Function EvaluateFeaturesImpl

catboost/libs/train_lib/eval_feature.cpp:833–1097

}

static void EvaluateFeaturesImpl(
    const NCatboostOptions::TCatBoostOptions& catBoostOptions,
    const NCatboostOptions::TOutputFilesOptions& outputFileOptions,
    const NCatboostOptions::TFeatureEvalOptions& featureEvalOptions,
    const TMaybe<TCustomObjectiveDescriptor>& objectiveDescriptor,
    const TMaybe<TCustomMetricDescriptor>& evalMetricDescriptor,
    ui32 foldRangeBegin,
    const TCvDataPartitionParams& cvParams,
    TDataProviderPtr data,
    ui32 processedFoldCount,
    TFeatureEvaluationCallbacks* callbacks,
    TFeatureEvaluationSummary* results
) {
    const ui32 foldCount = cvParams.Initialized() ? cvParams.FoldCount : featureEvalOptions.FoldCount.Get();
    CB_ENSURE(data->ObjectsData->GetObjectCount() > foldCount, "Pool is too small to be split into folds");
    CB_ENSURE(data->ObjectsData->GetObjectCount() > featureEvalOptions.FoldSize.Get(), "Pool is too small to be split into folds");

    const ui64 cpuUsedRamLimit
        = ParseMemorySizeDescription(catBoostOptions.SystemOptions->CpuUsedRamLimit.Get());

    TRestorableFastRng64 rand(catBoostOptions.RandomSeed);

    if (cvParams.Shuffle) {
        auto objectsGroupingSubset = NCB::Shuffle(data->ObjectsGrouping, 1, &rand);
        data = data->GetSubset(objectsGroupingSubset, cpuUsedRamLimit, &NPar::LocalExecutor());
    }

    TLabelConverter labelConverter;
    TMaybe<float> targetBorder = catBoostOptions.DataProcessingOptions->TargetBorder;
    NCatboostOptions::TCatBoostOptions dataSpecificOptions(catBoostOptions);

    TString tmpDir;
    if (outputFileOptions.AllowWriteFiles()) {
        NCB::NPrivate::CreateTrainDirWithTmpDirIfNotExist(outputFileOptions.GetTrainDir(), &tmpDir);
    }

    TTrainingDataProviderPtr trainingData = GetTrainingData(
        std::move(data),
        /*dataCanBeEmpty*/ false,
        /*isLearnData*/ true,
        TStringBuf(),
        Nothing(), // TODO(akhropov): allow loading borders and nanModes in CV?
        /*unloadCatFeaturePerfectHashFromRam*/ outputFileOptions.AllowWriteFiles(),
        /*ensureConsecutiveLearnFeaturesDataForCpu*/ false,
        tmpDir,
        /*quantizedFeaturesInfo*/ nullptr,
        &dataSpecificOptions,
        &labelConverter,
        &targetBorder,
        &NPar::LocalExecutor(),
        &rand);

    CB_ENSURE(
        dynamic_cast<TQuantizedObjectsDataProvider*>(trainingData->ObjectsData.Get()),
        "Unable to quantize dataset (probably because it contains categorical features)"
    );

    UpdateYetiRankEvalMetric(trainingData->MetaInfo.TargetStats, Nothing(), &dataSpecificOptions);