/**
* Generates an artificial topographyGrid of size numRows x numCols if no
* topographic data is available. Results are dumped into topographyGrid.
* @param topographyGrid A pointer to a zero-initialized Grid of size
* numRows x numCols.
* @param numRows The desired number of non-border rows in the resulting matrix.
* @param numCols The desired number of non-border cols in the resulting matrix.
*/
void simulatetopographyGrid(Grid* topographyGrid, int numRows, int numCols) {
Eigen::VectorXd refx = refx.LinSpaced(numCols, -2*M_PI, 2*M_PI);
Eigen::VectorXd refy = refx.LinSpaced(numRows, -2*M_PI, 2*M_PI);
Eigen::MatrixXd X = refx.replicate(1, numRows);
X.transposeInPlace();
Eigen::MatrixXd Y = refy.replicate(1, numCols);
// Eigen can only deal with two matrices at a time,
// so split the computation:
// topographyGrid = sin(X) * sin(Y) * abs(X) * abs(Y) -pi
Eigen::MatrixXd absXY = X.cwiseAbs().cwiseProduct(Y.cwiseAbs());
Eigen::MatrixXd sins = X.array().sin().cwiseProduct(Y.array().sin());
Eigen::MatrixXd temp;
temp.resize(numRows, numCols);
temp = absXY.cwiseProduct(sins);
// All this work to create a matrix of pi...
Eigen::MatrixXd pi;
pi.resize(numRows, numCols);
pi.setConstant(M_PI);
temp = temp - pi;
// And the final result.
topographyGrid->data.block(border, border, numRows, numCols) =
temp.block(0, 0, numRows, numCols);
// Ignore positive values.
topographyGrid->data =
topographyGrid->data.unaryExpr(std::ptr_fun(validateDepth));
topographyGrid->clearNA();
}
Eigen::MatrixXd localWeighting( const Eigen::MatrixXd &W, bool isFull, bool isSymmetric)
{
int n = W.rows();
double Ls = (W.count()-n)/2; //count number of edges of the graph
Eigen::MatrixXd C = Eigen::MatrixXd::Zero(n,n);
double Ws = 0.5*W.sum();
Eigen::VectorXd D = W.colwise().sum();
if (isFull)
{
if (isSymmetric)
{
computeC_symmetric_full(const_cast<Eigen::MatrixXd &>(W),C,D,n);
}
else
{
// this is a trick to ensure vectorizatoin and no cache misses! some tranpositions have to be made though
const_cast<Eigen::MatrixXd &>(W).transposeInPlace();
computeC_symmetric_full(const_cast<Eigen::MatrixXd &>(W),C,D,n);
const_cast<Eigen::MatrixXd &>(W).transposeInPlace();
C.transposeInPlace();
// the original code use vertical access to rows, but is slower
//compute_C_asymmetric_full(const_cast<Eigen::MatrixXd &>(W),C,D,n);
}
}
else
{
if (isSymmetric)
{
computeC_symmetric_sparse(const_cast<Eigen::MatrixXd &>(W),C,D,n);
}
else
{
compute_C_asymmetric_sparse(const_cast<Eigen::MatrixXd &>(W),C,D,n);
}
}
Eigen::MatrixXd G = ((Ls/Ws)*W).cwiseProduct(C);
Eigen::VectorXd DG = G.colwise().sum();
for (int i=0; i<n; i++)
{
G.row(i)/=DG(i);
}
return G;
}
void Plsa::plsa_Train(Eigen::MatrixXd& data,int topics,int Max_Iterations, double tolerance){
//Variables
data.transposeInPlace();
srand((unsigned int) time(0));
M=data.rows(); // number of visual words
Ni=data.cols(); // number of images
K=topics; // number of Topics
map<int,double> table_likehood; // likelihood Values
double offset=1e-7; // small offset to avoid numerical problems in Pw_z
//incializacion//
P_z=Eigen::VectorXd::Ones(K)/K; // P(z);
Pd_z=Eigen::MatrixXd::Random(Ni,K); // P(d|z)
Pw_z=Eigen::MatrixXd::Random(M,K); //P(w|z)
Pd_z=Pd_z.cwiseAbs();
Pw_z=Pw_z.cwiseAbs();
// normalize
Normalize(Pd_z);
Normalize(Pw_z);
cout <<"P(d|z) "<< Pd_z.colwise().sum()<<endl;
cout <<"P(w|z) " <<Pw_z.colwise().sum()<<endl;
//EM algorithm
for (int i=0;i<Max_Iterations;i++)
{
double dli=0;
E_step();
M_step(data);
table_likehood[i]=loglikehood(data);
Pw_z=(Pw_z.array()+offset).matrix();
if (i>0)
{
dli=table_likehood[i]-table_likehood[i-1];
if (dli< tolerance)
{
cout<<" finalizo en: "<<i<<endl;
break;
}
}
}
}
/** solves A * X' = B' for X in place */
void chol_solve_t(Eigen::MatrixXd & A, Eigen::MatrixXd & B) {
if (A.rows() != B.cols()) {throw std::runtime_error("A.rows() must equal B.cols()");}
B.transposeInPlace();
chol_solve(A, B);
B.transposeInPlace();
}
