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

Prepare external memory support for hist. (#7638)

Browse Source 
This PR prepares the GHistIndexMatrix to host the column matrix which is used by the hist tree method by accepting sparse_threshold parameter.

Some cleanups are made to ensure the correct batch param is being passed into DMatrix along with some additional tests for correctness of SimpleDMatrix.
This commit is contained in:
Jiaming Yuan
2022-02-10 16:58:02 +08:00
committed by GitHub
parent 87c01f49d8
commit 2775c2a1ab
24 changed files with 368 additions and 201 deletions
Download Patch File Download Diff File Expand all files Collapse all files

50
src/data/gradient_index.h
View File

@@ -4,12 +4,14 @@
*/
#ifndef XGBOOST_DATA_GRADIENT_INDEX_H_
#define XGBOOST_DATA_GRADIENT_INDEX_H_
#include <memory>
#include <vector>
#include "xgboost/base.h"
#include "xgboost/data.h"
#include "../common/categorical.h"
#include "../common/hist_util.h"
#include "../common/threading_utils.h"
#include "xgboost/base.h"
#include "xgboost/data.h"
namespace xgboost {
/*!
@@ -32,20 +34,22 @@ class GHistIndexMatrix {
/*! \brief The corresponding cuts */
common::HistogramCuts cut;
DMatrix* p_fmat;
/*! \brief max_bin for each feature. */
size_t max_num_bins;
/*! \brief base row index for current page (used by external memory) */
size_t base_rowid{0};
GHistIndexMatrix() = default;
GHistIndexMatrix(DMatrix* x, int32_t max_bin, bool sorted_sketch, int32_t n_threads,
common::Span<float> hess = {}) {
this->Init(x, max_bin, sorted_sketch, n_threads, hess);
}
GHistIndexMatrix();
GHistIndexMatrix(DMatrix* x, int32_t max_bin, double sparse_thresh, bool sorted_sketch,
int32_t n_threads, common::Span<float> hess = {});
~GHistIndexMatrix();
// Create a global histogram matrix, given cut
void Init(DMatrix* p_fmat, int max_num_bins, bool sorted_sketch, int32_t n_threads,
common::Span<float> hess);
void Init(DMatrix* p_fmat, int max_bins, double sparse_thresh, bool sorted_sketch,
int32_t n_threads, common::Span<float> hess);
void Init(SparsePage const& page, common::Span<FeatureType const> ft,
common::HistogramCuts const& cuts, int32_t max_bins_per_feat, bool is_dense,
int32_t n_threads);
double sparse_thresh, int32_t n_threads);
// specific method for sparse data as no possibility to reduce allocated memory
template <typename BinIdxType, typename GetOffset>
@@ -74,7 +78,7 @@ class GHistIndexMatrix {
index_data[ibegin + j] = get_offset(bin_idx, j);
++hit_count_tloc_[tid * nbins + bin_idx];
} else {
uint32_t idx = cut.SearchBin(inst[j].fvalue, inst[j].index, ptrs, values);
uint32_t idx = cut.SearchBin(e.fvalue, e.index, ptrs, values);
index_data[ibegin + j] = get_offset(idx, j);
++hit_count_tloc_[tid * nbins + idx];
}
@@ -82,10 +86,9 @@ class GHistIndexMatrix {
});
}
void ResizeIndex(const size_t n_index,
const bool isDense);
void ResizeIndex(const size_t n_index, const bool isDense);
inline void GetFeatureCounts(size_t* counts) const {
void GetFeatureCounts(size_t* counts) const {
auto nfeature = cut.Ptrs().size() - 1;
for (unsigned fid = 0; fid < nfeature; ++fid) {
auto ibegin = cut.Ptrs()[fid];
@@ -95,7 +98,8 @@ class GHistIndexMatrix {
}
}
}
inline bool IsDense() const {
bool IsDense() const {
return isDense_;
}
void SetDense(bool is_dense) { isDense_ = is_dense; }
@@ -105,6 +109,8 @@ class GHistIndexMatrix {
}
private:
// unused at the moment: https://github.com/dmlc/xgboost/pull/7531
std::unique_ptr<common::ColumnMatrix> columns_;
std::vector<size_t> hit_count_tloc_;
bool isDense_;
};
@@ -117,7 +123,19 @@ class GHistIndexMatrix {
*/
inline bool RegenGHist(BatchParam old, BatchParam p) {
// parameter is renewed or caller requests a regen
return p.regen || (old.gpu_id != p.gpu_id || old.max_bin != p.max_bin);
if (p == BatchParam{}) {
// empty parameter is passed in, don't regenerate so that we can use gindex in
// predictor, which doesn't have any training parameter.
return false;
}
// Avoid comparing nan values.
bool l_nan = std::isnan(old.sparse_thresh);
bool r_nan = std::isnan(p.sparse_thresh);
// regenerate if parameter is changed.
bool st_chg = (l_nan != r_nan) || (!l_nan && !r_nan && (old.sparse_thresh != p.sparse_thresh));
bool param_chg = old.gpu_id != p.gpu_id || old.max_bin != p.max_bin;
return p.regen || param_chg || st_chg;
}
} // namespace xgboost
#endif // XGBOOST_DATA_GRADIENT_INDEX_H_