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

Testing hist_util (#5251)

Browse Source 
* Rank tests

* Remove categorical split specialisation

* Extend tests to multiple features, switch to WQSketch

* Add tests for SparseCuts

* Add external memory quantile tests, fix some existing tests
This commit is contained in:
Rory Mitchell
2020-02-14 14:36:43 +13:00
committed by GitHub
parent 911a902835
commit 24ad9dec0b
10 changed files with 354 additions and 93 deletions
Download Patch File Download Diff File Expand all files Collapse all files

47
tests/cpp/tree/test_gpu_hist.cu
View File

@@ -342,20 +342,17 @@ TEST(GpuHist, MinSplitLoss) {
delete dmat;
}
void UpdateTree(HostDeviceVector<GradientPair>* gpair,
DMatrix* dmat,
size_t gpu_page_size,
RegTree* tree,
HostDeviceVector<bst_float>* preds,
float subsample = 1.0f,
const std::string& sampling_method = "uniform") {
constexpr size_t kMaxBin = 2;
void UpdateTree(HostDeviceVector<GradientPair>* gpair, DMatrix* dmat,
size_t gpu_page_size, RegTree* tree,
HostDeviceVector<bst_float>* preds, float subsample = 1.0f,
const std::string& sampling_method = "uniform",
int max_bin = 2) {
if (gpu_page_size > 0) {
// Loop over the batches and count the records
int64_t batch_count = 0;
int64_t row_count = 0;
for (const auto& batch : dmat->GetBatches<EllpackPage>({0, kMaxBin, 0, gpu_page_size})) {
for (const auto& batch : dmat->GetBatches<EllpackPage>({0, max_bin, 0, gpu_page_size})) {
EXPECT_LT(batch.Size(), dmat->Info().num_row_);
batch_count++;
row_count += batch.Size();
@@ -366,7 +363,7 @@ void UpdateTree(HostDeviceVector<GradientPair>* gpair,
Args args{
{"max_depth", "2"},
{"max_bin", std::to_string(kMaxBin)},
{"max_bin", std::to_string(max_bin)},
{"min_child_weight", "0.0"},
{"reg_alpha", "0"},
{"reg_lambda", "0"},
@@ -386,7 +383,7 @@ void UpdateTree(HostDeviceVector<GradientPair>* gpair,
TEST(GpuHist, UniformSampling) {
constexpr size_t kRows = 4096;
constexpr size_t kCols = 2;
constexpr float kSubsample = 0.99;
constexpr float kSubsample = 0.9999;
common::GlobalRandom().seed(1994);
// Create an in-memory DMatrix.
@@ -397,25 +394,25 @@ TEST(GpuHist, UniformSampling) {
// Build a tree using the in-memory DMatrix.
RegTree tree;
HostDeviceVector<bst_float> preds(kRows, 0.0, 0);
UpdateTree(&gpair, dmat.get(), 0, &tree, &preds);
UpdateTree(&gpair, dmat.get(), 0, &tree, &preds, 1.0, "uniform", kRows);
// Build another tree using sampling.
RegTree tree_sampling;
HostDeviceVector<bst_float> preds_sampling(kRows, 0.0, 0);
UpdateTree(&gpair, dmat.get(), 0, &tree_sampling, &preds_sampling, kSubsample);
UpdateTree(&gpair, dmat.get(), 0, &tree_sampling, &preds_sampling, kSubsample,
"uniform", kRows);
// Make sure the predictions are the same.
auto preds_h = preds.ConstHostVector();
auto preds_sampling_h = preds_sampling.ConstHostVector();
for (int i = 0; i < kRows; i++) {
EXPECT_NEAR(preds_h[i], preds_sampling_h[i], 2e-3);
EXPECT_NEAR(preds_h[i], preds_sampling_h[i], 1e-8);
}
}
TEST(GpuHist, GradientBasedSampling) {
constexpr size_t kRows = 4096;
constexpr size_t kCols = 2;
constexpr float kSubsample = 0.99;
constexpr float kSubsample = 0.9999;
common::GlobalRandom().seed(1994);
// Create an in-memory DMatrix.
@@ -426,12 +423,13 @@ TEST(GpuHist, GradientBasedSampling) {
// Build a tree using the in-memory DMatrix.
RegTree tree;
HostDeviceVector<bst_float> preds(kRows, 0.0, 0);
UpdateTree(&gpair, dmat.get(), 0, &tree, &preds);
UpdateTree(&gpair, dmat.get(), 0, &tree, &preds, 1.0, "uniform", kRows);
// Build another tree using sampling.
RegTree tree_sampling;
HostDeviceVector<bst_float> preds_sampling(kRows, 0.0, 0);
UpdateTree(&gpair, dmat.get(), 0, &tree_sampling, &preds_sampling, kSubsample, "gradient_based");
UpdateTree(&gpair, dmat.get(), 0, &tree_sampling, &preds_sampling, kSubsample,
"gradient_based", kRows);
// Make sure the predictions are the same.
auto preds_h = preds.ConstHostVector();
@@ -459,18 +457,17 @@ TEST(GpuHist, ExternalMemory) {
// Build a tree using the in-memory DMatrix.
RegTree tree;
HostDeviceVector<bst_float> preds(kRows, 0.0, 0);
UpdateTree(&gpair, dmat.get(), 0, &tree, &preds);
UpdateTree(&gpair, dmat.get(), 0, &tree, &preds, 1.0, "uniform", kRows);
// Build another tree using multiple ELLPACK pages.
RegTree tree_ext;
HostDeviceVector<bst_float> preds_ext(kRows, 0.0, 0);
UpdateTree(&gpair, dmat_ext.get(), kPageSize, &tree_ext, &preds_ext);
UpdateTree(&gpair, dmat_ext.get(), kPageSize, &tree_ext, &preds_ext, 1.0, "uniform", kRows);
// Make sure the predictions are the same.
auto preds_h = preds.ConstHostVector();
auto preds_ext_h = preds_ext.ConstHostVector();
for (int i = 0; i < kRows; i++) {
EXPECT_NEAR(preds_h[i], preds_ext_h[i], 2e-6);
EXPECT_NEAR(preds_h[i], preds_ext_h[i], 1e-6);
}
}
@@ -495,12 +492,14 @@ TEST(GpuHist, ExternalMemoryWithSampling) {
// Build a tree using the in-memory DMatrix.
RegTree tree;
HostDeviceVector<bst_float> preds(kRows, 0.0, 0);
UpdateTree(&gpair, dmat.get(), 0, &tree, &preds, kSubsample, kSamplingMethod);
UpdateTree(&gpair, dmat.get(), 0, &tree, &preds, kSubsample, kSamplingMethod,
kRows);
// Build another tree using multiple ELLPACK pages.
RegTree tree_ext;
HostDeviceVector<bst_float> preds_ext(kRows, 0.0, 0);
UpdateTree(&gpair, dmat_ext.get(), kPageSize, &tree_ext, &preds_ext, kSubsample, kSamplingMethod);
UpdateTree(&gpair, dmat_ext.get(), kPageSize, &tree_ext, &preds_ext,
kSubsample, kSamplingMethod, kRows);
// Make sure the predictions are the same.
auto preds_h = preds.ConstHostVector();