Function CompareOneHotAndPartition

tests/cpp/tree/hist/test_evaluate_splits.cc:279–327

namespace {
auto CompareOneHotAndPartition(bool onehot) {
  Context ctx;
  static constexpr bst_idx_t kRows = 128, kCols = 1;
  std::vector<FeatureType> ft(kCols, FeatureType::kCategorical);

  TrainParam param;
  if (onehot) {
    // force use one-hot
    param.UpdateAllowUnknown(
        Args{{"min_child_weight", "0"}, {"reg_lambda", "0"}, {"max_cat_to_onehot", "100"}});
  } else {
    param.UpdateAllowUnknown(
        Args{{"min_child_weight", "0"}, {"reg_lambda", "0"}, {"max_cat_to_onehot", "1"}});
  }

  size_t n_cats{2};

  auto dmat =
      RandomDataGenerator(kRows, kCols, 0).Seed(3).Type(ft).MaxCategory(n_cats).GenerateDMatrix();

  auto sampler = std::make_shared<common::ColumnSampler>();
  auto evaluator = HistEvaluator{&ctx, &param, dmat->Info(), sampler};
  std::vector<CPUExpandEntry> entries(1);
  HistMakerTrainParam hist_param;

  for (auto const &gmat : dmat->GetBatches<GHistIndexMatrix>(&ctx, {32, param.sparse_threshold})) {
    BoundedHistCollection hist;

    entries.front().nid = 0;
    entries.front().depth = 0;

    hist.Reset(gmat.cut.TotalBins(), hist_param.MaxCachedHistNodes(ctx.Device()));
    hist.AllocateHistograms({0});
    auto node_hist = hist[0];

    CHECK_EQ(node_hist.size(), n_cats);
    CHECK_EQ(node_hist.size(), gmat.cut.Ptrs().back());

    GradientPairPrecise total_gpair;
    for (size_t i = 0; i < node_hist.size(); ++i) {
      node_hist[i] = {static_cast<double>(node_hist.size() - i), 1.0};
      total_gpair += node_hist[i];
    }
    RegTree tree;
    evaluator.InitRoot(GradStats{total_gpair});
    evaluator.EvaluateSplits(hist, gmat.cut, ft, tree, &entries);
  }
  return entries.front();
}
}  // anonymous namespace

TEST(HistEvaluator, Categorical) {