double EMFit<InitialClusteringType, CovarianceConstraintPolicy>::LogLikelihood(
const arma::mat& observations,
const std::vector<arma::vec>& means,
const std::vector<arma::mat>& covariances,
const arma::vec& weights) const
{
double logLikelihood = 0;
arma::vec phis;
arma::mat likelihoods(means.size(), observations.n_cols);
for (size_t i = 0; i < means.size(); ++i)
{
phi(observations, means[i], covariances[i], phis);
likelihoods.row(i) = weights(i) * trans(phis);
}
// Now sum over every point.
for (size_t j = 0; j < observations.n_cols; ++j)
{
if (accu(likelihoods.col(j)) == 0)
Log::Info << "Likelihood of point " << j << " is 0! It is probably an "
<< "outlier." << std::endl;
logLikelihood += log(accu(likelihoods.col(j)));
}
return logLikelihood;
}
double VarMGCTM::Infer(CorpusC &test, MGCTMC &m) {
double sum = 0.0;
VReal likelihoods(test.Len());
#pragma omp parallel for
for (size_t d = 0; d < test.Len(); d++) {
MGVar var;
likelihoods[d] = Infer(test.docs[d], m, &var);
}
for (size_t d = 0; d < test.Len(); d++) {
sum += likelihoods[d];
}
return exp(- sum / test.TWordsNum());
}
void VarMGCTM::RunEM(CorpusC &test, MGCTM* m) {
MGSS ss;
ss.CorpusInit(cor_, *m);
MStep(ss, m);
LOG(INFO) << m->pi.transpose();
for (int i = 0; i < converged_.em_max_iter_; i++) {
std::vector<MGVar> vars(cor_.Len());
VReal likelihoods(cor_.Len());
#pragma omp parallel for
for (size_t d = 0; d < cor_.Len(); d++) {
likelihoods[d] = Infer(cor_.docs[d], *m, &vars[d]);
}
double likelihood = 0;
VStr etas(cor_.Len());
ss.SetZero(m->GTopicNum(), m->LTopicNum1(), m->LTopicNum2(), m->TermNum());
for (size_t d = 0; d < cor_.Len(); d++) {
DocC &doc = cor_.docs[d];
for (size_t n = 0; n < doc.ULen(); n++) {
for (int k = 0; k < m->GTopicNum(); k++) {
ss.g_topic(k, doc.Word(n)) += doc.Count(n)*vars[d].g_z(k, n)*
(1 - vars[d].delta[n]);
ss.g_topic_sum[k] += doc.Count(n)*vars[d].g_z(k, n)*(1 - vars[d].delta[n]);
}
}
for (int j = 0; j < m->LTopicNum1(); j++) {
for (size_t n = 0; n < doc.ULen(); n++) {
for (int k = 0; k < m->LTopicNum2(); k++) {
ss.l_topic[j](k, doc.Word(n)) += doc.Count(n)*vars[d].l_z[j](k, n)
*vars[d].delta[n]*vars[d].eta[j];
ss.l_topic_sum(k, j) += doc.Count(n)*vars[d].l_z[j](k, n) *
vars[d].delta[n] * vars[d].eta[j];
}
}
}
for (int j = 0; j < m->LTopicNum1(); j++) {
ss.pi[j] += vars[d].eta[j];
}
etas[d] = EVecToStr(vars[d].eta);
likelihood += likelihoods[d];
}
MStep(ss, m);
LOG(INFO) << m->pi.transpose();
OutputFile(*m, Join(etas,"\n"), i);
// LOG(INFO) <<"perplexity: " <<Infer(test,*m);
}
}
double GMM<FittingType>::LogLikelihood(
const arma::mat& data,
const std::vector<arma::vec>& meansL,
const std::vector<arma::mat>& covariancesL,
const arma::vec& weightsL) const
{
double loglikelihood = 0;
arma::vec phis;
arma::mat likelihoods(gaussians, data.n_cols);
for (size_t i = 0; i < gaussians; i++)
{
phi(data, meansL[i], covariancesL[i], phis);
likelihoods.row(i) = weightsL(i) * trans(phis);
}
// Now sum over every point.
for (size_t j = 0; j < data.n_cols; j++)
loglikelihood += log(accu(likelihoods.col(j)));
return loglikelihood;
}
void _Estepbfactor(vector<double> &expected, vector<double> &r1, vector<double> &ri,
const NumericMatrix &itemtrace, const vector<double> &prior, const vector<double> &Priorbetween,
const vector<int> &r, const int &ncores, const IntegerMatrix &data, const IntegerMatrix &sitems,
const vector<double> &Prior)
{
#ifdef SUPPORT_OPENMP
omp_set_num_threads(ncores);
#endif
const int sfact = sitems.ncol();
const int nitems = data.ncol();
const int npquad = prior.size();
const int nbquad = Priorbetween.size();
const int nquad = nbquad * npquad;
const int npat = r.size();
vector<double> r1vec(nquad*nitems*sfact, 0.0);
#pragma omp parallel for
for (int pat = 0; pat < npat; ++pat){
vector<double> L(nquad), Elk(nbquad*sfact), posterior(nquad*sfact);
vector<double> likelihoods(nquad*sfact, 1.0);
for (int fact = 0; fact < sfact; ++fact){
for (int item = 0; item < nitems; ++item){
if (data(pat,item) && sitems(item,fact))
for (int k = 0; k < nquad; ++k)
likelihoods[k + nquad*fact] = likelihoods[k + nquad*fact] * itemtrace(k,item);
}
}
vector<double> Plk(nbquad*sfact);
for (int fact = 0; fact < sfact; ++fact){
int k = 0;
for (int q = 0; q < npquad; ++q){
for (int i = 0; i < nbquad; ++i){
L[k] = likelihoods[k + nquad*fact] * prior[q];
++k;
}
}
vector<double> tempsum(nbquad, 0.0);
for (int i = 0; i < npquad; ++i)
for (int q = 0; q < nbquad; ++q)
tempsum[q] += L[q + i*nbquad];
for (int i = 0; i < nbquad; ++i)
Plk[i + fact*nbquad] = tempsum[i];
}
vector<double> Pls(nbquad, 1.0);
for (int i = 0; i < nbquad; ++i){
for(int fact = 0; fact < sfact; ++fact)
Pls[i] = Pls[i] * Plk[i + fact*nbquad];
expected[pat] += Pls[i] * Priorbetween[i];
}
for (int fact = 0; fact < sfact; ++fact)
for (int i = 0; i < nbquad; ++i)
Elk[i + fact*nbquad] = Pls[i] / Plk[i + fact*nbquad];
for (int fact = 0; fact < sfact; ++fact)
for (int i = 0; i < nquad; ++i)
posterior[i + nquad*fact] = likelihoods[i + nquad*fact] * r[pat] * Elk[i % nbquad + fact*nbquad] /
expected[pat];
#pragma omp critical
for (int i = 0; i < nbquad; ++i)
ri[i] += Pls[i] * r[pat] * Priorbetween[i] / expected[pat];
for (int item = 0; item < nitems; ++item)
if (data(pat,item))
for (int fact = 0; fact < sfact; ++fact)
for(int q = 0; q < nquad; ++q)
r1vec[q + fact*nquad*nitems + nquad*item] += posterior[q + fact*nquad];
} //end main
for (int item = 0; item < nitems; ++item)
for (int fact = 0; fact < sfact; ++fact)
if(sitems(item, fact))
for(int q = 0; q < nquad; ++q)
r1[q + nquad*item] = r1vec[q + nquad*item + nquad*nitems*fact] * Prior[q];
}
