static void AddTreeCtrs(const TTrainData& data,
const TSplitTree& currentTree,
TFold* fold,
TLearnContext* ctx,
TStatsFromPrevTree* statsFromPrevTree,
TCandidateList* candList) {
using TSeenProjHash = THashSet<TProjection>;
TSeenProjHash seenProj;
// greedy construction
TProjection binAndOneHotFeaturesTree;
binAndOneHotFeaturesTree.BinFeatures = currentTree.GetBinFeatures();
binAndOneHotFeaturesTree.OneHotFeatures = currentTree.GetOneHotFeatures();
seenProj.insert(binAndOneHotFeaturesTree);
for (const auto& ctrSplit : currentTree.GetCtrSplits()) {
seenProj.insert(ctrSplit.Projection);
}
TSeenProjHash addedProjHash;
for (const auto& baseProj : seenProj) {
if (baseProj.IsEmpty()) {
continue;
}
for (int cf = 0; cf < data.AllFeatures.CatFeatures.ysize(); ++cf) {
if (data.AllFeatures.CatFeatures[cf].empty() ||
data.AllFeatures.IsOneHot[cf] ||
ctx->Rand.GenRandReal1() > ctx->Params.ObliviousTreeOptions->Rsm) {
continue;
}
TProjection proj = baseProj;
proj.AddCatFeature(cf);
if (proj.IsRedundant() || proj.GetFullProjectionLength() > ctx->Params.CatFeatureParams->MaxTensorComplexity) {
continue;
}
if (addedProjHash.has(proj)) {
continue;
}
addedProjHash.insert(proj);
AddCtrsToCandList(*fold, *ctx, proj, candList);
fold->GetCtrRef(proj);
}
}
THashSet<TSplitCandidate> candidatesToErase;
for (auto& splitCandidate : statsFromPrevTree->Stats) {
if (splitCandidate.first.Type == ESplitType::OnlineCtr) {
if (!addedProjHash.has(splitCandidate.first.Ctr.Projection)) {
candidatesToErase.insert(splitCandidate.first);
}
}
}
for (const auto& splitCandidate : candidatesToErase) {
statsFromPrevTree->Stats.erase(splitCandidate);
}
}
