//' @title Safe softmax function for computing weights
//'
//' @description Computes the weights \eqn{w_i = \frac{e^{p_i}}{\sum_{j=1}^k e^{p_j}}} from \eqn{p_i}, \eqn{i=1,\ldots,k}
//' in a safe way to avoid overflows and to truncate automatically to zero low values of \eqn{w_i}.
//'
//' @param logs matrix of logarithms where each row contains a set of \eqn{p_1,\ldots,p_k} to compute the weights from.
//' @param etrunc truncation for exponential: \code{exp(x)} with \code{x <= -etrunc} is set to zero. Defaults to \code{30}.
//' @return A matrix of the size as \code{logs} containing the weights for each row.
//' @author Eduardo Garcia-Portugues (\email{egarcia@@math.ku.dk})
//' @details The \code{logs} argument must be always a matrix.
//' @examples
//' # A matrix
//' safeSoftMax(rbind(1:10, 20:11))
//' rbind(exp(1:10) / sum(exp(1:10)), exp(20:11) / sum(exp(20:11)))
//'
//' # A row-matrix
//' safeSoftMax(rbind(-100:100), etrunc = 30)
//' exp(-100:100) / sum(exp(-100:100))
//' @export
// [[Rcpp::export]]
arma::mat safeSoftMax(arma::mat logs, double etrunc = 30) {
// Maximum of logs by rows to avoid overflows
arma::vec m = max(logs, 1);
// Recenter by columns
logs.each_col() -= m;
// Ratios by columns
logs.each_col() -= log(sum(exp(logs), 1));
// Truncate exponential by using a lambda function - requires C++ 11
logs.transform([etrunc](double val) { return (val < -etrunc) ? double(0) : double(exp(val)); });
return logs;
}
/**
* Construct the exact kernel matrix.
*
* @param data Input data points.
* @param transformedData Matrix to output results into.
* @param eigval KPCA eigenvalues will be written to this vector.
* @param eigvec KPCA eigenvectors will be written to this matrix.
* @param rank Rank to be used for matrix approximation.
* @param kernel Kernel to be used for computation.
*/
static void ApplyKernelMatrix(const arma::mat& data,
arma::mat& transformedData,
arma::vec& eigval,
arma::mat& eigvec,
const size_t /* unused */,
KernelType kernel = KernelType())
{
// Construct the kernel matrix.
arma::mat kernelMatrix;
// Resize the kernel matrix to the right size.
kernelMatrix.set_size(data.n_cols, data.n_cols);
// Note that we only need to calculate the upper triangular part of the
// kernel matrix, since it is symmetric. This helps minimize the number of
// kernel evaluations.
for (size_t i = 0; i < data.n_cols; ++i)
{
for (size_t j = i; j < data.n_cols; ++j)
{
// Evaluate the kernel on these two points.
kernelMatrix(i, j) = kernel.Evaluate(data.unsafe_col(i),
data.unsafe_col(j));
}
}
// Copy to the lower triangular part of the matrix.
for (size_t i = 1; i < data.n_cols; ++i)
for (size_t j = 0; j < i; ++j)
kernelMatrix(i, j) = kernelMatrix(j, i);
// For PCA the data has to be centered, even if the data is centered. But it
// is not guaranteed that the data, when mapped to the kernel space, is also
// centered. Since we actually never work in the feature space we cannot
// center the data. So, we perform a "psuedo-centering" using the kernel
// matrix.
arma::rowvec rowMean = arma::sum(kernelMatrix, 0) / kernelMatrix.n_cols;
kernelMatrix.each_col() -= arma::sum(kernelMatrix, 1) / kernelMatrix.n_cols;
kernelMatrix.each_row() -= rowMean;
kernelMatrix += arma::sum(rowMean) / kernelMatrix.n_cols;
// Eigendecompose the centered kernel matrix.
arma::eig_sym(eigval, eigvec, kernelMatrix);
// Swap the eigenvalues since they are ordered backwards (we need largest to
// smallest).
for (size_t i = 0; i < floor(eigval.n_elem / 2.0); ++i)
eigval.swap_rows(i, (eigval.n_elem - 1) - i);
// Flip the coefficients to produce the same effect.
eigvec = arma::fliplr(eigvec);
transformedData = eigvec.t() * kernelMatrix;
transformedData.each_col() /= arma::sqrt(eigval);
}
/** Initialization
*
* @return Estimated state \f$\{\tilde{\mathbf{x}}^{(i)}_0,\tilde{\omega}^{(i)}\}_{i=1}^{M}\f$
*/
CompeleteState initialize() {
state_par_.each_col([this](arma::vec &v) {
v = process_.template getProcess<0>().getInitialPDF().random();
});
unsigned long cnt = 0;
w_.for_each([this, &cnt](double &e) {
e = process_.template getProcess<0>().getInitialPDF().likelihood(
state_par_.col(cnt++));
});
normalizeWeights();
return std::make_tuple(state_par_, w_);
}
void predict(const Args &... args) {
state_par_.each_col([this, &args...](arma::vec &v, const Args &... args) {
v = process_.template getProcess<0>().getCPDF().random(v, args...);
});
}
Rcpp::List irls_binomial_cpp_fast_br(arma::mat A, arma::vec b, double maxit, double tol)
{
//Def
arma::vec x;
x.zeros(A.n_cols,1);
arma::vec xold;
arma::mat varmatrix;
double nobs;
nobs = A.n_rows;
double df;
df = A.n_cols;
double n;
double ll;
double aic;
double bic;
double mdl;
arma::vec W(nobs);
arma::vec unit(nobs);
unit.ones(nobs);
arma::vec eta(nobs);
arma::vec g(nobs);
arma::vec gprime(nobs);
arma::vec z(nobs);
//mod
arma::vec bprime(nobs);
//int k;
for (int i = 0; i < maxit; ++i) {
eta = A * x;
for (int j=0; j < nobs; ++j) {
g[j] = 1.0 / (1.0 + exp(-1.0 * eta[j]));
gprime[j] = exp (-1.0 * eta[j]) / ((1.0 + exp (-1.0 * eta[j])) * (1.0 + exp (-1.0 * eta[j])));
//mod
bprime[j] = (b[j]+(df/nobs)*(0.5))/(1+(df/nobs));
//bprime[j] = b[j]+(sum(b)/nobs)*(0.5);
}
//z = eta+(b-g)/gprime;
z = eta+(bprime-g)/gprime;
W = (gprime % gprime);
W /= (g % (unit-g));
//W += unit;
xold = x;
//coefficients
//x = arma::solve(A.t()*(W % A.each_col()), A.t()*(W % z), arma::solve_opts::no_approx);
varmatrix = A.t()*(W % A.each_col());
x = arma::solve(varmatrix, A.t()*(W % z), arma::solve_opts::no_approx);
//k = i;
if(sqrt(arma::dot(x-xold,x-xold)) < tol){
break;
}}
n = A.n_rows;
//arma::vec e;
//double ssr;
//e = (b - A*x);
//e = (bprime - A*x);
//ssr = accu(e.t()*e);
//scores
//ll = arma::accu(-arma::dot(b,log(unit + exp(-(A*x)))) - arma::dot((unit-b),log(unit + exp(A*x))));
ll = arma::accu(-arma::dot(b,log(unit + exp(-(A*x)))) - arma::dot((unit-b),log(unit + exp(A*x))));
aic = - 2 * ll + 2 * df;
bic = - 2 * ll + log(nobs) * df;
//mdl
// arma::mat xz;
// xz.zeros(size(x));
//
// arma::vec ez;
// double ssrz;
// double ssrtot;
// double RR;
// double F;
// double mdl;
// arma::vec yaverage(n);
//
// ez = (b - A*xz);
// ssrz = accu(ez.t()*ez);
// F = (((ssrz - ssr)/df)/(ssr/((n-(df + 1)))));
//
// for (int j=0; j < n; ++j) {
// yaverage[j] = b[j] - arma::mean(b);
// }
//
// ssrtot = accu(yaverage.t()*yaverage);
//
// RR = 1-(ssr/ssrtot);
//
// if (RR > (df/n)) {
// mdl = (n/2) * log(ssr/(n-df)) + (df/2) * log(F) + log(n);
// } else {
// mdl = (n/2) * log((accu(b.t()*b))/n) + 0.5 * log(n);
// }
//return
return Rcpp::List::create(
Rcpp::Named("loglik") = ll,
Rcpp::Named("aic") = aic,
Rcpp::Named("bic") = bic,
Rcpp::Named("mdl") = mdl
);
}
//[[Rcpp::export(.emplik_intern)]]
List emplik_intern(arma::mat z, arma::colvec mu, arma::vec lam, double eps,
double M = 1e30, double thresh = 1e-12, int itermax = 1000){
//(arma::mat z, arma::vec mu = vec(z.n_cols,fill::zeros), double eps = 1/z.nrows, double M = datum::inf);
// # Backtracking line search parameters [Tweak only with extreme caution.]
// # See Boyd and Vandenberghe, pp 464-466.
double ALPHA = 0.3; // seems better than 0.01 on some 2d test data (sometimes fewer iters)
double BETA = 0.8;
// # We need 0 < ALPHA < 1/2 and 0 < BETA < 1
// # Backtrack threshold: you can miss by this much.
double BACKEPS = 0;
// # Consider replacing 0 by 1e-10 if backtracking seems to be
// # failing due to round off.
int n = z.n_rows;
int d = z.n_cols;
z.each_row() -= trans(mu);
// Use lam = 0 or initial lam, whichever is best
arma::vec onen = arma::vec(n,arma::fill::ones);
arma::mat init0 = mllog(onen, eps, M, 2);
arma::mat init(init0.n_rows, init0.n_cols);
if(!any(lam)){
init = init0;
} else {
init = mllog(onen+z*lam, eps, M, 2);
if(sum(init0.col(0) < sum(init.col(0)))){
lam = arma::rowvec(z.n_cols,arma::fill::zeros);
init = init0;
}
}
// # Initial f, g
double fold = sum(init.col(0));
arma::rowvec gold = sum(z.each_col() % init.col(1),0);
bool converged = false;
int iter = 0;
arma::mat oldvals = init;
arma::mat newvals(oldvals.n_rows, oldvals.n_cols);
arma::vec rootllpp;
arma::mat zt(z.n_rows, z.n_cols);
arma::vec yt(z.n_rows);
arma::vec wts;
double fnew; double targ;
double logelr;
arma::vec step(z.n_cols);
int s; double ndec; double gradnorm;
bool backtrack = false;
while(!converged){
iter += 1;
// # Get Newton Step
rootllpp = sqrt(oldvals.col(2)); //# sqrt 2nd deriv of -llog lik
zt = z;
for(int j=0; j<d; j++){
zt.col(j) %= rootllpp;
}
yt = oldvals.col(1) / rootllpp;
step = -svdlm(zt,yt);
backtrack = false;
s = 1;
while(!backtrack){
newvals = mllog(onen + z * (lam+s*step), eps, M, 2);
fnew = sum(newvals.col(0));
targ = fold + ALPHA * s * sum(trans(gold) % step) + BACKEPS; // (BACKEPS for roundoff, should not be needed)
if(fnew <= targ){ // backtracking has converged
backtrack = true;
oldvals = newvals;
fold = fnew;
gold = sum(z.each_col() % oldvals.col(1),0);
lam = lam + s*step;
} else{
s = s * BETA;
}
}
// # Newton decrement and gradient norm
ndec = sqrt(sum(square(step % trans(gold))));
gradnorm = sqrt(sum(square(gold)));
converged = (ndec * ndec <= thresh);
if( iter > itermax )
break;
}
wts = pow(1 + z * lam, -1)/n;
logelr = sum(mllog(onen + z * lam, eps, M, 0).col(0));
return Rcpp::List::create(Named("logelr")=logelr, Named("lam") = lam, Named("wts") = wts,
Named("conv") = converged, Named("niter") = iter,Named("ndec") = ndec, Named("gradnorm") = gradnorm);
}
