/**
* Unpacks the gain matrix for a specific time t.
*
* This could be pre-computed if it becomes a performance bottleneck.
*
* @param t time along the trajectory
*
* @retval gain matrix at that time with dimension: state_dimension x u_dimension
*/
Eigen::MatrixXd Trajectory::GetGainMatrix(double t) {
int index = GetIndexFromTime(t);
Eigen::VectorXd k_row = kpoints_.row(index);
Eigen::MatrixXd k_mat(udimension_, dimension_);
for (int i = 0; i < udimension_; i++) {
// +1 because the column index is time
int start_pos = i * dimension_ + 1;
k_mat.row(i) = k_row.segment(start_pos, dimension_);
}
return k_mat;
}
void NuSVC::train_binary(const DataSet &dataset, int i, int j, SyncArray<float_type> &alpha, float_type &rho) {
DataSet::node2d ins = dataset.instances(i, j);//get instances of class i and j
int n_pos = dataset.count()[i];
int n_neg = dataset.count()[j];
SyncArray<int> y(ins.size());
alpha.resize(ins.size());
SyncArray<float_type> f_val(ins.size());
alpha.mem_set(0);
f_val.mem_set(0);
float_type sum_pos = param.nu * ins.size() / 2;
float_type sum_neg = sum_pos;
int *y_data = y.host_data();
float_type *alpha_data = alpha.host_data();
for (int l = 0; l < n_pos; ++l) {
y_data[l] = +1;
alpha_data[l] = min(1., sum_pos);
sum_pos -= alpha_data[l];
}
for (int l = 0; l < n_neg; ++l) {
y_data[n_pos + l] = -1;
alpha_data[n_pos + l] = min(1., sum_neg);
sum_neg -= alpha_data[n_pos + l];
}
vector<int> ori = dataset.original_index(i, j);
KernelMatrix k_mat(ins, param);
int ws_size = get_working_set_size(ins.size(), k_mat.n_features());
NuSMOSolver solver(false);
solver.solve(k_mat, y, alpha, rho, f_val, param.epsilon, 1, 1, ws_size, max_iter);
LOG(INFO)<<"rho = "<<rho;
int n_sv = 0;
alpha_data = alpha.host_data();
for (int l = 0; l < alpha.size(); ++l) {
alpha_data[l] *= y_data[l];
if (alpha_data[l] != 0) n_sv++;
}
LOG(INFO)<<"#sv = "<<n_sv;
}
