// --------------------------------------------------
double
RngBoost::betaSample(double alpha, double beta) const
{
static boost::uniform_01<boost::mt19937> zeroone(m_rng);
boost::math::beta_distribution<double> beta_dist(alpha, beta);
return quantile(beta_dist, zeroone());
}
RealType skewed_beta<RealType>::operator()(real_type parent_val,
RNG &rng) const {
real_type beta = find_beta(parent_val);
real_type u = std::uniform_real_distribution<real_type>()(rng);
boost::math::beta_distribution<real_type> beta_dist(m_alpha, beta);
real_type x = boost::math::quantile(beta_dist, u);
return x;
}
int MHSampler::construct_alias_table() {
size_t vocab_size = _model->vocab_size();
_topic_indexes = std::vector<TopicIndex>(vocab_size);
_alias_tables = std::vector<VoseAlias>(vocab_size);
_prob_sum = std::vector<double>(vocab_size);
// 构建每个词的alias table (不包含先验部分)
std::vector<double> dist;
for (size_t i = 0; i < vocab_size; ++i) {
dist.clear();
double prob_sum = 0;
for (auto& iter : _model->word_topic(i)) {
int topic_id = iter.first; // topic index
int word_topic_count = iter.second; // topic count
size_t topic_sum = _model->topic_sum(topic_id); // topic sum
_topic_indexes[i].push_back(topic_id);
double q = word_topic_count / (topic_sum + _model->beta_sum());
dist.push_back(q);
prob_sum += q;
}
_prob_sum[i] = prob_sum;
if (dist.size() > 0) {
_alias_tables[i].initialize(dist);
}
}
// 构建先验参数beta的alias table
_beta_prior_sum = 0;
std::vector<double> beta_dist(_model->num_topics(), 0);
for (int i = 0; i < _model->num_topics(); ++i) {
beta_dist[i] = _model->beta() / (_model->topic_sum(i) + _model->beta_sum());
_beta_prior_sum += beta_dist[i];
}
_beta_alias.initialize(beta_dist);
return 0;
}
// --------------------------------------------------
double
BasicPdfsBoost::betaPdfActualValue(double x, double alpha, double beta) const
{
boost::math::beta_distribution<> beta_dist(alpha, beta);
return boost::math::pdf(beta_dist, x);
}
