Function CollectLeavesStatistics

catboost/libs/fstr/util.cpp:19–87

TVector<double> CollectLeavesStatistics(
    const TDataProvider& dataset,
    const TFullModel& model,
    NPar::ILocalExecutor* localExecutor) {

    TConstArrayRef<float> weights;

    TTargetDataProviderPtr targetData; // needed to own weights data

    if (const auto* modelInfoParams = MapFindPtr(model.ModelInfo, "params")) {
        NJson::TJsonValue paramsJson = ReadTJsonValue(*modelInfoParams);
        if (paramsJson.Has("loss_function")) {
            TRestorableFastRng64 rand(0);

            targetData = CreateModelCompatibleProcessedDataProvider(
                dataset,
                {},
                model,
                GetMonopolisticFreeCpuRam(),
                &rand,
                localExecutor,
                /*metricsThatRequireTargetCanBeSkipped*/true,
                /*skipMinMaxPairsCheck*/true,
                /*skipTargetConsistencyCheck*/true
            ).TargetData;

            weights = GetWeights(*targetData);
        }
    }

    // If it is impossible to get properly adjusted weights use raw weights from RawTargetData
    if (weights.empty()) {
        const TWeights<float>& rawWeights = dataset.RawTargetData.GetWeights();
        if (!rawWeights.IsTrivial()) {
            weights = rawWeights.GetNonTrivialData();
        }
    }

    size_t treeCount = model.GetTreeCount();
    const int approxDimension = model.ModelTrees->GetDimensionsCount();
    TVector<double> leavesStatistics(
        model.ModelTrees->GetModelTreeData()->GetLeafValues().size() / approxDimension
    );

    auto binFeatures = MakeQuantizedFeaturesForEvaluator(model, *dataset.ObjectsData.Get());

    const auto documentsCount = dataset.GetObjectCount();
    auto applyData = model.ModelTrees->GetApplyData();
    for (size_t treeIdx = 0; treeIdx < treeCount; ++treeIdx) {
        TVector<TIndexType> indices = BuildIndicesForBinTree(model, binFeatures.Get(), treeIdx);
        const int offset = applyData->TreeFirstLeafOffsets[treeIdx] / approxDimension;
        if (indices.empty()) {
            continue;
        }

        if (weights.empty()) {
            for (size_t doc = 0; doc < documentsCount; ++doc) {
                const TIndexType valueIndex = indices[doc];