Function CalcErrorsDistributed

catboost/private/libs/distributed/master.cpp:455–568

}

void CalcErrorsDistributed(
    const NCB::TTrainingDataProviders& trainData,
    const TVector<THolder<IMetric>>& metrics,
    bool calcAllMetrics,
    bool calcErrorTrackerMetric,
    TLearnContext* ctx) {

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

    // Calc additive stats in distributed manner and calc non-additive stats locally

    bool haveNonAdditiveTrainMetrics = false;
    bool haveNonAdditiveTestMetrics = false; // TODO(akhropov): do we need per test dataset flags?

    IterateOverMetrics(
        trainData,
        metrics,
        calcAllMetrics,
        calcErrorTrackerMetric,
        /*calcAdditiveMetrics*/ false,
        /*calcNonAdditiveMetrics*/ true,
        /*onLearn*/ [&] (TConstArrayRef<const IMetric*> trainMetrics) {
            haveNonAdditiveTrainMetrics = !trainMetrics.empty();
        },
        /*onTest*/ [&] (
            size_t /*testIdx*/,
            TConstArrayRef<const IMetric*> testMetrics,
            TMaybe<int> /*filteredTrackerIdx*/
        ) {
            haveNonAdditiveTestMetrics |= (!testMetrics.empty());
        }
    );


    // Compute non-additive metrics locally and distributed additive stats in parallel

    TVector<std::function<void()>> tasks;

    if (haveNonAdditiveTrainMetrics || haveNonAdditiveTestMetrics) {
        MapGetApprox(
            trainData,
            *ctx,
            /*useBestModel*/ false,
            haveNonAdditiveTrainMetrics ? &(ctx->LearnProgress->AvrgApprox) : nullptr,
            haveNonAdditiveTestMetrics ? &(ctx->LearnProgress->TestApprox) : nullptr
        );

        tasks.push_back(
            [&] () {
                CalcErrorsLocally(
                    trainData,
                    metrics,
                    calcAllMetrics,
                    calcErrorTrackerMetric,
                    /*calcNonAdditiveMetricsOnly*/true,
                    ctx
                );
            }