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

Improve multi-threaded performance (#2104)

Browse Source 
* Add UpdatePredictionCache() option to updaters

Some updaters (e.g. fast_hist) has enough information to quickly compute
prediction cache for the training data. Each updater may override
UpdaterPredictionCache() method to update the prediction cache. Note: this
trick does not apply to validation data.

* Respond to code review

* Disable some debug messages by default
* Document UpdatePredictionCache() interface
* Remove base_margin logic from UpdatePredictionCache() implementation
* Do not take pointer to cfg, as reference may get stale

* Improve multi-threaded performance

* Use columnwise accessor to accelerate ApplySplit() step,
  with support for a compressed representation
* Parallel sort for evaluation step
* Inline BuildHist() function
* Cache gradient pairs when building histograms in BuildHist()

* Add missing #if macro

* Respond to code review

* Use wrapper to enable parallel sort on Linux

* Fix C++ compatibility issues

* MSVC doesn't support unsigned in OpenMP loops
* gcc 4.6 doesn't support using keyword

* Fix lint issues

* Respond to code review

* Fix bug in ApplySplitSparseData()

* Attempting to read beyond the end of a sparse column
* Mishandling the case where an entire range of rows have missing values

* Fix training continuation bug

Disable UpdatePredictionCache() in the first iteration. This way, we can
accomodate the scenario where we build off of an existing (nonempty) ensemble.

* Add regression test for fast_hist

* Respond to code review

* Add back old version of ApplySplitSparseData
This commit is contained in:
Philip Cho
2017-03-25 10:35:01 -07:00
committed by Tianqi Chen
parent 332aea26a3
commit 14fba01b5a
14 changed files with 719 additions and 171 deletions
Download Patch File Download Diff File Expand all files Collapse all files

74
src/metric/rank_metric.cc
View File

@@ -97,44 +97,40 @@ struct EvalAuc : public Metric {
// sum statistics
bst_float sum_auc = 0.0f;
int auc_error = 0;
#pragma omp parallel reduction(+:sum_auc)
{
// each thread takes a local rec
std::vector< std::pair<bst_float, unsigned> > rec;
#pragma omp for schedule(static)
for (bst_omp_uint k = 0; k < ngroup; ++k) {
rec.clear();
for (unsigned j = gptr[k]; j < gptr[k + 1]; ++j) {
rec.push_back(std::make_pair(preds[j], j));
}
std::sort(rec.begin(), rec.end(), common::CmpFirst);
// calculate AUC
double sum_pospair = 0.0;
double sum_npos = 0.0, sum_nneg = 0.0, buf_pos = 0.0, buf_neg = 0.0;
for (size_t j = 0; j < rec.size(); ++j) {
const bst_float wt = info.GetWeight(rec[j].second);
const bst_float ctr = info.labels[rec[j].second];
// keep bucketing predictions in same bucket
if (j != 0 && rec[j].first != rec[j - 1].first) {
sum_pospair += buf_neg * (sum_npos + buf_pos *0.5);
sum_npos += buf_pos;
sum_nneg += buf_neg;
buf_neg = buf_pos = 0.0f;
}
buf_pos += ctr * wt;
buf_neg += (1.0f - ctr) * wt;
}
sum_pospair += buf_neg * (sum_npos + buf_pos *0.5);
sum_npos += buf_pos;
sum_nneg += buf_neg;
// check weird conditions
if (sum_npos <= 0.0 || sum_nneg <= 0.0) {
auc_error = 1;
continue;
}
// this is the AUC
sum_auc += sum_pospair / (sum_npos*sum_nneg);
// each thread takes a local rec
std::vector< std::pair<bst_float, unsigned> > rec;
for (bst_omp_uint k = 0; k < ngroup; ++k) {
rec.clear();
for (unsigned j = gptr[k]; j < gptr[k + 1]; ++j) {
rec.push_back(std::make_pair(preds[j], j));
}
XGBOOST_PARALLEL_SORT(rec.begin(), rec.end(), common::CmpFirst);
// calculate AUC
double sum_pospair = 0.0;
double sum_npos = 0.0, sum_nneg = 0.0, buf_pos = 0.0, buf_neg = 0.0;
for (size_t j = 0; j < rec.size(); ++j) {
const bst_float wt = info.GetWeight(rec[j].second);
const bst_float ctr = info.labels[rec[j].second];
// keep bucketing predictions in same bucket
if (j != 0 && rec[j].first != rec[j - 1].first) {
sum_pospair += buf_neg * (sum_npos + buf_pos *0.5);
sum_npos += buf_pos;
sum_nneg += buf_neg;
buf_neg = buf_pos = 0.0f;
}
buf_pos += ctr * wt;
buf_neg += (1.0f - ctr) * wt;
}
sum_pospair += buf_neg * (sum_npos + buf_pos *0.5);
sum_npos += buf_pos;
sum_nneg += buf_neg;
// check weird conditions
if (sum_npos <= 0.0 || sum_nneg <= 0.0) {
auc_error = 1;
continue;
}
// this is the AUC
sum_auc += sum_pospair / (sum_npos*sum_nneg);
}
CHECK(!auc_error)
<< "AUC: the dataset only contains pos or neg samples";
@@ -262,9 +258,9 @@ struct EvalNDCG : public EvalRankList{
return sumdcg;
}
virtual bst_float EvalMetric(std::vector<std::pair<bst_float, unsigned> > &rec) const { // NOLINT(*)
std::stable_sort(rec.begin(), rec.end(), common::CmpFirst);
XGBOOST_PARALLEL_STABLE_SORT(rec.begin(), rec.end(), common::CmpFirst);
bst_float dcg = this->CalcDCG(rec);
std::stable_sort(rec.begin(), rec.end(), common::CmpSecond);
XGBOOST_PARALLEL_STABLE_SORT(rec.begin(), rec.end(), common::CmpSecond);
bst_float idcg = this->CalcDCG(rec);
if (idcg == 0.0f) {
if (minus_) {