gurls::RLSPegasos< T > Class Template Reference

RLSPegasos is the sub-class of Optimizer that implements the Pegaosos algorithm.

#include <rlspegasos.h>

Inheritance diagram for gurls::RLSPegasos< T >:

Collaboration diagram for gurls::RLSPegasos< T >:

List of all members.

Public Member Functions
GurlsOptionsList *	execute (const gMat2D< T > &X, const gMat2D< T > &Y, const GurlsOptionsList &opt)
	Computes a classifier for the primal formulation of RLS.
Static Public Member Functions
static Optimizer< T > *	factory (const std::string &id) throw (BadOptimizerCreation)
	Factory function returning a pointer to the newly created object.

---

Detailed Description

template<typename T>
class gurls::RLSPegasos< T >

Definition at line 59 of file rlspegasos.h.

---

Member Function Documentation

template<typename T >

GurlsOptionsList * gurls::RLSPegasos< T >::execute	(	const gMat2D< T > &	X,
		const gMat2D< T > &	Y,
		const GurlsOptionsList &	opt
	)		[virtual]

The optimization is carried out using a stochastic gradient descent algorithm. The regularization parameter is set to the one found in the field paramsel of opt. In case of multiclass problems, the regularizers need to be combined with the function specified inthe field singlelambda of opt

Parameters:

X	input data matrix
Y	labels matrix
opt	options with the following: singlelambda (default) epochs (default) paramsel (settable with the class ParamSelection and its subclasses) Xte (test input data matrix, needed for accuracy evaluation) yte (test labels matrix, needed for accuracy evaluation)

Returns:

adds to opt the field optimizer which is a list containing the following fields:

• W = matrix of coefficient vectors of rls estimator for each class
• W_sum = sum of the classifiers across iterations
• t0 = stepsize parameter
• count = number of iterations
• acc_last = accuracy of the solution computed in the last iteration
• acc_avg = average accuracy across iterations

Implements gurls::Optimizer< T >.

Definition at line 91 of file rlspegasos.h.

{
    //  lambda = opt.singlelambda(opt.paramsel.lambdas);
    const gMat2D<T> &ll = opt.getOptValue<OptMatrix<gMat2D<T> > >("paramsel.lambdas");
    T lambda = opt.getOptAs<OptFunction>("singlelambda")->getValue(ll.getData(), ll.getSize());


    //   [n,d] = size(X);
    const unsigned long n = X.rows();
    const unsigned long d = X.cols();

    //   T = size(bY,2);
    const unsigned long t = Y.cols();


    GurlsOptionsList* optimizer = new GurlsOptionsList("optimizer");

    //   opt.cfr.W = zeros(d,T);
    gMat2D<T>* W = new gMat2D<T>(d,t);
    set(W->getData(), (T)0.0, d*t);
    optimizer->addOpt("W", new OptMatrix<gMat2D<T> >(*W));

    //   opt.cfr.W_sum = zeros(d,T);
    gMat2D<T>* W_sum = new gMat2D<T>(d,t);
    copy(W_sum->getData(), W->getData(), d*t);
    optimizer->addOpt("W_sum", new OptMatrix<gMat2D<T> >(*W_sum));

    optimizer->addOpt("count", new OptNumber(0.0));


    //   opt.cfr.acc_last = [];
    //   opt.cfr.acc_avg = [];

    //           opt.cfr.t0 = ceil(norm(X(1,:))/sqrt(opt.singlelambda(opt.paramsel.lambdas)));
    T* row = new T[d];
    getRow(X.getData(), n, d, 0, row);
    optimizer->addOpt("t0", new OptNumber( ceil( nrm2(d, row, 1)/sqrt(lambda))));

    delete[] row;


    //   % Run mulitple epochs
    //   for i = 1:opt.epochs,
    int epochs = static_cast<int>(opt.getOptAsNumber("epochs"));

    GurlsOptionsList* tmp_opt = new GurlsOptionsList("opt");

    GurlsOptionsList* tmp_paramsel = new GurlsOptionsList("paramsel");
    tmp_opt->addOpt("paramsel", tmp_paramsel);

    gMat2D<T>* ret_lambdas = new gMat2D<T>(ll);
    tmp_paramsel->addOpt("lambdas", new OptMatrix<gMat2D<T> >(*ret_lambdas));


    OptFunction* tmp_singlelambda = new OptFunction(opt.getOptAs<OptFunction>("singlelambda")->getName());
    tmp_opt->addOpt("singlelambda", tmp_singlelambda);

    tmp_opt->addOpt("optimizer", optimizer);

    for(int i=0; i<epochs; ++i)
    {
        //       if opt.cfr.count == 0
        //           opt.cfr.t0 = ceil(norm(X(1,:))/sqrt(opt.singlelambda(opt.paramsel.lambdas)));
        //           fprintf('\n\tt0 is set to : %f\n', opt.cfr.t0);
        //       end

        //       opt.cfr = rls_pegasos_singlepass(X, bY, opt);
        GurlsOptionsList* result = rls_pegasos_driver(X.getData(), Y.getData(), *tmp_opt, n, d, Y.rows(), t);

        tmp_opt->removeOpt("optimizer");
        tmp_opt->addOpt("optimizer", result);
    }

    optimizer = tmp_opt->getOptAs<GurlsOptionsList>("optimizer");
    tmp_opt->removeOpt("optimizer", false);
    delete tmp_opt;

    //   cfr = opt.cfr;

    //   cfr.W = opt.cfr.W_sum/opt.cfr.count;

    T count = static_cast<T>(optimizer->getOptAsNumber("count"));
    if(eq(count, (T)0.0))
        throw gException(Exception_Illegal_Argument_Value);

    W = &(optimizer->getOptValue<OptMatrix<gMat2D<T> > >("W"));
    W_sum = &(optimizer->getOptValue<OptMatrix<gMat2D<T> > >("W_sum"));

    set(W->getData(), (T)0.0, W->getSize());
    axpy(W->getSize(), (T)(1.0/count), W_sum->getData(), 1, W->getData(), 1);

    return optimizer;
}

template<typename T>

static Optimizer<T>* gurls::Optimizer< T >::factory

(

const std::string &

id

)

throw (BadOptimizerCreation) [inline, static, inherited]

Warning:

The returned pointer is a plain, un-managed pointer. The calling function is responsible of deallocating the object.

Definition at line 130 of file optimization.h.

    {
      if(id == "rlsauto")
        return new RLSAuto<T>;
      if(id == "rlsprimal")
        return new RLSPrimal<T>;
      if(id == "rlsprimalr")
        return new RLSPrimalr<T>;
      if(id == "rlsdual")
        return new RLSDual<T>;
      if(id == "rlsdualr")
        return new RLSDualr<T>;
      if(id == "rlspegasos")
        return new RLSPegasos<T>;
      if(id == "rlsgpregr")
        return new RLSGPRegr<T>;
      if(id == "rlsprimalrecinit")
        return new RLSPrimalRecInit<T>;
      if(id == "rlsprimalrecupdate")
        return new RLSPrimalRecUpdate<T>;
      if(id == "rlsrandfeats")
        return new RLSRandFeats<T>;

        throw BadOptimizerCreation(id);
    }

---

The documentation for this class was generated from the following file:

• rlspegasos.h