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

Accept histogram cut instead gradient index in evaluation. (#7336)

Browse Source
This commit is contained in:
Jiaming Yuan 2021-10-20 18:04:46 +08:00 committed by GitHub
parent 15685996fc
commit 8d7c6366d7
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 13 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
src/tree/hist/evaluate_splits.h
View File

@ -56,15 +56,15 @@ template <typename GradientSumT, typename ExpandEntry> class HistEvaluator {
// a non-missing value for the particular feature fid. // a non-missing value for the particular feature fid.
template <int d_step> template <int d_step>
GradStats EnumerateSplit( GradStats EnumerateSplit(
const GHistIndexMatrix &gmat, const common::GHistRow<GradientSumT> &hist, common::HistogramCuts const &cut, const common::GHistRow<GradientSumT> &hist,
const NodeEntry &snode, SplitEntry *p_best, bst_feature_t fidx, const NodeEntry &snode, SplitEntry *p_best, bst_feature_t fidx,
bst_node_t nidx, bst_node_t nidx,
TreeEvaluator::SplitEvaluator<TrainParam> const &evaluator) const { TreeEvaluator::SplitEvaluator<TrainParam> const &evaluator) const {
static_assert(d_step == +1 || d_step == -1, "Invalid step."); static_assert(d_step == +1 || d_step == -1, "Invalid step.");
// aliases // aliases
const std::vector<uint32_t> &cut_ptr = gmat.cut.Ptrs(); const std::vector<uint32_t> &cut_ptr = cut.Ptrs();
const std::vector<bst_float> &cut_val = gmat.cut.Values(); const std::vector<bst_float> &cut_val = cut.Values();
// statistics on both sides of split // statistics on both sides of split
GradStats c; GradStats c;
@ -116,7 +116,7 @@ template <typename GradientSumT, typename ExpandEntry> class HistEvaluator {
snode.root_gain); snode.root_gain);
if (i == imin) { if (i == imin) {
// for leftmost bin, left bound is the smallest feature value // for leftmost bin, left bound is the smallest feature value
split_pt = gmat.cut.MinValues()[fidx]; split_pt = cut.MinValues()[fidx];
} else { } else {
split_pt = cut_val[i - 1]; split_pt = cut_val[i - 1];
} }
@ -132,7 +132,7 @@ template <typename GradientSumT, typename ExpandEntry> class HistEvaluator {
public: public:
void EvaluateSplits(const common::HistCollection<GradientSumT> &hist, void EvaluateSplits(const common::HistCollection<GradientSumT> &hist,
GHistIndexMatrix const &gidx, const RegTree &tree, common::HistogramCuts const &cut, const RegTree &tree,
std::vector<ExpandEntry>* p_entries) { std::vector<ExpandEntry>* p_entries) {
auto& entries = *p_entries; auto& entries = *p_entries;
// All nodes are on the same level, so we can store the shared ptr. // All nodes are on the same level, so we can store the shared ptr.
@ -168,10 +168,10 @@ template <typename GradientSumT, typename ExpandEntry> class HistEvaluator {
for (auto fidx_in_set = r.begin(); fidx_in_set < r.end(); fidx_in_set++) { for (auto fidx_in_set = r.begin(); fidx_in_set < r.end(); fidx_in_set++) {
auto fidx = features_set[fidx_in_set]; auto fidx = features_set[fidx_in_set];
if (interaction_constraints_.Query(nidx, fidx)) { if (interaction_constraints_.Query(nidx, fidx)) {
auto grad_stats = EnumerateSplit<+1>(gidx, histogram, snode_[nidx], auto grad_stats = EnumerateSplit<+1>(cut, histogram, snode_[nidx],
best, fidx, nidx, evaluator); best, fidx, nidx, evaluator);
if (SplitContainsMissingValues(grad_stats, snode_[nidx])) { if (SplitContainsMissingValues(grad_stats, snode_[nidx])) {
EnumerateSplit<-1>(gidx, histogram, snode_[nidx], best, fidx, nidx, EnumerateSplit<-1>(cut, histogram, snode_[nidx], best, fidx, nidx,
evaluator); evaluator);
} }
} }

5
src/tree/updater_quantile_hist.cc
View File

@ -170,7 +170,8 @@ void QuantileHistMaker::Builder<GradientSumT>::InitRoot(
builder_monitor_.Start("EvaluateSplits"); builder_monitor_.Start("EvaluateSplits");
for (auto const &gmat : p_fmat->GetBatches<GHistIndexMatrix>( for (auto const &gmat : p_fmat->GetBatches<GHistIndexMatrix>(
BatchParam{GenericParameter::kCpuId, param_.max_bin})) { BatchParam{GenericParameter::kCpuId, param_.max_bin})) {
evaluator_->EvaluateSplits(histogram_builder_->Histogram(), gmat, *p_tree, &entries); evaluator_->EvaluateSplits(histogram_builder_->Histogram(), gmat.cut, *p_tree, &entries);
break;
} }
builder_monitor_.Stop("EvaluateSplits"); builder_monitor_.Stop("EvaluateSplits");
node = entries.front(); node = entries.front();
@ -271,7 +272,7 @@ void QuantileHistMaker::Builder<GradientSumT>::ExpandTree(
} }
builder_monitor_.Start("EvaluateSplits"); builder_monitor_.Start("EvaluateSplits");
evaluator_->EvaluateSplits(this->histogram_builder_->Histogram(), gmat, evaluator_->EvaluateSplits(this->histogram_builder_->Histogram(), gmat.cut,
*p_tree, &nodes_to_evaluate); *p_tree, &nodes_to_evaluate);
builder_monitor_.Stop("EvaluateSplits"); builder_monitor_.Stop("EvaluateSplits");

3
tests/cpp/data/test_gradient_index.cc
View File

@ -13,7 +13,8 @@ TEST(GradientIndex, ExternalMemory) {
std::unique_ptr<DMatrix> dmat = CreateSparsePageDMatrix(10000); std::unique_ptr<DMatrix> dmat = CreateSparsePageDMatrix(10000);
std::vector<size_t> base_rowids; std::vector<size_t> base_rowids;
std::vector<float> hessian(dmat->Info().num_row_, 1); std::vector<float> hessian(dmat->Info().num_row_, 1);
for (auto const& page : dmat->GetBatches<GHistIndexMatrix>({0, 64, hessian})) { for (auto const &page : dmat->GetBatches<GHistIndexMatrix>(
{GenericParameter::kCpuId, 64, hessian})) {
base_rowids.push_back(page.base_rowid); base_rowids.push_back(page.base_rowid);
} }
size_t i = 0; size_t i = 0;

2
tests/cpp/tree/hist/test_evaluate_splits.cc
View File

@ -58,7 +58,7 @@ template <typename GradientSumT> void TestEvaluateSplits() {
entries.front().depth = 0; entries.front().depth = 0;
evaluator.InitRoot(GradStats{total_gpair}); evaluator.InitRoot(GradStats{total_gpair});
evaluator.EvaluateSplits(hist, gmat, tree, &entries); evaluator.EvaluateSplits(hist, gmat.cut, tree, &entries);
auto best_loss_chg = auto best_loss_chg =
evaluator.Evaluator().CalcSplitGain( evaluator.Evaluator().CalcSplitGain(