Function ModelBasedEval

catboost/libs/train_lib/train_model.cpp:1421–1531

}

static void ModelBasedEval(
    const NJson::TJsonValue& trainOptionsJson,
    const NCatboostOptions::TOutputFilesOptions& outputOptions,
    TQuantizedFeaturesInfoPtr quantizedFeaturesInfo,
    TDataProviders pools,
    NPar::ILocalExecutor* const executor)
{
    CB_ENSURE(pools.Learn != nullptr, "Train data must be provided");

    const ETaskType taskType = NCatboostOptions::GetTaskType(trainOptionsJson);

    CB_ENSURE(taskType == ETaskType::GPU, "Model based eval is not implemented for CPU");

    CB_ENSURE(pools.Test.size() <= 1, "Multiple eval sets not supported for GPU");

    NJson::TJsonValue updatedTrainOptionsJson = trainOptionsJson;

    CB_ENSURE(TTrainerFactory::Has(ETaskType::GPU),
        "Can't load GPU learning library. Module was not compiled or driver is incompatible with package. Please install latest NVDIA driver and check again.");

    THolder<IModelTrainer> modelTrainerHolder(TTrainerFactory::Construct(ETaskType::GPU));
    if (outputOptions.SaveSnapshot()) {
        UpdateUndefinedRandomSeed(ETaskType::GPU, outputOptions, &updatedTrainOptionsJson, [&](IInputStream* in, TString& params) {
            ::Load(in, params);
        });
    }

    const auto learnFeaturesLayout = pools.Learn->MetaInfo.FeaturesLayout;
    NCatboostOptions::TCatBoostOptions catBoostOptions(taskType);
    catBoostOptions.Load(updatedTrainOptionsJson);

    ValidateFeaturesToEvaluate(trainOptionsJson, pools.Learn->MetaInfo.GetFeatureCount());

    if (!quantizedFeaturesInfo) {
        quantizedFeaturesInfo = MakeIntrusive<TQuantizedFeaturesInfo>(
            *learnFeaturesLayout,
            catBoostOptions.DataProcessingOptions.Get().IgnoredFeatures.Get(),
            catBoostOptions.DataProcessingOptions->FloatFeaturesBinarization.Get(),
            catBoostOptions.DataProcessingOptions->PerFloatFeatureQuantization.Get(),
            catBoostOptions.DataProcessingOptions->TextProcessingOptions.Get(),
            catBoostOptions.DataProcessingOptions->EmbeddingProcessingOptions.Get(),
            /*allowNansInTestOnly*/true
        );
    }

    for (auto testPoolIdx : xrange(pools.Test.size())) {
        const auto& testPool = *pools.Test[testPoolIdx];
        if (testPool.GetObjectCount() == 0) {
            continue;
        }
        CheckCompatibleForApply(
            *learnFeaturesLayout,
            *testPool.MetaInfo.FeaturesLayout,
            TStringBuilder() << "test dataset #" << testPoolIdx);
    }

    TSetLogging inThisScope(catBoostOptions.LoggingLevel);