void do_batch(size_t num) {
eigen_wsvec_list_t vs;
for (size_t i = 0; i < num; ++i) {
eigen_wsvec_t wsvec_a, wsvec_b;
wsvec_a.data = eigen_svec_t(2);
wsvec_b.data = eigen_svec_t(2);
wsvec_a.data.coeffRef(0) = r_.next_gaussian() + 2.0;
wsvec_a.data.coeffRef(1) = r_.next_gaussian() + 2.0;
wsvec_b.data.coeffRef(0) = r_.next_gaussian() - 2.0;
wsvec_b.data.coeffRef(1) = r_.next_gaussian() - 2.0;
wsvec_a.weight = 1.0;
wsvec_b.weight = 1.0;
vs.push_back(wsvec_a);
vs.push_back(wsvec_b);
}
gmm_->batch(vs, k_, d_);
}
void gmm::batch(const eigen_wsvec_list_t& data, int d, int k) {
if (data.empty()) {
*this = gmm();
return;
}
typedef eigen_wsvec_list_t::const_iterator data_iter;
initialize(data, d, k);
eigen_svec_list_t old_means;
eigen_smat_list_t old_covs;
eigen_solver_list_t old_solvers;
double old_obj = 0, obj = 0;
vector<double> weights(k);
bool converged = false;
int64_t niter = 1;
while (!converged) {
old_covs = covs_;
old_means = means_;
old_solvers = cov_solvers_;
old_obj = obj;
obj = 0;
fill(weights.begin(), weights.end(), 0);
fill(means_.begin(), means_.end(), eigen_svec_t(d));
fill(covs_.begin(), covs_.end(), eigen_smat_t(d, d));
for (data_iter i = data.begin(); i != data.end(); ++i) {
eigen_svec_t cps =
cluster_probs(i->data, old_means, old_covs, old_solvers);
for (int c = 0; c < k; ++c) {
double cp = i->weight * cps.coeff(c);
means_[c] += cp * i->data;
covs_[c] += i->data * (i->data.transpose()) * cp;
weights[c] += cp;
obj -= std::log(std::max(cp, std::numeric_limits<double>::min()));
}
}
for (int c = 0; c < k; ++c) {
means_[c] /= weights[c];
covs_[c] /= weights[c];
double eps = 0.1;
covs_[c] -= means_[c] * means_[c].transpose();
covs_[c] += eps * eye_;
cov_solvers_[c] =
shared_ptr<eigen_solver_t>(new eigen_solver_t(covs_[c]));
}
converged = is_converged(niter++, means_, old_means, obj, old_obj);
}
}