double fMat44::operator () (int i, int j) const
{
#ifndef NDEBUG
if(i < 0 || i >= 4 || j < 0 || j >= 4)
{
cerr << "matrix size error at operator ()" << endl;
return temp;
}
#endif
if(i<3 && j<3) return m_mat(i,j);
else if(i<3 && j==3) return m_vec(i);
else if(i==3 && j==3) return m_scalar;
else return temp;
}
void image::resize(int32_t width, int32_t height, bool cover_crop /* = true */)
{
if (m_mat.data == nullptr)
return;
cv::Size size(width, height);
if (cover_crop == true)
{
cv::Size mat_size(m_mat.cols, m_mat.rows);
// This code assumes the cover looks like the following
// -------------------------
// | front | | back |
// | | | |
// | | | |
// | | | |
// | | | |
// | | | |
// | | | |
// -------------------------
// We want the front
if (m_mat.cols > m_mat.rows)
{
// Using cv::Rect on the stack always results in bogus values, for whatever reason :/
std::unique_ptr<cv::Rect> region_front(new cv::Rect(0, 0, get_width() / 2, get_height()));
cv::Mat cropped_mat;
m_mat(*region_front).copyTo(cropped_mat);
cv::Size new_size(region_front->width, region_front->height);
m_mat = cropped_mat;
resize(width, height, false);
/*cv::resize(cropped_mat, m_mat, new_size, 0.0, 0.0, cv::INTER_AREA);*/
return;
}
}
if (m_mat.data != nullptr)
cv::resize(m_mat, m_mat, size, 0.0, 0.0, cv::INTER_AREA);
}
int main(int argc, char **argv){
Predictor my_predictor;
std::vector<signed char> categs;
cv::Mat features;
assert(argc == 3 || argc == 4 );
/*training*/
if(argv[1][0] == 't'){
std::cout<<"\n~~~~~~~~~~~~~~TRAINING PREDICTOR~~~~~~~~~~~~~\n "<<std::endl;
DataMaker dm(TRAINING_SET_IMG,TRAINING_SET_IMG_PS);
if(argc == 3)
dm.makeData(features,categs);
else
dm.makeDataPS(features,categs);
my_predictor.train(features,categs);
my_predictor.save(argv[2]);
std::cout<<"\n~~~~~~~~~~~~~~PREDICTOR TRAINED~~~~~~~~~~~~~\n "<<std::endl;
}
else{
DataMaker dm(TRAINING_SET_IMG,TRAINING_SET_IMG_PS);
if(argc == 3)
dm.makeData(features,categs);
else
dm.makeDataPS(features,categs);
my_predictor.loadTrainData(argv[2]);
std::vector<signed char> pred;
my_predictor.predict(features,pred);
//'
//'
int n_false(0);
cv::Mat m_mat(3,3,CV_32F,cv::Scalar(0));
std::map<char,int> lut;
lut['N'] = 0;
lut['S'] = 1;
lut['M'] = 2;
char revLut[3] = {'N','S','M'};
std::vector<int> counts(3,0);
for(unsigned int i = 0; i<pred.size(); i++){
char p =pred[i];
char r =categs[i];
int real = lut[r];
int pred = lut[p];
counts[real] +=1;
if(real != pred){
n_false++;
}
m_mat.at<float>(pred,real) = m_mat.at<float>(pred,real) + 1;
}
cv::Mat sum_cols;
cv::reduce(m_mat, sum_cols, 0, CV_REDUCE_SUM);
cv::repeat(sum_cols, 3, 1, sum_cols);
m_mat = 100*m_mat/sum_cols;
float accur = 1 - ((float) n_false/(float) pred.size());
std::cout<<"\n~~~~~~~~~~~~~~ASSESSING ACCURACY OF THE PREDICTOR~~~~~~~~~~~~~\n "<<std::endl;
std::cout<<"ACCURACY = "<<accur<<std::endl;
std::cout<<"\nTRAINNING CLASS BALANCE:"<<std::endl;
std::cout.setf(std::ios::fixed);
// std::cout.precision(1);
for(unsigned int i = 0; i < counts.size();i++)
std::cout<<"N(class =="<<revLut[i]<<") "<<counts[i]<<std::endl;
std::cout<<"\nConfusion Matrix = \n REAL(top) -> PRED(left) \n"<<cv::format(m_mat,"CSV")<<"\n"<<std::endl;
// std::cout<<cv::format(features,"CSV")<<std::endl;
}
return 0;
}
void operator()(const Uint node_idx)
{
const Uint nb_moments = 2*m_nb_phases;
m_rhs.setZero();
for(Uint alpha = 0; alpha != m_nb_phases; ++alpha) // For all phases
{
const Real n_alpha = (*m_concentration_fields[alpha])[node_idx][0];
const Real vol_alpha = (*m_weighted_volume_fields[alpha])[node_idx][0] / n_alpha;
for(int ki = 0; ki != nb_moments; ++ki) // For all moments
{
const Real k = static_cast<Real>(ki);
m_mat(ki, alpha) = (1.-k) * pow_int(vol_alpha, ki);
m_mat(ki, alpha+m_nb_phases) = k * pow_int(vol_alpha, ki-1);
}
for(Uint gamma = 0; gamma != m_nb_phases; ++gamma)
{
const Real n_gamma = (*m_concentration_fields[gamma])[node_idx][0];
const Real vol_gamma = (*m_weighted_volume_fields[gamma])[node_idx][0] / n_gamma;
const Real beta = m_beta->apply(node_idx, alpha, gamma);
// std::cout << "computing for vol_alpha " << vol_alpha << ", vol_gamma: " << vol_gamma << ", n_alpha " << n_alpha << ", n_gamma " << n_gamma << ", beta " << beta << std::endl;
// std::cout << "collision value: " << (beta*n_alpha*n_gamma) << std::endl;
for(int k = 0; k != nb_moments; ++k ) // For all moments
{
m_rhs[k] += (0.5*pow_int(vol_alpha + vol_gamma, k ) - pow_int(vol_alpha, k )) * (beta*n_alpha*n_gamma);
}
if(is_not_null(m_collision_rate_field))
{
const Uint var_idx = alpha*m_nb_phases + gamma;
cf3_assert(var_idx < m_collision_rate_field->row_size());
(*m_collision_rate_field)[node_idx][var_idx] = n_alpha*n_gamma*beta;
}
}
}
Eigen::Map<RealVector> x(&(*m_source_field)[node_idx][0], nb_moments);
// Check for NaN
for(Uint k = 0; k != nb_moments; ++k)
{
if(!std::isfinite(m_rhs[k]))
{
x.setZero();
return;
}
}
Eigen::FullPivLU<RealMatrix> lu(m_mat);
if(!lu.isInvertible())
{
Eigen::JacobiSVD<RealMatrix> svd(m_mat, Eigen::ComputeThinU | Eigen::ComputeThinV);
x = svd.solve(m_rhs);
}
else
{
x = lu.solve(m_rhs);
}
// std::cout << "matrix:\n" << m_mat << std::endl;
// std::cout << "rhs: " << m_rhs.transpose() << std::endl;
// std::cout << "x: " << x.transpose() << std::endl;
}