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

Extract Sketch Entry from hist maker. (#7503)

Browse Source 
* Extract Sketch Entry from hist maker.

* Add a new sketch container for sorted inputs.
* Optimize bin search.
This commit is contained in:
Jiaming Yuan
2021-12-18 05:36:56 +08:00
committed by GitHub
parent b4a1236cfc
commit 9ab73f737e
15 changed files with 393 additions and 217 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
tests/cpp/common/test_column_matrix.cc
View File

@@ -14,7 +14,7 @@ TEST(DenseColumn, Test) {
static_cast<uint64_t>(std::numeric_limits<uint16_t>::max()) + 2};
for (size_t max_num_bin : max_num_bins) {
auto dmat = RandomDataGenerator(100, 10, 0.0).GenerateDMatrix();
GHistIndexMatrix gmat(dmat.get(), max_num_bin);
GHistIndexMatrix gmat(dmat.get(), max_num_bin, false);
ColumnMatrix column_matrix;
column_matrix.Init(gmat, 0.2);
@@ -61,7 +61,7 @@ TEST(SparseColumn, Test) {
static_cast<uint64_t>(std::numeric_limits<uint16_t>::max()) + 2};
for (size_t max_num_bin : max_num_bins) {
auto dmat = RandomDataGenerator(100, 1, 0.85).GenerateDMatrix();
GHistIndexMatrix gmat(dmat.get(), max_num_bin);
GHistIndexMatrix gmat(dmat.get(), max_num_bin, false);
ColumnMatrix column_matrix;
column_matrix.Init(gmat, 0.5);
switch (column_matrix.GetTypeSize()) {
@@ -101,7 +101,7 @@ TEST(DenseColumnWithMissing, Test) {
static_cast<uint64_t>(std::numeric_limits<uint16_t>::max()) + 2 };
for (size_t max_num_bin : max_num_bins) {
auto dmat = RandomDataGenerator(100, 1, 0.5).GenerateDMatrix();
GHistIndexMatrix gmat(dmat.get(), max_num_bin);
GHistIndexMatrix gmat(dmat.get(), max_num_bin, false);
ColumnMatrix column_matrix;
column_matrix.Init(gmat, 0.2);
switch (column_matrix.GetTypeSize()) {
@@ -130,7 +130,7 @@ void TestGHistIndexMatrixCreation(size_t nthreads) {
/* This should create multiple sparse pages */
std::unique_ptr<DMatrix> dmat{ CreateSparsePageDMatrix(kEntries) };
omp_set_num_threads(nthreads);
GHistIndexMatrix gmat(dmat.get(), 256);
GHistIndexMatrix gmat(dmat.get(), 256, false);
}
TEST(HistIndexCreationWithExternalMemory, Test) {

25
tests/cpp/common/test_hist_util.cc
View File

@@ -223,13 +223,19 @@ TEST(HistUtil, DenseCutsAccuracyTestWeights) {
auto w = GenerateRandomWeights(num_rows);
dmat->Info().weights_.HostVector() = w;
for (auto num_bins : bin_sizes) {
HistogramCuts cuts = SketchOnDMatrix(dmat.get(), num_bins);
ValidateCuts(cuts, dmat.get(), num_bins);
{
HistogramCuts cuts = SketchOnDMatrix(dmat.get(), num_bins, true);
ValidateCuts(cuts, dmat.get(), num_bins);
}
{
HistogramCuts cuts = SketchOnDMatrix(dmat.get(), num_bins, false);
ValidateCuts(cuts, dmat.get(), num_bins);
}
}
}
}
TEST(HistUtil, QuantileWithHessian) {
void TestQuantileWithHessian(bool use_sorted) {
int bin_sizes[] = {2, 16, 256, 512};
int sizes[] = {1000, 1500};
int num_columns = 5;
@@ -243,13 +249,13 @@ TEST(HistUtil, QuantileWithHessian) {
dmat->Info().weights_.HostVector() = w;
for (auto num_bins : bin_sizes) {
HistogramCuts cuts_hess = SketchOnDMatrix(dmat.get(), num_bins, hessian);
HistogramCuts cuts_hess = SketchOnDMatrix(dmat.get(), num_bins, use_sorted, hessian);
for (size_t i = 0; i < w.size(); ++i) {
dmat->Info().weights_.HostVector()[i] = w[i] * hessian[i];
}
ValidateCuts(cuts_hess, dmat.get(), num_bins);
HistogramCuts cuts_wh = SketchOnDMatrix(dmat.get(), num_bins);
HistogramCuts cuts_wh = SketchOnDMatrix(dmat.get(), num_bins, use_sorted);
ValidateCuts(cuts_wh, dmat.get(), num_bins);
ASSERT_EQ(cuts_hess.Values().size(), cuts_wh.Values().size());
@@ -262,6 +268,11 @@ TEST(HistUtil, QuantileWithHessian) {
}
}
TEST(HistUtil, QuantileWithHessian) {
TestQuantileWithHessian(true);
TestQuantileWithHessian(false);
}
TEST(HistUtil, DenseCutsExternalMemory) {
int bin_sizes[] = {2, 16, 256, 512};
int sizes[] = {100, 1000, 1500};
@@ -292,7 +303,7 @@ TEST(HistUtil, IndexBinBound) {
for (auto max_bin : bin_sizes) {
auto p_fmat = RandomDataGenerator(kRows, kCols, 0).GenerateDMatrix();
GHistIndexMatrix hmat(p_fmat.get(), max_bin);
GHistIndexMatrix hmat(p_fmat.get(), max_bin, false);
EXPECT_EQ(hmat.index.Size(), kRows*kCols);
EXPECT_EQ(expected_bin_type_sizes[bin_id++], hmat.index.GetBinTypeSize());
}
@@ -315,7 +326,7 @@ TEST(HistUtil, IndexBinData) {
for (auto max_bin : kBinSizes) {
auto p_fmat = RandomDataGenerator(kRows, kCols, 0).GenerateDMatrix();
GHistIndexMatrix hmat(p_fmat.get(), max_bin);
GHistIndexMatrix hmat(p_fmat.get(), max_bin, false);
uint32_t* offsets = hmat.index.Offset();
EXPECT_EQ(hmat.index.Size(), kRows*kCols);
switch (max_bin) {

76
tests/cpp/common/test_quantile.cc
View File

@@ -19,7 +19,22 @@ TEST(Quantile, LoadBalance) {
}
CHECK_EQ(n_cols, kCols);
}
namespace {
template <bool use_column>
using ContainerType = std::conditional_t<use_column, SortedSketchContainer, HostSketchContainer>;
// Dispatch for push page.
void PushPage(SortedSketchContainer* container, SparsePage const& page, MetaInfo const& info,
Span<float const> hessian) {
container->PushColPage(page, info, hessian);
}
void PushPage(HostSketchContainer* container, SparsePage const& page, MetaInfo const& info,
Span<float const> hessian) {
container->PushRowPage(page, info, hessian);
}
} // anonymous namespace
template <bool use_column>
void TestDistributedQuantile(size_t rows, size_t cols) {
std::string msg {"Skipping AllReduce test"};
int32_t constexpr kWorkers = 4;
@@ -48,12 +63,23 @@ void TestDistributedQuantile(size_t rows, size_t cols) {
.Lower(.0f)
.Upper(1.0f)
.GenerateDMatrix();
HostSketchContainer sketch_distributed(
column_size, n_bins, m->Info().feature_types.ConstHostSpan(), false,
OmpGetNumThreads(0));
for (auto const &page : m->GetBatches<SparsePage>()) {
sketch_distributed.PushRowPage(page, m->Info());
std::vector<float> hessian(rows, 1.0);
auto hess = Span<float const>{hessian};
ContainerType<use_column> sketch_distributed(n_bins, m->Info(), column_size, false, hess,
OmpGetNumThreads(0));
if (use_column) {
for (auto const& page : m->GetBatches<SortedCSCPage>()) {
PushPage(&sketch_distributed, page, m->Info(), hess);
}
} else {
for (auto const& page : m->GetBatches<SparsePage>()) {
PushPage(&sketch_distributed, page, m->Info(), hess);
}
}
HistogramCuts distributed_cuts;
sketch_distributed.MakeCuts(&distributed_cuts);
@@ -61,17 +87,25 @@ void TestDistributedQuantile(size_t rows, size_t cols) {
rabit::Finalize();
CHECK_EQ(rabit::GetWorldSize(), 1);
std::for_each(column_size.begin(), column_size.end(), [=](auto& size) { size *= world; });
HostSketchContainer sketch_on_single_node(
column_size, n_bins, m->Info().feature_types.ConstHostSpan(), false,
OmpGetNumThreads(0));
m->Info().num_row_ = world * rows;
ContainerType<use_column> sketch_on_single_node(n_bins, m->Info(), column_size, false, hess,
OmpGetNumThreads(0));
m->Info().num_row_ = rows;
for (auto rank = 0; rank < world; ++rank) {
auto m = RandomDataGenerator{rows, cols, sparsity}
.Seed(rank)
.Lower(.0f)
.Upper(1.0f)
.GenerateDMatrix();
for (auto const &page : m->GetBatches<SparsePage>()) {
sketch_on_single_node.PushRowPage(page, m->Info());
if (use_column) {
for (auto const& page : m->GetBatches<SortedCSCPage>()) {
PushPage(&sketch_on_single_node, page, m->Info(), hess);
}
} else {
for (auto const& page : m->GetBatches<SparsePage>()) {
PushPage(&sketch_on_single_node, page, m->Info(), hess);
}
}
}
@@ -87,7 +121,7 @@ void TestDistributedQuantile(size_t rows, size_t cols) {
ASSERT_EQ(sptrs.size(), dptrs.size());
for (size_t i = 0; i < sptrs.size(); ++i) {
ASSERT_EQ(sptrs[i], dptrs[i]);
ASSERT_EQ(sptrs[i], dptrs[i]) << i;
}
ASSERT_EQ(svals.size(), dvals.size());
@@ -104,14 +138,28 @@ void TestDistributedQuantile(size_t rows, size_t cols) {
TEST(Quantile, DistributedBasic) {
#if defined(__unix__)
constexpr size_t kRows = 10, kCols = 10;
TestDistributedQuantile(kRows, kCols);
TestDistributedQuantile<false>(kRows, kCols);
#endif
}
TEST(Quantile, Distributed) {
#if defined(__unix__)
constexpr size_t kRows = 1000, kCols = 200;
TestDistributedQuantile(kRows, kCols);
constexpr size_t kRows = 4000, kCols = 200;
TestDistributedQuantile<false>(kRows, kCols);
#endif
}
TEST(Quantile, SortedDistributedBasic) {
#if defined(__unix__)
constexpr size_t kRows = 10, kCols = 10;
TestDistributedQuantile<true>(kRows, kCols);
#endif
}
TEST(Quantile, SortedDistributed) {
#if defined(__unix__)
constexpr size_t kRows = 4000, kCols = 200;
TestDistributedQuantile<true>(kRows, kCols);
#endif
}