void HDP::doc_state_update(DocState* doc_state, int i, int update) {
int d = doc_state->doc_id_;
int w = doc_state->words_[i].word_;
int c = doc_state->words_[i].count_;
int k = doc_state->words_[i].topic_assignment_;
//assert(k >= 0); // we must have it assigned before or assigned to a new one.
if (update > 0) {
if (hdp_state_->topic_lambda_[k][w] == 0)
unique_topic_by_word_[w].insert(k);
if (word_counts_by_topic_doc_[k][d] == 0)
unique_topic_by_doc_[d].insert(k);
}
update *= c;
// Update HDP state
smoothing_prob_sum_ -= smoothing_prob_[k];
hdp_state_->word_counts_by_topic_[k] += update;
hdp_state_->topic_lambda_[k][w] += update;
doc_prob_sum_[d] -= doc_prob_[k][d];
word_counts_by_topic_doc_[k][d] += update;
if (update < 0 ) {
if (hdp_state_->topic_lambda_[k][w] == 0)
unique_topic_by_word_[w].erase(k);
if (word_counts_by_topic_doc_[k][d] == 0)
unique_topic_by_doc_[d].erase(k);
}
if (update > 0 && k == hdp_state_->num_topics_) { // a new topic is generated.
RNGScope scope;
hdp_state_->num_topics_ ++;
double new_stick = Rf_rbeta(1.0, hdp_state_->gamma_) * hdp_state_->pi_left_;
hdp_state_->pi_left_ = hdp_state_->pi_left_ - new_stick;
hdp_state_->pi_[k] = new_stick;
if ((int)hdp_state_->word_counts_by_topic_.size() < hdp_state_->num_topics_ + 1) {
int new_size = 2 * hdp_state_->num_topics_ + 1;
vct_ptr_resize(&hdp_state_->topic_lambda_, new_size, hdp_state_->size_vocab_);
hdp_state_->word_counts_by_topic_.resize(new_size, 0);
hdp_state_->beta_u_.resize(new_size, 0);
//hdp_state_->beta_v_.resize(new_size, 0.0);
hdp_state_->pi_.resize(new_size, 0.0);
vct_ptr_resize(&word_counts_by_topic_doc_, new_size, num_docs_);
vct_ptr_resize(&table_counts_by_topic_doc_, new_size, num_docs_);
smoothing_prob_.resize(new_size, 0.0);
vct_ptr_resize(&doc_prob_, new_size, num_docs_);
}
}
double etaW = hdp_state_->size_vocab_ * hdp_state_->eta_;
smoothing_prob_[k] = hdp_state_->alpha_ * hdp_state_->pi_[k] / (hdp_state_->word_counts_by_topic_[k] + etaW);
smoothing_prob_sum_ += smoothing_prob_[k];
doc_prob_[k][d] = word_counts_by_topic_doc_[k][d] / (hdp_state_->word_counts_by_topic_[k] + etaW);
doc_prob_sum_[d] += doc_prob_[k][d];
}
void HDP::sample_first_level_concentration(double gamma_a, double gamma_b) {
/// (p 585 in escobar and west)
double shape = gamma_a;
double scale = gamma_b;
int n = 0;
RNGScope scope;
for (int k = 0; k < hdp_state_->num_topics_; ++k) {
n += hdp_state_->beta_u_[k];
}
double eta = Rf_rbeta(hdp_state_->gamma_ + 1, n);
double pi = shape + hdp_state_->num_topics_ - 1;
double rate = 1.0 / scale - log(eta);
pi = pi / (pi + rate * n);
unsigned int cc = Rf_rbinom(1,pi);
if (cc == 1)
hdp_state_->gamma_ = Rf_rgamma(shape + hdp_state_->num_topics_, 1.0 / rate);
else
hdp_state_->gamma_ = Rf_rgamma(shape + hdp_state_->num_topics_ - 1, 1.0 / rate);
}
void HDP::sample_second_level_concentration(double alpha_a, double alpha_b) {
double shape = alpha_a;
double scale = alpha_b;
RNGScope scope;
int n = 0;
for (int k = 0; k < hdp_state_->num_topics_; ++k) {
n += hdp_state_->beta_u_[k];
}
double rate, sum_log_w, sum_s;
for (int step = 0; step < 20; ++step) {
sum_log_w = 0.0;
sum_s = 0.0;
for (int d = 0; d < num_docs_; ++d) {
sum_log_w += log(Rf_rbeta(hdp_state_->alpha_ + 1, doc_states_[d]->doc_length_));
sum_s += (double)Rf_rbinom(1,doc_states_[d]->doc_length_ / (doc_states_[d]->doc_length_ + hdp_state_->alpha_));
}
rate = 1.0 / scale - sum_log_w;
hdp_state_->alpha_ = Rf_rgamma(shape + n - sum_s, 1.0 / rate);
}
}
Type rbeta(Type shape1, Type shape2)
{
return Rf_rbeta(asDouble(shape1), asDouble(shape2));
}
extern "C" void glm_gibbs(double * rZ, double * rxo, double * rlam, int * rmodelprior, double * rpriorprob, double * rbeta1, double * rbeta2, int * rburnin, int * rniter, int * rscalemixture, double * ralpha, int * rno, int * rna, int * rp, double * B_mcmc, double * prob_mcmc, int * gamma_mcmc, double * phi_mcmc, double * lam_mcmc, double * B_rb, double * prob_rb, double * intercept_mcmc, double * xo_scale)
{
GetRNGstate();
//MCMC Variables//
int burnin=*rburnin;
int niter=*rniter;
//Dimensions//
int p=*rp;
int no=*rno;
int na=*rna;
//Phi Variables//
double phi=1.0;
//Yo Variables//
std::vector<double> Z(rZ, rZ+no);
std::vector<double> xo(rxo, rxo+no*p);
standardize_xo(xo,xo_scale,no,p);
std::vector<double> xoyo(p);
double yobar=0;
std::vector<double> xoxo(p*p);
dgemm_( &transT, &transN, &p, &p, &no, &unity, &*xo.begin(), &no, &*xo.begin(), &no, &inputscale0, &*xoxo.begin(), &p );
//Construct Xa//
std::vector<double> xa(p*(p+1)/2); //Triangular Packed Storage
std::vector<double> d(p);
chol_xa(xa,xoxo,d,p);
//Reserve Memory for Submatrices//
std::vector<double> xog; xog.reserve(no*p);
std::vector<double> xogyo; xogyo.reserve(p);
std::vector<double> xogxog_Lamg; xogxog_Lamg.reserve(p*p);
std::vector<double> xag; xag.reserve(na*p);
//Ya Variables//
std::vector<double> xaya(p);
//Beta Variables//
double intercept=0;
std::vector<double> Bols(p);
std::vector<double> B(p,0.0);
std::vector<double> Bg; Bg.reserve(p);
//Lambda Variables//
int scalemixture=*rscalemixture;
double alpha=*ralpha;
std::vector<double> lam(rlam,rlam+p);
std::vector<double> lamg; lamg.reserve(p); //vector instead of diagonal pxp matrix
//Gamma Variables//
std::vector<int> gamma(p,1);
int p_gamma=std::accumulate(gamma.begin(),gamma.end(),0);
bool gamma_diff=true;
int modelprior=*rmodelprior;
//Probability Variables//
std::vector<double> prob(p);
std::vector<double> odds(p);
std::vector<double> priorprob(rpriorprob,rpriorprob+p);
//Theta Variables//
double theta=0.5;
double beta1=*rbeta1;
double beta2=*rbeta2;
//Store Initial Values//
std::copy(B.begin(),B.end(),B_mcmc);
std::copy(prob.begin(),prob.end(),prob_mcmc);
std::copy(gamma.begin(),gamma.end(),gamma_mcmc);
std::copy(lam.begin(),lam.end(),lam_mcmc);
//Run Gibbs Sampler//
for (int t = 1; t < niter; ++t)
{
//Form Submatrices//
if(p_gamma) submatrices_uncollapsed(gamma_diff,B,xog,xag,lamg,Bg,gamma,lam,xo,xa,p_gamma,p,no,na);
//Draw xoyo//
draw_xoyo(Z,xoyo,yobar,xo,xog,Bg,phi,no,p,p_gamma,intercept);
//Draw xaya//
draw_uncollapsed_xaya(xaya,xa,xag,Bg,phi,na,p,p_gamma);
//Compute Probabilities//
if(modelprior==1)
{
bernoulli_probabilities(prob,odds,Bols,d,xoyo,xaya,priorprob,lam,phi);
}else if(modelprior==2)
{
betabinomial_probabilities(prob,odds,Bols,d,xoyo,xaya,theta,lam,phi);
}else
{
uniform_probabilities(prob,odds,Bols,d,xoyo,xaya,lam,phi);
}
//Draw Gamma//
draw_gamma(gamma,p_gamma,prob);
//Draw Theta//
if(modelprior==2) theta=Rf_rbeta(beta1+p_gamma,p-p_gamma+beta2);
//Draw Beta//
draw_beta(gamma,B,Bols,d,lam,phi);
//Draw Intercept//
intercept=yobar+sqrt(1/(no*phi))*Rf_rnorm(0,1);
//Draw Lambda//
if(scalemixture) draw_lambda_t(lam,gamma,alpha,B,phi);
//Store Draws//
intercept_mcmc[t]=intercept;
std::copy(gamma.begin(),gamma.end(),(gamma_mcmc+p*t));
std::copy(prob.begin(),prob.end(),(prob_mcmc+p*t));
std::copy(B.begin(),B.end(),(B_mcmc+p*t));
std::copy(lam.begin(),lam.end(),(lam_mcmc+p*t));
//Rao Blackwell//
if(t>=burnin) rao_blackwell(B_rb,prob_rb,B,prob,burnin,niter);
//Has Gamma Changed?//
gamma_diff=gamma_change(gamma_mcmc,t,p);
}
PutRNGstate();
}
