xgboost/gbm/gbtree-inl.hpp
2014-08-15 20:15:58 -07:00

366 lines
13 KiB
C++

#ifndef XGBOOST_GBM_GBTREE_INL_HPP_
#define XGBOOST_GBM_GBTREE_INL_HPP_
/*!
* \file gbtree-inl.hpp
* \brief gradient boosted tree implementation
* \author Tianqi Chen
*/
#include <vector>
#include <utility>
#include <string>
#include "./gbm.h"
#include "../tree/updater.h"
namespace xgboost {
namespace gbm {
/*!
* \brief gradient boosted tree
* \tparam FMatrix the data type updater taking
*/
template<typename FMatrix>
class GBTree : public IGradBooster<FMatrix> {
public:
virtual ~GBTree(void) {
this->Clear();
}
virtual void SetParam(const char *name, const char *val) {
if (!strncmp(name, "bst:", 4)) {
cfg.push_back(std::make_pair(std::string(name+4), std::string(val)));
// set into updaters, if already intialized
for (size_t i = 0; i < updaters.size(); ++i) {
updaters[i]->SetParam(name+4, val);
}
}
if (!strcmp(name, "silent")) {
this->SetParam("bst:silent", val);
}
tparam.SetParam(name, val);
if (trees.size() == 0) mparam.SetParam(name, val);
}
virtual void LoadModel(utils::IStream &fi) {
this->Clear();
utils::Check(fi.Read(&mparam, sizeof(ModelParam)) != 0,
"GBTree: invalid model file");
trees.resize(mparam.num_trees);
for (size_t i = 0; i < trees.size(); ++i) {
trees[i] = new tree::RegTree();
trees[i]->LoadModel(fi);
}
tree_info.resize(mparam.num_trees);
if (mparam.num_trees != 0) {
utils::Check(fi.Read(&tree_info[0], sizeof(int) * mparam.num_trees) != 0,
"GBTree: invalid model file");
}
if (mparam.num_pbuffer != 0) {
pred_buffer.resize(mparam.PredBufferSize());
pred_counter.resize(mparam.PredBufferSize());
utils::Check(fi.Read(&pred_buffer[0], pred_buffer.size() * sizeof(float)) != 0,
"GBTree: invalid model file");
utils::Check(fi.Read(&pred_counter[0], pred_counter.size() * sizeof(unsigned)) != 0,
"GBTree: invalid model file");
}
}
virtual void SaveModel(utils::IStream &fo) const {
utils::Assert(mparam.num_trees == static_cast<int>(trees.size()), "GBTree");
fo.Write(&mparam, sizeof(ModelParam));
for (size_t i = 0; i < trees.size(); ++i) {
trees[i]->SaveModel(fo);
}
if (tree_info.size() != 0) {
fo.Write(&tree_info[0], sizeof(int) * tree_info.size());
}
if (mparam.num_pbuffer != 0) {
fo.Write(&pred_buffer[0], pred_buffer.size() * sizeof(float));
fo.Write(&pred_counter[0], pred_counter.size() * sizeof(unsigned));
}
}
// initialize the predic buffer
virtual void InitModel(void) {
pred_buffer.clear(); pred_counter.clear();
pred_buffer.resize(mparam.PredBufferSize(), 0.0f);
pred_counter.resize(mparam.PredBufferSize(), 0);
utils::Assert(mparam.num_trees == 0, "GBTree: model already initialized");
utils::Assert(trees.size() == 0, "GBTree: model already initialized");
}
virtual void DoBoost(const std::vector<bst_gpair> &gpair,
FMatrix &fmat,
const std::vector<unsigned> &root_index) {
if (mparam.num_output_group == 1) {
this->BoostNewTrees(gpair, fmat, root_index, 0);
} else {
const int ngroup = mparam.num_output_group;
utils::Check(gpair.size() % ngroup == 0,
"must have exactly ngroup*nrow gpairs");
std::vector<bst_gpair> tmp(gpair.size()/ngroup);
for (int gid = 0; gid < ngroup; ++gid) {
#pragma omp parallel for schedule(static)
for (size_t i = 0; i < tmp.size(); ++i) {
tmp[i] = gpair[i * ngroup + gid];
}
this->BoostNewTrees(tmp, fmat, root_index, gid);
}
}
}
virtual void Predict(const FMatrix &fmat,
int64_t buffer_offset,
const std::vector<unsigned> &root_index,
std::vector<float> *out_preds) {
int nthread;
#pragma omp parallel
{
nthread = omp_get_num_threads();
}
this->InitThreadTemp(nthread);
std::vector<float> &preds = *out_preds;
preds.resize(0);
// start collecting the prediction
utils::IIterator<SparseBatch> *iter = fmat.RowIterator();
iter->BeforeFirst();
while (iter->Next()) {
const SparseBatch &batch = iter->Value();
utils::Assert(batch.base_rowid * mparam.num_output_group == preds.size(),
"base_rowid is not set correctly");
// output convention: nrow * k, where nrow is number of rows
// k is number of group
preds.resize(preds.size() + batch.size * mparam.num_output_group);
// parallel over local batch
const unsigned nsize = static_cast<unsigned>(batch.size);
#pragma omp parallel for schedule(static)
for (unsigned i = 0; i < nsize; ++i) {
const int tid = omp_get_thread_num();
std::vector<float> &feats = thread_temp[tid];
const size_t ridx = batch.base_rowid + i;
const unsigned root_idx = root_index.size() == 0 ? 0 : root_index[ridx];
// loop over output groups
for (int gid = 0; gid < mparam.num_output_group; ++gid) {
preds[ridx * mparam.num_output_group + gid] =
this->Pred(batch[i],
buffer_offset < 0 ? -1 : buffer_offset+ridx,
gid, root_idx, &feats);
}
}
}
}
protected:
// clear the model
inline void Clear(void) {
for (size_t i = 0; i < trees.size(); ++i) {
delete trees[i];
}
trees.clear();
pred_buffer.clear();
pred_counter.clear();
}
// initialize updater before using them
inline void InitUpdater(void) {
if (tparam.updater_initialized != 0) return;
for (size_t i = 0; i < updaters.size(); ++i) {
delete updaters[i];
}
updaters.clear();
std::string tval = tparam.updater_seq;
char *saveptr, *pstr;
pstr = strtok_r(&tval[0], ",", &saveptr);
while (pstr != NULL) {
updaters.push_back(tree::CreateUpdater<FMatrix>(pstr));
for (size_t j = 0; j < cfg.size(); ++j) {
// set parameters
updaters.back()->SetParam(cfg[j].first.c_str(), cfg[j].second.c_str());
}
pstr = strtok_r(NULL, ",", &saveptr);
}
tparam.updater_initialized = 1;
}
// do group specific group
inline void BoostNewTrees(const std::vector<bst_gpair> &gpair,
FMatrix &fmat,
const std::vector<unsigned> &root_index,
int bst_group) {
this->InitUpdater();
// create the trees
std::vector<tree::RegTree *> new_trees;
for (int i = 0; i < tparam.num_parallel_tree; ++i) {
new_trees.push_back(new tree::RegTree());
for (size_t j = 0; j < cfg.size(); ++j) {
new_trees.back()->param.SetParam(cfg[j].first.c_str(), cfg[j].second.c_str());
}
new_trees.back()->InitModel();
}
// update the trees
for (size_t i = 0; i < updaters.size(); ++i) {
updaters[i]->Update(gpair, fmat, root_index, new_trees);
}
// push back to model
for (size_t i = 0; i < new_trees.size(); ++i) {
trees.push_back(new_trees[i]);
tree_info.push_back(bst_group);
}
mparam.num_trees += tparam.num_parallel_tree;
}
// make a prediction for a single instance
inline float Pred(const SparseBatch::Inst &inst,
int64_t buffer_index,
int bst_group,
unsigned root_index,
std::vector<float> *p_feats) {
size_t itop = 0;
float psum = 0.0f;
const int bid = mparam.BufferOffset(buffer_index, bst_group);
// load buffered results if any
if (bid >= 0) {
itop = pred_counter[bid];
psum = pred_buffer[bid];
}
if (itop != trees.size()) {
FillThreadTemp(inst, p_feats);
for (size_t i = itop; i < trees.size(); ++i) {
if (tree_info[i] == bst_group) {
psum += trees[i]->Predict(*p_feats, root_index);
}
}
DropThreadTemp(inst, p_feats);
}
// updated the buffered results
if (bid >= 0) {
pred_counter[bid] = static_cast<unsigned>(trees.size());
pred_buffer[bid] = psum;
}
return psum;
}
// initialize thread local space for prediction
inline void InitThreadTemp(int nthread) {
thread_temp.resize(nthread);
for (size_t i = 0; i < thread_temp.size(); ++i) {
thread_temp[i].resize(mparam.num_feature);
std::fill(thread_temp[i].begin(), thread_temp[i].end(), NAN);
}
}
// fill in a thread local dense vector using a sparse instance
inline static void FillThreadTemp(const SparseBatch::Inst &inst,
std::vector<float> *p_feats) {
std::vector<float> &feats = *p_feats;
for (bst_uint i = 0; i < inst.length; ++i) {
feats[inst[i].findex] = inst[i].fvalue;
}
}
// clear up a thread local dense vector
inline static void DropThreadTemp(const SparseBatch::Inst &inst,
std::vector<float> *p_feats) {
std::vector<float> &feats = *p_feats;
for (bst_uint i = 0; i < inst.length; ++i) {
feats[inst[i].findex] = NAN;
}
}
// --- data structure ---
/*! \brief training parameters */
struct TrainParam {
/*! \brief number of threads */
int nthread;
/*!
* \brief number of parallel trees constructed each iteration
* use this option to support boosted random forest
*/
int num_parallel_tree;
/*! \brief whether updater is already initialized */
int updater_initialized;
/*! \brief tree updater sequence */
std::string updater_seq;
// construction
TrainParam(void) {
nthread = 0;
updater_seq = "grow_colmaker,prune";
num_parallel_tree = 1;
updater_initialized = 0;
}
inline void SetParam(const char *name, const char *val){
if (!strcmp(name, "updater") &&
strcmp(updater_seq.c_str(), val) != 0) {
updater_seq = val;
updater_initialized = 0;
}
if (!strcmp(name, "nthread")) {
omp_set_num_threads(nthread);
nthread = atoi(val);
}
if (!strcmp(name, "num_parallel_tree")) {
num_parallel_tree = atoi(val);
}
}
};
/*! \brief model parameters */
struct ModelParam {
/*! \brief number of trees */
int num_trees;
/*! \brief number of root: default 0, means single tree */
int num_roots;
/*! \brief number of features to be used by trees */
int num_feature;
/*! \brief size of predicton buffer allocated used for buffering */
int64_t num_pbuffer;
/*!
* \brief how many output group a single instance can produce
* this affects the behavior of number of output we have:
* suppose we have n instance and k group, output will be k*n
*/
int num_output_group;
/*! \brief reserved parameters */
int reserved[32];
/*! \brief constructor */
ModelParam(void) {
num_trees = 0;
num_roots = num_feature = 0;
num_pbuffer = 0;
num_output_group = 1;
memset(reserved, 0, sizeof(reserved));
}
/*!
* \brief set parameters from outside
* \param name name of the parameter
* \param val value of the parameter
*/
inline void SetParam(const char *name, const char *val) {
if (!strcmp("num_pbuffer", name)) num_pbuffer = atol(val);
if (!strcmp("num_output_group", name)) num_output_group = atol(val);
if (!strcmp("bst:num_roots", name)) num_roots = atoi(val);
if (!strcmp("bst:num_feature", name)) num_feature = atoi(val);
}
/*! \return size of prediction buffer actually needed */
inline size_t PredBufferSize(void) const {
return num_output_group * num_pbuffer;
}
/*!
* \brief get the buffer offset given a buffer index and group id
* \return calculated buffer offset
*/
inline size_t BufferOffset(int64_t buffer_index, int bst_group) const {
if (buffer_index < 0) return -1;
utils::Check(buffer_index < num_pbuffer, "buffer_index exceed num_pbuffer");
return buffer_index + num_pbuffer * bst_group;
}
};
// training parameter
TrainParam tparam;
// model parameter
ModelParam mparam;
/*! \brief vector of trees stored in the model */
std::vector<tree::RegTree*> trees;
/*! \brief some information indicator of the tree, reserved */
std::vector<int> tree_info;
/*! \brief prediction buffer */
std::vector<float> pred_buffer;
/*! \brief prediction buffer counter, remember the prediction */
std::vector<unsigned> pred_counter;
// ----training fields----
// configurations for tree
std::vector< std::pair<std::string, std::string> > cfg;
// temporal storage for per thread
std::vector< std::vector<float> > thread_temp;
// the updaters that can be applied to each of tree
std::vector< tree::IUpdater<FMatrix>* > updaters;
};
} // namespace gbm
} // namespace xgboost
#endif // XGBOOST_GBM_GBTREE_INL_HPP_