predictor_type operator()(const DataIterator& begin, const DataIterator& end,
const InputOf& input_of, const OutputOf& label_of) const {
// Let us apply a variable selection process.
GenericLearner generic_learner;
// We need a real risk estimator for evaluating the algorithms when variables are selected.
auto real_risk_estimator = gaml::risk::cross_validation(gaml::loss::Quadratic<double>(),
gaml::partition::kfold(10), false);
// Wrapping evaluator
auto evaluator = gaml::varsel::make_wrapper_evaluator(generic_learner, real_risk_estimator,
begin, end, input_of, label_of);
// Make the evaluator verbose.
if(verbosity) evaluator.verbose();
// The result of a search is a subset of variables with its risk.
std::vector<int> variable_subset;
double best_risk = gaml::varsel::SFFS(evaluator, variable_subset, verbosity);
if(verbosity) {
std::cout << std::endl
<< "Risk is : " << best_risk << ", with the following variable selection:" << std::endl
<< " {";
for(auto attr : variable_subset) std::cout << ' ' << toString(attr);
std::cout << " }" << std::endl << std::endl;
}
// Let us set up the predictor from the projection of the data.
auto projection = gaml::project(begin, end,
variable_subset.begin(), variable_subset.end(),
input_of, label_of);
return projection.teach(generic_learner); // The predictor is computed here.
}
decltype(auto) async_make_next_batch() {
return std::async(std::launch::async, [this](){
cpu_vector cpu_train_data;
cpu_vector cpu_teach_data;
for(auto i = 0; i < label_num_; ++i) {
std::shuffle(data_[i].begin(), data_[i].end(), g_);
for(auto j = 0; j < data_num_per_label_; ++j) {
cpu_train_data.insert(cpu_train_data.end(),
data_[i][j].begin(), data_[i][j].end());
}
cpu_vector teach(label_num_, 0.f); teach[i] = 1.f;
for(auto j = 0; j < data_num_per_label_; ++j) {
cpu_teach_data.insert(cpu_teach_data.end(),
teach.begin(), teach.end());
}
}
boost::compute::copy(
cpu_train_data.begin(), cpu_train_data.end(),
batch_.train_data.begin());
boost::compute::copy(
cpu_teach_data.begin(), cpu_teach_data.end(),
batch_.teach_data.begin());
});
}
