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

now support distributed evaluation

Browse Source
This commit is contained in:
tqchen 2014-12-29 19:24:08 -08:00
parent c395c5bed3
commit 5ad100b5a3
6 changed files with 59 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
multi-node/row-split/machine-row-rabit-mock.sh
View File

@ -17,4 +17,4 @@ cd -
python splitrows.py ../../demo/regression/machine.txt.train train-machine $k
# run xgboost mpi
../../rabit/tracker/rabit_mpi.py $k local ../../rabit/test/keepalive.sh ../../xgboost-mock machine-row.conf dsplit=row num_round=3 mock=1,1,1,0
../../rabit/tracker/rabit_mpi.py -n $k ../../rabit/test/keepalive.sh ../../xgboost-mock machine-row.conf dsplit=row num_round=3 mock=1,1,1,0

6
multi-node/row-split/machine-row-rabit.sh
View File

@ -17,8 +17,8 @@ cd -
python splitrows.py ../../demo/regression/machine.txt.train train-machine $k
# run xgboost mpi
../../rabit/tracker/rabit_mpi.py $k local ../../xgboost machine-row.conf dsplit=row num_round=3
../../rabit/tracker/rabit_mpi.py -n $k ../../xgboost machine-row.conf dsplit=row num_round=3 eval_train=1
# run xgboost-mpi save model 0001, continue to run from existing model
../../rabit/tracker/rabit_mpi.py $k local ../../xgboost machine-row.conf dsplit=row num_round=1
../../rabit/tracker/rabit_mpi.py $k local ../../xgboost machine-row.conf dsplit=row num_round=2 model_in=0001.model
../../rabit/tracker/rabit_mpi.py -n $k ../../xgboost machine-row.conf dsplit=row num_round=1
../../rabit/tracker/rabit_mpi.py -n $k ../../xgboost machine-row.conf dsplit=row num_round=2 model_in=0001.model

44
src/learner/evaluation-inl.hpp
View File

@ -11,6 +11,8 @@
#include <cmath>
#include <climits>
#include <algorithm>
// rabit library for synchronization
#include <rabit.h>
#include "./evaluation.h"
#include "./helper_utils.h"
@ -23,7 +25,8 @@ namespace learner {
template<typename Derived>
struct EvalEWiseBase : public IEvaluator {
virtual float Eval(const std::vector<float> &preds,
const MetaInfo &info) const {
const MetaInfo &info,
bool distributed) const {
utils::Check(info.labels.size() != 0, "label set cannot be empty");
utils::Check(preds.size() % info.labels.size() == 0,
"label and prediction size not match");
@ -37,7 +40,11 @@ struct EvalEWiseBase : public IEvaluator {
sum += Derived::EvalRow(info.labels[i], preds[i]) * wt;
wsum += wt;
}
return Derived::GetFinal(sum, wsum);
float dat[2]; dat[0] = sum, dat[1] = wsum;
if (distributed) {
rabit::Allreduce<rabit::op::Sum>(dat, 2);
}
return Derived::GetFinal(dat[0], dat[1]);
}
/*!
* \brief to be implemented by subclass,
@ -113,7 +120,9 @@ struct EvalCTest: public IEvaluator {
return name_.c_str();
}
virtual float Eval(const std::vector<float> &preds,
const MetaInfo &info) const {
const MetaInfo &info,
bool distributed) const {
utils::Check(!distributed, "metric %s do not support distributed evaluation", name_.c_str());
utils::Check(preds.size() % info.labels.size() == 0,
"label and prediction size not match");
size_t ngroup = preds.size() / info.labels.size() - 1;
@ -150,7 +159,9 @@ struct EvalAMS : public IEvaluator {
utils::Check(std::sscanf(name, "ams@%f", &ratio_) == 1, "invalid ams format");
}
virtual float Eval(const std::vector<float> &preds,
const MetaInfo &info) const {
const MetaInfo &info,
bool distributed) const {
utils::Check(!distributed, "metric AMS do not support distributed evaluation");
using namespace std;
const bst_omp_uint ndata = static_cast<bst_omp_uint>(info.labels.size());
@ -212,7 +223,9 @@ struct EvalPrecisionRatio : public IEvaluator{
}
}
virtual float Eval(const std::vector<float> &preds,
const MetaInfo &info) const {
const MetaInfo &info,
bool distributed) const {
utils::Check(!distributed, "metric %s do not support distributed evaluation", Name());
utils::Check(info.labels.size() != 0, "label set cannot be empty");
utils::Assert(preds.size() % info.labels.size() == 0,
"label size predict size not match");
@ -252,7 +265,8 @@ struct EvalPrecisionRatio : public IEvaluator{
/*! \brief Area under curve, for both classification and rank */
struct EvalAuc : public IEvaluator {
virtual float Eval(const std::vector<float> &preds,
const MetaInfo &info) const {
const MetaInfo &info,
bool distributed) const {
utils::Check(info.labels.size() != 0, "label set cannot be empty");
utils::Check(preds.size() % info.labels.size() == 0,
"label size predict size not match");
@ -299,9 +313,15 @@ struct EvalAuc : public IEvaluator {
sum_auc += sum_pospair / (sum_npos*sum_nneg);
}
}
// return average AUC over list
if (distributed) {
float dat[2]; dat[0] = sum_auc; dat[1] = ngroup;
// approximately estimate auc using mean
rabit::Allreduce<rabit::op::Sum>(dat, 2);
return dat[0] / dat[1];
} else {
return static_cast<float>(sum_auc) / ngroup;
}
}
virtual const char *Name(void) const {
return "auc";
}
@ -311,7 +331,8 @@ struct EvalAuc : public IEvaluator {
struct EvalRankList : public IEvaluator {
public:
virtual float Eval(const std::vector<float> &preds,
const MetaInfo &info) const {
const MetaInfo &info,
bool distributed) const {
utils::Check(preds.size() == info.labels.size(),
"label size predict size not match");
// quick consistency when group is not available
@ -336,8 +357,15 @@ struct EvalRankList : public IEvaluator {
sum_metric += this->EvalMetric(rec);
}
}
if (distributed) {
float dat[2]; dat[0] = sum_metric; dat[1] = ngroup;
// approximately estimate auc using mean
rabit::Allreduce<rabit::op::Sum>(dat, 2);
return dat[0] / dat[1];
} else {
return static_cast<float>(sum_metric) / ngroup;
}
}
virtual const char *Name(void) const {
return name_.c_str();
}

11
src/learner/evaluation.h
View File

@ -19,9 +19,13 @@ struct IEvaluator{
* \brief evaluate a specific metric
* \param preds prediction
* \param info information, including label etc.
* \param distributed whether a call to Allreduce is needed to gather
* the average statistics across all the node,
* this is only supported by some metrics
*/
virtual float Eval(const std::vector<float> &preds,
const MetaInfo &info) const = 0;
const MetaInfo &info,
bool distributed = false) const = 0;
/*! \return name of metric */
virtual const char *Name(void) const = 0;
/*! \brief virtual destructor */
@ -70,10 +74,11 @@ class EvalSet{
}
inline std::string Eval(const char *evname,
const std::vector<float> &preds,
const MetaInfo &info) const {
const MetaInfo &info,
bool distributed = false) {
std::string result = "";
for (size_t i = 0; i < evals_.size(); ++i) {
float res = evals_[i]->Eval(preds, info);
float res = evals_[i]->Eval(preds, info, distributed);
char tmp[1024];
utils::SPrintf(tmp, sizeof(tmp), "\t%s-%s:%f", evname, evals_[i]->Name(), res);
result += tmp;

2
src/learner/learner-inl.hpp
View File

@ -287,7 +287,7 @@ class BoostLearner : public rabit::ISerializable {
for (size_t i = 0; i < evals.size(); ++i) {
this->PredictRaw(*evals[i], &preds_);
obj_->EvalTransform(&preds_);
res += evaluator_.Eval(evname[i].c_str(), preds_, evals[i]->info);
res += evaluator_.Eval(evname[i].c_str(), preds_, evals[i]->info, distributed_mode == 2);
}
return res;
}

6
src/xgboost_main.cpp
View File

@ -181,9 +181,15 @@ class BoostLearnTask {
if (!silent) printf("boosting round %d, %lu sec elapsed\n", i, elapsed);
learner.UpdateOneIter(i, *data);
std::string res = learner.EvalOneIter(i, devalall, eval_data_names);
if (rabit::IsDistributed()){
if (rabit::GetRank() == 0) {
rabit::TrackerPrintf("%s\n", res.c_str());
}
} else {
if (silent < 2) {
fprintf(stderr, "%s\n", res.c_str());
}
}
if (save_period != 0 && (i + 1) % save_period == 0) {
this->SaveModel(i);
}