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

37
src/learner.cc
View File

@@ -6,6 +6,7 @@
*/
#include <xgboost/logging.h>
#include <xgboost/learner.h>
#include <dmlc/timer.h>
#include <dmlc/io.h>
#include <algorithm>
#include <vector>
@@ -83,6 +84,8 @@ struct LearnerTrainParam
// number of threads to use if OpenMP is enabled
// if equals 0, use system default
int nthread;
// flag to print out detailed breakdown of runtime
int debug_verbose;
// declare parameters
DMLC_DECLARE_PARAMETER(LearnerTrainParam) {
DMLC_DECLARE_FIELD(seed).set_default(0)
@@ -109,6 +112,10 @@ struct LearnerTrainParam
.describe("maximum row per batch.");
DMLC_DECLARE_FIELD(nthread).set_default(0)
.describe("Number of threads to use.");
DMLC_DECLARE_FIELD(debug_verbose)
.set_lower_bound(0)
.set_default(0)
.describe("flag to print out detailed breakdown of runtime");
}
};
@@ -170,28 +177,9 @@ class LearnerImpl : public Learner {
if (tparam.tree_method == 3) {
/* histogram-based algorithm */
if (cfg_.count("updater") == 0) {
LOG(CONSOLE) << "Tree method is selected to be \'hist\', "
<< "which uses histogram aggregation for faster training. "
<< "Using default sequence of updaters: grow_fast_histmaker,prune";
cfg_["updater"] = "grow_fast_histmaker,prune";
} else {
const std::string first_str = "grow_fast_histmaker";
if (first_str.length() <= cfg_["updater"].length()
&& std::equal(first_str.begin(), first_str.end(), cfg_["updater"].begin())) {
// updater sequence starts with "grow_fast_histmaker"
LOG(CONSOLE) << "Tree method is selected to be \'hist\', "
<< "which uses histogram aggregation for faster training. "
<< "Using custom sequence of updaters: " << cfg_["updater"];
} else {
// updater sequence does not start with "grow_fast_histmaker"
LOG(CONSOLE) << "Tree method is selected to be \'hist\', but the given "
<< "sequence of updaters is not compatible; "
<< "grow_fast_histmaker must run first. "
<< "Using default sequence of updaters: grow_fast_histmaker,prune";
cfg_["updater"] = "grow_fast_histmaker,prune";
}
}
LOG(CONSOLE) << "Tree method is selected to be \'hist\', which uses a single updater "
<< "grow_fast_histmaker.";
cfg_["updater"] = "grow_fast_histmaker";
} else if (cfg_.count("updater") == 0) {
if (tparam.dsplit == 1) {
cfg_["updater"] = "distcol";
@@ -333,6 +321,7 @@ class LearnerImpl : public Learner {
std::string EvalOneIter(int iter,
const std::vector<DMatrix*>& data_sets,
const std::vector<std::string>& data_names) override {
double tstart = dmlc::GetTime();
std::ostringstream os;
os << '[' << iter << ']'
<< std::setiosflags(std::ios::fixed);
@@ -347,6 +336,10 @@ class LearnerImpl : public Learner {
<< ev->Eval(preds_, data_sets[i]->info(), tparam.dsplit == 2);
}
}
if (tparam.debug_verbose > 0) {
LOG(INFO) << "EvalOneIter(): " << dmlc::GetTime() - tstart << " sec";
}
return os.str();
}