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

Support hist in the partition builder under column split (#9120)

Browse Source
This commit is contained in:
Rong Ou 2023-05-10 14:24:29 -07:00 committed by GitHub
parent 52311dcec9
commit 603f8ce2fa
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
3 changed files with 160 additions and 16 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

50
src/common/partition_builder.h
View File

@ -183,14 +183,28 @@ class PartitionBuilder {
SetNRightElems(node_in_set, range.begin(), n_right);
}
template <bool any_missing, typename ColumnType, typename Predicate>
void MaskKernel(ColumnType* p_column, common::Span<const size_t> row_indices, size_t base_rowid,
BitVector* decision_bits, BitVector* missing_bits, Predicate&& pred) {
auto& column = *p_column;
for (auto const row_id : row_indices) {
auto const bin_id = column[row_id - base_rowid];
if (any_missing && bin_id == ColumnType::kMissingId) {
missing_bits->Set(row_id - base_rowid);
} else if (pred(row_id, bin_id)) {
decision_bits->Set(row_id - base_rowid);
}
}
}
/**
* @brief When data is split by column, we don't have all the features locally on the current
* worker, so we go through all the rows and mark the bit vectors on whether the decision is made
* to go right, or if the feature value used for the split is missing.
*/
template <typename ExpandEntry>
template <typename BinIdxType, bool any_missing, bool any_cat, typename ExpandEntry>
void MaskRows(const size_t node_in_set, std::vector<ExpandEntry> const& nodes,
const common::Range1d range, GHistIndexMatrix const& gmat,
const common::Range1d range, bst_bin_t split_cond, GHistIndexMatrix const& gmat,
const common::ColumnMatrix& column_matrix, const RegTree& tree, const size_t* rid,
BitVector* decision_bits, BitVector* missing_bits) {
common::Span<const size_t> rid_span(rid + range.begin(), rid + range.end());
@ -204,7 +218,7 @@ class PartitionBuilder {
for (auto row_id : rid_span) {
auto gidx = gmat.GetGindex(row_id, fid);
if (gidx > -1) {
bool go_left = false;
bool go_left;
if (is_cat) {
go_left = Decision(node_cats, cut_values[gidx]);
} else {
@ -218,7 +232,27 @@ class PartitionBuilder {
}
}
} else {
LOG(FATAL) << "Column data split is only supported for the `approx` tree method";
auto pred_hist = [&](auto ridx, auto bin_id) {
if (any_cat && is_cat) {
auto gidx = gmat.GetGindex(ridx, fid);
CHECK_GT(gidx, -1);
return Decision(node_cats, cut_values[gidx]);
} else {
return bin_id <= split_cond;
}
};
if (column_matrix.GetColumnType(fid) == xgboost::common::kDenseColumn) {
auto column = column_matrix.DenseColumn<BinIdxType, any_missing>(fid);
MaskKernel<any_missing>(&column, rid_span, gmat.base_rowid, decision_bits, missing_bits,
pred_hist);
} else {
CHECK_EQ(any_missing, true);
auto column =
column_matrix.SparseColumn<BinIdxType>(fid, rid_span.front() - gmat.base_rowid);
MaskKernel<any_missing>(&column, rid_span, gmat.base_rowid, decision_bits, missing_bits,
pred_hist);
}
}
}
@ -238,7 +272,7 @@ class PartitionBuilder {
std::size_t nid = nodes[node_in_set].nid;
bool default_left = tree[nid].DefaultLeft();
auto pred_approx = [&](auto ridx) {
auto pred = [&](auto ridx) {
bool go_left = default_left;
bool is_missing = missing_bits.Check(ridx - gmat.base_rowid);
if (!is_missing) {
@ -248,11 +282,7 @@ class PartitionBuilder {
};
std::pair<size_t, size_t> child_nodes_sizes;
if (!column_matrix.IsInitialized()) {
child_nodes_sizes = PartitionRangeKernel(rid_span, left, right, pred_approx);
} else {
LOG(FATAL) << "Column data split is only supported for the `approx` tree method";
}
child_nodes_sizes = PartitionRangeKernel(rid_span, left, right, pred);
const size_t n_left = child_nodes_sizes.first;
const size_t n_right = child_nodes_sizes.second;

15
src/tree/common_row_partitioner.h
View File

@ -38,19 +38,21 @@ class ColumnSplitHelper {
missing_bits_ = BitVector(common::Span<BitVector::value_type>(missing_storage_));
}
template <typename ExpandEntry>
template <typename BinIdxType, bool any_missing, bool any_cat, typename ExpandEntry>
void Partition(common::BlockedSpace2d const& space, std::int32_t n_threads,
GHistIndexMatrix const& gmat, common::ColumnMatrix const& column_matrix,
std::vector<ExpandEntry> const& nodes, RegTree const* p_tree) {
std::vector<ExpandEntry> const& nodes,
std::vector<int32_t> const& split_conditions, RegTree const* p_tree) {
// When data is split by column, we don't have all the feature values in the local worker, so
// we first collect all the decisions and whether the feature is missing into bit vectors.
std::fill(decision_storage_.begin(), decision_storage_.end(), 0);
std::fill(missing_storage_.begin(), missing_storage_.end(), 0);
common::ParallelFor2d(space, n_threads, [&](size_t node_in_set, common::Range1d r) {
const int32_t nid = nodes[node_in_set].nid;
partition_builder_->MaskRows(node_in_set, nodes, r, gmat, column_matrix, *p_tree,
(*row_set_collection_)[nid].begin, &decision_bits_,
&missing_bits_);
bst_bin_t split_cond = column_matrix.IsInitialized() ? split_conditions[node_in_set] : 0;
partition_builder_->MaskRows<BinIdxType, any_missing, any_cat>(
node_in_set, nodes, r, split_cond, gmat, column_matrix, *p_tree,
(*row_set_collection_)[nid].begin, &decision_bits_, &missing_bits_);
});
// Then aggregate the bit vectors across all the workers.
@ -217,7 +219,8 @@ class CommonRowPartitioner {
// 2.3 Split elements of row_set_collection_ to left and right child-nodes for each node
// Store results in intermediate buffers from partition_builder_
if (is_col_split_) {
column_split_helper_.Partition(space, ctx->Threads(), gmat, column_matrix, nodes, p_tree);
column_split_helper_.Partition<BinIdxType, any_missing, any_cat>(
space, ctx->Threads(), gmat, column_matrix, nodes, split_conditions, p_tree);
} else {
common::ParallelFor2d(space, ctx->Threads(), [&](size_t node_in_set, common::Range1d r) {
size_t begin = r.begin();

111
tests/cpp/tree/test_quantile_hist.cc
View File

@ -19,6 +19,8 @@
#include "xgboost/data.h"
namespace xgboost::tree {
namespace {
template <typename ExpandEntry>
void TestPartitioner(bst_target_t n_targets) {
std::size_t n_samples = 1024, base_rowid = 0;
@ -86,8 +88,117 @@ void TestPartitioner(bst_target_t n_targets) {
}
}
}
} // anonymous namespace
TEST(QuantileHist, Partitioner) { TestPartitioner<CPUExpandEntry>(1); }
TEST(QuantileHist, MultiPartitioner) { TestPartitioner<MultiExpandEntry>(3); }
namespace {
template <typename ExpandEntry>
void VerifyColumnSplitPartitioner(bst_target_t n_targets, size_t n_samples,
bst_feature_t n_features, size_t base_rowid,
std::shared_ptr<DMatrix> Xy, float min_value, float mid_value,
CommonRowPartitioner const& expected_mid_partitioner) {
auto dmat =
std::unique_ptr<DMatrix>{Xy->SliceCol(collective::GetWorldSize(), collective::GetRank())};
Context ctx;
ctx.InitAllowUnknown(Args{});
std::vector<ExpandEntry> candidates{{0, 0}};
candidates.front().split.loss_chg = 0.4;
auto cuts = common::SketchOnDMatrix(&ctx, dmat.get(), 64);
for (auto const& page : Xy->GetBatches<SparsePage>()) {
GHistIndexMatrix gmat(page, {}, cuts, 64, true, 0.5, ctx.Threads());
bst_feature_t const split_ind = 0;
common::ColumnMatrix column_indices;
column_indices.InitFromSparse(page, gmat, 0.5, ctx.Threads());
{
RegTree tree{n_targets, n_features};
CommonRowPartitioner partitioner{&ctx, n_samples, base_rowid, true};
if constexpr (std::is_same<ExpandEntry, CPUExpandEntry>::value) {
GetSplit(&tree, min_value, &candidates);
} else {
GetMultiSplitForTest(&tree, min_value, &candidates);
}
partitioner.UpdatePosition<false, true>(&ctx, gmat, column_indices, candidates, &tree);
ASSERT_EQ(partitioner.Size(), 3);
ASSERT_EQ(partitioner[1].Size(), 0);
ASSERT_EQ(partitioner[2].Size(), n_samples);
}
{
RegTree tree{n_targets, n_features};
CommonRowPartitioner partitioner{&ctx, n_samples, base_rowid, true};
if constexpr (std::is_same<ExpandEntry, CPUExpandEntry>::value) {
GetSplit(&tree, mid_value, &candidates);
} else {
GetMultiSplitForTest(&tree, mid_value, &candidates);
}
auto left_nidx = tree.LeftChild(RegTree::kRoot);
partitioner.UpdatePosition<false, true>(&ctx, gmat, column_indices, candidates, &tree);
auto elem = partitioner[left_nidx];
ASSERT_LT(elem.Size(), n_samples);
ASSERT_GT(elem.Size(), 1);
auto expected_elem = expected_mid_partitioner[left_nidx];
ASSERT_EQ(elem.Size(), expected_elem.Size());
for (auto it = elem.begin, eit = expected_elem.begin; it != elem.end; ++it, ++eit) {
ASSERT_EQ(*it, *eit);
}
auto right_nidx = tree.RightChild(RegTree::kRoot);
elem = partitioner[right_nidx];
expected_elem = expected_mid_partitioner[right_nidx];
ASSERT_EQ(elem.Size(), expected_elem.Size());
for (auto it = elem.begin, eit = expected_elem.begin; it != elem.end; ++it, ++eit) {
ASSERT_EQ(*it, *eit);
}
}
}
}
template <typename ExpandEntry>
void TestColumnSplitPartitioner(bst_target_t n_targets) {
std::size_t n_samples = 1024, base_rowid = 0;
bst_feature_t n_features = 16;
auto Xy = RandomDataGenerator{n_samples, n_features, 0}.GenerateDMatrix(true);
std::vector<ExpandEntry> candidates{{0, 0}};
candidates.front().split.loss_chg = 0.4;
Context ctx;
ctx.InitAllowUnknown(Args{});
auto cuts = common::SketchOnDMatrix(&ctx, Xy.get(), 64);
float min_value, mid_value;
CommonRowPartitioner mid_partitioner{&ctx, n_samples, base_rowid, false};
for (auto const& page : Xy->GetBatches<SparsePage>()) {
GHistIndexMatrix gmat(page, {}, cuts, 64, true, 0.5, ctx.Threads());
bst_feature_t const split_ind = 0;
common::ColumnMatrix column_indices;
column_indices.InitFromSparse(page, gmat, 0.5, ctx.Threads());
min_value = gmat.cut.MinValues()[split_ind];
auto ptr = gmat.cut.Ptrs()[split_ind + 1];
mid_value = gmat.cut.Values().at(ptr / 2);
RegTree tree{n_targets, n_features};
if constexpr (std::is_same<ExpandEntry, CPUExpandEntry>::value) {
GetSplit(&tree, mid_value, &candidates);
} else {
GetMultiSplitForTest(&tree, mid_value, &candidates);
}
mid_partitioner.UpdatePosition<false, true>(&ctx, gmat, column_indices, candidates, &tree);
}
auto constexpr kWorkers = 4;
RunWithInMemoryCommunicator(kWorkers, VerifyColumnSplitPartitioner<ExpandEntry>, n_targets,
n_samples, n_features, base_rowid, Xy, min_value, mid_value, mid_partitioner);
}
} // anonymous namespace
TEST(QuantileHist, PartitionerColSplit) { TestColumnSplitPartitioner<CPUExpandEntry>(1); }
TEST(QuantileHist, MultiPartitionerColSplit) { TestColumnSplitPartitioner<MultiExpandEntry>(3); }
} // namespace xgboost::tree