VectorXd CurvatureError::operator()(const VectorXd& a) const {
DblVec curvatures;
DblVec Deltas;
curvatures = toDblVec(a.head(pi->curvature_vars.size()));
Deltas = DblVec(pi->curvature_vars.size(), a(a.size() - 1));
VectorXd ret(1);
ret(0) = -this->total_curvature_limit;
for (int i = 0; i < curvatures.size(); ++i) {
ret(0) += fabs(curvatures[i]) * Deltas[i];
}
return ret;
}
double CurvatureCost::operator()(const VectorXd& a) const {
DblVec curvatures;
DblVec Deltas;
curvatures = toDblVec(a.head(pi->curvature_vars.size()));
Deltas = DblVec(pi->curvature_vars.size(), a(a.size() - 1));
double ret = 0;
for (int i = 0; i < curvatures.size(); ++i)
{
double tmp = curvatures[i] * Deltas[i];
ret += tmp * tmp;
}
return ret * coeff;
}
void populateSimpleDataFrame(DataFrame& df)
{
// create a very simple data frame were the first factor is perfect and the others are
// garbage
df.addDataVector("a", DblVec(0, 1, 3));
df.addDataVector("a", DblVec(0, 2, 2));
df.addDataVector("a", DblVec(0, 3, 1));
df.addDataVector("b", DblVec(1, 1, 3));
df.addDataVector("b", DblVec(1, 2, 2));
df.addDataVector("b", DblVec(1, 3, 1));
std::vector<std::string> factorLabels;
factorLabels.push_back("f1");
factorLabels.push_back("f2");
factorLabels.push_back("f3");
df.setFactorLabels(factorLabels);
}
double LogisticRegressionObjective::EvalLocalMultiThread(const DblVec& input, DblVec& gradient){
/*
create 24 thread;
each thread calculate a loss and gradient
*/
int threadNum = 24;
double lossList[threadNum];
DblVec *gradList = new DblVec[threadNum];
for(int i = 0; i < threadNum; i++){
gradList[i] = DblVec(gradient.size());
}
threadList = new pthread_t[threadNum];
for(int i = 0; i < threadNum; i++){
Parameter*p = new Parameter(*this, input, gradList[i], lossList[i], i, threadNum);
pthread_create(&threadList[i], NULL, ThreadEvalLocal, p);
}
for(int i = 0; i < threadNum; i++){
pthread_join(threadList[i], NULL);
}
double loss = 0.0;
for(int j = 0; j < gradient.size(); j++){
gradient[j] = 0.0;
}
for(int i = 0; i < threadNum; i++){
loss += lossList[i];
for(int j = 0; j < gradient.size(); j++){
gradient[j] += gradList[i][j];
}
}
delete []gradList;
delete []threadList;
return loss;
}
/*!
* A const function that calculates the matrix exponential and returns a copy.
* The exponential is calculated by finding the eigenvalues and eigenvectors
* of the matrix, exponentiating the eigenvalues. The eigenvalues is stored in
* a matrix V, eigenvectors is stored in a matrix A, inv(A) is calculated.
* The product A*V*inv(A) is returned.
* @return A copy of the exponentiated matrix
*/
Matrix Matrix::expm() const{
return expm(DblVec(1,1))[0];
}
