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

Use Booster context in DMatrix. (#8896)

Browse Source 
- Pass context from booster to DMatrix.
- Use context instead of integer for `n_threads`.
- Check the consistency configuration for `max_bin`.
- Test for all combinations of initialization options.
This commit is contained in:
Jiaming Yuan
2023-04-28 21:47:14 +08:00
committed by GitHub
parent 1f9a57d17b
commit 08ce495b5d
67 changed files with 1283 additions and 935 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -1,12 +1,13 @@
/*!
* Copyright 2020-2021 by XGBoost Contributors
/**
* Copyright 2020-2023, XGBoost Contributors
*/
#include <gtest/gtest.h>
#include "../../../../src/data/ellpack_page.cuh"
#include "../../../../src/tree/gpu_hist/gradient_based_sampler.cuh"
#include "../../../../src/tree/param.h"
#include "../../filesystem.h" // dmlc::TemporaryDirectory
#include "../../../../src/tree/param.h" // TrainParam
#include "../../filesystem.h" // dmlc::TemporaryDirectory
#include "../../helpers.h"
namespace xgboost {
@@ -31,14 +32,15 @@ void VerifySampling(size_t page_size,
}
gpair.SetDevice(0);
BatchParam param{0, 256};
auto page = (*dmat->GetBatches<EllpackPage>(param).begin()).Impl();
Context ctx{MakeCUDACtx(0)};
auto param = BatchParam{256, tree::TrainParam::DftSparseThreshold()};
auto page = (*dmat->GetBatches<EllpackPage>(&ctx, param).begin()).Impl();
if (page_size != 0) {
EXPECT_NE(page->n_rows, kRows);
}
GradientBasedSampler sampler(page, kRows, param, subsample, sampling_method);
auto sample = sampler.Sample(gpair.DeviceSpan(), dmat.get());
GradientBasedSampler sampler(&ctx, page, kRows, param, subsample, sampling_method);
auto sample = sampler.Sample(&ctx, gpair.DeviceSpan(), dmat.get());
if (fixed_size_sampling) {
EXPECT_EQ(sample.sample_rows, kRows);
@@ -86,12 +88,13 @@ TEST(GradientBasedSampler, NoSamplingExternalMemory) {
auto gpair = GenerateRandomGradients(kRows);
gpair.SetDevice(0);
BatchParam param{0, 256};
auto page = (*dmat->GetBatches<EllpackPage>(param).begin()).Impl();
Context ctx{MakeCUDACtx(0)};
auto param = BatchParam{256, tree::TrainParam::DftSparseThreshold()};
auto page = (*dmat->GetBatches<EllpackPage>(&ctx, param).begin()).Impl();
EXPECT_NE(page->n_rows, kRows);
GradientBasedSampler sampler(page, kRows, param, kSubsample, TrainParam::kUniform);
auto sample = sampler.Sample(gpair.DeviceSpan(), dmat.get());
GradientBasedSampler sampler(&ctx, page, kRows, param, kSubsample, TrainParam::kUniform);
auto sample = sampler.Sample(&ctx, gpair.DeviceSpan(), dmat.get());
auto sampled_page = sample.page;
EXPECT_EQ(sample.sample_rows, kRows);
EXPECT_EQ(sample.gpair.size(), gpair.Size());
@@ -103,7 +106,7 @@ TEST(GradientBasedSampler, NoSamplingExternalMemory) {
ci(buffer.data(), sampled_page->NumSymbols());
size_t offset = 0;
for (auto& batch : dmat->GetBatches<EllpackPage>(param)) {
for (auto& batch : dmat->GetBatches<EllpackPage>(&ctx, param)) {
auto page = batch.Impl();
std::vector<common::CompressedByteT> page_buffer(page->gidx_buffer.HostVector());
common::CompressedIterator<common::CompressedByteT>

19
tests/cpp/tree/gpu_hist/test_histogram.cu
View File

@@ -1,9 +1,14 @@
/**
* Copyright 2020-2023, XGBoost Contributors
*/
#include <gtest/gtest.h>
#include <vector>
#include "../../../../src/common/categorical.h"
#include "../../../../src/tree/gpu_hist/histogram.cuh"
#include "../../../../src/tree/gpu_hist/row_partitioner.cuh"
#include "../../../../src/tree/param.h" // TrainParam
#include "../../categorical_helpers.h"
#include "../../helpers.h"
@@ -11,15 +16,15 @@ namespace xgboost {
namespace tree {
void TestDeterministicHistogram(bool is_dense, int shm_size) {
Context ctx = CreateEmptyGenericParam(0);
Context ctx = MakeCUDACtx(0);
size_t constexpr kBins = 256, kCols = 120, kRows = 16384, kRounds = 16;
float constexpr kLower = -1e-2, kUpper = 1e2;
float sparsity = is_dense ? 0.0f : 0.5f;
auto matrix = RandomDataGenerator(kRows, kCols, sparsity).GenerateDMatrix();
BatchParam batch_param{0, static_cast<int32_t>(kBins)};
auto batch_param = BatchParam{kBins, tree::TrainParam::DftSparseThreshold()};
for (auto const& batch : matrix->GetBatches<EllpackPage>(batch_param)) {
for (auto const& batch : matrix->GetBatches<EllpackPage>(&ctx, batch_param)) {
auto* page = batch.Impl();
tree::RowPartitioner row_partitioner(0, kRows);
@@ -114,13 +119,13 @@ void ValidateCategoricalHistogram(size_t n_categories, common::Span<GradientPair
// Test 1 vs rest categorical histogram is equivalent to one hot encoded data.
void TestGPUHistogramCategorical(size_t num_categories) {
auto ctx = CreateEmptyGenericParam(0);
auto ctx = MakeCUDACtx(0);
size_t constexpr kRows = 340;
size_t constexpr kBins = 256;
auto x = GenerateRandomCategoricalSingleColumn(kRows, num_categories);
auto cat_m = GetDMatrixFromData(x, kRows, 1);
cat_m->Info().feature_types.HostVector().push_back(FeatureType::kCategorical);
BatchParam batch_param{0, static_cast<int32_t>(kBins)};
auto batch_param = BatchParam{kBins, tree::TrainParam::DftSparseThreshold()};
tree::RowPartitioner row_partitioner(0, kRows);
auto ridx = row_partitioner.GetRows(0);
dh::device_vector<GradientPairInt64> cat_hist(num_categories);
@@ -130,7 +135,7 @@ void TestGPUHistogramCategorical(size_t num_categories) {
/**
* Generate hist with cat data.
*/
for (auto const &batch : cat_m->GetBatches<EllpackPage>(batch_param)) {
for (auto const &batch : cat_m->GetBatches<EllpackPage>(&ctx, batch_param)) {
auto* page = batch.Impl();
FeatureGroups single_group(page->Cuts());
BuildGradientHistogram(ctx.CUDACtx(), page->GetDeviceAccessor(0),
@@ -144,7 +149,7 @@ void TestGPUHistogramCategorical(size_t num_categories) {
auto x_encoded = OneHotEncodeFeature(x, num_categories);
auto encode_m = GetDMatrixFromData(x_encoded, kRows, num_categories);
dh::device_vector<GradientPairInt64> encode_hist(2 * num_categories);
for (auto const &batch : encode_m->GetBatches<EllpackPage>(batch_param)) {
for (auto const &batch : encode_m->GetBatches<EllpackPage>(&ctx, batch_param)) {
auto* page = batch.Impl();
FeatureGroups single_group(page->Cuts());
BuildGradientHistogram(ctx.CUDACtx(), page->GetDeviceAccessor(0),