Function MapGenericCalcScore

catboost/private/libs/distributed/master.cpp:251–303

template <typename TScoreCalcMapper, typename TGetScore>
void MapGenericCalcScore(
    TGetScore getScore,
    double scoreStDev,
    TCandidatesContext* candidatesContext,
    TLearnContext* ctx) {

    Y_ASSERT(ctx->Params.SystemOptions->IsMaster());

    auto scoreDistribution = GetScoreDistribution(ctx->Params.ObliviousTreeOptions->RandomScoreType);

    auto& candidateList = candidatesContext->CandidateList;

    const int workerCount = TMasterEnvironment::GetRef().RootEnvironment->GetSlaveCount();
    auto allStatsFromAllWorkers = ApplyMapper<TScoreCalcMapper>(
        workerCount,
        TMasterEnvironment::GetRef().SharedTrainData,
        candidateList);

    // some workers may return empty result because they don't have any learn subset
    const TVector<size_t> validWorkers = GetNonEmptyElementsIndices(allStatsFromAllWorkers);
    auto validWorkersSize = validWorkers.size();
    CB_ENSURE_INTERNAL(validWorkersSize, "No workers returned score stats");

    const int candidateCount = candidateList.ysize();
    const ui64 randSeed = ctx->LearnProgress->Rand.GenRand();
    // set best split for each candidate
    NPar::ParallelFor(
        *ctx->LocalExecutor,
        0,
        candidateCount,
        [&] (int candidateIdx) {
            auto& subCandidates = candidateList[candidateIdx].Candidates;
            const int subcandidateCount = subCandidates.ysize();
            TVector<TVector<double>> allScores(subcandidateCount);
            for (int subcandidateIdx = 0; subcandidateIdx < subcandidateCount; ++subcandidateIdx) {
                // reduce across workers
                auto& reducedStats = allStatsFromAllWorkers[validWorkers[0]][candidateIdx][subcandidateIdx];
                for (size_t validWorkerIdx = 1; validWorkerIdx < validWorkersSize; ++validWorkerIdx) {
                    const auto& stats = allStatsFromAllWorkers[validWorkers[validWorkerIdx]][candidateIdx][subcandidateIdx];
                    reducedStats.Add(stats);
                }
                const auto& splitInfo = subCandidates[subcandidateIdx];
                allScores[subcandidateIdx] = getScore(reducedStats, splitInfo);
            }
            SetBestScore(
                randSeed + candidateIdx,
                allScores,
                scoreDistribution,
                scoreStDev,
                *candidatesContext,
                &subCandidates);
        });
}

// TODO(espetrov): Remove unused code.
void MapCalcScore(