test_working_newms_der.cpp

 
#include <iostream>
#include <Pds/Ra>
#include <Pds/Ml>
#include <Pds/Sp>
 
 
Pds::Vector GradientCostNewMS(Pds::Vector W,Pds::Matrix X,Pds::Vector Y,double K=10.0)
{
    unsigned int L=X.Nlin();
    unsigned int N=X.Ncol()+1;
    double Norm=W.Norm();
    
    double S=2.0*K/(L*Norm);
    Pds::Vector Wn=W/Norm;
    
    Pds::Matrix R=Pds::RegressorMatrix(X);
    Pds::Vector F=Pds::Sigmoid(K*R*Wn);
    Pds::Vector b=S*(F-Y).Product(F.Product(1.0-F));
    Pds::Matrix I=Pds::Eye(N);
    
    Pds::Vector DW=(b.T())*R*(I-Wn.MMt());
    
    return DW;
}
 
 
// ./test/test_working_newhb_der
int main(void)
{
    Pds::IterationConf Conf;
    Pds::Ra::Randomize();
    
    Pds::Octave::DAspect=true;
    Pds::Octave::YLimits=true;
    
    // Generating data
    // Data 1
    Pds::Matrix X(Pds::Ra::TextFormat,"../test/hard-data-x.txt");
    Pds::Vector Y(Pds::Ra::TextFormat,"../test/hard-data-y.txt");
    
    // Data 2
    //unsigned int L=1000;    Pds::Matrix X;  Pds::Vector Y;
    //Pds::DataSet::Classification::LoadDataBand(L,X,Y);
    
    
    //double d=Pds::DataSet::InterDistance(X);
    
    
    // Seleciono punto de inicio
    Pds::Vector W;
    //W=Pds::LogisticModel::GetW0MeanMethod(X);
    W=Pds::LogisticModel::FittingLogitMeanSquare(Conf,X,Y,0.01);
    W=Pds::LogisticModel::FittingGradientCrossEntropy(Conf,X,Y,W);
    W.Normalize();
    
    Pds::Octave::Plot::PointsX2DYW(X,Y,W,"testando_w.m","test_working_steps_0.png");
    
    
    Pds::Vector DW;
    Pds::Vector Yo;
    double E1;
    double K=10;
    for(unsigned int n=1;n<=60;n++)
    {
        std::cout<<"\n n: "<<n<<"\n";
        
        Yo=Pds::LogisticModel::Classify(W,X).Geq(0.5);
        if(Pds::Hb(Yo.Mean())<=0.0)
        {
            W=Pds::LogisticModel::GetW0MeanMethod(X);
            W.Normalize();
        }
        
        DW=GradientCostNewMS(W,X,Y,K);
        //DW.Normalize();
        //DW=DW*W.Norm()*0.1;
        
        DW.T().Print("DW: ");
        W=W-0.25*DW;
        W.T().Print(" W: ");
        
        //W.Normalize();
        
        E1=Pds::LogisticModel::CostMeanSquare(K*W/W.Norm(),X,Y);
        Pds::Octave::Title="NMS:"+std::to_string(E1);
        Pds::Octave::Plot::PointsX2DYW(X,Y,W,"testando_w.m","test_working_steps_"+std::to_string(n)+".png");
        
    }
 
    return 0;
}