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

Remove feature grouping (#7018)

Browse Source 
Co-authored-by: Kirill Shvets <kirill.shvets@intel.com>
This commit is contained in:
ShvetsKS
2021-06-02 23:35:26 +03:00
committed by GitHub
parent 05db6a6c29
commit 5cdaac00c1
6 changed files with 22 additions and 420 deletions
Download Patch File Download Diff File Expand all files Collapse all files

326
src/common/hist_util.cc
View File

@@ -187,267 +187,6 @@ void GHistIndexMatrix::Init(DMatrix* p_fmat, int max_bins) {
}
}
template <typename BinIdxType>
static size_t GetConflictCount(const std::vector<bool>& mark,
const Column<BinIdxType>& column_input,
size_t max_cnt) {
size_t ret = 0;
if (column_input.GetType() == xgboost::common::kDenseColumn) {
const DenseColumn<BinIdxType>& column
= static_cast<const DenseColumn<BinIdxType>& >(column_input);
for (size_t i = 0; i < column.Size(); ++i) {
if ((!column.IsMissing(i)) && mark[i]) {
++ret;
if (ret > max_cnt) {
return max_cnt + 1;
}
}
}
} else {
const SparseColumn<BinIdxType>& column
= static_cast<const SparseColumn<BinIdxType>& >(column_input);
for (size_t i = 0; i < column.Size(); ++i) {
if (mark[column.GetRowIdx(i)]) {
++ret;
if (ret > max_cnt) {
return max_cnt + 1;
}
}
}
}
return ret;
}
template <typename BinIdxType>
inline void
MarkUsed(std::vector<bool>* p_mark, const Column<BinIdxType>& column_input) {
std::vector<bool>& mark = *p_mark;
if (column_input.GetType() == xgboost::common::kDenseColumn) {
const DenseColumn<BinIdxType>& column
= static_cast<const DenseColumn<BinIdxType>& >(column_input);
for (size_t i = 0; i < column.Size(); ++i) {
if (!column.IsMissing(i)) {
mark[i] = true;
}
}
} else {
const SparseColumn<BinIdxType>& column
= static_cast<const SparseColumn<BinIdxType>& >(column_input);
for (size_t i = 0; i < column.Size(); ++i) {
mark[column.GetRowIdx(i)] = true;
}
}
}
template <typename BinIdxType>
inline void SetGroup(const unsigned fid, const Column<BinIdxType>& column,
const size_t max_conflict_cnt, const std::vector<size_t>& search_groups,
std::vector<size_t>* p_group_conflict_cnt,
std::vector<std::vector<bool>>* p_conflict_marks,
std::vector<std::vector<unsigned>>* p_groups,
std::vector<size_t>* p_group_nnz, const size_t cur_fid_nnz, const size_t nrow) {
bool need_new_group = true;
std::vector<size_t>& group_conflict_cnt = *p_group_conflict_cnt;
std::vector<std::vector<bool>>& conflict_marks = *p_conflict_marks;
std::vector<std::vector<unsigned>>& groups = *p_groups;
std::vector<size_t>& group_nnz = *p_group_nnz;
// examine each candidate group: is it okay to insert fid?
for (auto gid : search_groups) {
const size_t rest_max_cnt = max_conflict_cnt - group_conflict_cnt[gid];
const size_t cnt = GetConflictCount(conflict_marks[gid], column, rest_max_cnt);
if (cnt <= rest_max_cnt) {
need_new_group = false;
groups[gid].push_back(fid);
group_conflict_cnt[gid] += cnt;
group_nnz[gid] += cur_fid_nnz - cnt;
MarkUsed(&conflict_marks[gid], column);
break;
}
}
// create new group if necessary
if (need_new_group) {
groups.emplace_back();
groups.back().push_back(fid);
group_conflict_cnt.push_back(0);
conflict_marks.emplace_back(nrow, false);
MarkUsed(&conflict_marks.back(), column);
group_nnz.emplace_back(cur_fid_nnz);
}
}
inline std::vector<std::vector<unsigned>>
FindGroups(const std::vector<unsigned>& feature_list,
const std::vector<size_t>& feature_nnz,
const ColumnMatrix& colmat,
size_t nrow,
const tree::TrainParam& param) {
/* Goal: Bundle features together that has little or no "overlap", i.e.
only a few data points should have nonzero values for
member features.
Note that one-hot encoded features will be grouped together. */
std::vector<std::vector<unsigned>> groups;
std::vector<std::vector<bool>> conflict_marks;
std::vector<size_t> group_nnz;
std::vector<size_t> group_conflict_cnt;
const auto max_conflict_cnt
= static_cast<size_t>(param.max_conflict_rate * nrow);
for (auto fid : feature_list) {
const size_t cur_fid_nnz = feature_nnz[fid];
// randomly choose some of existing groups as candidates
std::vector<size_t> search_groups;
for (size_t gid = 0; gid < groups.size(); ++gid) {
if (group_nnz[gid] + cur_fid_nnz <= nrow + max_conflict_cnt) {
search_groups.push_back(gid);
}
}
std::shuffle(search_groups.begin(), search_groups.end(), common::GlobalRandom());
if (param.max_search_group > 0 && search_groups.size() > param.max_search_group) {
search_groups.resize(param.max_search_group);
}
BinTypeSize bins_type_size = colmat.GetTypeSize();
if (bins_type_size == kUint8BinsTypeSize) {
const auto column = colmat.GetColumn<uint8_t>(fid);
SetGroup(fid, *(column.get()), max_conflict_cnt, search_groups,
&group_conflict_cnt, &conflict_marks, &groups, &group_nnz, cur_fid_nnz, nrow);
} else if (bins_type_size == kUint16BinsTypeSize) {
const auto column = colmat.GetColumn<uint16_t>(fid);
SetGroup(fid, *(column.get()), max_conflict_cnt, search_groups,
&group_conflict_cnt, &conflict_marks, &groups, &group_nnz, cur_fid_nnz, nrow);
} else {
CHECK_EQ(bins_type_size, kUint32BinsTypeSize);
const auto column = colmat.GetColumn<uint32_t>(fid);
SetGroup(fid, *(column.get()), max_conflict_cnt, search_groups,
&group_conflict_cnt, &conflict_marks, &groups, &group_nnz, cur_fid_nnz, nrow);
}
}
return groups;
}
inline std::vector<std::vector<unsigned>>
FastFeatureGrouping(const GHistIndexMatrix& gmat,
const ColumnMatrix& colmat,
const tree::TrainParam& param) {
const size_t nrow = gmat.row_ptr.size() - 1;
const size_t nfeature = gmat.cut.Ptrs().size() - 1;
std::vector<unsigned> feature_list(nfeature);
std::iota(feature_list.begin(), feature_list.end(), 0);
// sort features by nonzero counts, descending order
std::vector<size_t> feature_nnz(nfeature);
std::vector<unsigned> features_by_nnz(feature_list);
gmat.GetFeatureCounts(&feature_nnz[0]);
std::sort(features_by_nnz.begin(), features_by_nnz.end(),
[&feature_nnz](unsigned a, unsigned b) {
return feature_nnz[a] > feature_nnz[b];
});
auto groups_alt1 = FindGroups(feature_list, feature_nnz, colmat, nrow, param);
auto groups_alt2 = FindGroups(features_by_nnz, feature_nnz, colmat, nrow, param);
auto& groups = (groups_alt1.size() > groups_alt2.size()) ? groups_alt2 : groups_alt1;
// take apart small, sparse groups, as it won't help speed
{
std::vector<std::vector<unsigned>> ret;
for (const auto& group : groups) {
if (group.size() <= 1 || group.size() >= 5) {
ret.push_back(group); // keep singleton groups and large (5+) groups
} else {
size_t nnz = 0;
for (auto fid : group) {
nnz += feature_nnz[fid];
}
double nnz_rate = static_cast<double>(nnz) / nrow;
// take apart small sparse group, due it will not gain on speed
if (nnz_rate <= param.sparse_threshold) {
for (auto fid : group) {
ret.emplace_back();
ret.back().push_back(fid);
}
} else {
ret.push_back(group);
}
}
}
groups = std::move(ret);
}
// shuffle groups
std::shuffle(groups.begin(), groups.end(), common::GlobalRandom());
return groups;
}
void GHistIndexBlockMatrix::Init(const GHistIndexMatrix& gmat,
const ColumnMatrix& colmat,
const tree::TrainParam& param) {
cut_ = &gmat.cut;
const size_t nrow = gmat.row_ptr.size() - 1;
const uint32_t nbins = gmat.cut.Ptrs().back();
/* step 1: form feature groups */
auto groups = FastFeatureGrouping(gmat, colmat, param);
const auto nblock = static_cast<uint32_t>(groups.size());
/* step 2: build a new CSR matrix for each feature group */
std::vector<uint32_t> bin2block(nbins); // lookup table [bin id] => [block id]
for (uint32_t group_id = 0; group_id < nblock; ++group_id) {
for (auto& fid : groups[group_id]) {
const uint32_t bin_begin = gmat.cut.Ptrs()[fid];
const uint32_t bin_end = gmat.cut.Ptrs()[fid + 1];
for (uint32_t bin_id = bin_begin; bin_id < bin_end; ++bin_id) {
bin2block[bin_id] = group_id;
}
}
}
std::vector<std::vector<uint32_t>> index_temp(nblock);
std::vector<std::vector<size_t>> row_ptr_temp(nblock);
for (uint32_t block_id = 0; block_id < nblock; ++block_id) {
row_ptr_temp[block_id].push_back(0);
}
for (size_t rid = 0; rid < nrow; ++rid) {
const size_t ibegin = gmat.row_ptr[rid];
const size_t iend = gmat.row_ptr[rid + 1];
for (size_t j = ibegin; j < iend; ++j) {
const uint32_t bin_id = gmat.index[j];
const uint32_t block_id = bin2block[bin_id];
index_temp[block_id].push_back(bin_id);
}
for (uint32_t block_id = 0; block_id < nblock; ++block_id) {
row_ptr_temp[block_id].push_back(index_temp[block_id].size());
}
}
/* step 3: concatenate CSR matrices into one (index, row_ptr) pair */
std::vector<size_t> index_blk_ptr;
std::vector<size_t> row_ptr_blk_ptr;
index_blk_ptr.push_back(0);
row_ptr_blk_ptr.push_back(0);
for (uint32_t block_id = 0; block_id < nblock; ++block_id) {
index_.insert(index_.end(), index_temp[block_id].begin(), index_temp[block_id].end());
row_ptr_.insert(row_ptr_.end(), row_ptr_temp[block_id].begin(), row_ptr_temp[block_id].end());
index_blk_ptr.push_back(index_.size());
row_ptr_blk_ptr.push_back(row_ptr_.size());
}
// save shortcut for each block
for (uint32_t block_id = 0; block_id < nblock; ++block_id) {
Block blk;
blk.index_begin = &index_[index_blk_ptr[block_id]];
blk.row_ptr_begin = &row_ptr_[row_ptr_blk_ptr[block_id]];
blk.index_end = &index_[index_blk_ptr[block_id + 1]];
blk.row_ptr_end = &row_ptr_[row_ptr_blk_ptr[block_id + 1]];
blocks_.push_back(blk);
}
}
/*!
* \brief fill a histogram by zeros in range [begin, end)
*/
@@ -703,71 +442,6 @@ void GHistBuilder<double>::BuildHist(const std::vector<GradientPair>& gpair,
GHistRow<double> hist,
bool isDense);
template<typename GradientSumT>
void GHistBuilder<GradientSumT>::BuildBlockHist(const std::vector<GradientPair>& gpair,
const RowSetCollection::Elem row_indices,
const GHistIndexBlockMatrix& gmatb,
GHistRowT hist) {
static constexpr int kUnroll = 8; // loop unrolling factor
const size_t nblock = gmatb.GetNumBlock();
const size_t nrows = row_indices.end - row_indices.begin;
const size_t rest = nrows % kUnroll;
#if defined(_OPENMP)
const auto nthread = static_cast<bst_omp_uint>(this->nthread_); // NOLINT
#endif // defined(_OPENMP)
xgboost::detail::GradientPairInternal<GradientSumT>* p_hist = hist.data();
dmlc::OMPException exc;
#pragma omp parallel for num_threads(nthread) schedule(guided)
for (bst_omp_uint bid = 0; bid < nblock; ++bid) {
exc.Run([&]() {
auto gmat = gmatb[bid];
for (size_t i = 0; i < nrows - rest; i += kUnroll) {
size_t rid[kUnroll];
size_t ibegin[kUnroll];
size_t iend[kUnroll];
GradientPair stat[kUnroll];
for (int k = 0; k < kUnroll; ++k) {
rid[k] = row_indices.begin[i + k];
ibegin[k] = gmat.row_ptr[rid[k]];
iend[k] = gmat.row_ptr[rid[k] + 1];
stat[k] = gpair[rid[k]];
}
for (int k = 0; k < kUnroll; ++k) {
for (size_t j = ibegin[k]; j < iend[k]; ++j) {
const uint32_t bin = gmat.index[j];
p_hist[bin].Add(stat[k].GetGrad(), stat[k].GetHess());
}
}
}
for (size_t i = nrows - rest; i < nrows; ++i) {
const size_t rid = row_indices.begin[i];
const size_t ibegin = gmat.row_ptr[rid];
const size_t iend = gmat.row_ptr[rid + 1];
const GradientPair stat = gpair[rid];
for (size_t j = ibegin; j < iend; ++j) {
const uint32_t bin = gmat.index[j];
p_hist[bin].Add(stat.GetGrad(), stat.GetHess());
}
}
});
}
exc.Rethrow();
}
template
void GHistBuilder<float>::BuildBlockHist(const std::vector<GradientPair>& gpair,
const RowSetCollection::Elem row_indices,
const GHistIndexBlockMatrix& gmatb,
GHistRow<float> hist);
template
void GHistBuilder<double>::BuildBlockHist(const std::vector<GradientPair>& gpair,
const RowSetCollection::Elem row_indices,
const GHistIndexBlockMatrix& gmatb,
GHistRow<double> hist);
template<typename GradientSumT>
void GHistBuilder<GradientSumT>::SubtractionTrick(GHistRowT self,
GHistRowT sibling,

45
src/common/hist_util.h
View File

@@ -321,48 +321,8 @@ int32_t XGBOOST_HOST_DEV_INLINE BinarySearchBin(bst_uint begin, bst_uint end,
return -1;
}
struct GHistIndexBlock {
const size_t* row_ptr;
const uint32_t* index;
inline GHistIndexBlock(const size_t* row_ptr, const uint32_t* index)
: row_ptr(row_ptr), index(index) {}
// get i-th row
inline GHistIndexRow operator[](size_t i) const {
return {&index[0] + row_ptr[i], row_ptr[i + 1] - row_ptr[i]};
}
};
class ColumnMatrix;
class GHistIndexBlockMatrix {
public:
void Init(const GHistIndexMatrix& gmat,
const ColumnMatrix& colmat,
const tree::TrainParam& param);
inline GHistIndexBlock operator[](size_t i) const {
return {blocks_[i].row_ptr_begin, blocks_[i].index_begin};
}
inline size_t GetNumBlock() const {
return blocks_.size();
}
private:
std::vector<size_t> row_ptr_;
std::vector<uint32_t> index_;
const HistogramCuts* cut_;
struct Block {
const size_t* row_ptr_begin;
const size_t* row_ptr_end;
const uint32_t* index_begin;
const uint32_t* index_end;
};
std::vector<Block> blocks_;
};
template<typename GradientSumT>
using GHistRow = Span<xgboost::detail::GradientPairInternal<GradientSumT> >;
@@ -672,11 +632,6 @@ class GHistBuilder {
const GHistIndexMatrix& gmat,
GHistRowT hist,
bool isDense);
// same, with feature grouping
void BuildBlockHist(const std::vector<GradientPair>& gpair,
const RowSetCollection::Elem row_indices,
const GHistIndexBlockMatrix& gmatb,
GHistRowT hist);
// construct a histogram via subtraction trick
void SubtractionTrick(GHistRowT self,
GHistRowT sibling,