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

Fix CPU hist init for sparse dataset. (#4625)

Browse Source 
* Fix CPU hist init for sparse dataset.

* Implement sparse histogram cut.
* Allow empty features.

* Fix windows build, don't use sparse in distributed environment.

* Comments.

* Smaller threshold.

* Fix windows omp.

* Fix msvc lambda capture.

* Fix MSVC macro.

* Fix MSVC initialization list.

* Fix MSVC initialization list x2.

* Preserve categorical feature behavior.

* Rename matrix to sparse cuts.
* Reuse UseGroup.
* Check for categorical data when adding cut.

Co-Authored-By: Philip Hyunsu Cho <chohyu01@cs.washington.edu>

* Sanity check.

* Fix comments.

* Fix comment.
This commit is contained in:
Jiaming Yuan
2019-07-04 19:27:03 -04:00
committed by Philip Hyunsu Cho
parent b7a1f22d24
commit d9a47794a5
33 changed files with 681 additions and 299 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
tests/cpp/common/test_gpu_hist_util.cu
View File

@@ -48,11 +48,11 @@ void TestDeviceSketch(const GPUSet& devices, bool use_external_memory) {
int gpu_batch_nrows = 0;
// find quantiles on the CPU
HistCutMatrix hmat_cpu;
hmat_cpu.Init((*dmat).get(), p.max_bin);
HistogramCuts hmat_cpu;
hmat_cpu.Build((*dmat).get(), p.max_bin);
// find the cuts on the GPU
HistCutMatrix hmat_gpu;
HistogramCuts hmat_gpu;
size_t row_stride = DeviceSketch(p, CreateEmptyGenericParam(0, devices.Size()), gpu_batch_nrows,
dmat->get(), &hmat_gpu);
@@ -69,12 +69,12 @@ void TestDeviceSketch(const GPUSet& devices, bool use_external_memory) {
// compare the cuts
double eps = 1e-2;
ASSERT_EQ(hmat_gpu.min_val.size(), num_cols);
ASSERT_EQ(hmat_gpu.row_ptr.size(), num_cols + 1);
ASSERT_EQ(hmat_gpu.cut.size(), hmat_cpu.cut.size());
ASSERT_LT(fabs(hmat_cpu.min_val[0] - hmat_gpu.min_val[0]), eps * nrows);
for (int i = 0; i < hmat_gpu.cut.size(); ++i) {
ASSERT_LT(fabs(hmat_cpu.cut[i] - hmat_gpu.cut[i]), eps * nrows);
ASSERT_EQ(hmat_gpu.MinValues().size(), num_cols);
ASSERT_EQ(hmat_gpu.Ptrs().size(), num_cols + 1);
ASSERT_EQ(hmat_gpu.Values().size(), hmat_cpu.Values().size());
ASSERT_LT(fabs(hmat_cpu.MinValues()[0] - hmat_gpu.MinValues()[0]), eps * nrows);
for (int i = 0; i < hmat_gpu.Values().size(); ++i) {
ASSERT_LT(fabs(hmat_cpu.Values()[i] - hmat_gpu.Values()[i]), eps * nrows);
}
delete dmat;