Hendrik/xgboost
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Always use partition based categorical splits. (#7857)

Browse Source
This commit is contained in:
Jiaming Yuan
2022-05-03 22:30:32 +08:00
committed by GitHub
parent 90cce38236
commit 317d7be6ee
13 changed files with 79 additions and 104 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
tests/cpp/tree/gpu_hist/test_evaluate_splits.cu
View File

@@ -57,8 +57,7 @@ void TestEvaluateSingleSplit(bool is_categorical) {
GPUHistEvaluator<GradientPair> evaluator{
tparam, static_cast<bst_feature_t>(feature_min_values.size()), 0};
dh::device_vector<common::CatBitField::value_type> out_cats;
DeviceSplitCandidate result =
evaluator.EvaluateSingleSplit(input, 0, ObjInfo{ObjInfo::kRegression}).split;
DeviceSplitCandidate result = evaluator.EvaluateSingleSplit(input, 0).split;
EXPECT_EQ(result.findex, 1);
EXPECT_EQ(result.fvalue, 11.0);
@@ -101,8 +100,7 @@ TEST(GpuHist, EvaluateSingleSplitMissing) {
dh::ToSpan(feature_histogram)};
GPUHistEvaluator<GradientPair> evaluator(tparam, feature_set.size(), 0);
DeviceSplitCandidate result =
evaluator.EvaluateSingleSplit(input, 0, ObjInfo{ObjInfo::kRegression}).split;
DeviceSplitCandidate result = evaluator.EvaluateSingleSplit(input, 0).split;
EXPECT_EQ(result.findex, 0);
EXPECT_EQ(result.fvalue, 1.0);
@@ -114,10 +112,8 @@ TEST(GpuHist, EvaluateSingleSplitMissing) {
TEST(GpuHist, EvaluateSingleSplitEmpty) {
TrainParam tparam = ZeroParam();
GPUHistEvaluator<GradientPair> evaluator(tparam, 1, 0);
DeviceSplitCandidate result = evaluator
.EvaluateSingleSplit(EvaluateSplitInputs<GradientPair>{}, 0,
ObjInfo{ObjInfo::kRegression})
.split;
DeviceSplitCandidate result =
evaluator.EvaluateSingleSplit(EvaluateSplitInputs<GradientPair>{}, 0).split;
EXPECT_EQ(result.findex, -1);
EXPECT_LT(result.loss_chg, 0.0f);
}
@@ -152,8 +148,7 @@ TEST(GpuHist, EvaluateSingleSplitFeatureSampling) {
dh::ToSpan(feature_histogram)};
GPUHistEvaluator<GradientPair> evaluator(tparam, feature_min_values.size(), 0);
DeviceSplitCandidate result =
evaluator.EvaluateSingleSplit(input, 0, ObjInfo{ObjInfo::kRegression}).split;
DeviceSplitCandidate result = evaluator.EvaluateSingleSplit(input, 0).split;
EXPECT_EQ(result.findex, 1);
EXPECT_EQ(result.fvalue, 11.0);
@@ -191,8 +186,7 @@ TEST(GpuHist, EvaluateSingleSplitBreakTies) {
dh::ToSpan(feature_histogram)};
GPUHistEvaluator<GradientPair> evaluator(tparam, feature_min_values.size(), 0);
DeviceSplitCandidate result =
evaluator.EvaluateSingleSplit(input, 0, ObjInfo{ObjInfo::kRegression}).split;
DeviceSplitCandidate result = evaluator.EvaluateSingleSplit(input, 0).split;
EXPECT_EQ(result.findex, 0);
EXPECT_EQ(result.fvalue, 1.0);
@@ -243,8 +237,8 @@ TEST(GpuHist, EvaluateSplits) {
GPUHistEvaluator<GradientPair> evaluator{
tparam, static_cast<bst_feature_t>(feature_min_values.size()), 0};
evaluator.EvaluateSplits(input_left, input_right, ObjInfo{ObjInfo::kRegression},
evaluator.GetEvaluator(), dh::ToSpan(out_splits));
evaluator.EvaluateSplits(input_left, input_right, evaluator.GetEvaluator(),
dh::ToSpan(out_splits));
DeviceSplitCandidate result_left = out_splits[0];
EXPECT_EQ(result_left.findex, 1);
@@ -264,8 +258,7 @@ TEST_F(TestPartitionBasedSplit, GpuHist) {
cuts_.cut_values_.SetDevice(0);
cuts_.min_vals_.SetDevice(0);
ObjInfo task{ObjInfo::kRegression};
evaluator.Reset(cuts_, dh::ToSpan(ft), task, info_.num_col_, param_, 0);
evaluator.Reset(cuts_, dh::ToSpan(ft), info_.num_col_, param_, 0);
dh::device_vector<GradientPairPrecise> d_hist(hist_[0].size());
auto node_hist = hist_[0];
@@ -282,7 +275,7 @@ TEST_F(TestPartitionBasedSplit, GpuHist) {
cuts_.cut_values_.ConstDeviceSpan(),
cuts_.min_vals_.ConstDeviceSpan(),
dh::ToSpan(d_hist)};
auto split = evaluator.EvaluateSingleSplit(input, 0, ObjInfo{ObjInfo::kRegression}).split;
auto split = evaluator.EvaluateSingleSplit(input, 0).split;
ASSERT_NEAR(split.loss_chg, best_score_, 1e-16);
}
} // namespace tree

13
tests/cpp/tree/hist/test_evaluate_splits.cc
View File

@@ -24,8 +24,8 @@ template <typename GradientSumT> void TestEvaluateSplits() {
auto dmat = RandomDataGenerator(kRows, kCols, 0).Seed(3).GenerateDMatrix();
auto evaluator = HistEvaluator<GradientSumT, CPUExpandEntry>{
param, dmat->Info(), n_threads, sampler, ObjInfo{ObjInfo::kRegression}};
auto evaluator =
HistEvaluator<GradientSumT, CPUExpandEntry>{param, dmat->Info(), n_threads, sampler};
common::HistCollection<GradientSumT> hist;
std::vector<GradientPair> row_gpairs = {
{1.23f, 0.24f}, {0.24f, 0.25f}, {0.26f, 0.27f}, {2.27f, 0.28f},
@@ -97,8 +97,7 @@ TEST(HistEvaluator, Apply) {
param.UpdateAllowUnknown(Args{{"min_child_weight", "0"}, {"reg_lambda", "0.0"}});
auto dmat = RandomDataGenerator(kNRows, kNCols, 0).Seed(3).GenerateDMatrix();
auto sampler = std::make_shared<common::ColumnSampler>();
auto evaluator_ = HistEvaluator<float, CPUExpandEntry>{param, dmat->Info(), 4, sampler,
ObjInfo{ObjInfo::kRegression}};
auto evaluator_ = HistEvaluator<float, CPUExpandEntry>{param, dmat->Info(), 4, sampler};
CPUExpandEntry entry{0, 0, 10.0f};
entry.split.left_sum = GradStats{0.4, 0.6f};
@@ -125,7 +124,7 @@ TEST_F(TestPartitionBasedSplit, CPUHist) {
std::vector<FeatureType> ft{FeatureType::kCategorical};
auto sampler = std::make_shared<common::ColumnSampler>();
HistEvaluator<double, CPUExpandEntry> evaluator{param_, info_, common::OmpGetNumThreads(0),
sampler, ObjInfo{ObjInfo::kRegression}};
sampler};
evaluator.InitRoot(GradStats{total_gpair_});
RegTree tree;
std::vector<CPUExpandEntry> entries(1);
@@ -156,8 +155,8 @@ auto CompareOneHotAndPartition(bool onehot) {
int32_t n_threads = 16;
auto sampler = std::make_shared<common::ColumnSampler>();
auto evaluator = HistEvaluator<GradientSumT, CPUExpandEntry>{
param, dmat->Info(), n_threads, sampler, ObjInfo{ObjInfo::kRegression}};
auto evaluator =
HistEvaluator<GradientSumT, CPUExpandEntry>{param, dmat->Info(), n_threads, sampler};
std::vector<CPUExpandEntry> entries(1);
for (auto const &gmat : dmat->GetBatches<GHistIndexMatrix>({32, param.sparse_threshold})) {

6
tests/cpp/tree/test_gpu_hist.cu
View File

@@ -264,7 +264,7 @@ TEST(GpuHist, EvaluateRootSplit) {
info.num_col_ = kNCols;
DeviceSplitCandidate res =
maker.EvaluateRootSplit({6.4f, 12.8f}, 0, ObjInfo{ObjInfo::kRegression}).split;
maker.EvaluateRootSplit({6.4f, 12.8f}, 0).split;
ASSERT_EQ(res.findex, 7);
ASSERT_NEAR(res.fvalue, 0.26, xgboost::kRtEps);
@@ -303,11 +303,11 @@ void TestHistogramIndexImpl() {
const auto &maker = hist_maker.maker;
auto grad = GenerateRandomGradients(kNRows);
grad.SetDevice(0);
maker->Reset(&grad, hist_maker_dmat.get(), kNCols, ObjInfo{ObjInfo::kRegression});
maker->Reset(&grad, hist_maker_dmat.get(), kNCols);
std::vector<common::CompressedByteT> h_gidx_buffer(maker->page->gidx_buffer.HostVector());
const auto &maker_ext = hist_maker_ext.maker;
maker_ext->Reset(&grad, hist_maker_ext_dmat.get(), kNCols, ObjInfo{ObjInfo::kRegression});
maker_ext->Reset(&grad, hist_maker_ext_dmat.get(), kNCols);
std::vector<common::CompressedByteT> h_gidx_buffer_ext(maker_ext->page->gidx_buffer.HostVector());
ASSERT_EQ(maker->page->Cuts().TotalBins(), maker_ext->page->Cuts().TotalBins());