vector<double> sampleAlpha( std::mt19937& rng )
{
vector<double> res(3);
res[0] = (double)rng()/rng.max();
res[1] = (double)rng()/rng.max();
std::sort(res.begin(), res.begin() + 2);
res[2] = 1.0 - res[1];
res[1] = res[1] - res[0];
res[0] = res[0] - 0.0;
return res;
}
void SeedFeatureFactory::train( const std::vector< std::shared_ptr<ImageOverSegmentation> > &ios, const std::vector<VectorXs> & lbl ) {
printf(" * Training SeedFeature\n");
static std::mt19937 rand;
const int N_SAMPLES = 5000;
int n_pos=0, n_neg=0;
for( VectorXs l: lbl ) {
n_pos += (l.array()>=0).cast<int>().sum();
n_neg += (l.array()==-1).cast<int>().sum();
}
// Collect training examples
float sampling_freq[] = {0.5f*N_SAMPLES / n_neg, 0.5f*N_SAMPLES / n_pos};
std::vector<RowVectorXf> f;
std::vector<float> l;
#pragma omp parallel for
for( int i=0; i<ios.size(); i++ ) {
RMatrixXf ftr = SeedFeature::computeObjFeatures( *ios[i] );
for( int j=0; j<ios[i]->Ns(); j++ )
if( lbl[i][j] >= -1 && rand() < rand.max()*sampling_freq[ lbl[i][j]>=0 ] ) {
#pragma omp critical
{
l.push_back( lbl[i][j]>=0 );
f.push_back( ftr.row(j) );
}
}
}
printf(" - Computing parameters\n");
// Fit the ranking functions
RMatrixXf A( f.size(), f[0].size() );
VectorXf b( l.size() );
for( int i=0; i<f.size(); i++ ) {
A.row(i) = f[i];
b[i] = l[i];
}
// Solve A*x = b
param_ = A.colPivHouseholderQr().solve(b);
printf(" - done %f\n",(A*param_-b).array().abs().mean());
}
inline double rndf()
{
return fabs(rng())/rng.max();
}
