Function RunComputeTest

catboost/cuda/methods/ut/test_multistat_histograms.cpp:473–532

    //
    void RunComputeTest(const TDocParallelDataSet& dataSet,
                        const ui32 numStats,
                        const ui32 maxLeaves,
                        double sampleRate,
                        const TComputeByBlocksConfig& byBlocksConfig,
                        const TBinarizedFeaturesManager& featuresManager) {
        TRandom rand(10);

        TComputeSplitPropertiesByBlocksHelper computeSplitPropertiesByBlocksHelper(dataSet,
                                                                                   byBlocksConfig);

        TSplitPropertiesHelper splitPropertiesHelper(dataSet,
                                                     featuresManager,
                                                     computeSplitPropertiesByBlocksHelper);

        TVector<float> featureWeights(featuresManager.GetFeatureCount(), 1.0f);
        auto subsets = splitPropertiesHelper.CreateInitialSubsets(
            CreateTestTarget(dataSet, numStats, sampleRate),
            maxLeaves,
            featureWeights);

        while (subsets.Leaves.size() < maxLeaves) {
            CATBOOST_DEBUG_LOG << "Leaves count #" << subsets.Leaves.size() << Endl;

            splitPropertiesHelper.BuildNecessaryHistograms(&subsets);
            CheckHistograms(dataSet, subsets);

            TVector<ui32> leavesToSplit;
            for (ui32 i = 0; i < subsets.Leaves.size(); ++i) {
                if (subsets.Leaves[i].Size) {
                    leavesToSplit.push_back(i);
                }
            }
            UNIT_ASSERT(leavesToSplit.size() > 0);

            Shuffle(leavesToSplit.begin(), leavesToSplit.end(), rand);
            leavesToSplit.resize(Max<ui32>(leavesToSplit.size() * 0.5, 1));
            ui32 maxLeavesToSplit = leavesToSplit.size() + subsets.Leaves.size() > maxLeaves ? maxLeaves - subsets.Leaves.size() : leavesToSplit.size();
            leavesToSplit.resize(maxLeavesToSplit);
            //            leavesToSplit.resize(1);

            for (ui32 leafId : leavesToSplit) {
                TBinarySplit bestSplit;
                auto featureIds = dataSet.GetFeatures().GetFeatures();
                auto localIdx = rand.NextUniformL() % featureIds.size();
                bestSplit.FeatureId = featureIds[localIdx];
                bestSplit.BinIdx = (featuresManager.GetBinCount(bestSplit.FeatureId) - 1) / 2;
                bestSplit.SplitType = featuresManager.IsCat(localIdx) ? EBinSplitType::TakeBin
                                                                      : EBinSplitType::TakeGreater;

                subsets.Leaves[leafId].BestSplit.Score = static_cast<float>(-rand.NextUniform());
                subsets.Leaves[leafId].BestSplit.FeatureId = bestSplit.FeatureId;
                subsets.Leaves[leafId].BestSplit.BinId = bestSplit.BinIdx;
            }

            CheckAndMakeSplit(leavesToSplit,
                              splitPropertiesHelper,
                              subsets);