xgboost/src/learner/objective-inl.hpp
2015-05-04 10:48:25 -07:00

641 lines
23 KiB
C++

#ifndef XGBOOST_LEARNER_OBJECTIVE_INL_HPP_
#define XGBOOST_LEARNER_OBJECTIVE_INL_HPP_
/*!
* \file objective-inl.hpp
* \brief objective function implementations
* \author Tianqi Chen, Kailong Chen
*/
#include <vector>
#include <algorithm>
#include <utility>
#include <cmath>
#include <functional>
#include "../data.h"
#include "./objective.h"
#include "./helper_utils.h"
#include "../utils/random.h"
#include "../utils/omp.h"
namespace xgboost {
namespace learner {
/*! \brief defines functions to calculate some commonly used functions */
struct LossType {
/*! \brief indicate which type we are using */
int loss_type;
// list of constants
static const int kLinearSquare = 0;
static const int kLogisticNeglik = 1;
static const int kLogisticClassify = 2;
static const int kLogisticRaw = 3;
/*!
* \brief transform the linear sum to prediction
* \param x linear sum of boosting ensemble
* \return transformed prediction
*/
inline float PredTransform(float x) const {
switch (loss_type) {
case kLogisticRaw:
case kLinearSquare: return x;
case kLogisticClassify:
case kLogisticNeglik: return 1.0f / (1.0f + std::exp(-x));
default: utils::Error("unknown loss_type"); return 0.0f;
}
}
/*!
* \brief check if label range is valid
*/
inline bool CheckLabel(float x) const {
if (loss_type != kLinearSquare) {
return x >= 0.0f && x <= 1.0f;
}
return true;
}
/*!
* \brief error message displayed when check label fail
*/
inline const char * CheckLabelErrorMsg(void) const {
if (loss_type != kLinearSquare) {
return "label must be in [0,1] for logistic regression";
} else {
return "";
}
}
/*!
* \brief calculate first order gradient of loss, given transformed prediction
* \param predt transformed prediction
* \param label true label
* \return first order gradient
*/
inline float FirstOrderGradient(float predt, float label) const {
switch (loss_type) {
case kLinearSquare: return predt - label;
case kLogisticRaw: predt = 1.0f / (1.0f + std::exp(-predt));
case kLogisticClassify:
case kLogisticNeglik: return predt - label;
default: utils::Error("unknown loss_type"); return 0.0f;
}
}
/*!
* \brief calculate second order gradient of loss, given transformed prediction
* \param predt transformed prediction
* \param label true label
* \return second order gradient
*/
inline float SecondOrderGradient(float predt, float label) const {
// cap second order gradient to postive value
const float eps = 1e-16f;
switch (loss_type) {
case kLinearSquare: return 1.0f;
case kLogisticRaw: predt = 1.0f / (1.0f + std::exp(-predt));
case kLogisticClassify:
case kLogisticNeglik: return std::max(predt * (1.0f - predt), eps);
default: utils::Error("unknown loss_type"); return 0.0f;
}
}
/*!
* \brief transform probability value back to margin
*/
inline float ProbToMargin(float base_score) const {
if (loss_type == kLogisticRaw ||
loss_type == kLogisticClassify ||
loss_type == kLogisticNeglik ) {
utils::Check(base_score > 0.0f && base_score < 1.0f,
"base_score must be in (0,1) for logistic loss");
base_score = -std::log(1.0f / base_score - 1.0f);
}
return base_score;
}
/*! \brief get default evaluation metric for the objective */
inline const char *DefaultEvalMetric(void) const {
if (loss_type == kLogisticClassify) return "error";
if (loss_type == kLogisticRaw) return "auc";
return "rmse";
}
};
/*! \brief objective function that only need to */
class RegLossObj : public IObjFunction {
public:
explicit RegLossObj(int loss_type) {
loss.loss_type = loss_type;
scale_pos_weight = 1.0f;
}
virtual ~RegLossObj(void) {}
virtual void SetParam(const char *name, const char *val) {
using namespace std;
if (!strcmp("scale_pos_weight", name)) {
scale_pos_weight = static_cast<float>(atof(val));
}
}
virtual void GetGradient(const std::vector<float> &preds,
const MetaInfo &info,
int iter,
std::vector<bst_gpair> *out_gpair) {
utils::Check(info.labels.size() != 0, "label set cannot be empty");
utils::Check(preds.size() % info.labels.size() == 0,
"labels are not correctly provided");
std::vector<bst_gpair> &gpair = *out_gpair;
gpair.resize(preds.size());
// check if label in range
bool label_correct = true;
// start calculating gradient
const unsigned nstep = static_cast<unsigned>(info.labels.size());
const bst_omp_uint ndata = static_cast<bst_omp_uint>(preds.size());
#pragma omp parallel for schedule(static)
for (bst_omp_uint i = 0; i < ndata; ++i) {
const unsigned j = i % nstep;
float p = loss.PredTransform(preds[i]);
float w = info.GetWeight(j);
if (info.labels[j] == 1.0f) w *= scale_pos_weight;
if (!loss.CheckLabel(info.labels[j])) label_correct = false;
gpair[i] = bst_gpair(loss.FirstOrderGradient(p, info.labels[j]) * w,
loss.SecondOrderGradient(p, info.labels[j]) * w);
}
utils::Check(label_correct, loss.CheckLabelErrorMsg());
}
virtual const char* DefaultEvalMetric(void) const {
return loss.DefaultEvalMetric();
}
virtual void PredTransform(std::vector<float> *io_preds) {
std::vector<float> &preds = *io_preds;
const bst_omp_uint ndata = static_cast<bst_omp_uint>(preds.size());
#pragma omp parallel for schedule(static)
for (bst_omp_uint j = 0; j < ndata; ++j) {
preds[j] = loss.PredTransform(preds[j]);
}
}
virtual float ProbToMargin(float base_score) const {
return loss.ProbToMargin(base_score);
}
protected:
float scale_pos_weight;
LossType loss;
};
// poisson regression for count
class PoissonRegression : public IObjFunction {
public:
explicit PoissonRegression(void) {
max_delta_step = 0.0f;
}
virtual ~PoissonRegression(void) {}
virtual void SetParam(const char *name, const char *val) {
using namespace std;
if (!strcmp( "max_delta_step", name )) {
max_delta_step = static_cast<float>(atof(val));
}
}
virtual void GetGradient(const std::vector<float> &preds,
const MetaInfo &info,
int iter,
std::vector<bst_gpair> *out_gpair) {
utils::Check(max_delta_step != 0.0f,
"PoissonRegression: need to set max_delta_step");
utils::Check(info.labels.size() != 0, "label set cannot be empty");
utils::Check(preds.size() == info.labels.size(),
"labels are not correctly provided");
std::vector<bst_gpair> &gpair = *out_gpair;
gpair.resize(preds.size());
// check if label in range
bool label_correct = true;
// start calculating gradient
const long ndata = static_cast<bst_omp_uint>(preds.size());
#pragma omp parallel for schedule(static)
for (long i = 0; i < ndata; ++i) {
float p = preds[i];
float w = info.GetWeight(i);
float y = info.labels[i];
if (y >= 0.0f) {
gpair[i] = bst_gpair((std::exp(p) - y) * w,
std::exp(p + max_delta_step) * w);
} else {
label_correct = false;
}
}
utils::Check(label_correct,
"PoissonRegression: label must be nonnegative");
}
virtual void PredTransform(std::vector<float> *io_preds) {
std::vector<float> &preds = *io_preds;
const long ndata = static_cast<long>(preds.size());
#pragma omp parallel for schedule(static)
for (long j = 0; j < ndata; ++j) {
preds[j] = std::exp(preds[j]);
}
}
virtual void EvalTransform(std::vector<float> *io_preds) {
PredTransform(io_preds);
}
virtual float ProbToMargin(float base_score) const {
return std::log(base_score);
}
virtual const char* DefaultEvalMetric(void) const {
return "poisson-nloglik";
}
private:
float max_delta_step;
};
// softmax multi-class classification
class SoftmaxMultiClassObj : public IObjFunction {
public:
explicit SoftmaxMultiClassObj(int output_prob)
: output_prob(output_prob) {
nclass = 0;
}
virtual ~SoftmaxMultiClassObj(void) {}
virtual void SetParam(const char *name, const char *val) {
using namespace std;
if (!strcmp( "num_class", name )) nclass = atoi(val);
}
virtual void GetGradient(const std::vector<float> &preds,
const MetaInfo &info,
int iter,
std::vector<bst_gpair> *out_gpair) {
utils::Check(nclass != 0, "must set num_class to use softmax");
utils::Check(info.labels.size() != 0, "label set cannot be empty");
utils::Check(preds.size() % (static_cast<size_t>(nclass) * info.labels.size()) == 0,
"SoftmaxMultiClassObj: label size and pred size does not match");
std::vector<bst_gpair> &gpair = *out_gpair;
gpair.resize(preds.size());
const unsigned nstep = static_cast<unsigned>(info.labels.size() * nclass);
const bst_omp_uint ndata = static_cast<bst_omp_uint>(preds.size() / nclass);
int label_error = 0;
#pragma omp parallel
{
std::vector<float> rec(nclass);
#pragma omp for schedule(static)
for (bst_omp_uint i = 0; i < ndata; ++i) {
for (int k = 0; k < nclass; ++k) {
rec[k] = preds[i * nclass + k];
}
Softmax(&rec);
const unsigned j = i % nstep;
int label = static_cast<int>(info.labels[j]);
if (label < 0 || label >= nclass) {
label_error = label; label = 0;
}
const float wt = info.GetWeight(j);
for (int k = 0; k < nclass; ++k) {
float p = rec[k];
const float h = 2.0f * p * (1.0f - p) * wt;
if (label == k) {
gpair[i * nclass + k] = bst_gpair((p - 1.0f) * wt, h);
} else {
gpair[i * nclass + k] = bst_gpair(p* wt, h);
}
}
}
}
utils::Check(label_error >= 0 && label_error < nclass,
"SoftmaxMultiClassObj: label must be in [0, num_class),"\
" num_class=%d but found %d in label", nclass, label_error);
}
virtual void PredTransform(std::vector<float> *io_preds) {
this->Transform(io_preds, output_prob);
}
virtual void EvalTransform(std::vector<float> *io_preds) {
this->Transform(io_preds, 1);
}
virtual const char* DefaultEvalMetric(void) const {
return "merror";
}
private:
inline void Transform(std::vector<float> *io_preds, int prob) {
utils::Check(nclass != 0, "must set num_class to use softmax");
std::vector<float> &preds = *io_preds;
std::vector<float> tmp;
const bst_omp_uint ndata = static_cast<bst_omp_uint>(preds.size()/nclass);
if (prob == 0) tmp.resize(ndata);
#pragma omp parallel
{
std::vector<float> rec(nclass);
#pragma omp for schedule(static)
for (bst_omp_uint j = 0; j < ndata; ++j) {
for (int k = 0; k < nclass; ++k) {
rec[k] = preds[j * nclass + k];
}
if (prob == 0) {
tmp[j] = static_cast<float>(FindMaxIndex(rec));
} else {
Softmax(&rec);
for (int k = 0; k < nclass; ++k) {
preds[j * nclass + k] = rec[k];
}
}
}
}
if (prob == 0) preds = tmp;
}
// data field
int nclass;
int output_prob;
};
/*! \brief objective for lambda rank */
class LambdaRankObj : public IObjFunction {
public:
LambdaRankObj(void) {
loss.loss_type = LossType::kLogisticRaw;
fix_list_weight = 0.0f;
num_pairsample = 1;
}
virtual ~LambdaRankObj(void) {}
virtual void SetParam(const char *name, const char *val) {
using namespace std;
if (!strcmp( "loss_type", name )) loss.loss_type = atoi(val);
if (!strcmp( "fix_list_weight", name)) fix_list_weight = static_cast<float>(atof(val));
if (!strcmp( "num_pairsample", name)) num_pairsample = atoi(val);
}
virtual void GetGradient(const std::vector<float> &preds,
const MetaInfo &info,
int iter,
std::vector<bst_gpair> *out_gpair) {
utils::Check(preds.size() == info.labels.size(), "label size predict size not match");
std::vector<bst_gpair> &gpair = *out_gpair;
gpair.resize(preds.size());
// quick consistency when group is not available
std::vector<unsigned> tgptr(2, 0); tgptr[1] = static_cast<unsigned>(info.labels.size());
const std::vector<unsigned> &gptr = info.group_ptr.size() == 0 ? tgptr : info.group_ptr;
utils::Check(gptr.size() != 0 && gptr.back() == info.labels.size(),
"group structure not consistent with #rows");
const bst_omp_uint ngroup = static_cast<bst_omp_uint>(gptr.size() - 1);
#pragma omp parallel
{
// parall construct, declare random number generator here, so that each
// thread use its own random number generator, seed by thread id and current iteration
random::Random rnd; rnd.Seed(iter* 1111 + omp_get_thread_num());
std::vector<LambdaPair> pairs;
std::vector<ListEntry> lst;
std::vector< std::pair<float, unsigned> > rec;
#pragma omp for schedule(static)
for (bst_omp_uint k = 0; k < ngroup; ++k) {
lst.clear(); pairs.clear();
for (unsigned j = gptr[k]; j < gptr[k+1]; ++j) {
lst.push_back(ListEntry(preds[j], info.labels[j], j));
gpair[j] = bst_gpair(0.0f, 0.0f);
}
std::sort(lst.begin(), lst.end(), ListEntry::CmpPred);
rec.resize(lst.size());
for (unsigned i = 0; i < lst.size(); ++i) {
rec[i] = std::make_pair(lst[i].label, i);
}
std::sort(rec.begin(), rec.end(), CmpFirst);
// enumerate buckets with same label, for each item in the lst, grab another sample randomly
for (unsigned i = 0; i < rec.size(); ) {
unsigned j = i + 1;
while (j < rec.size() && rec[j].first == rec[i].first) ++j;
// bucket in [i,j), get a sample outside bucket
unsigned nleft = i, nright = static_cast<unsigned>(rec.size() - j);
if (nleft + nright != 0) {
int nsample = num_pairsample;
while (nsample --) {
for (unsigned pid = i; pid < j; ++pid) {
unsigned ridx = static_cast<unsigned>(rnd.RandDouble() * (nleft+nright));
if (ridx < nleft) {
pairs.push_back(LambdaPair(rec[ridx].second, rec[pid].second));
} else {
pairs.push_back(LambdaPair(rec[pid].second, rec[ridx+j-i].second));
}
}
}
}
i = j;
}
// get lambda weight for the pairs
this->GetLambdaWeight(lst, &pairs);
// rescale each gradient and hessian so that the lst have constant weighted
float scale = 1.0f / num_pairsample;
if (fix_list_weight != 0.0f) {
scale *= fix_list_weight / (gptr[k+1] - gptr[k]);
}
for (size_t i = 0; i < pairs.size(); ++i) {
const ListEntry &pos = lst[pairs[i].pos_index];
const ListEntry &neg = lst[pairs[i].neg_index];
const float w = pairs[i].weight * scale;
float p = loss.PredTransform(pos.pred - neg.pred);
float g = loss.FirstOrderGradient(p, 1.0f);
float h = loss.SecondOrderGradient(p, 1.0f);
// accumulate gradient and hessian in both pid, and nid
gpair[pos.rindex].grad += g * w;
gpair[pos.rindex].hess += 2.0f * w * h;
gpair[neg.rindex].grad -= g * w;
gpair[neg.rindex].hess += 2.0f * w * h;
}
}
}
}
virtual const char* DefaultEvalMetric(void) const {
return "map";
}
protected:
/*! \brief helper information in a list */
struct ListEntry {
/*! \brief the predict score we in the data */
float pred;
/*! \brief the actual label of the entry */
float label;
/*! \brief row index in the data matrix */
unsigned rindex;
// constructor
ListEntry(float pred, float label, unsigned rindex)
: pred(pred), label(label), rindex(rindex) {}
// comparator by prediction
inline static bool CmpPred(const ListEntry &a, const ListEntry &b) {
return a.pred > b.pred;
}
// comparator by label
inline static bool CmpLabel(const ListEntry &a, const ListEntry &b) {
return a.label > b.label;
}
};
/*! \brief a pair in the lambda rank */
struct LambdaPair {
/*! \brief positive index: this is a position in the list */
unsigned pos_index;
/*! \brief negative index: this is a position in the list */
unsigned neg_index;
/*! \brief weight to be filled in */
float weight;
// constructor
LambdaPair(unsigned pos_index, unsigned neg_index)
: pos_index(pos_index), neg_index(neg_index), weight(1.0f) {}
};
/*!
* \brief get lambda weight for existing pairs
* \param list a list that is sorted by pred score
* \param io_pairs record of pairs, containing the pairs to fill in weights
*/
virtual void GetLambdaWeight(const std::vector<ListEntry> &sorted_list,
std::vector<LambdaPair> *io_pairs) = 0;
private:
// loss function
LossType loss;
// number of samples peformed for each instance
int num_pairsample;
// fix weight of each elements in list
float fix_list_weight;
};
class PairwiseRankObj: public LambdaRankObj{
public:
virtual ~PairwiseRankObj(void) {}
protected:
virtual void GetLambdaWeight(const std::vector<ListEntry> &sorted_list,
std::vector<LambdaPair> *io_pairs) {}
};
// beta version: NDCG lambda rank
class LambdaRankObjNDCG : public LambdaRankObj {
public:
virtual ~LambdaRankObjNDCG(void) {}
protected:
virtual void GetLambdaWeight(const std::vector<ListEntry> &sorted_list,
std::vector<LambdaPair> *io_pairs) {
std::vector<LambdaPair> &pairs = *io_pairs;
float IDCG;
{
std::vector<float> labels(sorted_list.size());
for (size_t i = 0; i < sorted_list.size(); ++i) {
labels[i] = sorted_list[i].label;
}
std::sort(labels.begin(), labels.end(), std::greater<float>());
IDCG = CalcDCG(labels);
}
if (IDCG == 0.0) {
for (size_t i = 0; i < pairs.size(); ++i) {
pairs[i].weight = 0.0f;
}
} else {
IDCG = 1.0f / IDCG;
for (size_t i = 0; i < pairs.size(); ++i) {
unsigned pos_idx = pairs[i].pos_index;
unsigned neg_idx = pairs[i].neg_index;
float pos_loginv = 1.0f / std::log(pos_idx + 2.0f);
float neg_loginv = 1.0f / std::log(neg_idx + 2.0f);
int pos_label = static_cast<int>(sorted_list[pos_idx].label);
int neg_label = static_cast<int>(sorted_list[neg_idx].label);
float original =
((1 << pos_label) - 1) * pos_loginv + ((1 << neg_label) - 1) * neg_loginv;
float changed =
((1 << neg_label) - 1) * pos_loginv + ((1 << pos_label) - 1) * neg_loginv;
float delta = (original - changed) * IDCG;
if (delta < 0.0f) delta = - delta;
pairs[i].weight = delta;
}
}
}
inline static float CalcDCG(const std::vector<float> &labels) {
double sumdcg = 0.0;
for (size_t i = 0; i < labels.size(); ++i) {
const unsigned rel = static_cast<unsigned>(labels[i]);
if (rel != 0) {
sumdcg += ((1 << rel) - 1) / std::log(static_cast<float>(i + 2));
}
}
return static_cast<float>(sumdcg);
}
};
class LambdaRankObjMAP : public LambdaRankObj {
public:
virtual ~LambdaRankObjMAP(void) {}
protected:
struct MAPStats {
/*! \brief the accumulated precision */
float ap_acc;
/*!
* \brief the accumulated precision,
* assuming a positive instance is missing
*/
float ap_acc_miss;
/*!
* \brief the accumulated precision,
* assuming that one more positive instance is inserted ahead
*/
float ap_acc_add;
/* \brief the accumulated positive instance count */
float hits;
MAPStats(void) {}
MAPStats(float ap_acc, float ap_acc_miss, float ap_acc_add, float hits)
: ap_acc(ap_acc), ap_acc_miss(ap_acc_miss), ap_acc_add(ap_acc_add), hits(hits) {}
};
/*!
* \brief Obtain the delta MAP if trying to switch the positions of instances in index1 or index2
* in sorted triples
* \param sorted_list the list containing entry information
* \param index1,index2 the instances switched
* \param map_stats a vector containing the accumulated precisions for each position in a list
*/
inline float GetLambdaMAP(const std::vector<ListEntry> &sorted_list,
int index1, int index2,
std::vector<MAPStats> *p_map_stats) {
std::vector<MAPStats> &map_stats = *p_map_stats;
if (index1 == index2 || map_stats[map_stats.size() - 1].hits == 0) {
return 0.0f;
}
if (index1 > index2) std::swap(index1, index2);
float original = map_stats[index2].ap_acc;
if (index1 != 0) original -= map_stats[index1 - 1].ap_acc;
float changed = 0;
float label1 = sorted_list[index1].label > 0.0f ? 1.0f : 0.0f;
float label2 = sorted_list[index2].label > 0.0f ? 1.0f : 0.0f;
if (label1 == label2) {
return 0.0;
} else if (label1 < label2) {
changed += map_stats[index2 - 1].ap_acc_add - map_stats[index1].ap_acc_add;
changed += (map_stats[index1].hits + 1.0f) / (index1 + 1);
} else {
changed += map_stats[index2 - 1].ap_acc_miss - map_stats[index1].ap_acc_miss;
changed += map_stats[index2].hits / (index2 + 1);
}
float ans = (changed - original) / (map_stats[map_stats.size() - 1].hits);
if (ans < 0) ans = -ans;
return ans;
}
/*
* \brief obtain preprocessing results for calculating delta MAP
* \param sorted_list the list containing entry information
* \param map_stats a vector containing the accumulated precisions for each position in a list
*/
inline void GetMAPStats(const std::vector<ListEntry> &sorted_list,
std::vector<MAPStats> *p_map_acc) {
std::vector<MAPStats> &map_acc = *p_map_acc;
map_acc.resize(sorted_list.size());
float hit = 0, acc1 = 0, acc2 = 0, acc3 = 0;
for (size_t i = 1; i <= sorted_list.size(); ++i) {
if (sorted_list[i - 1].label > 0.0f) {
hit++;
acc1 += hit / i;
acc2 += (hit - 1) / i;
acc3 += (hit + 1) / i;
}
map_acc[i - 1] = MAPStats(acc1, acc2, acc3, hit);
}
}
virtual void GetLambdaWeight(const std::vector<ListEntry> &sorted_list,
std::vector<LambdaPair> *io_pairs) {
std::vector<LambdaPair> &pairs = *io_pairs;
std::vector<MAPStats> map_stats;
GetMAPStats(sorted_list, &map_stats);
for (size_t i = 0; i < pairs.size(); ++i) {
pairs[i].weight =
GetLambdaMAP(sorted_list, pairs[i].pos_index,
pairs[i].neg_index, &map_stats);
}
}
};
} // namespace learner
} // namespace xgboost
#endif // XGBOOST_LEARNER_OBJECTIVE_INL_HPP_