arma::vec cpp_UpdateBeta (arma::mat x_small, arma::vec weights_small,
arma::vec y_small, arma::mat x_big,
arma::vec weights_big, arma::vec y_big,
arma::vec gamma) {
// This is a simple helper function used within the solver for
// weighted linear regression with an l1 penalty.
//
// Args:
// gamma: The current estimate of gamma, used to update beta.
// ...: For all other arguments see cpp_SolveL1Linear.
// Returns:
// beta: The update of the parameter vector beta.
// Obtain the matrix W * X where W is the diagonal weight matrix
arma::mat weighted_X_small(x_small.begin(), x_small.n_rows, x_small.n_cols, true);
arma::mat weighted_X_big(x_big.begin(), x_big.n_rows, x_big.n_cols, true);
weighted_X_small.each_col() %= weights_small;
weighted_X_big.each_col() %= weights_big;
// Update beta by a simple least squares like update.
arma::vec beta;
beta = inv(x_small.t() * weighted_X_small + x_big.t() * weighted_X_big) *
(weighted_X_small.t() * y_small + weighted_X_big.t() * y_big -
x_big.t() * weighted_X_big * gamma);
return beta;
}
// this function sample the loading parameters considering constant
// observational variance.
// [[Rcpp::export]]
arma::mat SampleDfmLoads(arma::mat y, arma::mat factors, arma::vec psi,
int s, double c0)
{
/**
* The data matrix 'y' is (T \times q).
* Note that the factors matrix is (T+h \times k). In the models considered
* in the dissertation, h > s. This means that we always have the
* VAR order, 'h', greater than the number of dynamic factors, 's', in the
* observational equation.
* The argument 's' is the order of the dynamic factors present in the
* observational equation.
* The prior mean is set to zero.
*/
int T = y.n_rows;
int q = y.n_cols;;
int k = factors.n_cols;
int TpH = factors.n_rows;
int h = TpH - T;
if (h < s+1){
throw std::range_error(
"Argument 's' greater than or equal to factors' VAR order in SampleDfmLoads.cpp"
);
}
arma::mat I_q = eye(q, q);
// std::cout << "\nh = " << h << "\n";
arma::mat Xlambda;
Xlambda = factors.rows(h, TpH-1);
// std::cout << "\nNumber of rows in Xlambda is " << Xlambda.n_rows << "\n";
if (s > 0){
for (int i = 1; i <= s; i++){
Xlambda = arma::join_rows(Xlambda, factors.rows(h-i, TpH-i-1));
}
}
// posterior variance
arma::mat Eigvec;
arma::vec eigval;
arma::eig_sym(eigval, Eigvec, arma::symmatu(Xlambda.t() * Xlambda));
arma::mat kronEig = kron(I_q, Eigvec);
arma::vec L1eigval = 1.0/(arma::kron((1.0/psi), eigval) + 1/c0);
arma::mat L1 = kronEig * diagmat(L1eigval) * kronEig.t();
// posterior mean
arma::vec yStar(y.begin(), T*q);
arma::vec l1 = L1 * (arma::kron(arma::diagmat(1.0/psi), Xlambda.t()) * yStar);
arma::mat l1Mat(l1.begin(), k*(s+1), q); // mean in matrix form
// random generation
arma::vec vecnorm = sqrt(L1eigval) % randn(q*k*(s+1), 1);
arma::mat LambdaBar(vecnorm.begin(), k*(s+1), q, false);
LambdaBar = l1Mat + Eigvec * LambdaBar;
return LambdaBar.t();
}
//[[Rcpp::export]]
IntegerVector whichNA(arma::mat X) {
NumericVector x(X.begin(), X.end());
int n = x.size();
IntegerVector y = seq_len(n);//(x.begin(), x.end());
IntegerVector z = y[is_na(x)];
if(z.size()==0)
std::cout << "\nNenhum valor faltante.\n";
return z;
}
/** \brief Extracts the data and passes it to the kmeans algorithm
* \param[in] model An ESBTL Protein model
* \param[in] An armadillo matrix, will get overwritten
* \return A labeling vector
* */
arma::urowvec
operator() (const ESBTL::Default_system::Model& model, arma::mat & data) {
data = arma::mat(3, (unsigned int) std::distance(model.atoms_begin(), model.atoms_end()));
auto mat_it = data.begin();
for (ESBTL::Default_system::Model::Atoms_const_iterator it_atm=model.atoms_begin();it_atm!=model.atoms_end();++it_atm){
*(mat_it++) = it_atm->x();
*(mat_it++) = it_atm->y();
*(mat_it++) = it_atm->z();
}
data = transformation_(data);
arma::urowvec labels = kmeans(data, params_);
return labels;
}
// [[Rcpp::export]]
arma::mat cpp_GenerateWeightMat(double lambda, arma::mat x_small,
arma::mat x_big, arma::mat x_test,
arma::vec mu_small, arma::vec mu_big) {
// This function is the c++ implementation of the function
// GenerateWrightMat. For details see helper_functions.R
//
// Args:
// Note: The arguments remain unchanged. The notation is slightly changed
// because in C++ we cannot name objects with '.' and hence is replaced by
// '_' for object names.
int N_small = x_small.n_rows;
int N_big = x_big.n_rows;
// int N_test = x_test.n_rows;
int p = x_small.n_cols;
arma::mat cov_small(x_small.begin(), N_small, p, true);
cov_small.each_col() %= mu_small % (1 - mu_small);
cov_small = trans(x_small) * cov_small;
arma::mat cov_big(x_big.begin(), N_big, p, true);
cov_big.each_col() %= mu_big % (1 - mu_big);
cov_big = trans(x_big) * cov_big;
arma::mat cov_test = trans(x_test) * x_test;
arma::mat temp_matrix = cov_test * inv_sympd(cov_big) * cov_small;
arma::mat weight_matrix = cov_small + lambda * (cov_test + temp_matrix);
weight_matrix = inv(weight_matrix) * (cov_small + lambda * temp_matrix);
// In case we would like these quantities.
// arma::mat cov_beta_small = inv(cov_small);
// arma::mat cov_beta_lambda = inv(cov_small + cov_big -
// cov_big * inv_sympd(cov_test) * cov_big/lambda);
return weight_matrix;
}
arma::vec cpp_UpdateGamma (arma::mat x_big, arma::vec weights_big,
arma::vec y_big,
arma::vec gamma, arma::vec beta,
double lambda) {
// Another helper function used by the main algorithm.
// In this algorithm we loop through and do a coordinate wise update of
// gamma.
//
// Agrs:
// beta/gamma: The 'current' parameter vectors.
// x/weights/y_big: Design, weights and response for the big population.
// lambda: The value of the penalty parameter.
// Returns:
// gamma: The updated gamma vector.
// Obtain value of dimension.
int p = gamma.size();
// Initialize matrix X * beta.
arma::mat x_beta = x_big * beta;
// We need the diagonal entries of X^T * W * X.
// This is easy when W is a diagonal matrix.
arma::mat W_times_X(x_big.begin(), x_big.n_rows, x_big.n_cols, true);
W_times_X.each_col() %= weights_big;
// We obtain the diagonal entries of X^T * W * X for the denominators.
arma::mat Xt_W_X = x_big.t() * W_times_X;
arma::vec denoms = Xt_W_X.diag();
// Some temporary vectors which we use later.
arma::vec temp_gamma(gamma.begin(), gamma.size(), true);
double temp1;
// Loop through the different gamma values which need to be updated.
for(int i = 0; i < p; i++) {
// Instead of excluding a column we simply set gamma_j = 0 to get the fitted
// value without the effect of gamma_j.
temp_gamma(i) = 0;
temp1 = as_scalar(x_big.col(i).t() * (weights_big % (y_big - x_beta - x_big * temp_gamma)));
gamma(i) = (cpp_sign(temp1) * cpp_max(abs(temp1) - lambda, 0.0))/denoms(i);
temp_gamma(i) = gamma(i);
}
return gamma;
}
//[[Rcpp::export]]
Rcpp::List changeNA(arma::mat X) {
NumericVector x(X.begin(), X.end());
int n = x.size();
IntegerVector y = seq_len(n)-1;//(x.begin(), x.end());
IntegerVector z = y[is_na(x)];
if(z.size()==0){
std::cout << "Nenhum valor faltante.\n";
SEXP xNULL = R_NilValue;
return Rcpp::List::create(xNULL);
} else{
arma::uvec Xna = as<arma::uvec>(z);
NumericVector bx(Xna.n_rows, 100.0);
arma::vec b = as<arma::vec>(bx);
X.elem(Xna) = b;
return Rcpp::List::create(X);
}
}
/** \brief Extracts the data and passes it to the kmeans algorithm
* \param[in] model An ESBTL Protein model
* \param[in] An armadillo matrix, will get overwritten
* \return A labeling vector
*/
arma::urowvec
operator() (const ESBTL::Default_system::Model& model, arma::mat & data) {
data = arma::mat(3, (unsigned int) std::distance(model.atoms_begin(), model.atoms_end()));
arma::urowvec labels = arma::urowvec(data.n_cols);
unsigned int * label_it = labels.begin();
unsigned int label = 0;
auto mat_it = data.begin();
for (ESBTL::Default_system::Model::Chains_const_iterator it_cha=model.chains_begin();it_cha!=model.chains_end();++it_cha){
for (ESBTL::Default_system::Model::Chain::Atoms_const_iterator it_atm = it_cha->atoms_begin(); it_atm != it_cha->atoms_end(); ++it_atm) {
*(mat_it++) = it_atm->x();
*(mat_it++) = it_atm->y();
*(mat_it++) = it_atm->z();
*(label_it++) = label;
}
++label;
}
return labels;
}
//[[Rcpp::export]]
arma::vec cpp_SolveLogistic(arma::mat X, arma::vec y,
arma::vec initial_beta,
int max_iter = 100,
double tol = 1e-5) {
// The C++ implementation of the function SolveLogistic.
// One main difference is that I cannot figure out how to pass null values
// to c++. In this case an initial estimate of beta is required.
// Initialize some vectors
arma::vec beta(initial_beta.begin(), initial_beta.size(), false);
arma::vec beta_new = beta;
arma::vec mu_current = cpp_GetFittedValues(X, beta);
// Begin main loop.
for (int i = 1; i <= max_iter; i++) {
// Set current vector of fitted values.
mu_current = cpp_GetFittedValues(X, beta);
// Evaluate V * X where V is the diagonal variance matrix.
arma::mat var_times_X(X.begin(), X.n_rows, X.n_cols, true);
var_times_X.each_col() %= mu_current % (1 - mu_current);
// Update beta.
beta_new = beta + inv(trans(X) * var_times_X) * (trans(X) * (y - mu_current));
// Check for convergence.
if (sum(pow(beta_new - beta, 2)) < tol) {
return beta_new;
} else {
beta = beta_new;
}
}
// Throw warning if it did not converge.
Function warning("warning");
warning("The newton-raphson algorithm did not converge.");
// Return the non-converged beta vector.
return beta_new;
}
////> Main function ////
// [[Rcpp::export]]
List calcPotts_cpp (const S4& W_SR, const S4& W_LR, arma::mat sample, NumericVector rho, const arma::mat& coords,
const IntegerVector& site_order, int iter_max, double cv_criterion,
bool test_regional, NumericVector distance_ref, double threshold, double neutre, int nbGroup_min, bool multiV, bool last_vs_others,
bool prior, bool type_reg, int verbose){
//// preparation
int p = sample.n_cols;
int n = sample.n_rows;
IntegerVector W_SRi = W_SR.slot("i");
IntegerVector W_SRp = W_SR.slot("p");
NumericVector W_SRx = W_SR.slot("x");
arma::mat Wpred(n, p);
vector < vector < double > > pred_global(p);
for(int iter_p = 0 ; iter_p < p ; iter_p ++){
pred_global[iter_p].resize(n);
for(int iter_px = 0 ; iter_px < n ; iter_px++){
pred_global[iter_p][iter_px] = sample(iter_px, iter_p);
}
}
vector < vector < double > > pred_global_hist = pred_global;
if(prior == false){
std::fill(sample.begin(), sample.end(), 1);
}
arma::mat V(n, p);
IntegerVector rang;
bool no_site_order = (site_order[0] < 0);
if(no_site_order == false){
rang = site_order;
}
int index_px;
double norm;
double val_criterion = cv_criterion + 1;
double diff_criterion = 0, diff;
double cv_criterion2 = n * cv_criterion ;
int iter_updateV = 0 ;
vector < double > res_multipotentiel(n) ;
int iter = 0 ;
//// estimation
while(iter < iter_max && ( (val_criterion > cv_criterion) || (test_regional == true && iter_updateV != iter) ) ){
iter++;
pred_global_hist = pred_global;
if(no_site_order){
rang = rank_hpp(runif(n)) - 1; // tirer aleatoirement l ordre des sites
}
//// regional potential
if(test_regional == true){
if(iter == 1 || val_criterion <= cv_criterion || diff_criterion > cv_criterion2){
iter_updateV = iter ;
for(int iter_p = (p - 1) ; iter_p>= 0 ; iter_p --){
if(last_vs_others == false || iter_p == (p - 1)){
res_multipotentiel = calcMultiPotential_hpp(W_SR, W_LR, pred_global[iter_p],
threshold, coords, Rcpp::as < std::vector < double > >(distance_ref),
nbGroup_min, multiV,
neutre)[0];
for(int iter_px = 0 ; iter_px < n ; iter_px++){
V(iter_px, iter_p) = res_multipotentiel[iter_px];
if(type_reg){
V(iter_px, iter_p) = V(iter_px, iter_p) * (V(iter_px, iter_p) - pred_global[iter_p][iter_px]);
}
}
}else{
V.col(iter_p) = 1 - V.col(p - 1);
}
}
}else{
cv_criterion2 /= 1.1;
}
}
//// update site probabilities
val_criterion = 0;
diff_criterion = 0;
for(int iter_px = 0 ; iter_px < n ; iter_px++){ // pour chaque pixel
norm = 0.0;
index_px = rang(iter_px);
for(int iter_p = 0 ; iter_p < p ; iter_p++){ // pour chaque groupe
Wpred(index_px, iter_p) = 0.0;
for(int iter_vois = W_SRp[index_px]; iter_vois < W_SRp[index_px + 1]; iter_vois++){ // pour chaque voisin
if(type_reg){
Wpred(index_px, iter_p) += W_SRx[iter_vois]*pred_global[iter_p][W_SRi[iter_vois]]*max(0.0, pred_global[iter_p][W_SRi[iter_vois]]-sample(index_px, iter_p));
}else{
Wpred(index_px, iter_p) += W_SRx[iter_vois] *pred_global[iter_p][W_SRi[iter_vois]];
}
}
// exponentielle rho
if(test_regional == false){
pred_global[iter_p][index_px] = sample(index_px, iter_p) *exp(rho(0) *Wpred(index_px, iter_p)) ;
}else{
pred_global[iter_p][index_px] = sample(index_px, iter_p) *exp(rho(0) *Wpred(index_px, iter_p) + rho(1) *V(index_px, iter_p));
}
norm += pred_global[iter_p][index_px];
}
// normalisation
for(int iter_p = 0 ; iter_p < p ; iter_p++){
pred_global[iter_p][index_px] /= norm;
diff = abs(pred_global_hist[iter_p][index_px]-pred_global[iter_p][index_px]);
diff_criterion += diff;
val_criterion = max(val_criterion, diff);
}
}
if(verbose > 0){
if(verbose == 1){Rcout << "*" ;}
if(verbose == 2){Rcout << "iteration " << iter << " : totaldiff = " << diff_criterion << " | maxdiff = " << val_criterion << endl;}
}
}
if(verbose == 1){Rcout << endl;}
//// export
arma::mat spatialPrior(n, p);
for(int iter_px = 0 ; iter_px < n ; iter_px++){
index_px = rang(iter_px);
for(int iter_p = 0 ; iter_p < p ; iter_p++){
if(test_regional == false){
spatialPrior(index_px, iter_p) = exp(rho(0) *Wpred(index_px, iter_p)) ;
}else{
spatialPrior(index_px, iter_p) = exp(rho(0) *Wpred(index_px, iter_p) + rho(1) *V(index_px, iter_p));
}
}
}
return Rcpp::List::create(Rcpp::Named("predicted") = pred_global,
Rcpp::Named("spatialPrior") = spatialPrior,
Rcpp::Named("cv") = val_criterion <= cv_criterion,
Rcpp::Named("iter_max") = iter
);
}
//[[Rcpp::export]]
arma::vec cpp_SolvePenaltyL2One (arma::mat x_full, arma::mat x_fancy,
arma::vec y_full, arma::mat x_test,
arma::vec pars_full,
double lambda,
int max_iter = 100, double tol = 1e-5) {
// The cpp version of the penalized regression solver.
// In this case we use a slightly different notation.
// Here also we solve for only a single value of lambda.
// Args:
// x_full: The top four blocks of the matrix fancy X in notes/GLM.pdf.
// x_fancy: The fancy X matrix in notes/GLM.pdf.
// x_test: The test set design matrix.
// y_full: The non-zero part of the fancy y vector of notes/GLM.pdf.
// pars_full: The parameter vector (beta, gamma).
// ...: The other parameters remain unchanged.
//
// Returns:
// The full parameter vector (beta, gamma).
// Initialize p.
arma::uword p = x_test.n_cols;
// Initialize vectors for algorithm.
arma::vec pars_full_new;
arma::vec mu_full;
arma::vec mu_full_new;
// Begin main loop
for (int i = 1; i <= max_iter; i++) {
// Get current mean vector
mu_full = 1/(1 + exp(-x_full * pars_full));
// Create the y fancy vector.
arma::vec y_fancy = join_vert(y_full - mu_full,
-lambda * x_test * pars_full.tail(p));
// Create the diagonal of the fancy V matrix.
arma::vec v_fancy = join_vert(mu_full % (1 - mu_full),
lambda * ones<vec>(x_test.n_rows));
// As before, we first obtain the V_fancy * X_fancy.
arma::mat var_times_X(x_fancy.begin(), x_fancy.n_rows, x_fancy.n_cols, true);
var_times_X.each_col() %= v_fancy;
// Update the parameter vector.
pars_full_new = pars_full +
solve(trans(x_fancy) * var_times_X, trans(x_fancy) * y_fancy);
// Check for convergence.
if (sum(pow(pars_full_new - pars_full, 2)) < tol) {
return pars_full_new;
} else {
pars_full = pars_full_new;
}
}
// Throw warning if it did not converge.
Function warning("warning");
warning("The newton-raphson algorithm did not converge.");
// Return the non-converged beta vector.
return pars_full_new;
}
arma::vec mat2vec(arma::mat y)
{
arma::vec b(y.begin(), y.size(), /*copy_aux_mem*/false, /*strict*/true);
return b;
}
