void train(unsigned n_boosting_item, double sigma_noise = 0.1)
{
MatrixXd sample_weight(1, dataset.rows());
sample_weight = MatrixXd::Ones(1, dataset.rows());
sample_weight /= sample_weight.sum();
while (n_boosting_item --> 0)
{
WLMS* new_boosting_item = new WLMS(dataset, label, sample_weight, sigma_noise);
double error_rate = 1e-5;
for(auto i=0; i<dataset.rows(); ++i)
{
if (label(i) * new_boosting_item->infer(dataset.row(i)) < 0)
{
error_rate += sample_weight(i);
}
}
double update_rate = log((1-error_rate)/error_rate);
boosting_item.push_back(make_pair(new_boosting_item, update_rate));
for(auto i=0; i<dataset.rows(); ++i)
{
sample_weight(i) *= exp(-update_rate * label(i) * new_boosting_item->infer(dataset.row(i)));
}
sample_weight /= sample_weight.sum();
}
}
void ClassificationCriterion::init(Mat _y,
Mat _sample_weight,
double _weight_n_samples,
vector<int>& _samples,
int _start,
int _end)
{
y = _y;
sample_weight = _sample_weight;
weighted_n_samples = _weight_n_samples;
samples = _samples;
start = _start;
end = _end;
// Find how many classes in y
std::set<double> unique;
for (int i = 0; i < y.rows; i++)
{
if (unique.find(y.at<double>(i)) == unique.end())
unique.insert(y.at<double>(i));
}
n_classes = unique.size();
// Initialize
label_count_total.resize(n_classes);
label_count_left.resize(n_classes);
label_count_right.resize(n_classes);
double _weighted_n_node_samples = 0.0;
double w = 1.0;
int index;
for (int i = start; i < end; i++)
{
index = samples.at(i);
if (sample_weight.total() == 0)
{
w = sample_weight(index);
}
// Get count of every class
int c = (int)y.at<double>(index, 0);
label_count_total.at(c) += w;
_weighted_n_node_samples += w;
}
weighted_n_node_samples = _weighted_n_node_samples;
reset();
}
