.

#include <shark/Data/DataDistribution.h>
#include <shark/Models/Kernels/GaussianRbfKernel.h>
#include <shark/ObjectiveFunctions/RadiusMarginQuotient.h>
#include <shark/Algorithms/GradientDescent/Rprop.h>


using namespace shark;


int main(int argc, char** argv)
{
    // generate dataset
    Chessboard problem;
    ClassificationDataset data = problem.generateDataset(100);

    // brute force search in [1.0, 10000.0] on log scale
    GaussianRbfKernel<> kernel;
    RadiusMarginQuotient<RealVector> rm(data, &kernel);
    RealVector param(1);
    double best_value = 1e100;
    double best_gamma = 0.0;
    
    std::cout<<"Grid search in the range [1, 10000] on log scale:"<<std::endl;
    for (unsigned i=0; i<=400; i++)
    {
        double gamma = pow(10.0, i / 100.0);
        param(0) = gamma;
        double f = rm.eval(param);
        if (f < best_value)
        {
            best_value = f;
            best_gamma = gamma;
        }
    }
    std::cout<<"best gamma: "<< best_gamma<< "  radius margin quotient: "<<best_value<<std::endl;

    // gradient-based alternative
    IRpropPlus rprop;
    rprop.init(rm, RealVector(1, 100.0), 1.0);
    std::cout<<"\nGradient-based optimization (IRprop+, 50 steps):"<<std::endl;
    for (unsigned i=0; i<50; i++) rprop.step(rm);
    best_gamma = rprop.solution().point(0);
    best_value = rm.eval(RealVector(1, best_gamma));
    std::cout<<"best gamma: "<< best_gamma<< "  radius margin quotient: "<<best_value<<std::endl;
}