void load_sigma2(const fmat &V, fvec &sigma2){
int nsnp = V.size();
sigma2 = fvec (nsnp, .0f);
for(int i = 0; i < nsnp; ++i){
sigma2[i] = V[i][i];
}
}
// X es la matriz de datos ([m casos] X [n features + BIAS])
// X ya se supone normalizada, y con la columna de BIAS agregada.
// Y es la matriz de respuestas ([m casos] X [c categorias posibles])
fmat SGD(const fmat& X, const fmat& Y, double alpha) {
int m = X.n_rows; // Filas = casos
int n = X.n_cols; // Columnas = features + BIAS
int c = Y.n_cols; // Categorias posibles
double lambda = 4.0;
fmat Theta(n, c);
fmat reg(n, c);
fmat gradient(n, c);
Theta.fill(0.0);
int its = m/SGD_N;
fmat subX, subY;
double loss;
for (int i = 0; i < GD_IT; i++) {
// SGD. Debería modularizar un poco esto. Quizás con un define.
// cout << "iterancion " << i << endl;
for (int j = 0; j < its; j++) {
subX = X.rows(SGD_N*j, SGD_N*(j+1)-1);
subY = Y.rows(SGD_N*j, SGD_N*(j+1)-1);
Theta = gdStep(Theta, subX, subY, alpha, lambda);
}
// Tomo las filas que faltan.
subX = X.rows(its*SGD_N, m - 1);
subY = Y.rows(its*SGD_N, m - 1);
Theta = gdStep(Theta, subX, subY, alpha, lambda);
cout << "terminada la iteración: %d" << i;
#ifndef NDEBUG
if (i % 10 == 0) {
loss = logloss(predict(X, Theta), Y);
cout << " logloss %G" << loss;
}
#endif
cout << endl;
}
loss = logloss(predict(X, Theta), Y);
cout << "Logloss final: " << loss << endl;
return Theta;
}
//function [y,Y,P,Y1]=ut(f,X,Wm,Wc,n,R)
void BFilterUKF::utMeasurement(fmat X, fvec Wm, fvec Wc, unsigned int n, fmat R)
{
//Unscented Transformation
//Input:
// f: nonlinear map
// X: sigma points
// Wm: weights for mean
// Wc: weights for covraiance
// n: numer of outputs of f
// R: additive covariance
//Output:
// y: transformed mean
// Y: transformed smapling points
// P: transformed covariance
// Y1: transformed deviations
unsigned int L=X.n_cols;
z1=zeros<fvec>(n);
Z1=zeros<fmat>(n,L);
//for k=1:L
for (unsigned int k=0; k < L; ++k)
{
fmat XColK = X.col(k);
Z1.col(k)= process->ffun(&XColK);
z1=z1+Wm(k)*Z1.col(k);
}
//Z2=Z1-x1(:,ones(1,L));
Z2 = Z1;
for (unsigned int j = 0; j < L; ++j)
{
for (unsigned int i = 0; i < x1.n_rows; ++i)
{
Z2(i,j) -= x1(i);
}
}
P2=Z2*Wc.diag()*Z2.t()+R;
}
//function [y,Y,P,Y1]=ut(f,X,Wm,Wc,n,R)
void BFilterUKF::utProcess(fmat X,fvec Wm, fvec Wc, unsigned int n, fmat R)
{
//Unscented Transformation
//Input:
// f: nonlinear map
// X: sigma points
// Wm: weights for mean
// Wc: weights for covraiance
// n: numer of outputs of f
// R: additive covariance
//Output:
// y: transformed mean
// Y: transformed smapling points
// P: transformed covariance
// Y1: transformed deviations
unsigned int L=X.n_cols;
x1 = zeros<fvec>(n);
X1 = zeros<fmat>(n,L);
//for k=1:L
for (unsigned int k=0; k < L; ++k)
{
fmat XColK = X.col(k);
X1.col(k)= process->ffun(&XColK);
x1=x1+Wm(k)*X1.col(k);
}
//X2=X1-x1(:,ones(1,L)); // generate duplicates of vector x1
X2 = X1;
for (unsigned int j = 0; j < L; ++j)
{
for (unsigned int i = 0; i < x1.n_rows; ++i)
{
X2(i,j) -= x1(i);
}
}
P1=X2*Wc.diag()*X2.t()+R;
}
fmat gdStep(const fmat& Theta, const fmat& X, const fmat& Y, double alpha, double lambda) {
fmat gradient = (alpha / X.n_rows) * X.t() * (sigmoide(X * Theta) - Y);
fmat reg = (lambda / X.n_rows) * Theta;
reg.row(0) = zeros<frowvec>(Y.n_cols);
return Theta - gradient - reg;
}
fmat scaleFeatures(fmat X, fmat mu, fmat sigma, int columns) {
for (unsigned int i = 0; i < columns; ++i) {
X.col(i) = (X.col(i) - mu(i)) / sigma(i);
}
return X;
}
