CSvmGridSearchTutorial.cpp

Go to the documentation of this file.

#include <shark/Models/Kernels/GaussianRbfKernel.h>
#include <shark/ObjectiveFunctions/Loss/ZeroOneLoss.h>
#include <shark/ObjectiveFunctions/CrossValidationError.h>
#include <shark/Algorithms/Trainers/CSvmTrainer.h>
#include <shark/Algorithms/DirectSearch/GridSearch.h>
#include <shark/Algorithms/JaakkolaHeuristic.h>
#include <shark/Data/Dataset.h>
#include <shark/Data/DataDistribution.h>


using namespace shark;
using namespace std;


int main()
{
    // problem definition
    Chessboard prob;
    ClassificationDataset dataTrain = prob.generateDataset(200);
    ClassificationDataset dataTest= prob.generateDataset(10000);

    // SVM setup
    GaussianRbfKernel<> kernel(0.5, true); //unconstrained?
    KernelExpansion<RealVector> svm(true); //use offset?
    CSvmTrainer<RealVector> trainer(&kernel, 1.0, true); //unconstrained?

    // cross-validation error
    const unsigned int N= 5;  // number of folds

    ZeroOneLoss<unsigned int, RealVector> loss;
    CVFolds<ClassificationDataset> folds = createCVSameSizeBalanced(dataTrain, N);
    CrossValidationError<KernelExpansion<RealVector>, unsigned int> cvError(
        folds, &trainer, &svm, &trainer, &loss
    );


    // find best parameters

    // use Jaakkola's heuristic as a starting point for the grid-search
    JaakkolaHeuristic ja(dataTrain);
    double ljg = log(ja.gamma());
    cout << "Tommi Jaakkola says gamma = " << ja.gamma() << " and ln(gamma) = " << ljg << endl;


    GridSearch grid;
    vector<double> min(2);
    vector<double> max(2);
    vector<size_t> sections(2);
    min[0] = ljg-4.; max[0] = ljg+4; sections[0] = 17;  // kernel parameter gamma
    min[1] = 0.0; max[1] = 10.0; sections[1] = 11;     // regularization parameter C
    grid.configure(min, max, sections);
    grid.step(cvError);

    // train model on the full dataset
    trainer.setParameterVector(grid.solution().point);
    trainer.train(svm, dataTrain);
    cout << "C =\t" << trainer.C() << endl;
    cout << "gamma =\t" << kernel.gamma() << endl;

    // evaluate
    Data<RealVector> output = svm(dataTrain.inputs());
    double train_error = loss.eval(dataTrain.labels(), output);
    cout << "training error:\t" << train_error << endl;
    output = svm(dataTest.inputs());
    double test_error = loss.eval(dataTest.labels(), output);
    cout << "test error: \t" << test_error << endl;
}