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

Support categorical data in GPU sketching. (#6137)

Browse Source
This commit is contained in:
Jiaming Yuan
2020-09-21 13:53:06 +08:00
committed by GitHub
parent c932fb50a1
commit 210c131ce7
6 changed files with 196 additions and 62 deletions
Download Patch File Download Diff File Expand all files Collapse all files

51
tests/cpp/common/test_quantile.cu
View File

@@ -8,7 +8,8 @@ namespace xgboost {
namespace common {
TEST(GPUQuantile, Basic) {
constexpr size_t kRows = 1000, kCols = 100, kBins = 256;
SketchContainer sketch(kBins, kCols, kRows, 0);
HostDeviceVector<FeatureType> ft;
SketchContainer sketch(ft, kBins, kCols, kRows, 0);
dh::caching_device_vector<Entry> entries;
dh::device_vector<bst_row_t> cuts_ptr(kCols+1);
thrust::fill(cuts_ptr.begin(), cuts_ptr.end(), 0);
@@ -20,7 +21,8 @@ TEST(GPUQuantile, Basic) {
void TestSketchUnique(float sparsity) {
constexpr size_t kRows = 1000, kCols = 100;
RunWithSeedsAndBins(kRows, [kRows, kCols, sparsity](int32_t seed, size_t n_bins, MetaInfo const& info) {
SketchContainer sketch(n_bins, kCols, kRows, 0);
HostDeviceVector<FeatureType> ft;
SketchContainer sketch(ft, n_bins, kCols, kRows, 0);
HostDeviceVector<float> storage;
std::string interface_str = RandomDataGenerator{kRows, kCols, sparsity}
@@ -93,8 +95,10 @@ void TestQuantileElemRank(int32_t device, Span<SketchEntry const> in,
TEST(GPUQuantile, Prune) {
constexpr size_t kRows = 1000, kCols = 100;
RunWithSeedsAndBins(kRows, [=](int32_t seed, size_t n_bins, MetaInfo const& info) {
SketchContainer sketch(n_bins, kCols, kRows, 0);
RunWithSeedsAndBins(kRows, [=](int32_t seed, size_t n_bins,
MetaInfo const &info) {
HostDeviceVector<FeatureType> ft;
SketchContainer sketch(ft, n_bins, kCols, kRows, 0);
HostDeviceVector<float> storage;
std::string interface_str = RandomDataGenerator{kRows, kCols, 0}
@@ -111,8 +115,8 @@ TEST(GPUQuantile, Prune) {
if (n_bins <= kRows) {
ASSERT_EQ(sketch.Data().size(), n_bins * kCols);
} else {
// LE because kRows * kCols is pushed into sketch, after removing duplicated entries
// we might not have that much inputs for prune.
// LE because kRows * kCols is pushed into sketch, after removing
// duplicated entries we might not have that much inputs for prune.
ASSERT_LE(sketch.Data().size(), kRows * kCols);
}
// This is not necessarily true for all inputs without calling unique after
@@ -127,7 +131,8 @@ TEST(GPUQuantile, Prune) {
TEST(GPUQuantile, MergeEmpty) {
constexpr size_t kRows = 1000, kCols = 100;
size_t n_bins = 10;
SketchContainer sketch_0(n_bins, kCols, kRows, 0);
HostDeviceVector<FeatureType> ft;
SketchContainer sketch_0(ft, n_bins, kCols, kRows, 0);
HostDeviceVector<float> storage_0;
std::string interface_str_0 =
RandomDataGenerator{kRows, kCols, 0}.Device(0).GenerateArrayInterface(
@@ -166,7 +171,8 @@ TEST(GPUQuantile, MergeEmpty) {
TEST(GPUQuantile, MergeBasic) {
constexpr size_t kRows = 1000, kCols = 100;
RunWithSeedsAndBins(kRows, [=](int32_t seed, size_t n_bins, MetaInfo const& info) {
SketchContainer sketch_0(n_bins, kCols, kRows, 0);
HostDeviceVector<FeatureType> ft;
SketchContainer sketch_0(ft, n_bins, kCols, kRows, 0);
HostDeviceVector<float> storage_0;
std::string interface_str_0 = RandomDataGenerator{kRows, kCols, 0}
.Device(0)
@@ -176,7 +182,7 @@ TEST(GPUQuantile, MergeBasic) {
AdapterDeviceSketch(adapter_0.Value(), n_bins, info,
std::numeric_limits<float>::quiet_NaN(), &sketch_0);
SketchContainer sketch_1(n_bins, kCols, kRows * kRows, 0);
SketchContainer sketch_1(ft, n_bins, kCols, kRows * kRows, 0);
HostDeviceVector<float> storage_1;
std::string interface_str_1 = RandomDataGenerator{kRows, kCols, 0}
.Device(0)
@@ -212,7 +218,8 @@ TEST(GPUQuantile, MergeBasic) {
void TestMergeDuplicated(int32_t n_bins, size_t cols, size_t rows, float frac) {
MetaInfo info;
int32_t seed = 0;
SketchContainer sketch_0(n_bins, cols, rows, 0);
HostDeviceVector<FeatureType> ft;
SketchContainer sketch_0(ft, n_bins, cols, rows, 0);
HostDeviceVector<float> storage_0;
std::string interface_str_0 = RandomDataGenerator{rows, cols, 0}
.Device(0)
@@ -224,7 +231,7 @@ void TestMergeDuplicated(int32_t n_bins, size_t cols, size_t rows, float frac) {
&sketch_0);
size_t f_rows = rows * frac;
SketchContainer sketch_1(n_bins, cols, f_rows, 0);
SketchContainer sketch_1(ft, n_bins, cols, f_rows, 0);
HostDeviceVector<float> storage_1;
std::string interface_str_1 = RandomDataGenerator{f_rows, cols, 0}
.Device(0)
@@ -286,12 +293,14 @@ TEST(GPUQuantile, AllReduceBasic) {
}
constexpr size_t kRows = 1000, kCols = 100;
RunWithSeedsAndBins(kRows, [=](int32_t seed, size_t n_bins, MetaInfo const& info) {
// Set up single node version;
SketchContainer sketch_on_single_node(n_bins, kCols, kRows, 0);
RunWithSeedsAndBins(kRows, [=](int32_t seed, size_t n_bins,
MetaInfo const &info) {
// Set up single node version
HostDeviceVector<FeatureType> ft;
SketchContainer sketch_on_single_node(ft, n_bins, kCols, kRows, 0);
size_t intermediate_num_cuts =
std::min(kRows * world, static_cast<size_t>(n_bins * WQSketch::kFactor));
size_t intermediate_num_cuts = std::min(
kRows * world, static_cast<size_t>(n_bins * WQSketch::kFactor));
std::vector<SketchContainer> containers;
for (auto rank = 0; rank < world; ++rank) {
HostDeviceVector<float> storage;
@@ -300,12 +309,13 @@ TEST(GPUQuantile, AllReduceBasic) {
.Seed(rank + seed)
.GenerateArrayInterface(&storage);
data::CupyAdapter adapter(interface_str);
containers.emplace_back(n_bins, kCols, kRows, 0);
HostDeviceVector<FeatureType> ft;
containers.emplace_back(ft, n_bins, kCols, kRows, 0);
AdapterDeviceSketch(adapter.Value(), n_bins, info,
std::numeric_limits<float>::quiet_NaN(),
&containers.back());
}
for (auto& sketch : containers) {
for (auto &sketch : containers) {
sketch.Prune(intermediate_num_cuts);
sketch_on_single_node.Merge(sketch.ColumnsPtr(), sketch.Data());
sketch_on_single_node.FixError();
@@ -317,7 +327,7 @@ TEST(GPUQuantile, AllReduceBasic) {
// Set up distributed version. We rely on using rank as seed to generate
// the exact same copy of data.
auto rank = rabit::GetRank();
SketchContainer sketch_distributed(n_bins, kCols, kRows, 0);
SketchContainer sketch_distributed(ft, n_bins, kCols, kRows, 0);
HostDeviceVector<float> storage;
std::string interface_str = RandomDataGenerator{kRows, kCols, 0}
.Device(0)
@@ -376,7 +386,8 @@ TEST(GPUQuantile, SameOnAllWorkers) {
RunWithSeedsAndBins(kRows, [=](int32_t seed, size_t n_bins,
MetaInfo const &info) {
auto rank = rabit::GetRank();
SketchContainer sketch_distributed(n_bins, kCols, kRows, 0);
HostDeviceVector<FeatureType> ft;
SketchContainer sketch_distributed(ft, n_bins, kCols, kRows, 0);
HostDeviceVector<float> storage;
std::string interface_str = RandomDataGenerator{kRows, kCols, 0}
.Device(0)