real T_kirkpatrick(real x[], int n)
{
int n_neg = 0;
real sx_neg = 0;
for (int i=0; i<n; i++)
{
if (x[i] < 0)
{
sx_neg += -x[i];
n_neg++;
}
}
return - (sx_neg / n_neg) / logr(0.8);
}
real randn()
{
static bool prepped = false;
static real prepped_result = 0;
if (prepped)
{
prepped = false;
return prepped_result;
}
else
{
real u1 = randf(), u2 = randf();
real len = sqrtr(-2 * logr(u1));
prepped = true;
prepped_result = len * sinr(2*M_PI*u2);
return len * cosr(2*M_PI*u2);
}
}
MatrixXf EMclustering::expectation(MatrixXf x, gaussian_model model, double &llh)
{
//cerr<<"===="<<endl;
MatrixXf mu(model.mu.rows(),model.mu.cols());
mu = model.mu;
MatrixXf *sigma;
sigma = new MatrixXf[clusternum];
//for(int i=0;i<clusternum;i++)
// sigma[i] = model.sigma[i];
//cerr<<"1"<<endl;
sigma = model.sigma;
VectorXf w(model.weight.size());
w = model.weight;
//cerr<<mu<<endl;
//cerr<<w<<endl<<endl;
//for(int i=0;i<clusternum;i++)cerr<<sigma[i]<<endl;
//cerr<<endl;
int n = x.cols();
int k = mu.cols();
MatrixXf logrho(n,k);
logrho.setZero(n,k);
//cerr<<logrho<<endl;
//cerr<<logrho<<endl<<endl;
for(int i=0;i<k;i++)
{
//cerr<<i<<endl;
logrho.col(i) = loggausspdf(x,mu.col(i),sigma[i]);
//cerr<<mu.col(i)<<endl;
}
//cerr<<logrho<<endl<<endl;
w = w.array().log();//cerr<<w<<endl<<endl;
MatrixXf tmp1(logrho.rows(),logrho.cols());
tmp1 = logrho.rowwise() + w.adjoint();
logrho = tmp1;//cerr<<logrho<<endl<<endl;
VectorXf t(logrho.rows());
t = logsumexp(logrho,2);//cerr<<t<<endl<<endl;
llh = t.sum()/n;//cerr<<llh<<endl<<endl;
MatrixXf logr(logrho.rows(),logrho.cols());
logr = logrho.colwise() - t;//cerr<<logr<<endl<<endl;
MatrixXf r(logrho.rows(),logrho.cols());
r = logr.array().exp();//cerr<<r<<endl<<endl;
logrho.resize(0,0);
mu.resize(0,0);
w.resize(0);
//for(int i=0;i<clusternum;i++)//..................
// sigma[i].resize(0,0);
delete [] sigma;
tmp1.resize(0,0);
t.resize(0);
logr.resize(0,0);
//cerr<<r<<endl<<endl;
//cerr<<llh<<endl;
return r;
}
