Gaussians.cpp

Go to the documentation of this file.

/*
 * Gaussians.cpp
 *
 *  Created on: Nov 19, 2011
 *      Author: Seungsu KIM
 */
 
#include "Gaussians.h"
 
#include <math.h>
 
#include <fstream>
#include <iostream>
 
using namespace std;
/*
Gaussians::Gaussians(GMMs *model)
{
    this->model.nbStates = model->nbStates;
    this->model.nbDim    = model->nbDim;
 
    this->model.States = (GMMState  *)malloc(model->nbStates*sizeof(GMMState) );
 
    for(int s=0; s<model->nbStates; s++ ){
        this->model.States[s].Mu    = model->GMMState[s].Mu;
        this->model.States[s].Sigma = model->GMMState[s].Sigma;
        this->model.States[s].Prio  = model->GMMState[s].Prio;
    }
}
*/
 
Gaussians::Gaussians(const char* f_mu, const char* f_sigma, const char* f_prio)
{
    int s, i, j;
    int nbStates;
    int nbDim;
 
    MathLib::Matrix fMatrix;
    fMatrix.Load(f_prio);
    if (fMatrix(0, fMatrix.ColumnSize() - 1) > 0.0)
    {
        nbStates = fMatrix.ColumnSize();
    }
    else
    {
        nbStates = fMatrix.ColumnSize() - 1;
    }
 
    for (s = 0; s < nbStates; s++)
    {
        model.States[s].Prio = fMatrix(0, s);
    }
 
    fMatrix.Load(f_mu);
    nbDim = fMatrix.RowSize();
    model.nbStates = nbStates;
    model.nbDim = nbDim;
 
 
    for (s = 0; s < nbStates; s++)
    {
        model.States[s].Mu.Resize(nbDim);
        model.States[s].Sigma.Resize(nbDim, nbDim);
    }
 
    for (s = 0; s < nbStates; s++)
    {
        model.States[s].Mu = fMatrix.GetColumn(s);
    }
 
    fMatrix.Load(f_sigma);
    j = 0;
    for (s = 0; s < nbStates; s++)
    {
        for (i = 0; i < nbDim; i++)
        {
            model.States[s].Sigma.SetRow(fMatrix.GetRow(j), i);
            j++;
        }
    }
}
 
Gaussians::Gaussians(int nbStates,
                     int nbDim,
                     const char* f_mu,
                     const char* f_sigma,
                     const char* f_prio)
{
 
    int s, i, j;
 
    model.nbStates = nbStates;
    model.nbDim = nbDim;
 
    for (s = 0; s < nbStates; s++)
    {
        model.States[s].Mu.Resize(nbDim);
        model.States[s].Sigma.Resize(nbDim, nbDim);
    }
 
    MathLib::Matrix fMatrix(nbDim, nbStates);
    fMatrix.Load(f_mu);
    for (s = 0; s < nbStates; s++)
    {
        model.States[s].Mu = fMatrix.GetColumn(s);
    }
 
    fMatrix.Resize(nbStates * nbDim, nbDim);
    fMatrix.Load(f_sigma);
    j = 0;
    for (s = 0; s < nbStates; s++)
    {
        for (i = 0; i < nbDim; i++)
        {
            model.States[s].Sigma.SetRow(fMatrix.GetRow(j), i);
            j++;
        }
    }
 
    fMatrix.Resize(1, nbStates);
    fMatrix.Load(f_prio);
    MathLib::Vector fVector(nbStates);
    for (s = 0; s < nbStates; s++)
    {
        model.States[s].Prio = fMatrix(0, s);
    }
}
 
Gaussians::Gaussians(const int nbStates,
                     const int nbDim,
                     const vector<double> pri_vec,
                     const vector<double> mu_vec,
                     const vector<double> sig_vec)
{
 
    model.nbStates = nbStates;
    model.nbDim = nbDim;
 
    for (int s = 0; s < nbStates; s++)
    {
        model.States[s].Mu.Resize(nbDim);
        model.States[s].Sigma.Resize(nbDim, nbDim);
    }
 
    for (int s = 0; s < nbStates; s++)
    {
        model.States[s].Prio = pri_vec[s];
    }
    // cout << endl << "Printing the constructed Priors" << endl;
    // for ( int s = 0; s < nbStates; s++ ) {
    //  cout << model.States[s].Prio  << "\t";
    // }
    // cout << endl;
 
 
    for (int s = 0; s < nbStates; s++)
    {
        for (int d = 0; d < nbDim; d++)
        {
            model.States[s].Mu[d] = mu_vec[s * nbDim + d];
        }
    }
 
 
    // cout << endl << "Printing the constructed Mu" << endl;
    // for ( int s = 0; s < nbStates; s++ ) {
    //  for (int d = 0; d < nbDim; d++) {
    //      cout << model.States[s].Mu[d]  << "\t";
    //  }
    //  cout << endl;
    // }
 
    for (int s = 0; s < nbStates; s++)
    {
        for (int row = 0; row < nbDim; row++)
        {
            for (int col = 0; col < nbDim; col++)
            {
                int ind = s * nbDim * nbDim + row * nbDim + col;
                model.States[s].Sigma(row, col) = sig_vec[ind];
            }
        }
    }
 
 
    // cout << endl << "Printing the constructed Sigma" << endl;
    // for ( int s = 0; s < nbStates; s++ ) {
    //  for (int row = 0; row < nbDim; row++) {
    //      for (int col = 0; col < nbDim; col++) {
    //          cout << model.States[s].Sigma(row, col) << "\t";
    //      }
    //      cout <<endl;
    //  }
    //  cout << endl;
    // }
}
 
 
void
Gaussians::setGMMs(GMMs* model)
{
    for (unsigned int s = 0; s < model->nbStates; s++)
    {
        this->model.States[s].Mu = model->States[s].Mu;
        this->model.States[s].Sigma = model->States[s].Sigma;
        this->model.States[s].Prio = model->States[s].Prio;
    }
}
 
 
void
Gaussians::InitFastGaussians(int first_inindex, int last_inindex)
{
    double det;
    int nbIN = last_inindex - first_inindex + 1;
 
    for (unsigned int s = 0; s < model.nbStates; s++)
    {
        gmmpinv[s].MuI.Resize(nbIN);
        gmmpinv[s].SigmaII.Resize(nbIN, nbIN);
        gmmpinv[s].SigmaIIInv.Resize(nbIN, nbIN);
    }
 
    for (unsigned int s = 0; s < model.nbStates; s++)
    {
        for (int i = first_inindex; i <= last_inindex; i++)
        {
            gmmpinv[s].MuI(i - first_inindex) = model.States[s].Mu(i);
        }
        for (int i = first_inindex; i <= last_inindex; i++)
            for (int j = first_inindex; j <= last_inindex; j++)
            {
                gmmpinv[s].SigmaII(i - first_inindex, j - first_inindex) =
                    model.States[s].Sigma(i, j);
            }
 
        gmmpinv[s].SigmaII.Inverse(gmmpinv[s].SigmaIIInv, &det);
        //gmmpinv[s].SigmaIIInv = gmmpinv[s].SigmaII.Inverse();
        //gmmpinv[s].SigmaII.Inverse(&det);
        if (det < 0)
        {
            det = 1e-30;
        }
        gmmpinv[s].detSigmaII = det;
    }
 
    nbDimI = last_inindex - first_inindex + 1;
    gfDiff.Resize(nbDimI);
    gfDiffp.Resize(nbDimI);
    gDer.Resize(nbDimI);
}
 
double
Gaussians::GaussianPDFFast(int state, MathLib::Vector x)
{
    double p;
    gfDiff = x - gmmpinv[state].MuI;
    gfDiffp = gmmpinv[state].SigmaIIInv * gfDiff;
 
    p = exp(-0.5 * gfDiff.Dot(gfDiffp)) /
        sqrt(pow(2.0 * PI, nbDimI) * (gmmpinv[state].detSigmaII + 1e-30));
 
    return p;
}
 
double
Gaussians::GaussianProbFast(MathLib::Vector x)
{
    double totalP = 0;
    for (unsigned int s = 0; s < model.nbStates; s++)
    {
        totalP += model.States[s].Prio * GaussianPDFFast(s, x);
    }
    return totalP;
}
 
MathLib::Vector
Gaussians::GaussianDerProbFast(MathLib::Vector x)
{
    gDer.Zero();
    for (unsigned int s = 0; s < model.nbStates; s++)
    {
        gDer += (gmmpinv[s].SigmaIIInv * (x - gmmpinv[s].MuI)) * model.States[s].Prio *
                GaussianPDFFast(s, x);
    }
    return -gDer;
}
 
void
Gaussians::InitFastGMR(int first_inindex, int last_inindex, int first_outindex, int last_outindex)
{
    double det;
    int nbIN = last_inindex - first_inindex + 1;
    int nbOUT = last_outindex - first_outindex + 1;
 
    gPdf.Resize(model.nbStates);
 
    for (unsigned int s = 0; s < model.nbStates; s++)
    {
        gmmpinv[s].MuI.Resize(nbIN);
        gmmpinv[s].SigmaII.Resize(nbIN, nbIN);
        gmmpinv[s].SigmaIIInv.Resize(nbIN, nbIN);
 
        gmmpinv[s].muO.Resize(nbOUT);
        gmmpinv[s].SigmaIO.Resize(nbIN, nbOUT);
        gmmpinv[s].SigmaIOInv.Resize(nbOUT, nbOUT);
    }
 
    for (unsigned int s = 0; s < model.nbStates; s++)
    {
        for (int i = first_inindex; i <= last_inindex; i++)
        {
            gmmpinv[s].MuI(i - first_inindex) = model.States[s].Mu(i);
 
            for (int j = first_inindex; j <= last_inindex; j++)
            {
                gmmpinv[s].SigmaII(i - first_inindex, j - first_inindex) =
                    model.States[s].Sigma(i, j);
            }
            for (int j = first_outindex; j <= last_outindex; j++)
            {
                gmmpinv[s].SigmaIO(i - first_inindex, j - first_outindex) =
                    model.States[s].Sigma(i, j);
            }
        }
 
        for (int i = first_outindex; i <= last_outindex; i++)
        {
            gmmpinv[s].muO(i - first_outindex) = model.States[s].Mu(i);
        }
 
        gmmpinv[s].SigmaII.Inverse(gmmpinv[s].SigmaIIInv, &det);
        if (det < 0)
        {
            det = 1e-30;
        }
        gmmpinv[s].detSigmaII = det;
        (gmmpinv[s].SigmaIO).Transpose().Inverse(gmmpinv[s].SigmaIOInv, &det);
    }
 
    nbDimI = last_inindex - first_inindex + 1;
    gfDiff.Resize(nbDimI);
    gfDiffp.Resize(nbDimI);
    gDer.Resize(nbDimI);
}
 
void
Gaussians::Regression(const MathLib::Vector& indata,
                      MathLib::Vector& outdata,
                      MathLib::Matrix& derGMR)
{
    Regression(indata, outdata);
    cout << "derivative is not implemented yet " << endl;
}
 
void
Gaussians::Regression(const MathLib::Vector& indata, MathLib::Vector& outdata)
{
    double pdfall;
    MathLib::Vector h(model.nbStates);
    MathLib::Vector r_diff(outdata.Size());
 
    for (unsigned int s = 0; s < model.nbStates; s++)
    {
        gPdf(s) = model.States[s].Prio * GaussianPDFFast(s, indata);
    }
    pdfall = gPdf.Sum();
 
    outdata.Zero();
    for (unsigned int s = 0; s < model.nbStates; s++)
    {
        //h(s) = gPdf(s)/(pdfall + 1e-30 );
        h(s) = gPdf(s) / (pdfall);
        r_diff = gmmpinv[s].SigmaIO.Transpose() * gmmpinv[s].SigmaIIInv * (indata - gmmpinv[s].MuI);
 
        for (unsigned int i = 0; i < r_diff.Size(); i++)
        {
            outdata(i) += h(s) * (r_diff(i) + gmmpinv[s].muO(i));
        }
    }
}
 
MathLib::Vector
Gaussians::Regression(const MathLib::Vector& indata)
{
    MathLib::Vector outdata(indata.Size());
    Regression(indata, outdata);
    return outdata;
}
 
 
/*
#include <math.h>
 
#include "Gaussians.h"
#include "armadillo"
 
using namespace arma;
using namespace std;
 
Gaussians::Gaussians(GMMs *model)
{
    this->model.nbStates = model->nbStates;
    this->model.nbDim    = model->nbDim;
 
    this->model.States = (GMMState  *)malloc(model->nbStates*sizeof(GMMState) );
 
    for(int s=0; s<model->nbStates; s++ ){
        this->model.States[s].Mu    = model->GMMState[s].Mu;
        this->model.States[s].Sigma = model->GMMState[s].Sigma;
        this->model.States[s].Prio  = model->GMMState[s].Prio;
    }
}
 
Gaussians::Gaussians(int nbStates, int nbDim, char *f_mu, char *f_sigma, char *f_prio)
{
 
    int s, i, j;
 
    model.nbStates = nbStates;
    model.nbDim    = nbDim;
    model.States = (GMMState  *)malloc(nbStates*sizeof(GMMState) );
 
    for( s=0; s<nbStates; s++ ){
        model.States[s].Mu       =  zeros<vec>(nbDim);
        model.States[s].Sigma    =  zeros<mat>(nbDim, nbDim );
    }
 
    // f_mu
    ifstream fin;
    fin.open(f_mu);
    for( i=0; i<nbDim; i++ ){
        for( s=0; s<nbStates; s++ ){
            fin >> model.States[s].Mu(i);
        }
    }
    fin.close();
 
    // f_sigma
    fin.open(f_sigma);
    for( s=0; s<nbStates; s++ ){
        for( i=0; i<nbDim; i++ ){
            for( j=0; j<nbDim; j++ ){
                fin >> model.States[s].Sigma(i,j);
            }
        }
    }
    fin.close();
 
    // f_prio
    fin.open(f_prio);
    for( s=0; s<nbStates; s++ ){
        fin >>model.States[s].Prio;
    }
    fin.close();
}
 
void Gaussians::setGMMs(GMMs *model)
{
    for(int s=0; s<model->nbStates; s++ ){
        this->model.States[s].Mu    = model->GMMState[s].Mu;
        this->model.States[s].Sigma = model->GMMState[s].Sigma;
        this->model.States[s].Prio  = model->GMMState[s].Prio;
    }
}
 
 
void Gaussians::InitFastGaussians(int first_inindex, int last_inindex)
{
    gmmpinv = (GMMStateP *)malloc(model.nbStates*sizeof(GMMStateP) );
 
    for(int s=0; s<model.nbStates; s++ ){
        gmmpinv[s].MuI = model.States[s].Mu.subvec(first_inindex, last_inindex);
        gmmpinv[s].SigmaII = model.States[s].Sigma.submat(first_inindex, first_inindex, last_inindex, last_inindex);
        gmmpinv[s].SigmaIIInv = pinv(gmmpinv[s].SigmaII);
        gmmpinv[s].detSigmaII = det(gmmpinv[s].SigmaII);
    }
 
    nbDimI = last_inindex - first_inindex +1;
    gfDiff  = zeros<vec>(nbDimI);
    gfDiffp = zeros<vec>(nbDimI);
    gDer    = zeros<vec>(nbDimI);
}
 
double Gaussians::GaussianPDFFast(int state, vec x)
{
    double p;
    gfDiff  = x - gmmpinv[state].MuI;
    gfDiffp = gmmpinv[state].SigmaIIInv * gfDiff;
 
    p = exp(-0.5*dot(gfDiff, gfDiffp)) / sqrt(pow(2.0*math::pi(), nbDimI)*( gmmpinv[state].detSigmaII +1e-30));
 
    return p;
}
 
double Gaussians::GaussianProbFast(vec x)
{
    double totalP=0;
    for(int s=0; s<model.nbStates; s++ ){
        totalP += model.States[s].Prio*GaussianPDFFast(s,x);
    }
    return totalP;
}
 
vec Gaussians::GaussianDerProbFast(vec x)
{
    gDer.zeros();
    for(int s=0; s<model.nbStates; s++ ){
        gDer += model.States[s].Prio * gmmpinv[s].SigmaIIInv *(x-gmmpinv[s].MuI)*GaussianPDFFast(s,x);
    }
    return -gDer;
}
 
//-------------------------------------------------------------------------------------------------------
void AllocateGMMs(GMMs *model, int nbDim, int nbStates)
{
    model->nbDim = nbDim;
    model->nbStates = nbStates;
    model->GMMState = (GMMState  *)malloc(nbStates*sizeof(GMMState) );
 
    for(int s=0; s<nbStates; s++ ){
        model->GMMState[s].Mu       =  zeros<vec>(nbDim);
        model->GMMState[s].Sigma    =  zeros<mat>(nbDim, nbDim );
    }
}
 
 
double GaussianPDF(vec x, vec Mu, mat Sigma)
{
    double p;
    vec diff  = x - Mu;
    vec diffp = pinv( Sigma ) * diff;
    int nbDim = x.size();
 
    p = exp(-0.5*dot(diff, diffp)) / sqrt(pow(2.0*math::pi(), nbDim)*( abs(det(Sigma)) +1e-30));
 
    if(p < 1e-30){
        return 1e-30;
    }
    else{
        return p;
    }
}
 
void GaussianMux(GMMs *modelK, GMMs *modelL, GMMs *modelOut)
{
    int k,l,j;
    int K = modelK->nbStates;
    int L = modelL->nbStates;
    int J = K*L;
 
    //modelOut->nbDim = modelK->nbDim;
    //modelOut->nbStates = J;
    //modelOut->GMMState = (GMMState *)malloc(J*sizeof(GMMState) );
 
    j=0;
    for(k=0; k<K; k++){
        for(l=0; l<L; l++){
            modelOut->GMMState[j].Sigma = pinv( pinv(modelK->GMMState[k].Sigma) + pinv( modelL->GMMState[l].Sigma) );
            modelOut->GMMState[j].Mu    = modelOut->GMMState[j].Sigma *( pinv(modelK->GMMState[k].Sigma) * modelK->GMMState[k].Mu + pinv(modelL->GMMState[l].Sigma) * modelL->GMMState[l].Mu );
            modelOut->GMMState[j].Prio  = modelK->GMMState[k].Prio* modelL->GMMState[l].Prio * GaussianPDF( modelK->GMMState[k].Mu, modelL->GMMState[l].Mu, modelK->GMMState[k].Sigma + modelL->GMMState[l].Sigma);
            j++;
        }
    }
}
 
void GaussianRotate(GMMs *model, arma::vec P, arma::mat R, GMMs *modelOut)
{
    for(int s=0; s<model->nbStates; s++){
        modelOut->GMMState[s].Mu    = R*model->GMMState[s].Mu + P;
        modelOut->GMMState[s].Sigma = R*model->GMMState[s].Sigma*trans(R);
    }
    //modelOut->nbDim = model->nbDim;
    //modelOut->nbStates = model->nbStates;
}
*/