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

Initial support for external memory in gradient index. (#7183)

Browse Source 
* Add hessian to batch param in preparation of new approx impl.
* Extract a push method for gradient index matrix.
* Use span instead of vector ref for hessian in sketching.
* Create a binary format for gradient index.
This commit is contained in:
Jiaming Yuan
2021-09-13 12:40:56 +08:00
committed by GitHub
parent a0dcf6f5c1
commit 3515931305
26 changed files with 546 additions and 171 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
src/tree/gpu_hist/gradient_based_sampler.cu
View File

@@ -9,6 +9,7 @@
#include <algorithm>
#include <limits>
#include <utility>
#include "../../common/compressed_iterator.h"
#include "../../common/random.h"
@@ -185,10 +186,10 @@ GradientBasedSample UniformSampling::Sample(common::Span<GradientPair> gpair, DM
ExternalMemoryUniformSampling::ExternalMemoryUniformSampling(EllpackPageImpl const* page,
size_t n_rows,
const BatchParam& batch_param,
BatchParam batch_param,
float subsample)
: original_page_(page),
batch_param_(batch_param),
batch_param_(std::move(batch_param)),
subsample_(subsample),
sample_row_index_(n_rows) {}
@@ -259,10 +260,10 @@ GradientBasedSample GradientBasedSampling::Sample(common::Span<GradientPair> gpa
ExternalMemoryGradientBasedSampling::ExternalMemoryGradientBasedSampling(
EllpackPageImpl const* page,
size_t n_rows,
const BatchParam& batch_param,
BatchParam batch_param,
float subsample)
: original_page_(page),
batch_param_(batch_param),
batch_param_(std::move(batch_param)),
subsample_(subsample),
threshold_(n_rows + 1, 0.0f),
grad_sum_(n_rows, 0.0f),

4
src/tree/gpu_hist/gradient_based_sampler.cuh
View File

@@ -68,7 +68,7 @@ class ExternalMemoryUniformSampling : public SamplingStrategy {
public:
ExternalMemoryUniformSampling(EllpackPageImpl const* page,
size_t n_rows,
const BatchParam& batch_param,
BatchParam batch_param,
float subsample);
GradientBasedSample Sample(common::Span<GradientPair> gpair, DMatrix* dmat) override;
@@ -102,7 +102,7 @@ class ExternalMemoryGradientBasedSampling : public SamplingStrategy {
public:
ExternalMemoryGradientBasedSampling(EllpackPageImpl const* page,
size_t n_rows,
const BatchParam& batch_param,
BatchParam batch_param,
float subsample);
GradientBasedSample Sample(common::Span<GradientPair> gpair, DMatrix* dmat) override;

2
src/tree/updater_gpu_hist.cu
View File

@@ -209,7 +209,7 @@ struct GPUHistMakerDevice {
tree_evaluator(param, n_features, _device_id),
column_sampler(column_sampler_seed),
interaction_constraints(param, n_features),
batch_param(_batch_param) {
batch_param(std::move(_batch_param)) {
sampler.reset(new GradientBasedSampler(
page, _n_rows, batch_param, param.subsample, param.sampling_method));
if (!param.monotone_constraints.empty()) {

14
src/tree/updater_quantile_hist.cc
View File

@@ -69,13 +69,13 @@ void QuantileHistMaker::CallBuilderUpdate(const std::unique_ptr<Builder<Gradient
void QuantileHistMaker::Update(HostDeviceVector<GradientPair> *gpair,
DMatrix *dmat,
const std::vector<RegTree *> &trees) {
auto const &gmat =
*(dmat->GetBatches<GHistIndexMatrix>(
BatchParam{GenericParameter::kCpuId, param_.max_bin})
.begin());
auto it = dmat->GetBatches<GHistIndexMatrix>(
BatchParam{GenericParameter::kCpuId, param_.max_bin})
.begin();
auto p_gmat = it.Page();
if (dmat != p_last_dmat_ || is_gmat_initialized_ == false) {
updater_monitor_.Start("GmatInitialization");
column_matrix_.Init(gmat, param_.sparse_threshold);
column_matrix_.Init(*p_gmat, param_.sparse_threshold);
updater_monitor_.Stop("GmatInitialization");
// A proper solution is puting cut matrix in DMatrix, see:
// https://github.com/dmlc/xgboost/issues/5143
@@ -91,12 +91,12 @@ void QuantileHistMaker::Update(HostDeviceVector<GradientPair> *gpair,
if (!float_builder_) {
this->SetBuilder(n_trees, &float_builder_, dmat);
}
CallBuilderUpdate(float_builder_, gpair, dmat, gmat, trees);
CallBuilderUpdate(float_builder_, gpair, dmat, *p_gmat, trees);
} else {
if (!double_builder_) {
SetBuilder(n_trees, &double_builder_, dmat);
}
CallBuilderUpdate(double_builder_, gpair, dmat, gmat, trees);
CallBuilderUpdate(double_builder_, gpair, dmat, *p_gmat, trees);
}
param_.learning_rate = lr;