/*
* calculates the pca for the static code features or dynamic setup features
*/
size_t PcaSeparateExt::calcSpecializedPca(double toBeCovered, bool dynamic) {
Array<double> in;
Array<int64> ids;
if(dynamic) {
if(dynamicQuery.size() == 0)
genDefaultDynamicQuery();
readDatabase(in, ids, dynamicFeatures.size(), dynamicQuery);
} else {
if(query.size() == 0)
genDefaultQuery();
readDatabase(in, ids, staticFeatures.size(), query);
}
AffineLinearMap model(in.cols(), in.cols());
Array<double> eigenvalues;
genPCAmodel(model, in, eigenvalues);
double sum = 0, partSum = 0;
for(size_t i = 0; i < eigenvalues.nelem(); ++i) {
sum += eigenvalues(i);
}
size_t nPCs = 0;
toBeCovered /= 100.0;
for(size_t i = 0; i < model.getOutputDimension(); ++i) {
partSum += eigenvalues(i);
if(partSum / sum > toBeCovered) {
nPCs = i+1;
break;
}
}
AffineLinearMap reductionModel(in.cols(), nPCs);
genPCAmodel(reductionModel, in);
// (reductionModel.getOutputDimension(), dynamic ? dynamicFeatures.size() : staticFeatures.size());
LOG(INFO) << reductionModel.getOutputDimension() << " PCs cover " << (partSum/sum)*100.0 << "% of the static feature's total variance\n";
Array<double> out = genPCs(reductionModel, in);
// std::cout << "REsult: " << trans << std::endl;
// std::cout << "AFTER " << eigenvalues << std::endl;
// std::cout << "modeld " << out << std::endl;
if(dynamic)
writeToSetup(out, ids);
else
writeToCode(out, ids);
return out.cols();
}
/*
* calculates the principal components of static features based on the given query and stores them in the database
*/
double PcaSeparateExt::calcSpecializedPca(size_t nInFeatures, size_t nOutFeatures, bool dynamic) {
Array<double> in;
Array<int64> ids;
if(dynamic) {
if(dynamicQuery.size() == 0)
genDefaultDynamicQuery();
readDatabase(in, ids, dynamicFeatures.size(), dynamicQuery);
} else {
if(query.size() == 0)
genDefaultQuery();
readDatabase(in, ids, staticFeatures.size(), query);
}
AffineLinearMap model(nInFeatures, nOutFeatures);
Array<double> eigenvalues;
genPCAmodel(model, in, eigenvalues);
// calculate the percentage of covered variance
double sum = 0, partSum = 0;
size_t i = 0;
for(; i < nOutFeatures; ++i)
partSum += eigenvalues(i);
sum = partSum;
for(; i < eigenvalues.nelem(); ++i)
sum += eigenvalues(i);
double covered = (partSum / sum) * 100.0;
LOG(INFO) << nOutFeatures << " PCs cover " << covered << "% of the " << (dynamic ? "dynamic" : "static") << " feature's total variance\n";
Array<double> out = genPCs(model, in);
// std::cout << "REsult: " << trans << std::endl;
// std::cout << "AFTER " << eigenvalues << std::endl;
// std::cout << "modeld " << out << std::endl;
if(dynamic)
writeToSetup(out, ids);
else
writeToCode(out, ids);
return covered;
}
/*
* Reads values form the database and stores the features in in, the targets (mapped according to the set policy) in targets as one-of-n coding
*/
size_t Trainer::readDatabase(Array<double>& in, Array<double>& target) throw(Kompex::SQLiteException) {
// if no query has been set, use default query
if(query.size() == 0)
genDefaultQuery();
// read the maximum of the column in measurement for which to train
double max = 0.0, min = 0.0;
if(genOut != GenNNoutput::ML_KEEP_INT && genOut != GenNNoutput::ML_FUZZY_VECTOR)
max = getMaximum(trainForName), min = getMinimum(trainForName);
Kompex::SQLiteStatement *localStmt = new Kompex::SQLiteStatement(pDatabase);
unsigned int nClasses = model.getOutputDimension();
localStmt->Sql(query);
size_t nRows = localStmt->GetNumberOfRows();
in = Array<double>(nRows, nFeatures());
LOG(INFO) << "Queried Rows: " << nRows << ", Number of features: " << staticFeatures.size() << " + " << dynamicFeatures.size() <<
" + " << pcaFeatures.size() << std::endl;
if(nRows == 0)
throw MachineLearningException("No dataset for the requested features could be found");
std::list<std::pair<double, size_t> > measurements;
Array<double> oneOfN(nClasses);
for(Array<double>::iterator I = oneOfN.begin(); I != oneOfN.end() && model.usesOneOfNCoding(); ++I) {
*I = NEG;
}
//Train machine
size_t i = 0;
// fetch all results
while(localStmt->FetchRow()){
// std::cout << "Result: " << localStmt->GetColumnName(2) << " " << localStmt->GetColumnName(3) << " " << localStmt->GetColumnName(4) << std::endl;
// std::cout << "Data: " << localStmt->GetColumnInt(2) << " " << localStmt->GetColumnInt(3) << " " << localStmt->GetColumnInt(4) << std::endl;
//std::cout << "[";
// construct training vectors
for(size_t j = 0; j < nFeatures(); ++j) {
in(i, j) = localStmt->GetColumnDouble(j);
//std::cout << in(i, j) << " ";
}
// translate index to one-of-n coding
if(genOut == ML_MAP_TO_N_CLASSES)
measurements.push_back(std::make_pair(localStmt->GetColumnDouble(nFeatures()), i));
else
appendToTrainArray(target, localStmt, nFeatures(), max, min, oneOfN);
//std::cout << target(i) << "]\n";
++i;
}
if(genOut == ML_MAP_TO_N_CLASSES)
mapToNClasses(measurements, model.getOutputDimension(), NEG, POS, target);
// reset the prepared statement
localStmt->Reset();
// do not forget to clean-up
localStmt->FreeQuery();
delete localStmt;
FeaturePreconditioner fp;
featureNormalization = fp.normalize(in, -1, 1);
return nRows;
}