cleanup of evaluation metric, move c++11 codes into sample.h for backup, add lambda in a clean way latter
This commit is contained in:
tqchen
parent
cce96e8f41
commit
439d4725a0
@ -13,8 +13,6 @@
|
||||
#include "../utils/xgboost_omp.h"
|
||||
#include "../utils/xgboost_random.h"
|
||||
#include "xgboost_regrank_data.h"
|
||||
#include <functional>
|
||||
#include <tuple>
|
||||
|
||||
namespace xgboost{
|
||||
namespace regrank{
|
||||
@ -36,6 +34,9 @@ namespace xgboost{
|
||||
inline static bool CmpFirst(const std::pair<float, unsigned> &a, const std::pair<float, unsigned> &b){
|
||||
return a.first > b.first;
|
||||
}
|
||||
inline static bool CmpSecond(const std::pair<float, unsigned> &a, const std::pair<float, unsigned> &b){
|
||||
return a.second > b.second;
|
||||
}
|
||||
|
||||
/*! \brief RMSE */
|
||||
struct EvalRMSE : public IEvaluator{
|
||||
@ -43,7 +44,7 @@ namespace xgboost{
|
||||
const DMatrix::Info &info) const {
|
||||
const unsigned ndata = static_cast<unsigned>(preds.size());
|
||||
float sum = 0.0, wsum = 0.0;
|
||||
#pragma omp parallel for reduction(+:sum,wsum) schedule( static )
|
||||
#pragma omp parallel for reduction(+:sum,wsum) schedule( static )
|
||||
for (unsigned i = 0; i < ndata; ++i){
|
||||
const float wt = info.GetWeight(i);
|
||||
const float diff = info.labels[i] - preds[i];
|
||||
@ -63,7 +64,7 @@ namespace xgboost{
|
||||
const DMatrix::Info &info) const {
|
||||
const unsigned ndata = static_cast<unsigned>(preds.size());
|
||||
float sum = 0.0f, wsum = 0.0f;
|
||||
#pragma omp parallel for reduction(+:sum,wsum) schedule( static )
|
||||
#pragma omp parallel for reduction(+:sum,wsum) schedule( static )
|
||||
for (unsigned i = 0; i < ndata; ++i){
|
||||
const float y = info.labels[i];
|
||||
const float py = preds[i];
|
||||
@ -84,7 +85,7 @@ namespace xgboost{
|
||||
const DMatrix::Info &info) const {
|
||||
const unsigned ndata = static_cast<unsigned>(preds.size());
|
||||
float sum = 0.0f, wsum = 0.0f;
|
||||
#pragma omp parallel for reduction(+:sum,wsum) schedule( static )
|
||||
#pragma omp parallel for reduction(+:sum,wsum) schedule( static )
|
||||
for (unsigned i = 0; i < ndata; ++i){
|
||||
const float wt = info.GetWeight(i);
|
||||
if (preds[i] > 0.5f){
|
||||
@ -112,11 +113,11 @@ namespace xgboost{
|
||||
const unsigned ngroup = static_cast<unsigned>(gptr.size() - 1);
|
||||
|
||||
double sum_auc = 0.0f;
|
||||
#pragma omp parallel reduction(+:sum_auc)
|
||||
#pragma omp parallel reduction(+:sum_auc)
|
||||
{
|
||||
// each thread takes a local rec
|
||||
std::vector< std::pair<float, unsigned> > rec;
|
||||
#pragma omp for schedule(static)
|
||||
#pragma omp for schedule(static)
|
||||
for (unsigned k = 0; k < ngroup; ++k){
|
||||
rec.clear();
|
||||
for (unsigned j = gptr[k]; j < gptr[k + 1]; ++j){
|
||||
@ -153,142 +154,109 @@ namespace xgboost{
|
||||
}
|
||||
};
|
||||
|
||||
/*! \brief Precison at N, for both classification and rank */
|
||||
struct EvalPrecision : public IEvaluator{
|
||||
unsigned topn_;
|
||||
std::string name_;
|
||||
EvalPrecision(const char *name){
|
||||
name_ = name;
|
||||
utils::Assert(sscanf(name, "pre@%u", &topn_));
|
||||
}
|
||||
/*! \brief Evaluate rank list */
|
||||
struct EvalRankList : public IEvaluator{
|
||||
public:
|
||||
virtual float Eval(const std::vector<float> &preds,
|
||||
const DMatrix::Info &info) const {
|
||||
const std::vector<unsigned> &gptr = info.group_ptr;
|
||||
utils::Assert(gptr.size() != 0 && gptr.back() == preds.size(), "EvalAuc: group structure must match number of prediction");
|
||||
const unsigned ngroup = static_cast<unsigned>(gptr.size() - 1);
|
||||
|
||||
double sum_pre = 0.0f;
|
||||
#pragma omp parallel reduction(+:sum_pre)
|
||||
double sum_metric = 0.0f;
|
||||
#pragma omp parallel reduction(+:sum_metric)
|
||||
{
|
||||
// each thread takes a local rec
|
||||
std::vector< std::pair<float, unsigned> > rec;
|
||||
#pragma omp for schedule(static)
|
||||
#pragma omp for schedule(static)
|
||||
for (unsigned 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], (int)info.labels[j]));
|
||||
}
|
||||
std::sort(rec.begin(), rec.end(), CmpFirst);
|
||||
// calculate Preicsion
|
||||
unsigned nhit = 0;
|
||||
for (size_t j = 0; j < rec.size() && j < topn_; ++j){
|
||||
nhit += rec[j].second;
|
||||
}
|
||||
sum_pre += ((float)nhit) / topn_;
|
||||
sum_metric += this->EvalMetric( rec );
|
||||
}
|
||||
}
|
||||
return static_cast<float>(sum_pre) / ngroup;
|
||||
return static_cast<float>(sum_metric) / ngroup;
|
||||
}
|
||||
virtual const char *Name(void) const{
|
||||
return name_.c_str();
|
||||
}
|
||||
protected:
|
||||
EvalRankList(const char *name){
|
||||
name_ = name;
|
||||
if( sscanf(name, "%*[^@]@%u", &topn_) != 1 ){
|
||||
topn_ = UINT_MAX;
|
||||
}
|
||||
}
|
||||
/*! \return evaluation metric, given the pair_sort record, (pred,label) */
|
||||
virtual float EvalMetric( std::vector< std::pair<float, unsigned> > &pair_sort ) const = 0;
|
||||
protected:
|
||||
unsigned topn_;
|
||||
std::string name_;
|
||||
};
|
||||
|
||||
/*! \brief Normalized DCG */
|
||||
class EvalNDCG : public IEvaluator {
|
||||
/*! \brief Precison at N, for both classification and rank */
|
||||
struct EvalPrecision : public EvalRankList{
|
||||
public:
|
||||
virtual float Eval(const std::vector<float> &preds,
|
||||
const DMatrix::Info &info) const{
|
||||
if (info.group_ptr.size() <= 1) return 0;
|
||||
float acc = 0;
|
||||
std::vector< std::pair<float, float> > pairs_sort;
|
||||
for (int i = 0; i < info.group_ptr.size() - 1; i++){
|
||||
for (int j = info.group_ptr[i]; j < info.group_ptr[i + 1]; j++){
|
||||
pairs_sort.push_back(std::make_pair(preds[j], info.labels[j]));
|
||||
EvalPrecision(const char *name):EvalRankList(name){}
|
||||
protected:
|
||||
virtual float EvalMetric( std::vector< std::pair<float, unsigned> > &rec ) const {
|
||||
// calculate Preicsion
|
||||
std::sort(rec.begin(), rec.end(), CmpFirst);
|
||||
unsigned nhit = 0;
|
||||
for (size_t j = 0; j < rec.size() && j < this->topn_; ++j){
|
||||
nhit += (rec[j].second != 0 );
|
||||
}
|
||||
acc += NDCG(pairs_sort);
|
||||
}
|
||||
return acc / (info.group_ptr.size() - 1);
|
||||
}
|
||||
|
||||
static float DCG(const std::vector<float> &labels){
|
||||
float ans = 0.0;
|
||||
for (int i = 0; i < labels.size(); i++){
|
||||
ans += (pow(2, labels[i]) - 1) / log(i + 2);
|
||||
}
|
||||
return ans;
|
||||
}
|
||||
|
||||
virtual const char *Name(void) const {
|
||||
return "NDCG";
|
||||
}
|
||||
|
||||
private:
|
||||
/*\brief Obtain NDCG given the list of labels and predictions
|
||||
* \param pairs_sort the first field is prediction and the second is label
|
||||
*/
|
||||
float NDCG(std::vector< std::pair<float, float> > pairs_sort) const{
|
||||
std::sort(pairs_sort.begin(), pairs_sort.end(), [](std::pair<float, float> a, std::pair<float, float> b){
|
||||
return std::get<0>(a) > std::get<0>(b);
|
||||
});
|
||||
float dcg = DCG(pairs_sort);
|
||||
std::sort(pairs_sort.begin(), pairs_sort.end(), [](std::pair<float, float> a, std::pair<float, float> b){
|
||||
return std::get<1>(a) > std::get<1>(b);
|
||||
});
|
||||
float IDCG = DCG(pairs_sort);
|
||||
if (IDCG == 0) return 0;
|
||||
return dcg / IDCG;
|
||||
}
|
||||
|
||||
float DCG(std::vector< std::pair<float, float> > pairs_sort) const{
|
||||
std::vector<float> labels;
|
||||
for (int i = 1; i < pairs_sort.size(); i++){
|
||||
labels.push_back(std::get<1>(pairs_sort[i]));
|
||||
}
|
||||
return DCG(labels);
|
||||
return static_cast<float>( nhit ) / topn_;
|
||||
}
|
||||
};
|
||||
|
||||
|
||||
/*! \brief Mean Average Precision */
|
||||
class EvalMAP : public IEvaluator {
|
||||
/*! \brief NDCG */
|
||||
struct EvalNDCG : public EvalRankList{
|
||||
public:
|
||||
virtual float Eval(const std::vector<float> &preds,
|
||||
const DMatrix::Info &info) const{
|
||||
if (info.group_ptr.size() <= 1) return 0;
|
||||
float acc = 0;
|
||||
std::vector<std::pair<float,float>> pairs_sort;
|
||||
for (int i = 0; i < info.group_ptr.size() - 1; i++){
|
||||
for (int j = info.group_ptr[i]; j < info.group_ptr[i + 1]; j++){
|
||||
pairs_sort.push_back(std::make_pair(preds[j], info.labels[j]));
|
||||
EvalNDCG(const char *name):EvalRankList(name){}
|
||||
protected:
|
||||
inline float CalcDCG( const std::vector< std::pair<float,unsigned> > &rec ) const {
|
||||
double sumdcg = 0.0;
|
||||
for( size_t i = 0; i < rec.size() && i < this->topn_; i ++ ){
|
||||
const unsigned rel = rec[i].second;
|
||||
if( rel != 0 ){
|
||||
sumdcg += logf( 2.0f ) *((1<<rel)-1) / logf( i + 1 );
|
||||
}
|
||||
acc += average_precision(pairs_sort);
|
||||
}
|
||||
return acc / (info.group_ptr.size() - 1);
|
||||
return static_cast<float>(sumdcg);
|
||||
}
|
||||
virtual float EvalMetric( std::vector< std::pair<float, unsigned> > &rec ) const {
|
||||
std::sort(rec.begin(), rec.end(), CmpFirst);
|
||||
float idcg = this->CalcDCG(rec);
|
||||
std::sort(rec.begin(), rec.end(), CmpSecond);
|
||||
float dcg = this->CalcDCG(rec);
|
||||
if( idcg == 0.0f ) return 0.0f;
|
||||
else return dcg/idcg;
|
||||
}
|
||||
};
|
||||
|
||||
virtual const char *Name(void) const {
|
||||
return "MAP";
|
||||
}
|
||||
|
||||
private:
|
||||
/*\brief Obtain average precision given the list of labels and predictions
|
||||
* \param pairs_sort the first field is prediction and the second is label
|
||||
*/
|
||||
float average_precision(std::vector< std::pair<float,float> > pairs_sort) const{
|
||||
std::sort(pairs_sort.begin(), pairs_sort.end(), [](std::pair<float, float> a, std::pair<float, float> b){
|
||||
return std::get<0>(a) > std::get<0>(b);
|
||||
});
|
||||
float hits = 0;
|
||||
float average_precision = 0;
|
||||
for (int j = 0; j < pairs_sort.size(); j++){
|
||||
if (std::get<1>(pairs_sort[j]) == 1){
|
||||
hits++;
|
||||
average_precision += hits / (j + 1);
|
||||
/*! \brief Precison at N, for both classification and rank */
|
||||
struct EvalMAP : public EvalRankList{
|
||||
public:
|
||||
EvalMAP(const char *name):EvalRankList(name){}
|
||||
protected:
|
||||
virtual float EvalMetric( std::vector< std::pair<float, unsigned> > &rec ) const {
|
||||
std::sort(rec.begin(), rec.end(), CmpFirst);
|
||||
unsigned nhits = 0;
|
||||
double sumap = 0.0;
|
||||
for( size_t i = 0; i < rec.size(); ++i){
|
||||
if( rec[i].second != 0 ){
|
||||
nhits += 1;
|
||||
if( i < this->topn_ ){
|
||||
sumap += static_cast<float>(nhits) / (i+1);
|
||||
}
|
||||
}
|
||||
if (hits != 0) average_precision /= hits;
|
||||
return average_precision;
|
||||
}
|
||||
if (nhits != 0) sumap /= nhits;
|
||||
return static_cast<float>(sumap);
|
||||
}
|
||||
};
|
||||
};
|
||||
@ -306,6 +274,8 @@ namespace xgboost{
|
||||
if (!strcmp(name, "logloss")) evals_.push_back(new EvalLogLoss());
|
||||
if (!strcmp(name, "auc")) evals_.push_back(new EvalAuc());
|
||||
if (!strncmp(name, "pre@", 4)) evals_.push_back(new EvalPrecision(name));
|
||||
if (!strncmp(name, "map", 3)) evals_.push_back(new EvalMAP(name));
|
||||
if (!strncmp(name, "ndcg", 3)) evals_.push_back(new EvalNDCG(name));
|
||||
}
|
||||
~EvalSet(){
|
||||
for (size_t i = 0; i < evals_.size(); ++i){
|
||||
|
||||
@ -5,10 +5,9 @@
|
||||
* \brief implementation of objective functions
|
||||
* \author Tianqi Chen, Kailong Chen
|
||||
*/
|
||||
#include "xgboost_regrank_sample.h"
|
||||
#include <tuple>
|
||||
//#include "xgboost_regrank_sample.h"
|
||||
#include <vector>
|
||||
#include <functional>
|
||||
|
||||
namespace xgboost{
|
||||
namespace regrank{
|
||||
class RegressionObj : public IObjFunction{
|
||||
@ -206,208 +205,5 @@ namespace xgboost{
|
||||
LossType loss;
|
||||
};
|
||||
};
|
||||
|
||||
namespace regrank{
|
||||
// simple pairwise rank
|
||||
class LambdaRankObj : public IObjFunction{
|
||||
public:
|
||||
LambdaRankObj(void){}
|
||||
|
||||
virtual ~LambdaRankObj(){}
|
||||
|
||||
virtual void SetParam(const char *name, const char *val){
|
||||
if (!strcmp("loss_type", name)) loss_.loss_type = atoi(val);
|
||||
if (!strcmp("sampler", name)) sampler_.AssignSampler(atoi(val));
|
||||
if (!strcmp("lambda", name)) lambda_ = atoi(val);
|
||||
}
|
||||
|
||||
virtual void GetGradient(const std::vector<float>& preds,
|
||||
const DMatrix::Info &info,
|
||||
int iter,
|
||||
std::vector<float> &grad,
|
||||
std::vector<float> &hess) {
|
||||
grad.resize(preds.size()); hess.resize(preds.size());
|
||||
const std::vector<unsigned> &group_index = info.group_ptr;
|
||||
utils::Assert(group_index.size() != 0 && group_index.back() == preds.size(), "rank loss must have group file");
|
||||
|
||||
for (int i = 0; i < group_index.size() - 1; i++){
|
||||
sample::Pairs pairs = sampler_.GenPairs(preds, info.labels, group_index[i], group_index[i + 1]);
|
||||
//pairs.GetPairs()
|
||||
std::vector< std::tuple<float, float, int> > sorted_triple = GetSortedTuple(preds, info.labels, group_index, i);
|
||||
std::vector<int> index_remap = GetIndexMap(sorted_triple, group_index[i]);
|
||||
GetGroupGradient(preds, info.labels, group_index,
|
||||
grad, hess, sorted_triple, index_remap, pairs, i);
|
||||
}
|
||||
}
|
||||
|
||||
virtual const char* DefaultEvalMetric(void) {
|
||||
return "auc";
|
||||
}
|
||||
|
||||
private:
|
||||
int lambda_;
|
||||
const static int PAIRWISE = 0;
|
||||
const static int MAP = 1;
|
||||
const static int NDCG = 2;
|
||||
sample::PairSamplerWrapper sampler_;
|
||||
LossType loss_;
|
||||
/* \brief Sorted tuples of a group by the predictions, and
|
||||
* the fields in the return tuples successively are predicions,
|
||||
* labels, and the index of the instance
|
||||
*/
|
||||
inline std::vector< std::tuple<float, float, int> > GetSortedTuple(const std::vector<float> &preds,
|
||||
const std::vector<float> &labels,
|
||||
const std::vector<unsigned> &group_index,
|
||||
int group){
|
||||
std::vector< std::tuple<float, float, int> > sorted_triple;
|
||||
for (int j = group_index[group]; j < group_index[group + 1]; j++){
|
||||
sorted_triple.push_back(std::tuple<float, float, int>(preds[j], labels[j], j));
|
||||
}
|
||||
std::sort(sorted_triple.begin(), sorted_triple.end(),
|
||||
[](std::tuple<float, float, int> a, std::tuple<float, float, int> b){
|
||||
return std::get<0>(a) > std::get<0>(b);
|
||||
});
|
||||
return sorted_triple;
|
||||
}
|
||||
|
||||
inline std::vector<int> GetIndexMap(std::vector< std::tuple<float, float, int> > sorted_triple, int start){
|
||||
std::vector<int> index_remap;
|
||||
index_remap.resize(sorted_triple.size());
|
||||
for (int i = 0; i < sorted_triple.size(); i++){
|
||||
index_remap[std::get<2>(sorted_triple[i]) - start] = i;
|
||||
}
|
||||
return index_remap;
|
||||
}
|
||||
|
||||
inline float GetLambdaMAP(const std::vector< std::tuple<float, float, int> > sorted_triple,
|
||||
int index1, int index2,
|
||||
std::vector< std::tuple<float, float, float, float> > map_acc){
|
||||
if (index1 > index2) std::swap(index1, index2);
|
||||
float original = std::get<0>(map_acc[index2]);
|
||||
if (index1 != 0) original -= std::get<0>(map_acc[index1 - 1]);
|
||||
float changed = 0;
|
||||
if (std::get<1>(sorted_triple[index1]) < std::get<1>(sorted_triple[index2])){
|
||||
changed += std::get<2>(map_acc[index2 - 1]) - std::get<2>(map_acc[index1]);
|
||||
changed += (std::get<3>(map_acc[index1])+ 1.0f) / (index1 + 1);
|
||||
}
|
||||
else{
|
||||
changed += std::get<1>(map_acc[index2 - 1]) - std::get<1>(map_acc[index1]);
|
||||
changed += std::get<3>(map_acc[index2]) / (index2 + 1);
|
||||
}
|
||||
float ans = (changed - original) / (std::get<3>(map_acc[map_acc.size() - 1]));
|
||||
if (ans < 0) ans = -ans;
|
||||
return ans;
|
||||
}
|
||||
|
||||
inline float GetLambdaNDCG(const std::vector< std::tuple<float, float, int> > sorted_triple,
|
||||
int index1,
|
||||
int index2, float IDCG){
|
||||
float original = pow(2, std::get<1>(sorted_triple[index1])) / log(index1 + 2)
|
||||
+ pow(2, std::get<1>(sorted_triple[index2])) / log(index2 + 2);
|
||||
float changed = pow(2, std::get<1>(sorted_triple[index2])) / log(index1 + 2)
|
||||
+ pow(2, std::get<1>(sorted_triple[index1])) / log(index2 + 2);
|
||||
float ans = (original - changed) / IDCG;
|
||||
if (ans < 0) ans = -ans;
|
||||
return ans;
|
||||
}
|
||||
|
||||
|
||||
inline float GetIDCG(const std::vector< std::tuple<float, float, int> > sorted_triple){
|
||||
std::vector<float> labels;
|
||||
for (int i = 0; i < sorted_triple.size(); i++){
|
||||
labels.push_back(std::get<1>(sorted_triple[i]));
|
||||
}
|
||||
|
||||
std::sort(labels.begin(), labels.end(), std::greater<float>());
|
||||
return EvalNDCG::DCG(labels);
|
||||
}
|
||||
|
||||
inline std::vector< std::tuple<float, float, float, float> > GetMAPAcc(const std::vector< std::tuple<float, float, int> > sorted_triple){
|
||||
std::vector< std::tuple<float, float, float, float> > map_acc;
|
||||
float hit = 0, acc1 = 0, acc2 = 0, acc3 = 0;
|
||||
for (int i = 0; i < sorted_triple.size(); i++){
|
||||
if (std::get<1>(sorted_triple[i]) == 1) {
|
||||
hit++;
|
||||
acc1 += hit / (i + 1);
|
||||
acc2 += (hit - 1) / (i + 1);
|
||||
acc3 += (hit + 1) / (i + 1);
|
||||
}
|
||||
map_acc.push_back(std::make_tuple(acc1, acc2, acc3, hit));
|
||||
}
|
||||
return map_acc;
|
||||
|
||||
}
|
||||
|
||||
inline void GetGroupGradient(const std::vector<float> &preds,
|
||||
const std::vector<float> &labels,
|
||||
const std::vector<unsigned> &group_index,
|
||||
std::vector<float> &grad,
|
||||
std::vector<float> &hess,
|
||||
const std::vector< std::tuple<float, float, int> > sorted_triple,
|
||||
const std::vector<int> index_remap,
|
||||
const sample::Pairs& pairs,
|
||||
int group){
|
||||
bool j_better;
|
||||
float IDCG, pred_diff, pred_diff_exp, delta;
|
||||
float first_order_gradient, second_order_gradient;
|
||||
std::vector< std::tuple<float, float, float, float> > map_acc;
|
||||
|
||||
if (lambda_ == NDCG){
|
||||
IDCG = GetIDCG(sorted_triple);
|
||||
}
|
||||
else if (lambda_ == MAP){
|
||||
map_acc = GetMAPAcc(sorted_triple);
|
||||
}
|
||||
|
||||
for (int j = group_index[group]; j < group_index[group + 1]; j++){
|
||||
std::vector<int> pair_instance = pairs.GetPairs(j);
|
||||
for (int k = 0; k < pair_instance.size(); k++){
|
||||
j_better = labels[j] > labels[pair_instance[k]];
|
||||
if (j_better){
|
||||
switch (lambda_){
|
||||
case PAIRWISE: delta = 1.0; break;
|
||||
case MAP: delta = GetLambdaMAP(sorted_triple, index_remap[j - group_index[group]], index_remap[pair_instance[k] - group_index[group]], map_acc); break;
|
||||
case NDCG: delta = GetLambdaNDCG(sorted_triple, index_remap[j - group_index[group]], index_remap[pair_instance[k] - group_index[group]], IDCG); break;
|
||||
default: utils::Error("Cannot find the specified loss type");
|
||||
}
|
||||
|
||||
pred_diff = preds[preds[j] - pair_instance[k]];
|
||||
pred_diff_exp = j_better ? expf(-pred_diff) : expf(pred_diff);
|
||||
first_order_gradient = delta * FirstOrderGradient(pred_diff_exp);
|
||||
second_order_gradient = 2 * delta * SecondOrderGradient(pred_diff_exp);
|
||||
hess[j] += second_order_gradient;
|
||||
grad[j] += first_order_gradient;
|
||||
hess[pair_instance[k]] += second_order_gradient;
|
||||
grad[pair_instance[k]] += -first_order_gradient;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
/*!
|
||||
* \brief calculate first order gradient of pairwise loss function(f(x) = ln(1+exp(-x)),
|
||||
* given the exponential of the difference of intransformed pair predictions
|
||||
* \param the intransformed prediction of positive instance
|
||||
* \param the intransformed prediction of negative instance
|
||||
* \return first order gradient
|
||||
*/
|
||||
inline float FirstOrderGradient(float pred_diff_exp) const {
|
||||
return -pred_diff_exp / (1 + pred_diff_exp);
|
||||
}
|
||||
|
||||
/*!
|
||||
* \brief calculate second order gradient of pairwise loss function(f(x) = ln(1+exp(-x)),
|
||||
* given the exponential of the difference of intransformed pair predictions
|
||||
* \param the intransformed prediction of positive instance
|
||||
* \param the intransformed prediction of negative instance
|
||||
* \return second order gradient
|
||||
*/
|
||||
inline float SecondOrderGradient(float pred_diff_exp) const {
|
||||
return pred_diff_exp / pow(1 + pred_diff_exp, 2);
|
||||
}
|
||||
|
||||
|
||||
};
|
||||
};
|
||||
};
|
||||
#endif
|
||||
|
||||
@ -125,5 +125,191 @@ namespace xgboost {
|
||||
};
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
namespace regrank{
|
||||
// simple pairwise rank
|
||||
class LambdaRankObj : public IObjFunction{
|
||||
public:
|
||||
LambdaRankObj(void){}
|
||||
|
||||
virtual ~LambdaRankObj(){}
|
||||
|
||||
virtual void SetParam(const char *name, const char *val){
|
||||
if (!strcmp("loss_type", name)) loss_.loss_type = atoi(val);
|
||||
if (!strcmp("sampler", name)) sampler_.AssignSampler(atoi(val));
|
||||
if (!strcmp("lambda", name)) lambda_ = atoi(val);
|
||||
}
|
||||
|
||||
virtual void GetGradient(const std::vector<float>& preds,
|
||||
const DMatrix::Info &info,
|
||||
int iter,
|
||||
std::vector<float> &grad,
|
||||
std::vector<float> &hess) {
|
||||
grad.resize(preds.size()); hess.resize(preds.size());
|
||||
const std::vector<unsigned> &group_index = info.group_ptr;
|
||||
utils::Assert(group_index.size() != 0 && group_index.back() == preds.size(), "rank loss must have group file");
|
||||
|
||||
for (int i = 0; i < group_index.size() - 1; i++){
|
||||
sample::Pairs pairs = sampler_.GenPairs(preds, info.labels, group_index[i], group_index[i + 1]);
|
||||
//pairs.GetPairs()
|
||||
std::vector< std::tuple<float, float, int> > sorted_triple = GetSortedTuple(preds, info.labels, group_index, i);
|
||||
std::vector<int> index_remap = GetIndexMap(sorted_triple, group_index[i]);
|
||||
GetGroupGradient(preds, info.labels, group_index,
|
||||
grad, hess, sorted_triple, index_remap, pairs, i);
|
||||
}
|
||||
}
|
||||
|
||||
virtual const char* DefaultEvalMetric(void) {
|
||||
return "auc";
|
||||
}
|
||||
|
||||
private:
|
||||
int lambda_;
|
||||
const static int PAIRWISE = 0;
|
||||
const static int MAP = 1;
|
||||
const static int NDCG = 2;
|
||||
sample::PairSamplerWrapper sampler_;
|
||||
LossType loss_;
|
||||
/* \brief Sorted tuples of a group by the predictions, and
|
||||
* the fields in the return tuples successively are predicions,
|
||||
* labels, and the index of the instance
|
||||
*/
|
||||
inline std::vector< std::tuple<float, float, int> > GetSortedTuple(const std::vector<float> &preds,
|
||||
const std::vector<float> &labels,
|
||||
const std::vector<unsigned> &group_index,
|
||||
int group){
|
||||
std::vector< std::tuple<float, float, int> > sorted_triple;
|
||||
for (int j = group_index[group]; j < group_index[group + 1]; j++){
|
||||
sorted_triple.push_back(std::tuple<float, float, int>(preds[j], labels[j], j));
|
||||
}
|
||||
std::sort(sorted_triple.begin(), sorted_triple.end(),
|
||||
[](std::tuple<float, float, int> a, std::tuple<float, float, int> b){
|
||||
return std::get<0>(a) > std::get<0>(b);
|
||||
});
|
||||
return sorted_triple;
|
||||
}
|
||||
|
||||
inline std::vector<int> GetIndexMap(std::vector< std::tuple<float, float, int> > sorted_triple, int start){
|
||||
std::vector<int> index_remap;
|
||||
index_remap.resize(sorted_triple.size());
|
||||
for (int i = 0; i < sorted_triple.size(); i++){
|
||||
index_remap[std::get<2>(sorted_triple[i]) - start] = i;
|
||||
}
|
||||
return index_remap;
|
||||
}
|
||||
|
||||
inline float GetLambdaMAP(const std::vector< std::tuple<float, float, int> > sorted_triple,
|
||||
int index1, int index2,
|
||||
std::vector< std::tuple<float, float, float, float> > map_acc){
|
||||
if (index1 > index2) std::swap(index1, index2);
|
||||
float original = std::get<0>(map_acc[index2]);
|
||||
if (index1 != 0) original -= std::get<0>(map_acc[index1 - 1]);
|
||||
float changed = 0;
|
||||
if (std::get<1>(sorted_triple[index1]) < std::get<1>(sorted_triple[index2])){
|
||||
changed += std::get<2>(map_acc[index2 - 1]) - std::get<2>(map_acc[index1]);
|
||||
changed += (std::get<3>(map_acc[index1])+ 1.0f) / (index1 + 1);
|
||||
}
|
||||
else{
|
||||
changed += std::get<1>(map_acc[index2 - 1]) - std::get<1>(map_acc[index1]);
|
||||
changed += std::get<3>(map_acc[index2]) / (index2 + 1);
|
||||
}
|
||||
float ans = (changed - original) / (std::get<3>(map_acc[map_acc.size() - 1]));
|
||||
if (ans < 0) ans = -ans;
|
||||
return ans;
|
||||
}
|
||||
|
||||
inline float GetLambdaNDCG(const std::vector< std::tuple<float, float, int> > sorted_triple,
|
||||
int index1,
|
||||
int index2, float IDCG){
|
||||
float original = pow(2, std::get<1>(sorted_triple[index1])) / log(index1 + 2)
|
||||
+ pow(2, std::get<1>(sorted_triple[index2])) / log(index2 + 2);
|
||||
float changed = pow(2, std::get<1>(sorted_triple[index2])) / log(index1 + 2)
|
||||
+ pow(2, std::get<1>(sorted_triple[index1])) / log(index2 + 2);
|
||||
float ans = (original - changed) / IDCG;
|
||||
if (ans < 0) ans = -ans;
|
||||
return ans;
|
||||
}
|
||||
|
||||
|
||||
inline float GetIDCG(const std::vector< std::tuple<float, float, int> > sorted_triple){
|
||||
std::vector<float> labels;
|
||||
for (int i = 0; i < sorted_triple.size(); i++){
|
||||
labels.push_back(std::get<1>(sorted_triple[i]));
|
||||
}
|
||||
|
||||
std::sort(labels.begin(), labels.end(), std::greater<float>());
|
||||
return EvalNDCG::DCG(labels);
|
||||
}
|
||||
|
||||
inline std::vector< std::tuple<float, float, float, float> > GetMAPAcc(const std::vector< std::tuple<float, float, int> > sorted_triple){
|
||||
std::vector< std::tuple<float, float, float, float> > map_acc;
|
||||
float hit = 0, acc1 = 0, acc2 = 0, acc3 = 0;
|
||||
for (int i = 0; i < sorted_triple.size(); i++){
|
||||
if (std::get<1>(sorted_triple[i]) == 1) {
|
||||
hit++;
|
||||
acc1 += hit / (i + 1);
|
||||
acc2 += (hit - 1) / (i + 1);
|
||||
acc3 += (hit + 1) / (i + 1);
|
||||
}
|
||||
map_acc.push_back(std::make_tuple(acc1, acc2, acc3, hit));
|
||||
}
|
||||
return map_acc;
|
||||
|
||||
}
|
||||
|
||||
inline void GetGroupGradient(const std::vector<float> &preds,
|
||||
const std::vector<float> &labels,
|
||||
const std::vector<unsigned> &group_index,
|
||||
std::vector<float> &grad,
|
||||
std::vector<float> &hess,
|
||||
const std::vector< std::tuple<float, float, int> > sorted_triple,
|
||||
const std::vector<int> index_remap,
|
||||
const sample::Pairs& pairs,
|
||||
int group){
|
||||
bool j_better;
|
||||
float IDCG, pred_diff, pred_diff_exp, delta;
|
||||
float first_order_gradient, second_order_gradient;
|
||||
std::vector< std::tuple<float, float, float, float> > map_acc;
|
||||
|
||||
if (lambda_ == NDCG){
|
||||
IDCG = GetIDCG(sorted_triple);
|
||||
}
|
||||
else if (lambda_ == MAP){
|
||||
map_acc = GetMAPAcc(sorted_triple);
|
||||
}
|
||||
|
||||
for (int j = group_index[group]; j < group_index[group + 1]; j++){
|
||||
std::vector<int> pair_instance = pairs.GetPairs(j);
|
||||
for (int k = 0; k < pair_instance.size(); k++){
|
||||
j_better = labels[j] > labels[pair_instance[k]];
|
||||
if (j_better){
|
||||
switch (lambda_){
|
||||
case PAIRWISE: delta = 1.0; break;
|
||||
case MAP: delta = GetLambdaMAP(sorted_triple, index_remap[j - group_index[group]], index_remap[pair_instance[k] - group_index[group]], map_acc); break;
|
||||
case NDCG: delta = GetLambdaNDCG(sorted_triple, index_remap[j - group_index[group]], index_remap[pair_instance[k] - group_index[group]], IDCG); break;
|
||||
default: utils::Error("Cannot find the specified loss type");
|
||||
}
|
||||
|
||||
pred_diff = preds[preds[j] - pair_instance[k]];
|
||||
pred_diff_exp = j_better ? expf(-pred_diff) : expf(pred_diff);
|
||||
first_order_gradient = delta * FirstOrderGradient(pred_diff_exp);
|
||||
second_order_gradient = 2 * delta * SecondOrderGradient(pred_diff_exp);
|
||||
hess[j] += second_order_gradient;
|
||||
grad[j] += first_order_gradient;
|
||||
hess[pair_instance[k]] += second_order_gradient;
|
||||
grad[pair_instance[k]] += -first_order_gradient;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
inline float FirstOrderGradient(float pred_diff_exp) const {
|
||||
return -pred_diff_exp / (1 + pred_diff_exp);
|
||||
}
|
||||
inline float SecondOrderGradient(float pred_diff_exp) const {
|
||||
return pred_diff_exp / pow(1 + pred_diff_exp, 2);
|
||||
}
|
||||
};
|
||||
|
||||
}
|
||||
#endif
|
||||
@ -129,6 +129,7 @@ namespace xgboost{
|
||||
};
|
||||
|
||||
namespace random{
|
||||
/*! \brief random number generator with independent random number seed*/
|
||||
struct Random{
|
||||
/*! \brief set random number seed */
|
||||
inline void Seed( unsigned sd ){
|
||||
@ -136,10 +137,7 @@ namespace xgboost{
|
||||
}
|
||||
/*! \brief return a real number uniform in [0,1) */
|
||||
inline double RandDouble( void ){
|
||||
|
||||
// return static_cast<double>( rand_( &rseed ) ) / (static_cast<double>( RAND_MAX )+1.0);
|
||||
return static_cast<double>(rand()) / (static_cast<double>(RAND_MAX)+1.0);
|
||||
|
||||
return static_cast<double>( rand_r( &rseed ) ) / (static_cast<double>( RAND_MAX )+1.0);
|
||||
}
|
||||
// random number seed
|
||||
unsigned rseed;
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user