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

CPU evaluation for cat data. (#7393)

Browse Source 
* Implementation for one hot based.
* Implementation for partition based. (LightGBM)
This commit is contained in:
Jiaming Yuan
2021-11-06 14:41:35 +08:00
committed by GitHub
parent 6ede12412c
commit d7d1b6e3a6
15 changed files with 540 additions and 166 deletions
Download Patch File Download Diff File Expand all files Collapse all files

253
src/tree/hist/evaluate_splits.h
View File

@@ -6,13 +6,16 @@
#include <algorithm>
#include <memory>
#include <numeric>
#include <limits>
#include <utility>
#include <vector>
#include "xgboost/task.h"
#include "../param.h"
#include "../constraints.h"
#include "../split_evaluator.h"
#include "../../common/categorical.h"
#include "../../common/random.h"
#include "../../common/hist_util.h"
#include "../../data/gradient_index.h"
@@ -36,13 +39,13 @@ template <typename GradientSumT, typename ExpandEntry> class HistEvaluator {
int32_t n_threads_ {0};
FeatureInteractionConstraintHost interaction_constraints_;
std::vector<NodeEntry> snode_;
ObjInfo task_;
// if sum of statistics for non-missing values in the node
// is equal to sum of statistics for all values:
// then - there are no missing values
// else - there are missing values
bool static SplitContainsMissingValues(const GradStats e,
const NodeEntry &snode) {
bool static SplitContainsMissingValues(const GradStats e, const NodeEntry &snode) {
if (e.GetGrad() == snode.stats.GetGrad() &&
e.GetHess() == snode.stats.GetHess()) {
return false;
@@ -50,38 +53,40 @@ template <typename GradientSumT, typename ExpandEntry> class HistEvaluator {
return true;
}
}
enum SplitType { kNum = 0, kOneHot = 1, kPart = 2 };
// Enumerate/Scan the split values of specific feature
// Returns the sum of gradients corresponding to the data points that contains
// a non-missing value for the particular feature fid.
template <int d_step>
GradStats EnumerateSplit(
common::HistogramCuts const &cut, const common::GHistRow<GradientSumT> &hist,
const NodeEntry &snode, SplitEntry *p_best, bst_feature_t fidx,
bst_node_t nidx,
TreeEvaluator::SplitEvaluator<TrainParam> const &evaluator) const {
template <int d_step, SplitType split_type>
GradStats EnumerateSplit(common::HistogramCuts const &cut, common::Span<size_t const> sorted_idx,
const common::GHistRow<GradientSumT> &hist, bst_feature_t fidx,
bst_node_t nidx,
TreeEvaluator::SplitEvaluator<TrainParam> const &evaluator,
SplitEntry *p_best) const {
static_assert(d_step == +1 || d_step == -1, "Invalid step.");
// aliases
const std::vector<uint32_t> &cut_ptr = cut.Ptrs();
const std::vector<bst_float> &cut_val = cut.Values();
auto const &parent = snode_[nidx];
int32_t n_bins{static_cast<int32_t>(cut_ptr.at(fidx + 1) - cut_ptr[fidx])};
auto f_hist = hist.subspan(cut_ptr[fidx], n_bins);
// statistics on both sides of split
GradStats c;
GradStats e;
GradStats left_sum;
GradStats right_sum;
// best split so far
SplitEntry best;
// bin boundaries
CHECK_LE(cut_ptr[fidx],
static_cast<uint32_t>(std::numeric_limits<int32_t>::max()));
CHECK_LE(cut_ptr[fidx + 1],
static_cast<uint32_t>(std::numeric_limits<int32_t>::max()));
// imin: index (offset) of the minimum value for feature fid
// need this for backward enumeration
CHECK_LE(cut_ptr[fidx], static_cast<uint32_t>(std::numeric_limits<int32_t>::max()));
CHECK_LE(cut_ptr[fidx + 1], static_cast<uint32_t>(std::numeric_limits<int32_t>::max()));
// imin: index (offset) of the minimum value for feature fid need this for backward
// enumeration
const auto imin = static_cast<int32_t>(cut_ptr[fidx]);
// ibegin, iend: smallest/largest cut points for feature fid
// use int to allow for value -1
// ibegin, iend: smallest/largest cut points for feature fid use int to allow for
// value -1
int32_t ibegin, iend;
if (d_step > 0) {
ibegin = static_cast<int32_t>(cut_ptr[fidx]);
@@ -91,49 +96,118 @@ template <typename GradientSumT, typename ExpandEntry> class HistEvaluator {
iend = static_cast<int32_t>(cut_ptr[fidx]) - 1;
}
auto calc_bin_value = [&](auto i) {
switch (split_type) {
case kNum: {
left_sum.Add(hist[i].GetGrad(), hist[i].GetHess());
right_sum.SetSubstract(parent.stats, left_sum);
break;
}
case kOneHot: {
// not-chosen categories go to left
right_sum = GradStats{hist[i]};
left_sum.SetSubstract(parent.stats, right_sum);
break;
}
case kPart: {
auto j = d_step == 1 ? (i - ibegin) : (ibegin - i);
right_sum.Add(f_hist[sorted_idx[j]].GetGrad(), f_hist[sorted_idx[j]].GetHess());
left_sum.SetSubstract(parent.stats, right_sum);
break;
}
default: {
std::terminate();
}
}
};
int32_t best_thresh{-1};
for (int32_t i = ibegin; i != iend; i += d_step) {
// start working
// try to find a split
e.Add(hist[i].GetGrad(), hist[i].GetHess());
if (e.GetHess() >= param_.min_child_weight) {
c.SetSubstract(snode.stats, e);
if (c.GetHess() >= param_.min_child_weight) {
bst_float loss_chg;
bst_float split_pt;
if (d_step > 0) {
// forward enumeration: split at right bound of each bin
loss_chg = static_cast<bst_float>(
evaluator.CalcSplitGain(param_, nidx, fidx, GradStats{e},
GradStats{c}) -
snode.root_gain);
split_pt = cut_val[i];
best.Update(loss_chg, fidx, split_pt, d_step == -1, e, c);
} else {
// backward enumeration: split at left bound of each bin
loss_chg = static_cast<bst_float>(
evaluator.CalcSplitGain(param_, nidx, fidx, GradStats{c},
GradStats{e}) -
snode.root_gain);
if (i == imin) {
// for leftmost bin, left bound is the smallest feature value
split_pt = cut.MinValues()[fidx];
} else {
split_pt = cut_val[i - 1];
calc_bin_value(i);
bool improved{false};
if (left_sum.GetHess() >= param_.min_child_weight &&
right_sum.GetHess() >= param_.min_child_weight) {
bst_float loss_chg;
bst_float split_pt;
if (d_step > 0) {
// forward enumeration: split at right bound of each bin
loss_chg =
static_cast<float>(evaluator.CalcSplitGain(param_, nidx, fidx, GradStats{left_sum},
GradStats{right_sum}) -
parent.root_gain);
split_pt = cut_val[i];
improved = best.Update(loss_chg, fidx, split_pt, d_step == -1, split_type != kNum,
left_sum, right_sum);
} else {
// backward enumeration: split at left bound of each bin
loss_chg =
static_cast<float>(evaluator.CalcSplitGain(param_, nidx, fidx, GradStats{right_sum},
GradStats{left_sum}) -
parent.root_gain);
switch (split_type) {
case kNum: {
if (i == imin) {
split_pt = cut.MinValues()[fidx];
} else {
split_pt = cut_val[i - 1];
}
break;
}
case kOneHot: {
split_pt = cut_val[i];
break;
}
case kPart: {
split_pt = cut_val[i];
break;
}
best.Update(loss_chg, fidx, split_pt, d_step == -1, c, e);
}
improved = best.Update(loss_chg, fidx, split_pt, d_step == -1, split_type != kNum,
right_sum, left_sum);
}
if (improved) {
best_thresh = i;
}
}
}
if (split_type == kPart && best_thresh != -1) {
auto n = common::CatBitField::ComputeStorageSize(n_bins);
best.cat_bits.resize(n, 0);
common::CatBitField cat_bits{best.cat_bits};
if (d_step == 1) {
std::for_each(sorted_idx.begin(), sorted_idx.begin() + (best_thresh - ibegin + 1),
[&cat_bits](size_t c) { cat_bits.Set(c); });
} else {
std::for_each(sorted_idx.rbegin(), sorted_idx.rbegin() + (ibegin - best_thresh),
[&cat_bits](size_t c) { cat_bits.Set(c); });
}
}
p_best->Update(best);
return e;
switch (split_type) {
case kNum:
// Normal, accumulated to left
return left_sum;
case kOneHot:
// Doesn't matter, not accumulating.
return {};
case kPart:
// Accumulated to right due to chosen cats go to right.
return right_sum;
}
return left_sum;
}
public:
void EvaluateSplits(const common::HistCollection<GradientSumT> &hist,
common::HistogramCuts const &cut, const RegTree &tree,
std::vector<ExpandEntry>* p_entries) {
common::HistogramCuts const &cut,
common::Span<FeatureType const> feature_types,
const RegTree &tree,
std::vector<ExpandEntry> *p_entries) {
auto& entries = *p_entries;
// All nodes are on the same level, so we can store the shared ptr.
std::vector<std::shared_ptr<HostDeviceVector<bst_feature_t>>> features(
@@ -150,7 +224,7 @@ template <typename GradientSumT, typename ExpandEntry> class HistEvaluator {
return features[nidx_in_set]->Size();
}, grain_size);
std::vector<ExpandEntry> tloc_candidates(omp_get_max_threads() * entries.size());
std::vector<ExpandEntry> tloc_candidates(n_threads_ * entries.size());
for (size_t i = 0; i < entries.size(); ++i) {
for (decltype(n_threads_) j = 0; j < n_threads_; ++j) {
tloc_candidates[i * n_threads_ + j] = entries[i];
@@ -167,12 +241,37 @@ template <typename GradientSumT, typename ExpandEntry> class HistEvaluator {
auto features_set = features[nidx_in_set]->ConstHostSpan();
for (auto fidx_in_set = r.begin(); fidx_in_set < r.end(); fidx_in_set++) {
auto fidx = features_set[fidx_in_set];
if (interaction_constraints_.Query(nidx, fidx)) {
auto grad_stats = EnumerateSplit<+1>(cut, histogram, snode_[nidx],
best, fidx, nidx, evaluator);
bool is_cat = common::IsCat(feature_types, fidx);
if (!interaction_constraints_.Query(nidx, fidx)) {
continue;
}
if (is_cat) {
auto n_bins = cut.Ptrs().at(fidx + 1) - cut.Ptrs()[fidx];
if (common::UseOneHot(n_bins, param_.max_cat_to_onehot, task_)) {
EnumerateSplit<+1, kOneHot>(cut, {}, histogram, fidx, nidx, evaluator, best);
EnumerateSplit<-1, kOneHot>(cut, {}, histogram, fidx, nidx, evaluator, best);
} else {
auto const &cut_ptr = cut.Ptrs();
std::vector<size_t> sorted_idx(n_bins);
std::iota(sorted_idx.begin(), sorted_idx.end(), 0);
auto feat_hist = histogram.subspan(cut_ptr[fidx], n_bins);
std::stable_sort(sorted_idx.begin(), sorted_idx.end(), [&](size_t l, size_t r) {
auto ret = evaluator.CalcWeightCat(param_, feat_hist[l]) <
evaluator.CalcWeightCat(param_, feat_hist[r]);
static_assert(std::is_same<decltype(ret), bool>::value, "");
return ret;
});
auto grad_stats =
EnumerateSplit<+1, kPart>(cut, sorted_idx, histogram, fidx, nidx, evaluator, best);
if (SplitContainsMissingValues(grad_stats, snode_[nidx])) {
EnumerateSplit<-1, kPart>(cut, sorted_idx, histogram, fidx, nidx, evaluator, best);
}
}
} else {
auto grad_stats =
EnumerateSplit<+1, kNum>(cut, {}, histogram, fidx, nidx, evaluator, best);
if (SplitContainsMissingValues(grad_stats, snode_[nidx])) {
EnumerateSplit<-1>(cut, histogram, snode_[nidx], best, fidx, nidx,
evaluator);
EnumerateSplit<-1, kNum>(cut, {}, histogram, fidx, nidx, evaluator, best);
}
}
}
@@ -187,7 +286,7 @@ template <typename GradientSumT, typename ExpandEntry> class HistEvaluator {
}
}
// Add splits to tree, handles all statistic
void ApplyTreeSplit(ExpandEntry candidate, RegTree *p_tree) {
void ApplyTreeSplit(ExpandEntry const& candidate, RegTree *p_tree) {
auto evaluator = tree_evaluator_.GetEvaluator();
RegTree &tree = *p_tree;
@@ -201,13 +300,31 @@ template <typename GradientSumT, typename ExpandEntry> class HistEvaluator {
auto right_weight = evaluator.CalcWeight(
candidate.nid, param_, GradStats{candidate.split.right_sum});
tree.ExpandNode(candidate.nid, candidate.split.SplitIndex(),
candidate.split.split_value, candidate.split.DefaultLeft(),
base_weight, left_weight * param_.learning_rate,
right_weight * param_.learning_rate,
candidate.split.loss_chg, parent_sum.GetHess(),
candidate.split.left_sum.GetHess(),
candidate.split.right_sum.GetHess());
if (candidate.split.is_cat) {
std::vector<uint32_t> split_cats;
if (candidate.split.cat_bits.empty()) {
CHECK_LT(candidate.split.split_value, std::numeric_limits<bst_cat_t>::max())
<< "Categorical feature value too large.";
auto cat = common::AsCat(candidate.split.split_value);
split_cats.resize(LBitField32::ComputeStorageSize(std::max(cat + 1, 1)), 0);
LBitField32 cat_bits;
cat_bits = LBitField32(split_cats);
cat_bits.Set(cat);
} else {
split_cats = candidate.split.cat_bits;
}
tree.ExpandCategorical(
candidate.nid, candidate.split.SplitIndex(), split_cats, candidate.split.DefaultLeft(),
base_weight, left_weight, right_weight, candidate.split.loss_chg, parent_sum.GetHess(),
candidate.split.left_sum.GetHess(), candidate.split.right_sum.GetHess());
} else {
tree.ExpandNode(candidate.nid, candidate.split.SplitIndex(), candidate.split.split_value,
candidate.split.DefaultLeft(), base_weight,
left_weight * param_.learning_rate, right_weight * param_.learning_rate,
candidate.split.loss_chg, parent_sum.GetHess(),
candidate.split.left_sum.GetHess(), candidate.split.right_sum.GetHess());
}
// Set up child constraints
auto left_child = tree[candidate.nid].LeftChild();
@@ -249,14 +366,14 @@ template <typename GradientSumT, typename ExpandEntry> class HistEvaluator {
public:
// The column sampler must be constructed by caller since we need to preserve the rng
// for the entire training session.
explicit HistEvaluator(TrainParam const &param, MetaInfo const &info,
int32_t n_threads,
std::shared_ptr<common::ColumnSampler> sampler,
explicit HistEvaluator(TrainParam const &param, MetaInfo const &info, int32_t n_threads,
std::shared_ptr<common::ColumnSampler> sampler, ObjInfo task,
bool skip_0_index = false)
: param_{param}, column_sampler_{std::move(sampler)},
tree_evaluator_{param, static_cast<bst_feature_t>(info.num_col_),
GenericParameter::kCpuId},
n_threads_{n_threads} {
: param_{param},
column_sampler_{std::move(sampler)},
tree_evaluator_{param, static_cast<bst_feature_t>(info.num_col_), GenericParameter::kCpuId},
n_threads_{n_threads},
task_{task} {
interaction_constraints_.Configure(param, info.num_col_);
column_sampler_->Init(info.num_col_, info.feature_weigths.HostVector(),
param_.colsample_bynode, param_.colsample_bylevel,

44
src/tree/param.h
View File

@@ -1,5 +1,5 @@
/*!
* Copyright 2014-2019 by Contributors
* Copyright 2014-2021 by Contributors
* \file param.h
* \brief training parameters, statistics used to support tree construction.
* \author Tianqi Chen
@@ -7,6 +7,7 @@
#ifndef XGBOOST_TREE_PARAM_H_
#define XGBOOST_TREE_PARAM_H_
#include <algorithm>
#include <cmath>
#include <cstring>
#include <limits>
@@ -15,6 +16,7 @@
#include "xgboost/parameter.h"
#include "xgboost/data.h"
#include "../common/categorical.h"
#include "../common/math.h"
namespace xgboost {
@@ -36,6 +38,8 @@ struct TrainParam : public XGBoostParameter<TrainParam> {
enum TreeGrowPolicy { kDepthWise = 0, kLossGuide = 1 };
int grow_policy;
uint32_t max_cat_to_onehot{1};
//----- the rest parameters are less important ----
// minimum amount of hessian(weight) allowed in a child
float min_child_weight;
@@ -119,6 +123,10 @@ struct TrainParam : public XGBoostParameter<TrainParam> {
"Tree growing policy. 0: favor splitting at nodes closest to the node, "
"i.e. grow depth-wise. 1: favor splitting at nodes with highest loss "
"change. (cf. LightGBM)");
DMLC_DECLARE_FIELD(max_cat_to_onehot)
.set_default(4)
.set_lower_bound(1)
.describe("Maximum number of categories to use one-hot encoding based split.");
DMLC_DECLARE_FIELD(min_child_weight)
.set_lower_bound(0.0f)
.set_default(1.0f)
@@ -384,6 +392,8 @@ struct SplitEntryContainer {
/*! \brief split index */
bst_feature_t sindex{0};
bst_float split_value{0.0f};
std::vector<uint32_t> cat_bits;
bool is_cat{false};
GradientT left_sum;
GradientT right_sum;
@@ -433,6 +443,8 @@ struct SplitEntryContainer {
this->loss_chg = e.loss_chg;
this->sindex = e.sindex;
this->split_value = e.split_value;
this->is_cat = e.is_cat;
this->cat_bits = e.cat_bits;
this->left_sum = e.left_sum;
this->right_sum = e.right_sum;
return true;
@@ -449,9 +461,8 @@ struct SplitEntryContainer {
* \return whether the proposed split is better and can replace current split
*/
bool Update(bst_float new_loss_chg, unsigned split_index,
bst_float new_split_value, bool default_left,
const GradientT &left_sum,
const GradientT &right_sum) {
bst_float new_split_value, bool default_left, bool is_cat,
const GradientT &left_sum, const GradientT &right_sum) {
if (this->NeedReplace(new_loss_chg, split_index)) {
this->loss_chg = new_loss_chg;
if (default_left) {
@@ -459,6 +470,31 @@ struct SplitEntryContainer {
}
this->sindex = split_index;
this->split_value = new_split_value;
this->is_cat = is_cat;
this->left_sum = left_sum;
this->right_sum = right_sum;
return true;
} else {
return false;
}
}
/*!
* \brief Update with partition based categorical split.
*
* \return Whether the proposed split is better and can replace current split.
*/
bool Update(float new_loss_chg, bst_feature_t split_index, common::KCatBitField cats,
bool default_left, GradientT const &left_sum, GradientT const &right_sum) {
if (this->NeedReplace(new_loss_chg, split_index)) {
this->loss_chg = new_loss_chg;
if (default_left) {
split_index |= (1U << 31);
}
this->sindex = split_index;
cat_bits.resize(cats.Bits().size());
std::copy(cats.Bits().begin(), cats.Bits().end(), cat_bits.begin());
this->is_cat = true;
this->left_sum = left_sum;
this->right_sum = right_sum;
return true;

8
src/tree/split_evaluator.h
View File

@@ -92,7 +92,7 @@ class TreeEvaluator {
XGBOOST_DEVICE float CalcWeight(bst_node_t nodeid, const ParamT &param,
tree::GradStats const& stats) const {
float w = xgboost::tree::CalcWeight(param, stats);
float w = ::xgboost::tree::CalcWeight(param, stats);
if (!has_constraint) {
return w;
}
@@ -107,6 +107,12 @@ class TreeEvaluator {
return w;
}
}
template <typename GradientSumT>
XGBOOST_DEVICE double CalcWeightCat(ParamT const& param, GradientSumT const& stats) const {
return ::xgboost::tree::CalcWeight(param, stats);
}
XGBOOST_DEVICE float
CalcGainGivenWeight(ParamT const &p, tree::GradStats const& stats, float w) const {
if (stats.GetHess() <= 0) {

14
src/tree/updater_colmaker.cc
View File

@@ -336,10 +336,10 @@ class ColMaker: public TreeUpdater {
bst_float proposed_split = (fvalue + e.last_fvalue) * 0.5f;
if ( proposed_split == fvalue ) {
e.best.Update(loss_chg, fid, e.last_fvalue,
d_step == -1, c, e.stats);
d_step == -1, false, c, e.stats);
} else {
e.best.Update(loss_chg, fid, proposed_split,
d_step == -1, c, e.stats);
d_step == -1, false, c, e.stats);
}
} else {
loss_chg = static_cast<bst_float>(
@@ -348,10 +348,10 @@ class ColMaker: public TreeUpdater {
bst_float proposed_split = (fvalue + e.last_fvalue) * 0.5f;
if ( proposed_split == fvalue ) {
e.best.Update(loss_chg, fid, e.last_fvalue,
d_step == -1, e.stats, c);
d_step == -1, false, e.stats, c);
} else {
e.best.Update(loss_chg, fid, proposed_split,
d_step == -1, e.stats, c);
d_step == -1, false, e.stats, c);
}
}
}
@@ -430,14 +430,14 @@ class ColMaker: public TreeUpdater {
loss_chg = static_cast<bst_float>(
evaluator.CalcSplitGain(param_, nid, fid, c, e.stats) -
snode_[nid].root_gain);
e.best.Update(loss_chg, fid, e.last_fvalue + delta, d_step == -1, c,
e.stats);
e.best.Update(loss_chg, fid, e.last_fvalue + delta, d_step == -1,
false, c, e.stats);
} else {
loss_chg = static_cast<bst_float>(
evaluator.CalcSplitGain(param_, nid, fid, e.stats, c) -
snode_[nid].root_gain);
e.best.Update(loss_chg, fid, e.last_fvalue + delta, d_step == -1,
e.stats, c);
false, e.stats, c);
}
}
}

6
src/tree/updater_histmaker.cc
View File

@@ -173,7 +173,8 @@ class HistMaker: public BaseMaker {
if (c.sum_hess >= param_.min_child_weight) {
double loss_chg = CalcGain(param_, s.GetGrad(), s.GetHess()) +
CalcGain(param_, c.GetGrad(), c.GetHess()) - root_gain;
if (best->Update(static_cast<bst_float>(loss_chg), fid, hist.cut[i], false, s, c)) {
if (best->Update(static_cast<bst_float>(loss_chg), fid, hist.cut[i],
false, false, s, c)) {
*left_sum = s;
}
}
@@ -187,7 +188,8 @@ class HistMaker: public BaseMaker {
if (c.sum_hess >= param_.min_child_weight) {
double loss_chg = CalcGain(param_, s.GetGrad(), s.GetHess()) +
CalcGain(param_, c.GetGrad(), c.GetHess()) - root_gain;
if (best->Update(static_cast<bst_float>(loss_chg), fid, hist.cut[i-1], true, c, s)) {
if (best->Update(static_cast<bst_float>(loss_chg), fid,
hist.cut[i - 1], true, false, c, s)) {
*left_sum = c;
}
}

17
src/tree/updater_quantile_hist.cc
View File

@@ -168,9 +168,11 @@ void QuantileHistMaker::Builder<GradientSumT>::InitRoot(
std::vector<CPUExpandEntry> entries{node};
builder_monitor_.Start("EvaluateSplits");
auto ft = p_fmat->Info().feature_types.ConstHostSpan();
for (auto const &gmat : p_fmat->GetBatches<GHistIndexMatrix>(
BatchParam{GenericParameter::kCpuId, param_.max_bin})) {
evaluator_->EvaluateSplits(histogram_builder_->Histogram(), gmat.cut, *p_tree, &entries);
evaluator_->EvaluateSplits(histogram_builder_->Histogram(), gmat.cut, ft,
*p_tree, &entries);
break;
}
builder_monitor_.Stop("EvaluateSplits");
@@ -272,8 +274,9 @@ void QuantileHistMaker::Builder<GradientSumT>::ExpandTree(
}
builder_monitor_.Start("EvaluateSplits");
evaluator_->EvaluateSplits(this->histogram_builder_->Histogram(), gmat.cut,
*p_tree, &nodes_to_evaluate);
auto ft = p_fmat->Info().feature_types.ConstHostSpan();
evaluator_->EvaluateSplits(this->histogram_builder_->Histogram(),
gmat.cut, ft, *p_tree, &nodes_to_evaluate);
builder_monitor_.Stop("EvaluateSplits");
for (size_t i = 0; i < nodes_for_apply_split.size(); ++i) {
@@ -529,11 +532,11 @@ void QuantileHistMaker::Builder<GradientSumT>::InitData(
// store a pointer to the tree
p_last_tree_ = &tree;
if (data_layout_ == DataLayout::kDenseDataOneBased) {
evaluator_.reset(new HistEvaluator<GradientSumT, CPUExpandEntry>{param_, info, this->nthread_,
column_sampler_, true});
evaluator_.reset(new HistEvaluator<GradientSumT, CPUExpandEntry>{
param_, info, this->nthread_, column_sampler_, task_, true});
} else {
evaluator_.reset(new HistEvaluator<GradientSumT, CPUExpandEntry>{param_, info, this->nthread_,
column_sampler_, false});
evaluator_.reset(new HistEvaluator<GradientSumT, CPUExpandEntry>{
param_, info, this->nthread_, column_sampler_, task_, false});
}
if (data_layout_ == DataLayout::kDenseDataZeroBased