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

[TREE] Move the files to target refactor location

Browse Source
This commit is contained in:
tqchen
2016-01-01 05:01:22 -08:00
parent 3128e1705b
commit 4adc4cf0b9
14 changed files with 0 additions and 889 deletions
Download Patch File Download Diff File Expand all files Collapse all files

427
old_src/tree/updater_basemaker-inl.hpp
View File

@@ -1,427 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file updater_basemaker-inl.hpp
* \brief implement a common tree constructor
* \author Tianqi Chen
*/
#ifndef XGBOOST_TREE_UPDATER_BASEMAKER_INL_HPP_
#define XGBOOST_TREE_UPDATER_BASEMAKER_INL_HPP_
#include <vector>
#include <algorithm>
#include <string>
#include <limits>
#include "../sync/sync.h"
#include "../utils/random.h"
#include "../utils/quantile.h"
namespace xgboost {
namespace tree {
/*!
* \brief base tree maker class that defines common operation
* needed in tree making
*/
class BaseMaker: public IUpdater {
public:
// destructor
virtual ~BaseMaker(void) {}
// set training parameter
virtual void SetParam(const char *name, const char *val) {
param.SetParam(name, val);
}
protected:
// helper to collect and query feature meta information
struct FMetaHelper {
public:
/*! \brief find type of each feature, use column format */
inline void InitByCol(IFMatrix *p_fmat,
const RegTree &tree) {
fminmax.resize(tree.param.num_feature * 2);
std::fill(fminmax.begin(), fminmax.end(),
-std::numeric_limits<bst_float>::max());
// start accumulating statistics
utils::IIterator<ColBatch> *iter = p_fmat->ColIterator();
iter->BeforeFirst();
while (iter->Next()) {
const ColBatch &batch = iter->Value();
for (bst_uint i = 0; i < batch.size; ++i) {
const bst_uint fid = batch.col_index[i];
const ColBatch::Inst &c = batch[i];
if (c.length != 0) {
fminmax[fid * 2 + 0] = std::max(-c[0].fvalue, fminmax[fid * 2 + 0]);
fminmax[fid * 2 + 1] = std::max(c[c.length - 1].fvalue, fminmax[fid * 2 + 1]);
}
}
}
rabit::Allreduce<rabit::op::Max>(BeginPtr(fminmax), fminmax.size());
}
// get feature type, 0:empty 1:binary 2:real
inline int Type(bst_uint fid) const {
utils::Assert(fid * 2 + 1 < fminmax.size(),
"FeatHelper fid exceed query bound ");
bst_float a = fminmax[fid * 2];
bst_float b = fminmax[fid * 2 + 1];
if (a == -std::numeric_limits<bst_float>::max()) return 0;
if (-a == b) {
return 1;
} else {
return 2;
}
}
inline bst_float MaxValue(bst_uint fid) const {
return fminmax[fid *2 + 1];
}
inline void SampleCol(float p, std::vector<bst_uint> *p_findex) const {
std::vector<bst_uint> &findex = *p_findex;
findex.clear();
for (size_t i = 0; i < fminmax.size(); i += 2) {
const bst_uint fid = static_cast<bst_uint>(i / 2);
if (this->Type(fid) != 0) findex.push_back(fid);
}
unsigned n = static_cast<unsigned>(p * findex.size());
random::Shuffle(findex);
findex.resize(n);
// sync the findex if it is subsample
std::string s_cache;
utils::MemoryBufferStream fc(&s_cache);
utils::IStream &fs = fc;
if (rabit::GetRank() == 0) {
fs.Write(findex);
}
rabit::Broadcast(&s_cache, 0);
fs.Read(&findex);
}
private:
std::vector<bst_float> fminmax;
};
// ------static helper functions ------
// helper function to get to next level of the tree
/*! \brief this is helper function for row based data*/
inline static int NextLevel(const RowBatch::Inst &inst, const RegTree &tree, int nid) {
const RegTree::Node &n = tree[nid];
bst_uint findex = n.split_index();
for (unsigned i = 0; i < inst.length; ++i) {
if (findex == inst[i].index) {
if (inst[i].fvalue < n.split_cond()) {
return n.cleft();
} else {
return n.cright();
}
}
}
return n.cdefault();
}
/*! \brief get number of omp thread in current context */
inline static int get_nthread(void) {
int nthread;
#pragma omp parallel
{
nthread = omp_get_num_threads();
}
return nthread;
}
// ------class member helpers---------
/*! \brief initialize temp data structure */
inline void InitData(const std::vector<bst_gpair> &gpair,
const IFMatrix &fmat,
const std::vector<unsigned> &root_index,
const RegTree &tree) {
utils::Assert(tree.param.num_nodes == tree.param.num_roots,
"TreeMaker: can only grow new tree");
{
// setup position
position.resize(gpair.size());
if (root_index.size() == 0) {
std::fill(position.begin(), position.end(), 0);
} else {
for (size_t i = 0; i < position.size(); ++i) {
position[i] = root_index[i];
utils::Assert(root_index[i] < (unsigned)tree.param.num_roots,
"root index exceed setting");
}
}
// mark delete for the deleted datas
for (size_t i = 0; i < position.size(); ++i) {
if (gpair[i].hess < 0.0f) position[i] = ~position[i];
}
// mark subsample
if (param.subsample < 1.0f) {
for (size_t i = 0; i < position.size(); ++i) {
if (gpair[i].hess < 0.0f) continue;
if (random::SampleBinary(param.subsample) == 0) position[i] = ~position[i];
}
}
}
{
// expand query
qexpand.reserve(256); qexpand.clear();
for (int i = 0; i < tree.param.num_roots; ++i) {
qexpand.push_back(i);
}
this->UpdateNode2WorkIndex(tree);
}
}
/*! \brief update queue expand add in new leaves */
inline void UpdateQueueExpand(const RegTree &tree) {
std::vector<int> newnodes;
for (size_t i = 0; i < qexpand.size(); ++i) {
const int nid = qexpand[i];
if (!tree[nid].is_leaf()) {
newnodes.push_back(tree[nid].cleft());
newnodes.push_back(tree[nid].cright());
}
}
// use new nodes for qexpand
qexpand = newnodes;
this->UpdateNode2WorkIndex(tree);
}
// return decoded position
inline int DecodePosition(bst_uint ridx) const {
const int pid = position[ridx];
return pid < 0 ? ~pid : pid;
}
// encode the encoded position value for ridx
inline void SetEncodePosition(bst_uint ridx, int nid) {
if (position[ridx] < 0) {
position[ridx] = ~nid;
} else {
position[ridx] = nid;
}
}
/*!
* \brief this is helper function uses column based data structure,
* reset the positions to the lastest one
* \param nodes the set of nodes that contains the split to be used
* \param p_fmat feature matrix needed for tree construction
* \param tree the regression tree structure
*/
inline void ResetPositionCol(const std::vector<int> &nodes,
IFMatrix *p_fmat, const RegTree &tree) {
// set the positions in the nondefault
this->SetNonDefaultPositionCol(nodes, p_fmat, tree);
// set rest of instances to default position
const std::vector<bst_uint> &rowset = p_fmat->buffered_rowset();
// set default direct nodes to default
// for leaf nodes that are not fresh, mark then to ~nid,
// so that they are ignored in future statistics collection
const bst_omp_uint ndata = static_cast<bst_omp_uint>(rowset.size());
#pragma omp parallel for schedule(static)
for (bst_omp_uint i = 0; i < ndata; ++i) {
const bst_uint ridx = rowset[i];
const int nid = this->DecodePosition(ridx);
if (tree[nid].is_leaf()) {
// mark finish when it is not a fresh leaf
if (tree[nid].cright() == -1) {
position[ridx] = ~nid;
}
} else {
// push to default branch
if (tree[nid].default_left()) {
this->SetEncodePosition(ridx, tree[nid].cleft());
} else {
this->SetEncodePosition(ridx, tree[nid].cright());
}
}
}
}
/*!
* \brief this is helper function uses column based data structure,
* update all positions into nondefault branch, if any, ignore the default branch
* \param nodes the set of nodes that contains the split to be used
* \param p_fmat feature matrix needed for tree construction
* \param tree the regression tree structure
*/
virtual void SetNonDefaultPositionCol(const std::vector<int> &nodes,
IFMatrix *p_fmat, const RegTree &tree) {
// step 1, classify the non-default data into right places
std::vector<unsigned> fsplits;
for (size_t i = 0; i < nodes.size(); ++i) {
const int nid = nodes[i];
if (!tree[nid].is_leaf()) {
fsplits.push_back(tree[nid].split_index());
}
}
std::sort(fsplits.begin(), fsplits.end());
fsplits.resize(std::unique(fsplits.begin(), fsplits.end()) - fsplits.begin());
utils::IIterator<ColBatch> *iter = p_fmat->ColIterator(fsplits);
while (iter->Next()) {
const ColBatch &batch = iter->Value();
for (size_t i = 0; i < batch.size; ++i) {
ColBatch::Inst col = batch[i];
const bst_uint fid = batch.col_index[i];
const bst_omp_uint ndata = static_cast<bst_omp_uint>(col.length);
#pragma omp parallel for schedule(static)
for (bst_omp_uint j = 0; j < ndata; ++j) {
const bst_uint ridx = col[j].index;
const float fvalue = col[j].fvalue;
const int nid = this->DecodePosition(ridx);
// go back to parent, correct those who are not default
if (!tree[nid].is_leaf() && tree[nid].split_index() == fid) {
if (fvalue < tree[nid].split_cond()) {
this->SetEncodePosition(ridx, tree[nid].cleft());
} else {
this->SetEncodePosition(ridx, tree[nid].cright());
}
}
}
}
}
}
/*! \brief helper function to get statistics from a tree */
template<typename TStats>
inline void GetNodeStats(const std::vector<bst_gpair> &gpair,
const IFMatrix &fmat,
const RegTree &tree,
const BoosterInfo &info,
std::vector< std::vector<TStats> > *p_thread_temp,
std::vector<TStats> *p_node_stats) {
std::vector< std::vector<TStats> > &thread_temp = *p_thread_temp;
thread_temp.resize(this->get_nthread());
p_node_stats->resize(tree.param.num_nodes);
#pragma omp parallel
{
const int tid = omp_get_thread_num();
thread_temp[tid].resize(tree.param.num_nodes, TStats(param));
for (size_t i = 0; i < qexpand.size(); ++i) {
const unsigned nid = qexpand[i];
thread_temp[tid][nid].Clear();
}
}
const std::vector<bst_uint> &rowset = fmat.buffered_rowset();
// setup position
const bst_omp_uint ndata = static_cast<bst_omp_uint>(rowset.size());
#pragma omp parallel for schedule(static)
for (bst_omp_uint i = 0; i < ndata; ++i) {
const bst_uint ridx = rowset[i];
const int nid = position[ridx];
const int tid = omp_get_thread_num();
if (nid >= 0) {
thread_temp[tid][nid].Add(gpair, info, ridx);
}
}
// sum the per thread statistics together
for (size_t j = 0; j < qexpand.size(); ++j) {
const int nid = qexpand[j];
TStats &s = (*p_node_stats)[nid];
s.Clear();
for (size_t tid = 0; tid < thread_temp.size(); ++tid) {
s.Add(thread_temp[tid][nid]);
}
}
}
/*! \brief common helper data structure to build sketch */
struct SketchEntry {
/*! \brief total sum of amount to be met */
double sum_total;
/*! \brief statistics used in the sketch */
double rmin, wmin;
/*! \brief last seen feature value */
bst_float last_fvalue;
/*! \brief current size of sketch */
double next_goal;
// pointer to the sketch to put things in
utils::WXQuantileSketch<bst_float, bst_float> *sketch;
// initialize the space
inline void Init(unsigned max_size) {
next_goal = -1.0f;
rmin = wmin = 0.0f;
sketch->temp.Reserve(max_size + 1);
sketch->temp.size = 0;
}
/*!
* \brief push a new element to sketch
* \param fvalue feature value, comes in sorted ascending order
* \param w weight
* \param max_size
*/
inline void Push(bst_float fvalue, bst_float w, unsigned max_size) {
if (next_goal == -1.0f) {
next_goal = 0.0f;
last_fvalue = fvalue;
wmin = w;
return;
}
if (last_fvalue != fvalue) {
double rmax = rmin + wmin;
if (rmax >= next_goal && sketch->temp.size != max_size) {
if (sketch->temp.size == 0 ||
last_fvalue > sketch->temp.data[sketch->temp.size-1].value) {
// push to sketch
sketch->temp.data[sketch->temp.size] =
utils::WXQuantileSketch<bst_float, bst_float>::
Entry(static_cast<bst_float>(rmin),
static_cast<bst_float>(rmax),
static_cast<bst_float>(wmin), last_fvalue);
utils::Assert(sketch->temp.size < max_size,
"invalid maximum size max_size=%u, stemp.size=%lu\n",
max_size, sketch->temp.size);
++sketch->temp.size;
}
if (sketch->temp.size == max_size) {
next_goal = sum_total * 2.0f + 1e-5f;
} else {
next_goal = static_cast<bst_float>(sketch->temp.size * sum_total / max_size);
}
} else {
if (rmax >= next_goal) {
rabit::TrackerPrintf("INFO: rmax=%g, sum_total=%g, next_goal=%g, size=%lu\n",
rmax, sum_total, next_goal, sketch->temp.size);
}
}
rmin = rmax;
wmin = w;
last_fvalue = fvalue;
} else {
wmin += w;
}
}
/*! \brief push final unfinished value to the sketch */
inline void Finalize(unsigned max_size) {
double rmax = rmin + wmin;
if (sketch->temp.size == 0 || last_fvalue > sketch->temp.data[sketch->temp.size-1].value) {
utils::Assert(sketch->temp.size <= max_size,
"Finalize: invalid maximum size, max_size=%u, stemp.size=%lu",
sketch->temp.size, max_size);
// push to sketch
sketch->temp.data[sketch->temp.size] =
utils::WXQuantileSketch<bst_float, bst_float>::
Entry(static_cast<bst_float>(rmin),
static_cast<bst_float>(rmax),
static_cast<bst_float>(wmin), last_fvalue);
++sketch->temp.size;
}
sketch->PushTemp();
}
};
/*! \brief training parameter of tree grower */
TrainParam param;
/*! \brief queue of nodes to be expanded */
std::vector<int> qexpand;
/*!
* \brief map active node to is working index offset in qexpand,
* can be -1, which means the node is node actively expanding
*/
std::vector<int> node2workindex;
/*!
* \brief position of each instance in the tree
* can be negative, which means this position is no longer expanding
* see also Decode/EncodePosition
*/
std::vector<int> position;
private:
inline void UpdateNode2WorkIndex(const RegTree &tree) {
// update the node2workindex
std::fill(node2workindex.begin(), node2workindex.end(), -1);
node2workindex.resize(tree.param.num_nodes);
for (size_t i = 0; i < qexpand.size(); ++i) {
node2workindex[qexpand[i]] = static_cast<int>(i);
}
}
};
} // namespace tree
} // namespace xgboost
#endif // XGBOOST_TREE_UPDATER_BASEMAKER_INL_HPP_

769
old_src/tree/updater_histmaker-inl.hpp
View File

@@ -1,769 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file updater_histmaker-inl.hpp
* \brief use histogram counting to construct a tree
* \author Tianqi Chen
*/
#ifndef XGBOOST_TREE_UPDATER_HISTMAKER_INL_HPP_
#define XGBOOST_TREE_UPDATER_HISTMAKER_INL_HPP_
#include <vector>
#include <algorithm>
#include "../sync/sync.h"
#include "../utils/quantile.h"
#include "../utils/group_data.h"
#include "./updater_basemaker-inl.hpp"
namespace xgboost {
namespace tree {
template<typename TStats>
class HistMaker: public BaseMaker {
public:
virtual ~HistMaker(void) {}
virtual void Update(const std::vector<bst_gpair> &gpair,
IFMatrix *p_fmat,
const BoosterInfo &info,
const std::vector<RegTree*> &trees) {
TStats::CheckInfo(info);
// rescale learning rate according to size of trees
float lr = param.learning_rate;
param.learning_rate = lr / trees.size();
// build tree
for (size_t i = 0; i < trees.size(); ++i) {
this->Update(gpair, p_fmat, info, trees[i]);
}
param.learning_rate = lr;
}
protected:
/*! \brief a single histogram */
struct HistUnit {
/*! \brief cutting point of histogram, contains maximum point */
const bst_float *cut;
/*! \brief content of statistics data */
TStats *data;
/*! \brief size of histogram */
unsigned size;
// default constructor
HistUnit(void) {}
// constructor
HistUnit(const bst_float *cut, TStats *data, unsigned size)
: cut(cut), data(data), size(size) {}
/*! \brief add a histogram to data */
inline void Add(bst_float fv,
const std::vector<bst_gpair> &gpair,
const BoosterInfo &info,
const bst_uint ridx) {
unsigned i = std::upper_bound(cut, cut + size, fv) - cut;
utils::Assert(size != 0, "try insert into size=0");
utils::Assert(i < size,
"maximum value must be in cut, fv = %g, cutmax=%g", fv, cut[size-1]);
data[i].Add(gpair, info, ridx);
}
};
/*! \brief a set of histograms from different index */
struct HistSet {
/*! \brief the index pointer of each histunit */
const unsigned *rptr;
/*! \brief cutting points in each histunit */
const bst_float *cut;
/*! \brief data in different hist unit */
std::vector<TStats> data;
/*! \brief */
inline HistUnit operator[](size_t fid) {
return HistUnit(cut + rptr[fid],
&data[0] + rptr[fid],
rptr[fid+1] - rptr[fid]);
}
};
// thread workspace
struct ThreadWSpace {
/*! \brief actual unit pointer */
std::vector<unsigned> rptr;
/*! \brief cut field */
std::vector<bst_float> cut;
// per thread histset
std::vector<HistSet> hset;
// initialize the hist set
inline void Init(const TrainParam &param, int nthread) {
hset.resize(nthread);
// cleanup statistics
for (int tid = 0; tid < nthread; ++tid) {
for (size_t i = 0; i < hset[tid].data.size(); ++i) {
hset[tid].data[i].Clear();
}
hset[tid].rptr = BeginPtr(rptr);
hset[tid].cut = BeginPtr(cut);
hset[tid].data.resize(cut.size(), TStats(param));
}
}
// aggregate all statistics to hset[0]
inline void Aggregate(void) {
bst_omp_uint nsize = static_cast<bst_omp_uint>(cut.size());
#pragma omp parallel for schedule(static)
for (bst_omp_uint i = 0; i < nsize; ++i) {
for (size_t tid = 1; tid < hset.size(); ++tid) {
hset[0].data[i].Add(hset[tid].data[i]);
}
}
}
/*! \brief clear the workspace */
inline void Clear(void) {
cut.clear(); rptr.resize(1); rptr[0] = 0;
}
/*! \brief total size */
inline size_t Size(void) const {
return rptr.size() - 1;
}
};
// workspace of thread
ThreadWSpace wspace;
// reducer for histogram
rabit::Reducer<TStats, TStats::Reduce> histred;
// set of working features
std::vector<bst_uint> fwork_set;
// update function implementation
virtual void Update(const std::vector<bst_gpair> &gpair,
IFMatrix *p_fmat,
const BoosterInfo &info,
RegTree *p_tree) {
this->InitData(gpair, *p_fmat, info.root_index, *p_tree);
this->InitWorkSet(p_fmat, *p_tree, &fwork_set);
for (int depth = 0; depth < param.max_depth; ++depth) {
// reset and propose candidate split
this->ResetPosAndPropose(gpair, p_fmat, info, fwork_set, *p_tree);
// create histogram
this->CreateHist(gpair, p_fmat, info, fwork_set, *p_tree);
// find split based on histogram statistics
this->FindSplit(depth, gpair, p_fmat, info, fwork_set, p_tree);
// reset position after split
this->ResetPositionAfterSplit(p_fmat, *p_tree);
this->UpdateQueueExpand(*p_tree);
// if nothing left to be expand, break
if (qexpand.size() == 0) break;
}
for (size_t i = 0; i < qexpand.size(); ++i) {
const int nid = qexpand[i];
(*p_tree)[nid].set_leaf(p_tree->stat(nid).base_weight * param.learning_rate);
}
}
// this function does two jobs
// (1) reset the position in array position, to be the latest leaf id
// (2) propose a set of candidate cuts and set wspace.rptr wspace.cut correctly
virtual void ResetPosAndPropose(const std::vector<bst_gpair> &gpair,
IFMatrix *p_fmat,
const BoosterInfo &info,
const std::vector <bst_uint> &fset,
const RegTree &tree) = 0;
// initialize the current working set of features in this round
virtual void InitWorkSet(IFMatrix *p_fmat,
const RegTree &tree,
std::vector<bst_uint> *p_fset) {
p_fset->resize(tree.param.num_feature);
for (size_t i = 0; i < p_fset->size(); ++i) {
(*p_fset)[i] = static_cast<unsigned>(i);
}
}
// reset position after split, this is not a must, depending on implementation
virtual void ResetPositionAfterSplit(IFMatrix *p_fmat,
const RegTree &tree) {
}
virtual void CreateHist(const std::vector<bst_gpair> &gpair,
IFMatrix *p_fmat,
const BoosterInfo &info,
const std::vector <bst_uint> &fset,
const RegTree &tree) = 0;
private:
inline void EnumerateSplit(const HistUnit &hist,
const TStats &node_sum,
bst_uint fid,
SplitEntry *best,
TStats *left_sum) {
if (hist.size == 0) return;
double root_gain = node_sum.CalcGain(param);
TStats s(param), c(param);
for (bst_uint i = 0; i < hist.size; ++i) {
s.Add(hist.data[i]);
if (s.sum_hess >= param.min_child_weight) {
c.SetSubstract(node_sum, s);
if (c.sum_hess >= param.min_child_weight) {
double loss_chg = s.CalcGain(param) + c.CalcGain(param) - root_gain;
if (best->Update(static_cast<float>(loss_chg), fid, hist.cut[i], false)) {
*left_sum = s;
}
}
}
}
s.Clear();
for (bst_uint i = hist.size - 1; i != 0; --i) {
s.Add(hist.data[i]);
if (s.sum_hess >= param.min_child_weight) {
c.SetSubstract(node_sum, s);
if (c.sum_hess >= param.min_child_weight) {
double loss_chg = s.CalcGain(param) + c.CalcGain(param) - root_gain;
if (best->Update(static_cast<float>(loss_chg), fid, hist.cut[i-1], true)) {
*left_sum = c;
}
}
}
}
}
inline void FindSplit(int depth,
const std::vector<bst_gpair> &gpair,
IFMatrix *p_fmat,
const BoosterInfo &info,
const std::vector <bst_uint> &fset,
RegTree *p_tree) {
const size_t num_feature = fset.size();
// get the best split condition for each node
std::vector<SplitEntry> sol(qexpand.size());
std::vector<TStats> left_sum(qexpand.size());
bst_omp_uint nexpand = static_cast<bst_omp_uint>(qexpand.size());
#pragma omp parallel for schedule(dynamic, 1)
for (bst_omp_uint wid = 0; wid < nexpand; ++wid) {
const int nid = qexpand[wid];
utils::Assert(node2workindex[nid] == static_cast<int>(wid),
"node2workindex inconsistent");
SplitEntry &best = sol[wid];
TStats &node_sum = wspace.hset[0][num_feature + wid * (num_feature + 1)].data[0];
for (size_t i = 0; i < fset.size(); ++i) {
EnumerateSplit(this->wspace.hset[0][i + wid * (num_feature+1)],
node_sum, fset[i], &best, &left_sum[wid]);
}
}
// get the best result, we can synchronize the solution
for (bst_omp_uint wid = 0; wid < nexpand; ++wid) {
const int nid = qexpand[wid];
const SplitEntry &best = sol[wid];
const TStats &node_sum = wspace.hset[0][num_feature + wid * (num_feature + 1)].data[0];
this->SetStats(p_tree, nid, node_sum);
// set up the values
p_tree->stat(nid).loss_chg = best.loss_chg;
// now we know the solution in snode[nid], set split
if (best.loss_chg > rt_eps) {
p_tree->AddChilds(nid);
(*p_tree)[nid].set_split(best.split_index(),
best.split_value, best.default_left());
// mark right child as 0, to indicate fresh leaf
(*p_tree)[(*p_tree)[nid].cleft()].set_leaf(0.0f, 0);
(*p_tree)[(*p_tree)[nid].cright()].set_leaf(0.0f, 0);
// right side sum
TStats right_sum;
right_sum.SetSubstract(node_sum, left_sum[wid]);
this->SetStats(p_tree, (*p_tree)[nid].cleft(), left_sum[wid]);
this->SetStats(p_tree, (*p_tree)[nid].cright(), right_sum);
} else {
(*p_tree)[nid].set_leaf(p_tree->stat(nid).base_weight * param.learning_rate);
}
}
}
inline void SetStats(RegTree *p_tree, int nid, const TStats &node_sum) {
p_tree->stat(nid).base_weight = static_cast<float>(node_sum.CalcWeight(param));
p_tree->stat(nid).sum_hess = static_cast<float>(node_sum.sum_hess);
node_sum.SetLeafVec(param, p_tree->leafvec(nid));
}
};
template<typename TStats>
class CQHistMaker: public HistMaker<TStats> {
protected:
struct HistEntry {
typename HistMaker<TStats>::HistUnit hist;
unsigned istart;
/*!
* \brief add a histogram to data,
* do linear scan, start from istart
*/
inline void Add(bst_float fv,
const std::vector<bst_gpair> &gpair,
const BoosterInfo &info,
const bst_uint ridx) {
while (istart < hist.size && !(fv < hist.cut[istart])) ++istart;
utils::Assert(istart != hist.size, "the bound variable must be max");
hist.data[istart].Add(gpair, info, ridx);
}
/*!
* \brief add a histogram to data,
* do linear scan, start from istart
*/
inline void Add(bst_float fv,
bst_gpair gstats) {
while (istart < hist.size && !(fv < hist.cut[istart])) ++istart;
utils::Assert(istart != hist.size, "the bound variable must be max");
hist.data[istart].Add(gstats);
}
};
// sketch type used for this
typedef utils::WXQuantileSketch<bst_float, bst_float> WXQSketch;
// initialize the work set of tree
virtual void InitWorkSet(IFMatrix *p_fmat,
const RegTree &tree,
std::vector<bst_uint> *p_fset) {
feat_helper.InitByCol(p_fmat, tree);
feat_helper.SampleCol(this->param.colsample_bytree, p_fset);
}
// code to create histogram
virtual void CreateHist(const std::vector<bst_gpair> &gpair,
IFMatrix *p_fmat,
const BoosterInfo &info,
const std::vector<bst_uint> &fset,
const RegTree &tree) {
// fill in reverse map
feat2workindex.resize(tree.param.num_feature);
std::fill(feat2workindex.begin(), feat2workindex.end(), -1);
for (size_t i = 0; i < fset.size(); ++i) {
feat2workindex[fset[i]] = static_cast<int>(i);
}
// start to work
this->wspace.Init(this->param, 1);
// if it is C++11, use lazy evaluation for Allreduce,
// to gain speedup in recovery
#if __cplusplus >= 201103L
auto lazy_get_hist = [&]()
#endif
{
thread_hist.resize(this->get_nthread());
// start accumulating statistics
utils::IIterator<ColBatch> *iter = p_fmat->ColIterator(fset);
iter->BeforeFirst();
while (iter->Next()) {
const ColBatch &batch = iter->Value();
// start enumeration
const bst_omp_uint nsize = static_cast<bst_omp_uint>(batch.size);
#pragma omp parallel for schedule(dynamic, 1)
for (bst_omp_uint i = 0; i < nsize; ++i) {
int offset = feat2workindex[batch.col_index[i]];
if (offset >= 0) {
this->UpdateHistCol(gpair, batch[i], info, tree,
fset, offset,
&thread_hist[omp_get_thread_num()]);
}
}
}
for (size_t i = 0; i < this->qexpand.size(); ++i) {
const int nid = this->qexpand[i];
const int wid = this->node2workindex[nid];
this->wspace.hset[0][fset.size() + wid * (fset.size()+1)]
.data[0] = node_stats[nid];
}
};
// sync the histogram
// if it is C++11, use lazy evaluation for Allreduce
#if __cplusplus >= 201103L
this->histred.Allreduce(BeginPtr(this->wspace.hset[0].data),
this->wspace.hset[0].data.size(), lazy_get_hist);
#else
this->histred.Allreduce(BeginPtr(this->wspace.hset[0].data), this->wspace.hset[0].data.size());
#endif
}
virtual void ResetPositionAfterSplit(IFMatrix *p_fmat,
const RegTree &tree) {
this->ResetPositionCol(this->qexpand, p_fmat, tree);
}
virtual void ResetPosAndPropose(const std::vector<bst_gpair> &gpair,
IFMatrix *p_fmat,
const BoosterInfo &info,
const std::vector<bst_uint> &fset,
const RegTree &tree) {
// fill in reverse map
feat2workindex.resize(tree.param.num_feature);
std::fill(feat2workindex.begin(), feat2workindex.end(), -1);
freal_set.clear();
for (size_t i = 0; i < fset.size(); ++i) {
if (feat_helper.Type(fset[i]) == 2) {
feat2workindex[fset[i]] = static_cast<int>(freal_set.size());
freal_set.push_back(fset[i]);
} else {
feat2workindex[fset[i]] = -2;
}
}
this->GetNodeStats(gpair, *p_fmat, tree, info,
&thread_stats, &node_stats);
sketchs.resize(this->qexpand.size() * freal_set.size());
for (size_t i = 0; i < sketchs.size(); ++i) {
sketchs[i].Init(info.num_row, this->param.sketch_eps);
}
// intitialize the summary array
summary_array.resize(sketchs.size());
// setup maximum size
unsigned max_size = this->param.max_sketch_size();
for (size_t i = 0; i < sketchs.size(); ++i) {
summary_array[i].Reserve(max_size);
}
// if it is C++11, use lazy evaluation for Allreduce
#if __cplusplus >= 201103L
auto lazy_get_summary = [&]()
#endif
{
// get smmary
thread_sketch.resize(this->get_nthread());
// number of rows in
const size_t nrows = p_fmat->buffered_rowset().size();
// start accumulating statistics
utils::IIterator<ColBatch> *iter = p_fmat->ColIterator(freal_set);
iter->BeforeFirst();
while (iter->Next()) {
const ColBatch &batch = iter->Value();
// start enumeration
const bst_omp_uint nsize = static_cast<bst_omp_uint>(batch.size);
#pragma omp parallel for schedule(dynamic, 1)
for (bst_omp_uint i = 0; i < nsize; ++i) {
int offset = feat2workindex[batch.col_index[i]];
if (offset >= 0) {
this->UpdateSketchCol(gpair, batch[i], tree,
node_stats,
freal_set, offset,
batch[i].length == nrows,
&thread_sketch[omp_get_thread_num()]);
}
}
}
for (size_t i = 0; i < sketchs.size(); ++i) {
utils::WXQuantileSketch<bst_float, bst_float>::SummaryContainer out;
sketchs[i].GetSummary(&out);
summary_array[i].SetPrune(out, max_size);
}
utils::Assert(summary_array.size() == sketchs.size(), "shape mismatch");
};
if (summary_array.size() != 0) {
size_t nbytes = WXQSketch::SummaryContainer::CalcMemCost(max_size);
#if __cplusplus >= 201103L
sreducer.Allreduce(BeginPtr(summary_array), nbytes, summary_array.size(), lazy_get_summary);
#else
sreducer.Allreduce(BeginPtr(summary_array), nbytes, summary_array.size());
#endif
}
// now we get the final result of sketch, setup the cut
this->wspace.cut.clear();
this->wspace.rptr.clear();
this->wspace.rptr.push_back(0);
for (size_t wid = 0; wid < this->qexpand.size(); ++wid) {
for (size_t i = 0; i < fset.size(); ++i) {
int offset = feat2workindex[fset[i]];
if (offset >= 0) {
const WXQSketch::Summary &a = summary_array[wid * freal_set.size() + offset];
for (size_t i = 1; i < a.size; ++i) {
bst_float cpt = a.data[i].value - rt_eps;
if (i == 1 || cpt > this->wspace.cut.back()) {
this->wspace.cut.push_back(cpt);
}
}
// push a value that is greater than anything
if (a.size != 0) {
bst_float cpt = a.data[a.size - 1].value;
// this must be bigger than last value in a scale
bst_float last = cpt + fabs(cpt) + rt_eps;
this->wspace.cut.push_back(last);
}
this->wspace.rptr.push_back(static_cast<unsigned>(this->wspace.cut.size()));
} else {
utils::Assert(offset == -2, "BUG in mark");
bst_float cpt = feat_helper.MaxValue(fset[i]);
this->wspace.cut.push_back(cpt + fabs(cpt) + rt_eps);
this->wspace.rptr.push_back(static_cast<unsigned>(this->wspace.cut.size()));
}
}
// reserve last value for global statistics
this->wspace.cut.push_back(0.0f);
this->wspace.rptr.push_back(static_cast<unsigned>(this->wspace.cut.size()));
}
utils::Assert(this->wspace.rptr.size() ==
(fset.size() + 1) * this->qexpand.size() + 1,
"cut space inconsistent");
}
private:
inline void UpdateHistCol(const std::vector<bst_gpair> &gpair,
const ColBatch::Inst &c,
const BoosterInfo &info,
const RegTree &tree,
const std::vector<bst_uint> &fset,
bst_uint fid_offset,
std::vector<HistEntry> *p_temp) {
if (c.length == 0) return;
// initialize sbuilder for use
std::vector<HistEntry> &hbuilder = *p_temp;
hbuilder.resize(tree.param.num_nodes);
for (size_t i = 0; i < this->qexpand.size(); ++i) {
const unsigned nid = this->qexpand[i];
const unsigned wid = this->node2workindex[nid];
hbuilder[nid].istart = 0;
hbuilder[nid].hist = this->wspace.hset[0][fid_offset + wid * (fset.size()+1)];
}
if (TStats::kSimpleStats != 0 && this->param.cache_opt != 0) {
const bst_uint kBuffer = 32;
bst_uint align_length = c.length / kBuffer * kBuffer;
int buf_position[kBuffer];
bst_gpair buf_gpair[kBuffer];
for (bst_uint j = 0; j < align_length; j += kBuffer) {
for (bst_uint i = 0; i < kBuffer; ++i) {
bst_uint ridx = c[j + i].index;
buf_position[i] = this->position[ridx];
buf_gpair[i] = gpair[ridx];
}
for (bst_uint i = 0; i < kBuffer; ++i) {
const int nid = buf_position[i];
if (nid >= 0) {
hbuilder[nid].Add(c[j + i].fvalue, buf_gpair[i]);
}
}
}
for (bst_uint j = align_length; j < c.length; ++j) {
const bst_uint ridx = c[j].index;
const int nid = this->position[ridx];
if (nid >= 0) {
hbuilder[nid].Add(c[j].fvalue, gpair[ridx]);
}
}
} else {
for (bst_uint j = 0; j < c.length; ++j) {
const bst_uint ridx = c[j].index;
const int nid = this->position[ridx];
if (nid >= 0) {
hbuilder[nid].Add(c[j].fvalue, gpair, info, ridx);
}
}
}
}
inline void UpdateSketchCol(const std::vector<bst_gpair> &gpair,
const ColBatch::Inst &c,
const RegTree &tree,
const std::vector<TStats> &nstats,
const std::vector<bst_uint> &frealset,
bst_uint offset,
bool col_full,
std::vector<BaseMaker::SketchEntry> *p_temp) {
if (c.length == 0) return;
// initialize sbuilder for use
std::vector<BaseMaker::SketchEntry> &sbuilder = *p_temp;
sbuilder.resize(tree.param.num_nodes);
for (size_t i = 0; i < this->qexpand.size(); ++i) {
const unsigned nid = this->qexpand[i];
const unsigned wid = this->node2workindex[nid];
sbuilder[nid].sum_total = 0.0f;
sbuilder[nid].sketch = &sketchs[wid * frealset.size() + offset];
}
if (!col_full) {
// first pass, get sum of weight, TODO, optimization to skip first pass
for (bst_uint j = 0; j < c.length; ++j) {
const bst_uint ridx = c[j].index;
const int nid = this->position[ridx];
if (nid >= 0) {
sbuilder[nid].sum_total += gpair[ridx].hess;
}
}
} else {
for (size_t i = 0; i < this->qexpand.size(); ++i) {
const unsigned nid = this->qexpand[i];
sbuilder[nid].sum_total = static_cast<bst_float>(nstats[nid].sum_hess);
}
}
// if only one value, no need to do second pass
if (c[0].fvalue == c[c.length-1].fvalue) {
for (size_t i = 0; i < this->qexpand.size(); ++i) {
const int nid = this->qexpand[i];
sbuilder[nid].sketch->Push(c[0].fvalue, static_cast<bst_float>(sbuilder[nid].sum_total));
}
return;
}
// two pass scan
unsigned max_size = this->param.max_sketch_size();
for (size_t i = 0; i < this->qexpand.size(); ++i) {
const int nid = this->qexpand[i];
sbuilder[nid].Init(max_size);
}
// second pass, build the sketch
if (TStats::kSimpleStats != 0 && this->param.cache_opt != 0) {
const bst_uint kBuffer = 32;
bst_uint align_length = c.length / kBuffer * kBuffer;
int buf_position[kBuffer];
bst_float buf_hess[kBuffer];
for (bst_uint j = 0; j < align_length; j += kBuffer) {
for (bst_uint i = 0; i < kBuffer; ++i) {
bst_uint ridx = c[j + i].index;
buf_position[i] = this->position[ridx];
buf_hess[i] = gpair[ridx].hess;
}
for (bst_uint i = 0; i < kBuffer; ++i) {
const int nid = buf_position[i];
if (nid >= 0) {
sbuilder[nid].Push(c[j + i].fvalue, buf_hess[i], max_size);
}
}
}
for (bst_uint j = align_length; j < c.length; ++j) {
const bst_uint ridx = c[j].index;
const int nid = this->position[ridx];
if (nid >= 0) {
sbuilder[nid].Push(c[j].fvalue, gpair[ridx].hess, max_size);
}
}
} else {
for (bst_uint j = 0; j < c.length; ++j) {
const bst_uint ridx = c[j].index;
const int nid = this->position[ridx];
if (nid >= 0) {
sbuilder[nid].Push(c[j].fvalue, gpair[ridx].hess, max_size);
}
}
}
for (size_t i = 0; i < this->qexpand.size(); ++i) {
const int nid = this->qexpand[i];
sbuilder[nid].Finalize(max_size);
}
}
// feature helper
BaseMaker::FMetaHelper feat_helper;
// temp space to map feature id to working index
std::vector<int> feat2workindex;
// set of index from fset that are real
std::vector<bst_uint> freal_set;
// thread temp data
std::vector< std::vector<BaseMaker::SketchEntry> > thread_sketch;
// used to hold statistics
std::vector< std::vector<TStats> > thread_stats;
// used to hold start pointer
std::vector< std::vector<HistEntry> > thread_hist;
// node statistics
std::vector<TStats> node_stats;
// summary array
std::vector<WXQSketch::SummaryContainer> summary_array;
// reducer for summary
rabit::SerializeReducer<WXQSketch::SummaryContainer> sreducer;
// per node, per feature sketch
std::vector< utils::WXQuantileSketch<bst_float, bst_float> > sketchs;
};
template<typename TStats>
class QuantileHistMaker: public HistMaker<TStats> {
protected:
typedef utils::WXQuantileSketch<bst_float, bst_float> WXQSketch;
virtual void ResetPosAndPropose(const std::vector<bst_gpair> &gpair,
IFMatrix *p_fmat,
const BoosterInfo &info,
const std::vector <bst_uint> &fset,
const RegTree &tree) {
// initialize the data structure
int nthread = BaseMaker::get_nthread();
sketchs.resize(this->qexpand.size() * tree.param.num_feature);
for (size_t i = 0; i < sketchs.size(); ++i) {
sketchs[i].Init(info.num_row, this->param.sketch_eps);
}
// start accumulating statistics
utils::IIterator<RowBatch> *iter = p_fmat->RowIterator();
iter->BeforeFirst();
while (iter->Next()) {
const RowBatch &batch = iter->Value();
// parallel convert to column major format
utils::ParallelGroupBuilder<SparseBatch::Entry> builder(&col_ptr, &col_data, &thread_col_ptr);
builder.InitBudget(tree.param.num_feature, nthread);
const bst_omp_uint nbatch = static_cast<bst_omp_uint>(batch.size);
#pragma omp parallel for schedule(static)
for (bst_omp_uint i = 0; i < nbatch; ++i) {
RowBatch::Inst inst = batch[i];
const bst_uint ridx = static_cast<bst_uint>(batch.base_rowid + i);
int nid = this->position[ridx];
if (nid >= 0) {
if (!tree[nid].is_leaf()) {
this->position[ridx] = nid = HistMaker<TStats>::NextLevel(inst, tree, nid);
}
if (this->node2workindex[nid] < 0) {
this->position[ridx] = ~nid;
} else {
for (bst_uint j = 0; j < inst.length; ++j) {
builder.AddBudget(inst[j].index, omp_get_thread_num());
}
}
}
}
builder.InitStorage();
#pragma omp parallel for schedule(static)
for (bst_omp_uint i = 0; i < nbatch; ++i) {
RowBatch::Inst inst = batch[i];
const bst_uint ridx = static_cast<bst_uint>(batch.base_rowid + i);
const int nid = this->position[ridx];
if (nid >= 0) {
for (bst_uint j = 0; j < inst.length; ++j) {
builder.Push(inst[j].index,
SparseBatch::Entry(nid, inst[j].fvalue),
omp_get_thread_num());
}
}
}
// start putting things into sketch
const bst_omp_uint nfeat = col_ptr.size() - 1;
#pragma omp parallel for schedule(dynamic, 1)
for (bst_omp_uint k = 0; k < nfeat; ++k) {
for (size_t i = col_ptr[k]; i < col_ptr[k+1]; ++i) {
const SparseBatch::Entry &e = col_data[i];
const int wid = this->node2workindex[e.index];
sketchs[wid * tree.param.num_feature + k].Push(e.fvalue, gpair[e.index].hess);
}
}
}
// setup maximum size
unsigned max_size = this->param.max_sketch_size();
// synchronize sketch
summary_array.resize(sketchs.size());
for (size_t i = 0; i < sketchs.size(); ++i) {
utils::WQuantileSketch<bst_float, bst_float>::SummaryContainer out;
sketchs[i].GetSummary(&out);
summary_array[i].Reserve(max_size);
summary_array[i].SetPrune(out, max_size);
}
size_t nbytes = WXQSketch::SummaryContainer::CalcMemCost(max_size);
sreducer.Allreduce(BeginPtr(summary_array), nbytes, summary_array.size());
// now we get the final result of sketch, setup the cut
this->wspace.cut.clear();
this->wspace.rptr.clear();
this->wspace.rptr.push_back(0);
for (size_t wid = 0; wid < this->qexpand.size(); ++wid) {
for (int fid = 0; fid < tree.param.num_feature; ++fid) {
const WXQSketch::Summary &a = summary_array[wid * tree.param.num_feature + fid];
for (size_t i = 1; i < a.size; ++i) {
bst_float cpt = a.data[i].value - rt_eps;
if (i == 1 || cpt > this->wspace.cut.back()) {
this->wspace.cut.push_back(cpt);
}
}
// push a value that is greater than anything
if (a.size != 0) {
bst_float cpt = a.data[a.size - 1].value;
// this must be bigger than last value in a scale
bst_float last = cpt + fabs(cpt) + rt_eps;
this->wspace.cut.push_back(last);
}
this->wspace.rptr.push_back(this->wspace.cut.size());
}
// reserve last value for global statistics
this->wspace.cut.push_back(0.0f);
this->wspace.rptr.push_back(this->wspace.cut.size());
}
utils::Assert(this->wspace.rptr.size() ==
(tree.param.num_feature + 1) * this->qexpand.size() + 1,
"cut space inconsistent");
}
private:
// summary array
std::vector<WXQSketch::SummaryContainer> summary_array;
// reducer for summary
rabit::SerializeReducer<WXQSketch::SummaryContainer> sreducer;
// local temp column data structure
std::vector<size_t> col_ptr;
// local storage of column data
std::vector<SparseBatch::Entry> col_data;
std::vector< std::vector<size_t> > thread_col_ptr;
// per node, per feature sketch
std::vector< utils::WQuantileSketch<bst_float, bst_float> > sketchs;
};
} // namespace tree
} // namespace xgboost
#endif // XGBOOST_TREE_UPDATER_HISTMAKER_INL_HPP_

87
old_src/tree/updater_prune-inl.hpp
View File

@@ -1,87 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file updater_prune-inl.hpp
* \brief prune a tree given the statistics
* \author Tianqi Chen
*/
#ifndef XGBOOST_TREE_UPDATER_PRUNE_INL_HPP_
#define XGBOOST_TREE_UPDATER_PRUNE_INL_HPP_
#include <vector>
#include "./param.h"
#include "./updater.h"
#include "./updater_sync-inl.hpp"
namespace xgboost {
namespace tree {
/*! \brief pruner that prunes a tree after growing finishes */
class TreePruner: public IUpdater {
public:
virtual ~TreePruner(void) {}
// set training parameter
virtual void SetParam(const char *name, const char *val) {
using namespace std;
param.SetParam(name, val);
syncher.SetParam(name, val);
if (!strcmp(name, "silent")) silent = atoi(val);
}
// update the tree, do pruning
virtual void Update(const std::vector<bst_gpair> &gpair,
IFMatrix *p_fmat,
const BoosterInfo &info,
const std::vector<RegTree*> &trees) {
// rescale learning rate according to size of trees
float lr = param.learning_rate;
param.learning_rate = lr / trees.size();
for (size_t i = 0; i < trees.size(); ++i) {
this->DoPrune(*trees[i]);
}
param.learning_rate = lr;
syncher.Update(gpair, p_fmat, info, trees);
}
private:
// try to prune off current leaf
inline int TryPruneLeaf(RegTree &tree, int nid, int depth, int npruned) { // NOLINT(*)
if (tree[nid].is_root()) return npruned;
int pid = tree[nid].parent();
RegTree::NodeStat &s = tree.stat(pid);
++s.leaf_child_cnt;
if (s.leaf_child_cnt >= 2 && param.need_prune(s.loss_chg, depth - 1)) {
// need to be pruned
tree.ChangeToLeaf(pid, param.learning_rate * s.base_weight);
// tail recursion
return this->TryPruneLeaf(tree, pid, depth - 1, npruned+2);
} else {
return npruned;
}
}
/*! \brief do pruning of a tree */
inline void DoPrune(RegTree &tree) { // NOLINT(*)
int npruned = 0;
// initialize auxiliary statistics
for (int nid = 0; nid < tree.param.num_nodes; ++nid) {
tree.stat(nid).leaf_child_cnt = 0;
}
for (int nid = 0; nid < tree.param.num_nodes; ++nid) {
if (tree[nid].is_leaf()) {
npruned = this->TryPruneLeaf(tree, nid, tree.GetDepth(nid), npruned);
}
}
if (silent == 0) {
utils::Printf("tree pruning end, %d roots, %d extra nodes, %d pruned nodes, max_depth=%d\n",
tree.param.num_roots, tree.num_extra_nodes(), npruned, tree.MaxDepth());
}
}
private:
// synchronizer
TreeSyncher syncher;
// shutup
int silent;
// training parameter
TrainParam param;
};
} // namespace tree
} // namespace xgboost
#endif // XGBOOST_TREE_UPDATER_PRUNE_INL_HPP_

157
old_src/tree/updater_refresh-inl.hpp
View File

@@ -1,157 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file updater_refresh-inl.hpp
* \brief refresh the statistics and leaf value on the tree on the dataset
* \author Tianqi Chen
*/
#ifndef XGBOOST_TREE_UPDATER_REFRESH_INL_HPP_
#define XGBOOST_TREE_UPDATER_REFRESH_INL_HPP_
#include <vector>
#include <limits>
#include "../sync/sync.h"
#include "./param.h"
#include "./updater.h"
#include "../utils/omp.h"
namespace xgboost {
namespace tree {
/*! \brief pruner that prunes a tree after growing finishs */
template<typename TStats>
class TreeRefresher: public IUpdater {
public:
virtual ~TreeRefresher(void) {}
// set training parameter
virtual void SetParam(const char *name, const char *val) {
param.SetParam(name, val);
}
// update the tree, do pruning
virtual void Update(const std::vector<bst_gpair> &gpair,
IFMatrix *p_fmat,
const BoosterInfo &info,
const std::vector<RegTree*> &trees) {
if (trees.size() == 0) return;
// number of threads
// thread temporal space
std::vector< std::vector<TStats> > stemp;
std::vector<RegTree::FVec> fvec_temp;
// setup temp space for each thread
int nthread;
#pragma omp parallel
{
nthread = omp_get_num_threads();
}
fvec_temp.resize(nthread, RegTree::FVec());
stemp.resize(nthread, std::vector<TStats>());
#pragma omp parallel
{
int tid = omp_get_thread_num();
int num_nodes = 0;
for (size_t i = 0; i < trees.size(); ++i) {
num_nodes += trees[i]->param.num_nodes;
}
stemp[tid].resize(num_nodes, TStats(param));
std::fill(stemp[tid].begin(), stemp[tid].end(), TStats(param));
fvec_temp[tid].Init(trees[0]->param.num_feature);
}
// if it is C++11, use lazy evaluation for Allreduce,
// to gain speedup in recovery
#if __cplusplus >= 201103L
auto lazy_get_stats = [&]()
#endif
{
// start accumulating statistics
utils::IIterator<RowBatch> *iter = p_fmat->RowIterator();
iter->BeforeFirst();
while (iter->Next()) {
const RowBatch &batch = iter->Value();
utils::Check(batch.size < std::numeric_limits<unsigned>::max(),
"too large batch size ");
const bst_omp_uint nbatch = static_cast<bst_omp_uint>(batch.size);
#pragma omp parallel for schedule(static)
for (bst_omp_uint i = 0; i < nbatch; ++i) {
RowBatch::Inst inst = batch[i];
const int tid = omp_get_thread_num();
const bst_uint ridx = static_cast<bst_uint>(batch.base_rowid + i);
RegTree::FVec &feats = fvec_temp[tid];
feats.Fill(inst);
int offset = 0;
for (size_t j = 0; j < trees.size(); ++j) {
AddStats(*trees[j], feats, gpair, info, ridx,
BeginPtr(stemp[tid]) + offset);
offset += trees[j]->param.num_nodes;
}
feats.Drop(inst);
}
}
// aggregate the statistics
int num_nodes = static_cast<int>(stemp[0].size());
#pragma omp parallel for schedule(static)
for (int nid = 0; nid < num_nodes; ++nid) {
for (int tid = 1; tid < nthread; ++tid) {
stemp[0][nid].Add(stemp[tid][nid]);
}
}
};
#if __cplusplus >= 201103L
reducer.Allreduce(BeginPtr(stemp[0]), stemp[0].size(), lazy_get_stats);
#else
reducer.Allreduce(BeginPtr(stemp[0]), stemp[0].size());
#endif
// rescale learning rate according to size of trees
float lr = param.learning_rate;
param.learning_rate = lr / trees.size();
int offset = 0;
for (size_t i = 0; i < trees.size(); ++i) {
for (int rid = 0; rid < trees[i]->param.num_roots; ++rid) {
this->Refresh(BeginPtr(stemp[0]) + offset, rid, trees[i]);
}
offset += trees[i]->param.num_nodes;
}
// set learning rate back
param.learning_rate = lr;
}
private:
inline static void AddStats(const RegTree &tree,
const RegTree::FVec &feat,
const std::vector<bst_gpair> &gpair,
const BoosterInfo &info,
const bst_uint ridx,
TStats *gstats) {
// start from groups that belongs to current data
int pid = static_cast<int>(info.GetRoot(ridx));
gstats[pid].Add(gpair, info, ridx);
// tranverse tree
while (!tree[pid].is_leaf()) {
unsigned split_index = tree[pid].split_index();
pid = tree.GetNext(pid, feat.fvalue(split_index), feat.is_missing(split_index));
gstats[pid].Add(gpair, info, ridx);
}
}
inline void Refresh(const TStats *gstats,
int nid, RegTree *p_tree) {
RegTree &tree = *p_tree;
tree.stat(nid).base_weight = static_cast<float>(gstats[nid].CalcWeight(param));
tree.stat(nid).sum_hess = static_cast<float>(gstats[nid].sum_hess);
gstats[nid].SetLeafVec(param, tree.leafvec(nid));
if (tree[nid].is_leaf()) {
tree[nid].set_leaf(tree.stat(nid).base_weight * param.learning_rate);
} else {
tree.stat(nid).loss_chg = static_cast<float>(
gstats[tree[nid].cleft()].CalcGain(param) +
gstats[tree[nid].cright()].CalcGain(param) -
gstats[nid].CalcGain(param));
this->Refresh(gstats, tree[nid].cleft(), p_tree);
this->Refresh(gstats, tree[nid].cright(), p_tree);
}
}
// training parameter
TrainParam param;
// reducer
rabit::Reducer<TStats, TStats::Reduce> reducer;
};
} // namespace tree
} // namespace xgboost
#endif // XGBOOST_TREE_UPDATER_REFRESH_INL_HPP_

399
old_src/tree/updater_skmaker-inl.hpp
View File

@@ -1,399 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file updater_skmaker-inl.hpp
* \brief use approximation sketch to construct a tree,
a refresh is needed to make the statistics exactly correct
* \author Tianqi Chen
*/
#ifndef XGBOOST_TREE_UPDATER_SKMAKER_INL_HPP_
#define XGBOOST_TREE_UPDATER_SKMAKER_INL_HPP_
#include <vector>
#include <algorithm>
#include "../sync/sync.h"
#include "../utils/quantile.h"
#include "./updater_basemaker-inl.hpp"
namespace xgboost {
namespace tree {
class SketchMaker: public BaseMaker {
public:
virtual ~SketchMaker(void) {}
virtual void Update(const std::vector<bst_gpair> &gpair,
IFMatrix *p_fmat,
const BoosterInfo &info,
const std::vector<RegTree*> &trees) {
// rescale learning rate according to size of trees
float lr = param.learning_rate;
param.learning_rate = lr / trees.size();
// build tree
for (size_t i = 0; i < trees.size(); ++i) {
this->Update(gpair, p_fmat, info, trees[i]);
}
param.learning_rate = lr;
}
protected:
inline void Update(const std::vector<bst_gpair> &gpair,
IFMatrix *p_fmat,
const BoosterInfo &info,
RegTree *p_tree) {
this->InitData(gpair, *p_fmat, info.root_index, *p_tree);
for (int depth = 0; depth < param.max_depth; ++depth) {
this->GetNodeStats(gpair, *p_fmat, *p_tree, info,
&thread_stats, &node_stats);
this->BuildSketch(gpair, p_fmat, info, *p_tree);
this->SyncNodeStats();
this->FindSplit(depth, gpair, p_fmat, info, p_tree);
this->ResetPositionCol(qexpand, p_fmat, *p_tree);
this->UpdateQueueExpand(*p_tree);
// if nothing left to be expand, break
if (qexpand.size() == 0) break;
}
if (qexpand.size() != 0) {
this->GetNodeStats(gpair, *p_fmat, *p_tree, info,
&thread_stats, &node_stats);
this->SyncNodeStats();
}
// set all statistics correctly
for (int nid = 0; nid < p_tree->param.num_nodes; ++nid) {
this->SetStats(nid, node_stats[nid], p_tree);
if (!(*p_tree)[nid].is_leaf()) {
p_tree->stat(nid).loss_chg = static_cast<float>(
node_stats[(*p_tree)[nid].cleft()].CalcGain(param) +
node_stats[(*p_tree)[nid].cright()].CalcGain(param) -
node_stats[nid].CalcGain(param));
}
}
// set left leaves
for (size_t i = 0; i < qexpand.size(); ++i) {
const int nid = qexpand[i];
(*p_tree)[nid].set_leaf(p_tree->stat(nid).base_weight * param.learning_rate);
}
}
// define the sketch we want to use
typedef utils::WXQuantileSketch<bst_float, bst_float> WXQSketch;
private:
// statistics needed in the gradient calculation
struct SKStats {
/*! \brief sum of all positive gradient */
double pos_grad;
/*! \brief sum of all negative gradient */
double neg_grad;
/*! \brief sum of hessian statistics */
double sum_hess;
SKStats(void) {}
// constructor
explicit SKStats(const TrainParam &param) {
this->Clear();
}
/*! \brief clear the statistics */
inline void Clear(void) {
neg_grad = pos_grad = sum_hess = 0.0f;
}
// accumulate statistics
inline void Add(const std::vector<bst_gpair> &gpair,
const BoosterInfo &info,
bst_uint ridx) {
const bst_gpair &b = gpair[ridx];
if (b.grad >= 0.0f) {
pos_grad += b.grad;
} else {
neg_grad -= b.grad;
}
sum_hess += b.hess;
}
/*! \brief calculate gain of the solution */
inline double CalcGain(const TrainParam &param) const {
return param.CalcGain(pos_grad - neg_grad, sum_hess);
}
/*! \brief set current value to a - b */
inline void SetSubstract(const SKStats &a, const SKStats &b) {
pos_grad = a.pos_grad - b.pos_grad;
neg_grad = a.neg_grad - b.neg_grad;
sum_hess = a.sum_hess - b.sum_hess;
}
// calculate leaf weight
inline double CalcWeight(const TrainParam &param) const {
return param.CalcWeight(pos_grad - neg_grad, sum_hess);
}
/*! \brief add statistics to the data */
inline void Add(const SKStats &b) {
pos_grad += b.pos_grad;
neg_grad += b.neg_grad;
sum_hess += b.sum_hess;
}
/*! \brief same as add, reduce is used in All Reduce */
inline static void Reduce(SKStats &a, const SKStats &b) { // NOLINT(*)
a.Add(b);
}
/*! \brief set leaf vector value based on statistics */
inline void SetLeafVec(const TrainParam &param, bst_float *vec) const {
}
};
inline void BuildSketch(const std::vector<bst_gpair> &gpair,
IFMatrix *p_fmat,
const BoosterInfo &info,
const RegTree &tree) {
sketchs.resize(this->qexpand.size() * tree.param.num_feature * 3);
for (size_t i = 0; i < sketchs.size(); ++i) {
sketchs[i].Init(info.num_row, this->param.sketch_eps);
}
thread_sketch.resize(this->get_nthread());
// number of rows in
const size_t nrows = p_fmat->buffered_rowset().size();
// start accumulating statistics
utils::IIterator<ColBatch> *iter = p_fmat->ColIterator();
iter->BeforeFirst();
while (iter->Next()) {
const ColBatch &batch = iter->Value();
// start enumeration
const bst_omp_uint nsize = static_cast<bst_omp_uint>(batch.size);
#pragma omp parallel for schedule(dynamic, 1)
for (bst_omp_uint i = 0; i < nsize; ++i) {
this->UpdateSketchCol(gpair, batch[i], tree,
node_stats,
batch.col_index[i],
batch[i].length == nrows,
&thread_sketch[omp_get_thread_num()]);
}
}
// setup maximum size
unsigned max_size = param.max_sketch_size();
// synchronize sketch
summary_array.resize(sketchs.size());
for (size_t i = 0; i < sketchs.size(); ++i) {
utils::WXQuantileSketch<bst_float, bst_float>::SummaryContainer out;
sketchs[i].GetSummary(&out);
summary_array[i].Reserve(max_size);
summary_array[i].SetPrune(out, max_size);
}
size_t nbytes = WXQSketch::SummaryContainer::CalcMemCost(max_size);
sketch_reducer.Allreduce(BeginPtr(summary_array), nbytes, summary_array.size());
}
// update sketch information in column fid
inline void UpdateSketchCol(const std::vector<bst_gpair> &gpair,
const ColBatch::Inst &c,
const RegTree &tree,
const std::vector<SKStats> &nstats,
bst_uint fid,
bool col_full,
std::vector<SketchEntry> *p_temp) {
if (c.length == 0) return;
// initialize sbuilder for use
std::vector<SketchEntry> &sbuilder = *p_temp;
sbuilder.resize(tree.param.num_nodes * 3);
for (size_t i = 0; i < this->qexpand.size(); ++i) {
const unsigned nid = this->qexpand[i];
const unsigned wid = this->node2workindex[nid];
for (int k = 0; k < 3; ++k) {
sbuilder[3 * nid + k].sum_total = 0.0f;
sbuilder[3 * nid + k].sketch = &sketchs[(wid * tree.param.num_feature + fid) * 3 + k];
}
}
if (!col_full) {
for (bst_uint j = 0; j < c.length; ++j) {
const bst_uint ridx = c[j].index;
const int nid = this->position[ridx];
if (nid >= 0) {
const bst_gpair &e = gpair[ridx];
if (e.grad >= 0.0f) {
sbuilder[3 * nid + 0].sum_total += e.grad;
} else {
sbuilder[3 * nid + 1].sum_total -= e.grad;
}
sbuilder[3 * nid + 2].sum_total += e.hess;
}
}
} else {
for (size_t i = 0; i < this->qexpand.size(); ++i) {
const unsigned nid = this->qexpand[i];
sbuilder[3 * nid + 0].sum_total = static_cast<bst_float>(nstats[nid].pos_grad);
sbuilder[3 * nid + 1].sum_total = static_cast<bst_float>(nstats[nid].neg_grad);
sbuilder[3 * nid + 2].sum_total = static_cast<bst_float>(nstats[nid].sum_hess);
}
}
// if only one value, no need to do second pass
if (c[0].fvalue == c[c.length-1].fvalue) {
for (size_t i = 0; i < this->qexpand.size(); ++i) {
const int nid = this->qexpand[i];
for (int k = 0; k < 3; ++k) {
sbuilder[3 * nid + k].sketch->Push(c[0].fvalue,
static_cast<bst_float>(
sbuilder[3 * nid + k].sum_total));
}
}
return;
}
// two pass scan
unsigned max_size = param.max_sketch_size();
for (size_t i = 0; i < this->qexpand.size(); ++i) {
const int nid = this->qexpand[i];
for (int k = 0; k < 3; ++k) {
sbuilder[3 * nid + k].Init(max_size);
}
}
// second pass, build the sketch
for (bst_uint j = 0; j < c.length; ++j) {
const bst_uint ridx = c[j].index;
const int nid = this->position[ridx];
if (nid >= 0) {
const bst_gpair &e = gpair[ridx];
if (e.grad >= 0.0f) {
sbuilder[3 * nid + 0].Push(c[j].fvalue, e.grad, max_size);
} else {
sbuilder[3 * nid + 1].Push(c[j].fvalue, -e.grad, max_size);
}
sbuilder[3 * nid + 2].Push(c[j].fvalue, e.hess, max_size);
}
}
for (size_t i = 0; i < this->qexpand.size(); ++i) {
const int nid = this->qexpand[i];
for (int k = 0; k < 3; ++k) {
sbuilder[3 * nid + k].Finalize(max_size);
}
}
}
inline void SyncNodeStats(void) {
utils::Assert(qexpand.size() != 0, "qexpand must not be empty");
std::vector<SKStats> tmp(qexpand.size());
for (size_t i = 0; i < qexpand.size(); ++i) {
tmp[i] = node_stats[qexpand[i]];
}
stats_reducer.Allreduce(BeginPtr(tmp), tmp.size());
for (size_t i = 0; i < qexpand.size(); ++i) {
node_stats[qexpand[i]] = tmp[i];
}
}
inline void FindSplit(int depth,
const std::vector<bst_gpair> &gpair,
IFMatrix *p_fmat,
const BoosterInfo &info,
RegTree *p_tree) {
const bst_uint num_feature = p_tree->param.num_feature;
// get the best split condition for each node
std::vector<SplitEntry> sol(qexpand.size());
bst_omp_uint nexpand = static_cast<bst_omp_uint>(qexpand.size());
#pragma omp parallel for schedule(dynamic, 1)
for (bst_omp_uint wid = 0; wid < nexpand; ++wid) {
const int nid = qexpand[wid];
utils::Assert(node2workindex[nid] == static_cast<int>(wid),
"node2workindex inconsistent");
SplitEntry &best = sol[wid];
for (bst_uint fid = 0; fid < num_feature; ++fid) {
unsigned base = (wid * p_tree->param.num_feature + fid) * 3;
EnumerateSplit(summary_array[base + 0],
summary_array[base + 1],
summary_array[base + 2],
node_stats[nid], fid, &best);
}
}
// get the best result, we can synchronize the solution
for (bst_omp_uint wid = 0; wid < nexpand; ++wid) {
const int nid = qexpand[wid];
const SplitEntry &best = sol[wid];
// set up the values
p_tree->stat(nid).loss_chg = best.loss_chg;
this->SetStats(nid, node_stats[nid], p_tree);
// now we know the solution in snode[nid], set split
if (best.loss_chg > rt_eps) {
p_tree->AddChilds(nid);
(*p_tree)[nid].set_split(best.split_index(),
best.split_value, best.default_left());
// mark right child as 0, to indicate fresh leaf
(*p_tree)[(*p_tree)[nid].cleft()].set_leaf(0.0f, 0);
(*p_tree)[(*p_tree)[nid].cright()].set_leaf(0.0f, 0);
} else {
(*p_tree)[nid].set_leaf(p_tree->stat(nid).base_weight * param.learning_rate);
}
}
}
// set statistics on ptree
inline void SetStats(int nid, const SKStats &node_sum, RegTree *p_tree) {
p_tree->stat(nid).base_weight = static_cast<float>(node_sum.CalcWeight(param));
p_tree->stat(nid).sum_hess = static_cast<float>(node_sum.sum_hess);
node_sum.SetLeafVec(param, p_tree->leafvec(nid));
}
inline void EnumerateSplit(const WXQSketch::Summary &pos_grad,
const WXQSketch::Summary &neg_grad,
const WXQSketch::Summary &sum_hess,
const SKStats &node_sum,
bst_uint fid,
SplitEntry *best) {
if (sum_hess.size == 0) return;
double root_gain = node_sum.CalcGain(param);
std::vector<bst_float> fsplits;
for (size_t i = 0; i < pos_grad.size; ++i) {
fsplits.push_back(pos_grad.data[i].value);
}
for (size_t i = 0; i < neg_grad.size; ++i) {
fsplits.push_back(neg_grad.data[i].value);
}
for (size_t i = 0; i < sum_hess.size; ++i) {
fsplits.push_back(sum_hess.data[i].value);
}
std::sort(fsplits.begin(), fsplits.end());
fsplits.resize(std::unique(fsplits.begin(), fsplits.end()) - fsplits.begin());
// sum feature
SKStats feat_sum;
feat_sum.pos_grad = pos_grad.data[pos_grad.size - 1].rmax;
feat_sum.neg_grad = neg_grad.data[neg_grad.size - 1].rmax;
feat_sum.sum_hess = sum_hess.data[sum_hess.size - 1].rmax;
size_t ipos = 0, ineg = 0, ihess = 0;
for (size_t i = 1; i < fsplits.size(); ++i) {
WXQSketch::Entry pos = pos_grad.Query(fsplits[i], ipos);
WXQSketch::Entry neg = neg_grad.Query(fsplits[i], ineg);
WXQSketch::Entry hess = sum_hess.Query(fsplits[i], ihess);
SKStats s, c;
s.pos_grad = 0.5f * (pos.rmin + pos.rmax - pos.wmin);
s.neg_grad = 0.5f * (neg.rmin + neg.rmax - neg.wmin);
s.sum_hess = 0.5f * (hess.rmin + hess.rmax - hess.wmin);
c.SetSubstract(node_sum, s);
// forward
if (s.sum_hess >= param.min_child_weight &&
c.sum_hess >= param.min_child_weight) {
double loss_chg = s.CalcGain(param) + c.CalcGain(param) - root_gain;
best->Update(static_cast<bst_float>(loss_chg), fid, fsplits[i], false);
}
// backward
c.SetSubstract(feat_sum, s);
s.SetSubstract(node_sum, c);
if (s.sum_hess >= param.min_child_weight &&
c.sum_hess >= param.min_child_weight) {
double loss_chg = s.CalcGain(param) + c.CalcGain(param) - root_gain;
best->Update(static_cast<bst_float>(loss_chg), fid, fsplits[i], true);
}
}
{
// all including
SKStats s = feat_sum, c;
c.SetSubstract(node_sum, s);
if (s.sum_hess >= param.min_child_weight &&
c.sum_hess >= param.min_child_weight) {
bst_float cpt = fsplits.back();
double loss_chg = s.CalcGain(param) + c.CalcGain(param) - root_gain;
best->Update(static_cast<bst_float>(loss_chg), fid, cpt + fabsf(cpt) + 1.0f, false);
}
}
}
// thread temp data
// used to hold temporal sketch
std::vector< std::vector<SketchEntry> > thread_sketch;
// used to hold statistics
std::vector< std::vector<SKStats> > thread_stats;
// node statistics
std::vector<SKStats> node_stats;
// summary array
std::vector<WXQSketch::SummaryContainer> summary_array;
// reducer for summary
rabit::Reducer<SKStats, SKStats::Reduce> stats_reducer;
// reducer for summary
rabit::SerializeReducer<WXQSketch::SummaryContainer> sketch_reducer;
// per node, per feature sketch
std::vector< utils::WXQuantileSketch<bst_float, bst_float> > sketchs;
};
} // namespace tree
} // namespace xgboost
#endif // XGBOOST_TREE_UPDATER_SKMAKER_INL_HPP_

267
old_src/utils/base64-inl.h
View File

@@ -1,267 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file base64.h
* \brief data stream support to input and output from/to base64 stream
* base64 is easier to store and pass as text format in mapreduce
* \author Tianqi Chen
*/
#ifndef XGBOOST_UTILS_BASE64_INL_H_
#define XGBOOST_UTILS_BASE64_INL_H_
#include <cctype>
#include <cstdio>
#include <string>
#include "./io.h"
namespace xgboost {
namespace utils {
/*! \brief buffer reader of the stream that allows you to get */
class StreamBufferReader {
public:
explicit StreamBufferReader(size_t buffer_size)
:stream_(NULL),
read_len_(1), read_ptr_(1) {
buffer_.resize(buffer_size);
}
/*!
* \brief set input stream
*/
inline void set_stream(IStream *stream) {
stream_ = stream;
read_len_ = read_ptr_ = 1;
}
/*!
* \brief allows quick read using get char
*/
inline char GetChar(void) {
while (true) {
if (read_ptr_ < read_len_) {
return buffer_[read_ptr_++];
} else {
read_len_ = stream_->Read(&buffer_[0], buffer_.length());
if (read_len_ == 0) return EOF;
read_ptr_ = 0;
}
}
}
/*! \brief whether we are reaching the end of file */
inline bool AtEnd(void) const {
return read_len_ == 0;
}
private:
/*! \brief the underlying stream */
IStream *stream_;
/*! \brief buffer to hold data */
std::string buffer_;
/*! \brief length of valid data in buffer */
size_t read_len_;
/*! \brief pointer in the buffer */
size_t read_ptr_;
};
/*! \brief namespace of base64 decoding and encoding table */
namespace base64 {
const char DecodeTable[] = {
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
62, // '+'
0, 0, 0,
63, // '/'
52, 53, 54, 55, 56, 57, 58, 59, 60, 61, // '0'-'9'
0, 0, 0, 0, 0, 0, 0,
0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12,
13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, // 'A'-'Z'
0, 0, 0, 0, 0, 0,
26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38,
39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, // 'a'-'z'
};
static const char EncodeTable[] =
"ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/";
} // namespace base64
/*! \brief the stream that reads from base64, note we take from file pointers */
class Base64InStream: public IStream {
public:
explicit Base64InStream(IStream *fs) : reader_(256) {
reader_.set_stream(fs);
num_prev = 0; tmp_ch = 0;
}
/*!
* \brief initialize the stream position to beginning of next base64 stream
* call this function before actually start read
*/
inline void InitPosition(void) {
// get a character
do {
tmp_ch = reader_.GetChar();
} while (isspace(tmp_ch));
}
/*! \brief whether current position is end of a base64 stream */
inline bool IsEOF(void) const {
return num_prev == 0 && (tmp_ch == EOF || isspace(tmp_ch));
}
virtual size_t Read(void *ptr, size_t size) {
using base64::DecodeTable;
if (size == 0) return 0;
// use tlen to record left size
size_t tlen = size;
unsigned char *cptr = static_cast<unsigned char*>(ptr);
// if anything left, load from previous buffered result
if (num_prev != 0) {
if (num_prev == 2) {
if (tlen >= 2) {
*cptr++ = buf_prev[0];
*cptr++ = buf_prev[1];
tlen -= 2;
num_prev = 0;
} else {
// assert tlen == 1
*cptr++ = buf_prev[0]; --tlen;
buf_prev[0] = buf_prev[1];
num_prev = 1;
}
} else {
// assert num_prev == 1
*cptr++ = buf_prev[0]; --tlen; num_prev = 0;
}
}
if (tlen == 0) return size;
int nvalue;
// note: everything goes with 4 bytes in Base64
// so we process 4 bytes a unit
while (tlen && tmp_ch != EOF && !isspace(tmp_ch)) {
// first byte
nvalue = DecodeTable[tmp_ch] << 18;
{
// second byte
utils::Check((tmp_ch = reader_.GetChar(), tmp_ch != EOF && !isspace(tmp_ch)),
"invalid base64 format");
nvalue |= DecodeTable[tmp_ch] << 12;
*cptr++ = (nvalue >> 16) & 0xFF; --tlen;
}
{
// third byte
utils::Check((tmp_ch = reader_.GetChar(), tmp_ch != EOF && !isspace(tmp_ch)),
"invalid base64 format");
// handle termination
if (tmp_ch == '=') {
utils::Check((tmp_ch = reader_.GetChar(), tmp_ch == '='), "invalid base64 format");
utils::Check((tmp_ch = reader_.GetChar(), tmp_ch == EOF || isspace(tmp_ch)),
"invalid base64 format");
break;
}
nvalue |= DecodeTable[tmp_ch] << 6;
if (tlen) {
*cptr++ = (nvalue >> 8) & 0xFF; --tlen;
} else {
buf_prev[num_prev++] = (nvalue >> 8) & 0xFF;
}
}
{
// fourth byte
utils::Check((tmp_ch = reader_.GetChar(), tmp_ch != EOF && !isspace(tmp_ch)),
"invalid base64 format");
if (tmp_ch == '=') {
utils::Check((tmp_ch = reader_.GetChar(), tmp_ch == EOF || isspace(tmp_ch)),
"invalid base64 format");
break;
}
nvalue |= DecodeTable[tmp_ch];
if (tlen) {
*cptr++ = nvalue & 0xFF; --tlen;
} else {
buf_prev[num_prev ++] = nvalue & 0xFF;
}
}
// get next char
tmp_ch = reader_.GetChar();
}
if (kStrictCheck) {
utils::Check(tlen == 0, "Base64InStream: read incomplete");
}
return size - tlen;
}
virtual void Write(const void *ptr, size_t size) {
utils::Error("Base64InStream do not support write");
}
private:
StreamBufferReader reader_;
int tmp_ch;
int num_prev;
unsigned char buf_prev[2];
// whether we need to do strict check
static const bool kStrictCheck = false;
};
/*! \brief the stream that write to base64, note we take from file pointers */
class Base64OutStream: public IStream {
public:
explicit Base64OutStream(IStream *fp) : fp(fp) {
buf_top = 0;
}
virtual void Write(const void *ptr, size_t size) {
using base64::EncodeTable;
size_t tlen = size;
const unsigned char *cptr = static_cast<const unsigned char*>(ptr);
while (tlen) {
while (buf_top < 3 && tlen != 0) {
buf[++buf_top] = *cptr++; --tlen;
}
if (buf_top == 3) {
// flush 4 bytes out
PutChar(EncodeTable[buf[1] >> 2]);
PutChar(EncodeTable[((buf[1] << 4) | (buf[2] >> 4)) & 0x3F]);
PutChar(EncodeTable[((buf[2] << 2) | (buf[3] >> 6)) & 0x3F]);
PutChar(EncodeTable[buf[3] & 0x3F]);
buf_top = 0;
}
}
}
virtual size_t Read(void *ptr, size_t size) {
utils::Error("Base64OutStream do not support read");
return 0;
}
/*!
* \brief finish writing of all current base64 stream, do some post processing
* \param endch character to put to end of stream, if it is EOF, then nothing will be done
*/
inline void Finish(char endch = EOF) {
using base64::EncodeTable;
if (buf_top == 1) {
PutChar(EncodeTable[buf[1] >> 2]);
PutChar(EncodeTable[(buf[1] << 4) & 0x3F]);
PutChar('=');
PutChar('=');
}
if (buf_top == 2) {
PutChar(EncodeTable[buf[1] >> 2]);
PutChar(EncodeTable[((buf[1] << 4) | (buf[2] >> 4)) & 0x3F]);
PutChar(EncodeTable[(buf[2] << 2) & 0x3F]);
PutChar('=');
}
buf_top = 0;
if (endch != EOF) PutChar(endch);
this->Flush();
}
private:
IStream *fp;
int buf_top;
unsigned char buf[4];
std::string out_buf;
static const size_t kBufferSize = 256;
inline void PutChar(char ch) {
out_buf += ch;
if (out_buf.length() >= kBufferSize) Flush();
}
inline void Flush(void) {
if (out_buf.length() != 0) {
fp->Write(&out_buf[0], out_buf.length());
out_buf.clear();
}
}
};
} // namespace utils
} // namespace xgboost
#endif // XGBOOST_UTILS_BASE64_INL_H_

194
old_src/utils/config.h
View File

@@ -1,194 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file config.h
* \brief helper class to load in configures from file
* \author Tianqi Chen
*/
#ifndef XGBOOST_UTILS_CONFIG_H_
#define XGBOOST_UTILS_CONFIG_H_
#include <cstdio>
#include <cstring>
#include <string>
#include <istream>
#include <fstream>
#include "./utils.h"
namespace xgboost {
namespace utils {
/*!
* \brief base implementation of config reader
*/
class ConfigReaderBase {
public:
/*!
* \brief get current name, called after Next returns true
* \return current parameter name
*/
inline const char *name(void) const {
return s_name.c_str();
}
/*!
* \brief get current value, called after Next returns true
* \return current parameter value
*/
inline const char *val(void) const {
return s_val.c_str();
}
/*!
* \brief move iterator to next position
* \return true if there is value in next position
*/
inline bool Next(void) {
while (!this->IsEnd()) {
GetNextToken(&s_name);
if (s_name == "=") return false;
if (GetNextToken(&s_buf) || s_buf != "=") return false;
if (GetNextToken(&s_val) || s_val == "=") return false;
return true;
}
return false;
}
// called before usage
inline void Init(void) {
ch_buf = this->GetChar();
}
protected:
/*!
* \brief to be implemented by subclass,
* get next token, return EOF if end of file
*/
virtual char GetChar(void) = 0;
/*! \brief to be implemented by child, check if end of stream */
virtual bool IsEnd(void) = 0;
private:
char ch_buf;
std::string s_name, s_val, s_buf;
inline void SkipLine(void) {
do {
ch_buf = this->GetChar();
} while (ch_buf != EOF && ch_buf != '\n' && ch_buf != '\r');
}
inline void ParseStr(std::string *tok) {
while ((ch_buf = this->GetChar()) != EOF) {
switch (ch_buf) {
case '\\': *tok += this->GetChar(); break;
case '\"': return;
case '\r':
case '\n': Error("ConfigReader: unterminated string");
default: *tok += ch_buf;
}
}
Error("ConfigReader: unterminated string");
}
inline void ParseStrML(std::string *tok) {
while ((ch_buf = this->GetChar()) != EOF) {
switch (ch_buf) {
case '\\': *tok += this->GetChar(); break;
case '\'': return;
default: *tok += ch_buf;
}
}
Error("unterminated string");
}
// return newline
inline bool GetNextToken(std::string *tok) {
tok->clear();
bool new_line = false;
while (ch_buf != EOF) {
switch (ch_buf) {
case '#' : SkipLine(); new_line = true; break;
case '\"':
if (tok->length() == 0) {
ParseStr(tok); ch_buf = this->GetChar(); return new_line;
} else {
Error("ConfigReader: token followed directly by string");
}
case '\'':
if (tok->length() == 0) {
ParseStrML(tok); ch_buf = this->GetChar(); return new_line;
} else {
Error("ConfigReader: token followed directly by string");
}
case '=':
if (tok->length() == 0) {
ch_buf = this->GetChar();
*tok = '=';
}
return new_line;
case '\r':
case '\n':
if (tok->length() == 0) new_line = true;
case '\t':
case ' ' :
ch_buf = this->GetChar();
if (tok->length() != 0) return new_line;
break;
default:
*tok += ch_buf;
ch_buf = this->GetChar();
break;
}
}
if (tok->length() == 0) {
return true;
} else {
return false;
}
}
};
/*!
* \brief an iterator use stream base, allows use all types of istream
*/
class ConfigStreamReader: public ConfigReaderBase {
public:
/*!
* \brief constructor
* \param istream input stream
*/
explicit ConfigStreamReader(std::istream &fin) : fin(fin) {}
protected:
virtual char GetChar(void) {
return fin.get();
}
/*! \brief to be implemented by child, check if end of stream */
virtual bool IsEnd(void) {
return fin.eof();
}
private:
std::istream &fin;
};
/*!
* \brief an iterator that iterates over a configure file and gets the configures
*/
class ConfigIterator: public ConfigStreamReader {
public:
/*!
* \brief constructor
* \param fname name of configure file
*/
explicit ConfigIterator(const char *fname) : ConfigStreamReader(fi) {
fi.open(fname);
if (fi.fail()) {
utils::Error("cannot open file %s", fname);
}
ConfigReaderBase::Init();
}
/*! \brief destructor */
~ConfigIterator(void) {
fi.close();
}
private:
std::ifstream fi;
};
} // namespace utils
} // namespace xgboost
#endif // XGBOOST_UTILS_CONFIG_H_

42
old_src/utils/iterator.h
View File

@@ -1,42 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file iterator.h
* \brief itertator interface
* \author Tianqi Chen
*/
#ifndef XGBOOST_UTILS_ITERATOR_H_
#define XGBOOST_UTILS_ITERATOR_H_
#include <cstdio>
namespace xgboost {
namespace utils {
/*!
* \brief iterator interface
* \tparam DType data type
*/
template<typename DType>
class IIterator {
public:
/*!
* \brief set the parameter
* \param name name of parameter
* \param val value of parameter
*/
virtual void SetParam(const char *name, const char *val) {}
/*! \brief initialize the iterator so that we can use the iterator */
virtual void Init(void) {}
/*! \brief set before first of the item */
virtual void BeforeFirst(void) = 0;
/*! \brief move to next item */
virtual bool Next(void) = 0;
/*! \brief get current data */
virtual const DType &Value(void) const = 0;
public:
/*! \brief constructor */
virtual ~IIterator(void) {}
};
} // namespace utils
} // namespace xgboost
#endif // XGBOOST_UTILS_ITERATOR_H_

34
old_src/utils/omp.h
View File

@@ -1,34 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file omp.h
* \brief header to handle OpenMP compatibility issues
* \author Tianqi Chen
*/
#ifndef XGBOOST_UTILS_OMP_H_
#define XGBOOST_UTILS_OMP_H_
#if defined(_OPENMP) && !defined(DISABLE_OPENMP)
#include <omp.h>
#else
#if !defined(DISABLE_OPENMP) && !defined(_MSC_VER)
// use pragma message instead of warning
#pragma message("Warning: OpenMP is not available,"\
"xgboost will be compiled into single-thread code."\
"Use OpenMP-enabled compiler to get benefit of multi-threading")
#endif
inline int omp_get_thread_num() { return 0; }
inline int omp_get_num_threads() { return 1; }
inline void omp_set_num_threads(int nthread) {}
inline int omp_get_num_procs() { return 1; }
#endif
// loop variable used in openmp
namespace xgboost {
#ifdef _MSC_VER
typedef int bst_omp_uint;
#else
typedef unsigned bst_omp_uint;
#endif
} // namespace xgboost
#endif // XGBOOST_UTILS_OMP_H_

820
old_src/utils/quantile.h
View File

@@ -1,820 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file quantile.h
* \brief util to compute quantiles
* \author Tianqi Chen
*/
#ifndef XGBOOST_UTILS_QUANTILE_H_
#define XGBOOST_UTILS_QUANTILE_H_
#include <cmath>
#include <vector>
#include <cstring>
#include <algorithm>
#include <iostream>
#include "./io.h"
#include "./utils.h"
namespace xgboost {
namespace utils {
/*!
* \brief experimental wsummary
* \tparam DType type of data content
* \tparam RType type of rank
*/
template<typename DType, typename RType>
struct WQSummary {
/*! \brief an entry in the sketch summary */
struct Entry {
/*! \brief minimum rank */
RType rmin;
/*! \brief maximum rank */
RType rmax;
/*! \brief maximum weight */
RType wmin;
/*! \brief the value of data */
DType value;
// constructor
Entry(void) {}
// constructor
Entry(RType rmin, RType rmax, RType wmin, DType value)
: rmin(rmin), rmax(rmax), wmin(wmin), value(value) {}
/*!
* \brief debug function, check Valid
* \param eps the tolerate level for violating the relation
*/
inline void CheckValid(RType eps = 0) const {
utils::Assert(rmin >= 0 && rmax >= 0 && wmin >= 0, "nonneg constraint");
utils::Assert(rmax- rmin - wmin > -eps, "relation constraint: min/max");
}
/*! \return rmin estimation for v strictly bigger than value */
inline RType rmin_next(void) const {
return rmin + wmin;
}
/*! \return rmax estimation for v strictly smaller than value */
inline RType rmax_prev(void) const {
return rmax - wmin;
}
};
/*! \brief input data queue before entering the summary */
struct Queue {
// entry in the queue
struct QEntry {
// value of the instance
DType value;
// weight of instance
RType weight;
// default constructor
QEntry(void) {}
// constructor
QEntry(DType value, RType weight)
: value(value), weight(weight) {}
// comparator on value
inline bool operator<(const QEntry &b) const {
return value < b.value;
}
};
// the input queue
std::vector<QEntry> queue;
// end of the queue
size_t qtail;
// push data to the queue
inline void Push(DType x, RType w) {
if (qtail == 0 || queue[qtail - 1].value != x) {
queue[qtail++] = QEntry(x, w);
} else {
queue[qtail - 1].weight += w;
}
}
inline void MakeSummary(WQSummary *out) {
std::sort(queue.begin(), queue.begin() + qtail);
out->size = 0;
// start update sketch
RType wsum = 0;
// construct data with unique weights
for (size_t i = 0; i < qtail;) {
size_t j = i + 1;
RType w = queue[i].weight;
while (j < qtail && queue[j].value == queue[i].value) {
w += queue[j].weight; ++j;
}
out->data[out->size++] = Entry(wsum, wsum + w, w, queue[i].value);
wsum += w; i = j;
}
}
};
/*! \brief data field */
Entry *data;
/*! \brief number of elements in the summary */
size_t size;
// constructor
WQSummary(Entry *data, size_t size)
: data(data), size(size) {}
/*!
* \return the maximum error of the Summary
*/
inline RType MaxError(void) const {
RType res = data[0].rmax - data[0].rmin - data[0].wmin;
for (size_t i = 1; i < size; ++i) {
res = std::max(data[i].rmax_prev() - data[i - 1].rmin_next(), res);
res = std::max(data[i].rmax - data[i].rmin - data[i].wmin, res);
}
return res;
}
/*!
* \brief query qvalue, start from istart
* \param qvalue the value we query for
* \param istart starting position
*/
inline Entry Query(DType qvalue, size_t &istart) const { // NOLINT(*)
while (istart < size && qvalue > data[istart].value) {
++istart;
}
if (istart == size) {
RType rmax = data[size - 1].rmax;
return Entry(rmax, rmax, 0.0f, qvalue);
}
if (qvalue == data[istart].value) {
return data[istart];
} else {
if (istart == 0) {
return Entry(0.0f, 0.0f, 0.0f, qvalue);
} else {
return Entry(data[istart - 1].rmin_next(),
data[istart].rmax_prev(),
0.0f, qvalue);
}
}
}
/*! \return maximum rank in the summary */
inline RType MaxRank(void) const {
return data[size - 1].rmax;
}
/*!
* \brief copy content from src
* \param src source sketch
*/
inline void CopyFrom(const WQSummary &src) {
size = src.size;
std::memcpy(data, src.data, sizeof(Entry) * size);
}
/*!
* \brief debug function, validate whether the summary
* run consistency check to check if it is a valid summary
* \param eps the tolerate error level, used when RType is floating point and
* some inconsistency could occur due to rounding error
*/
inline void CheckValid(RType eps) const {
for (size_t i = 0; i < size; ++i) {
data[i].CheckValid(eps);
if (i != 0) {
utils::Assert(data[i].rmin >= data[i - 1].rmin + data[i - 1].wmin, "rmin range constraint");
utils::Assert(data[i].rmax >= data[i - 1].rmax + data[i].wmin, "rmax range constraint");
}
}
}
/*!
* \brief set current summary to be pruned summary of src
* assume data field is already allocated to be at least maxsize
* \param src source summary
* \param maxsize size we can afford in the pruned sketch
*/
inline void SetPrune(const WQSummary &src, size_t maxsize) {
if (src.size <= maxsize) {
this->CopyFrom(src); return;
}
const RType begin = src.data[0].rmax;
const RType range = src.data[src.size - 1].rmin - src.data[0].rmax;
const size_t n = maxsize - 1;
data[0] = src.data[0];
this->size = 1;
// lastidx is used to avoid duplicated records
size_t i = 1, lastidx = 0;
for (size_t k = 1; k < n; ++k) {
RType dx2 = 2 * ((k * range) / n + begin);
// find first i such that d < (rmax[i+1] + rmin[i+1]) / 2
while (i < src.size - 1
&& dx2 >= src.data[i + 1].rmax + src.data[i + 1].rmin) ++i;
utils::Assert(i != src.size - 1, "this cannot happen");
if (dx2 < src.data[i].rmin_next() + src.data[i + 1].rmax_prev()) {
if (i != lastidx) {
data[size++] = src.data[i]; lastidx = i;
}
} else {
if (i + 1 != lastidx) {
data[size++] = src.data[i + 1]; lastidx = i + 1;
}
}
}
if (lastidx != src.size - 1) {
data[size++] = src.data[src.size - 1];
}
}
/*!
* \brief set current summary to be merged summary of sa and sb
* \param sa first input summary to be merged
* \param sb second input summary to be merged
*/
inline void SetCombine(const WQSummary &sa,
const WQSummary &sb) {
if (sa.size == 0) {
this->CopyFrom(sb); return;
}
if (sb.size == 0) {
this->CopyFrom(sa); return;
}
utils::Assert(sa.size > 0 && sb.size > 0, "invalid input for merge");
const Entry *a = sa.data, *a_end = sa.data + sa.size;
const Entry *b = sb.data, *b_end = sb.data + sb.size;
// extended rmin value
RType aprev_rmin = 0, bprev_rmin = 0;
Entry *dst = this->data;
while (a != a_end && b != b_end) {
// duplicated value entry
if (a->value == b->value) {
*dst = Entry(a->rmin + b->rmin,
a->rmax + b->rmax,
a->wmin + b->wmin, a->value);
aprev_rmin = a->rmin_next();
bprev_rmin = b->rmin_next();
++dst; ++a; ++b;
} else if (a->value < b->value) {
*dst = Entry(a->rmin + bprev_rmin,
a->rmax + b->rmax_prev(),
a->wmin, a->value);
aprev_rmin = a->rmin_next();
++dst; ++a;
} else {
*dst = Entry(b->rmin + aprev_rmin,
b->rmax + a->rmax_prev(),
b->wmin, b->value);
bprev_rmin = b->rmin_next();
++dst; ++b;
}
}
if (a != a_end) {
RType brmax = (b_end - 1)->rmax;
do {
*dst = Entry(a->rmin + bprev_rmin, a->rmax + brmax, a->wmin, a->value);
++dst; ++a;
} while (a != a_end);
}
if (b != b_end) {
RType armax = (a_end - 1)->rmax;
do {
*dst = Entry(b->rmin + aprev_rmin, b->rmax + armax, b->wmin, b->value);
++dst; ++b;
} while (b != b_end);
}
this->size = dst - data;
const RType tol = 10;
RType err_mingap, err_maxgap, err_wgap;
this->FixError(&err_mingap, &err_maxgap, &err_wgap);
if (err_mingap > tol || err_maxgap > tol || err_wgap > tol) {
utils::Printf("INFO: mingap=%g, maxgap=%g, wgap=%g\n",
err_mingap, err_maxgap, err_wgap);
}
utils::Assert(size <= sa.size + sb.size, "bug in combine");
}
// helper function to print the current content of sketch
inline void Print() const {
for (size_t i = 0; i < this->size; ++i) {
utils::Printf("[%lu] rmin=%g, rmax=%g, wmin=%g, v=%g\n",
i, data[i].rmin, data[i].rmax,
data[i].wmin, data[i].value);
}
}
// try to fix rounding error
// and re-establish invariance
inline void FixError(RType *err_mingap,
RType *err_maxgap,
RType *err_wgap) const {
*err_mingap = 0;
*err_maxgap = 0;
*err_wgap = 0;
RType prev_rmin = 0, prev_rmax = 0;
for (size_t i = 0; i < this->size; ++i) {
if (data[i].rmin < prev_rmin) {
data[i].rmin = prev_rmin;
*err_mingap = std::max(*err_mingap, prev_rmin - data[i].rmin);
} else {
prev_rmin = data[i].rmin;
}
if (data[i].rmax < prev_rmax) {
data[i].rmax = prev_rmax;
*err_maxgap = std::max(*err_maxgap, prev_rmax - data[i].rmax);
}
RType rmin_next = data[i].rmin_next();
if (data[i].rmax < rmin_next) {
data[i].rmax = rmin_next;
*err_wgap = std::max(*err_wgap, data[i].rmax - rmin_next);
}
prev_rmax = data[i].rmax;
}
}
// check consistency of the summary
inline bool Check(const char *msg) const {
const float tol = 10.0f;
for (size_t i = 0; i < this->size; ++i) {
if (data[i].rmin + data[i].wmin > data[i].rmax + tol ||
data[i].rmin < -1e-6f || data[i].rmax < -1e-6f) {
utils::Printf("----%s: Check not Pass------\n", msg);
this->Print();
return false;
}
}
return true;
}
};
/*! \brief try to do efficient pruning */
template<typename DType, typename RType>
struct WXQSummary : public WQSummary<DType, RType> {
// redefine entry type
typedef typename WQSummary<DType, RType>::Entry Entry;
// constructor
WXQSummary(Entry *data, size_t size)
: WQSummary<DType, RType>(data, size) {}
// check if the block is large chunk
inline static bool CheckLarge(const Entry &e, RType chunk) {
return e.rmin_next() > e.rmax_prev() + chunk;
}
// set prune
inline void SetPrune(const WQSummary<DType, RType> &src, size_t maxsize) {
if (src.size <= maxsize) {
this->CopyFrom(src); return;
}
RType begin = src.data[0].rmax;
size_t n = maxsize - 1, nbig = 0;
RType range = src.data[src.size - 1].rmin - begin;
// prune off zero weights
if (range == 0.0f) {
// special case, contain only two effective data pts
this->data[0] = src.data[0];
this->data[1] = src.data[src.size - 1];
this->size = 2;
return;
} else {
range = std::max(range, static_cast<RType>(1e-3f));
}
const RType chunk = 2 * range / n;
// minimized range
RType mrange = 0;
{
// first scan, grab all the big chunk
// moving block index
size_t bid = 0;
for (size_t i = 1; i < src.size; ++i) {
if (CheckLarge(src.data[i], chunk)) {
if (bid != i - 1) {
mrange += src.data[i].rmax_prev() - src.data[bid].rmin_next();
}
bid = i; ++nbig;
}
}
if (bid != src.size - 2) {
mrange += src.data[src.size-1].rmax_prev() - src.data[bid].rmin_next();
}
}
if (nbig >= n - 1) {
// see what was the case
utils::Printf("LOG: check quantile stats, nbig=%lu, n=%lu\n", nbig, n);
utils::Printf("LOG: srcsize=%lu, maxsize=%lu, range=%g, chunk=%g\n",
src.size, maxsize, static_cast<double>(range),
static_cast<double>(chunk));
src.Print();
utils::Assert(nbig < n - 1, "quantile: too many large chunk");
}
this->data[0] = src.data[0];
this->size = 1;
// use smaller size
n = n - nbig;
// find the rest of point
size_t bid = 0, k = 1, lastidx = 0;
for (size_t end = 1; end < src.size; ++end) {
if (end == src.size - 1 || CheckLarge(src.data[end], chunk)) {
if (bid != end - 1) {
size_t i = bid;
RType maxdx2 = src.data[end].rmax_prev() * 2;
for (; k < n; ++k) {
RType dx2 = 2 * ((k * mrange) / n + begin);
if (dx2 >= maxdx2) break;
while (i < end &&
dx2 >= src.data[i + 1].rmax + src.data[i + 1].rmin) ++i;
if (i == end) break;
if (dx2 < src.data[i].rmin_next() + src.data[i + 1].rmax_prev()) {
if (i != lastidx) {
this->data[this->size++] = src.data[i]; lastidx = i;
}
} else {
if (i + 1 != lastidx) {
this->data[this->size++] = src.data[i + 1]; lastidx = i + 1;
}
}
}
}
if (lastidx != end) {
this->data[this->size++] = src.data[end];
lastidx = end;
}
bid = end;
// shift base by the gap
begin += src.data[bid].rmin_next() - src.data[bid].rmax_prev();
}
}
}
};
/*!
* \brief traditional GK summary
*/
template<typename DType, typename RType>
struct GKSummary {
/*! \brief an entry in the sketch summary */
struct Entry {
/*! \brief minimum rank */
RType rmin;
/*! \brief maximum rank */
RType rmax;
/*! \brief the value of data */
DType value;
// constructor
Entry(void) {}
// constructor
Entry(RType rmin, RType rmax, DType value)
: rmin(rmin), rmax(rmax), value(value) {}
};
/*! \brief input data queue before entering the summary */
struct Queue {
// the input queue
std::vector<DType> queue;
// end of the queue
size_t qtail;
// push data to the queue
inline void Push(DType x, RType w) {
queue[qtail++] = x;
}
inline void MakeSummary(GKSummary *out) {
std::sort(queue.begin(), queue.begin() + qtail);
out->size = qtail;
for (size_t i = 0; i < qtail; ++i) {
out->data[i] = Entry(i + 1, i + 1, queue[i]);
}
}
};
/*! \brief data field */
Entry *data;
/*! \brief number of elements in the summary */
size_t size;
GKSummary(Entry *data, size_t size)
: data(data), size(size) {}
/*! \brief the maximum error of the summary */
inline RType MaxError(void) const {
RType res = 0;
for (size_t i = 1; i < size; ++i) {
res = std::max(data[i].rmax - data[i-1].rmin, res);
}
return res;
}
/*! \return maximum rank in the summary */
inline RType MaxRank(void) const {
return data[size - 1].rmax;
}
/*!
* \brief copy content from src
* \param src source sketch
*/
inline void CopyFrom(const GKSummary &src) {
size = src.size;
std::memcpy(data, src.data, sizeof(Entry) * size);
}
inline void CheckValid(RType eps) const {
// assume always valid
}
/*! \brief used for debug purpose, print the summary */
inline void Print(void) const {
for (size_t i = 0; i < size; ++i) {
std::cout << "x=" << data[i].value << "\t"
<< "[" << data[i].rmin << "," << data[i].rmax << "]"
<< std::endl;
}
}
/*!
* \brief set current summary to be pruned summary of src
* assume data field is already allocated to be at least maxsize
* \param src source summary
* \param maxsize size we can afford in the pruned sketch
*/
inline void SetPrune(const GKSummary &src, size_t maxsize) {
if (src.size <= maxsize) {
this->CopyFrom(src); return;
}
const RType max_rank = src.MaxRank();
this->size = maxsize;
data[0] = src.data[0];
size_t n = maxsize - 1;
RType top = 1;
for (size_t i = 1; i < n; ++i) {
RType k = (i * max_rank) / n;
while (k > src.data[top + 1].rmax) ++top;
// assert src.data[top].rmin <= k
// because k > src.data[top].rmax >= src.data[top].rmin
if ((k - src.data[top].rmin) < (src.data[top+1].rmax - k)) {
data[i] = src.data[top];
} else {
data[i] = src.data[top + 1];
}
}
data[n] = src.data[src.size - 1];
}
inline void SetCombine(const GKSummary &sa,
const GKSummary &sb) {
if (sa.size == 0) {
this->CopyFrom(sb); return;
}
if (sb.size == 0) {
this->CopyFrom(sa); return;
}
utils::Assert(sa.size > 0 && sb.size > 0, "invalid input for merge");
const Entry *a = sa.data, *a_end = sa.data + sa.size;
const Entry *b = sb.data, *b_end = sb.data + sb.size;
this->size = sa.size + sb.size;
RType aprev_rmin = 0, bprev_rmin = 0;
Entry *dst = this->data;
while (a != a_end && b != b_end) {
if (a->value < b->value) {
*dst = Entry(bprev_rmin + a->rmin,
a->rmax + b->rmax - 1, a->value);
aprev_rmin = a->rmin;
++dst; ++a;
} else {
*dst = Entry(aprev_rmin + b->rmin,
b->rmax + a->rmax - 1, b->value);
bprev_rmin = b->rmin;
++dst; ++b;
}
}
if (a != a_end) {
RType bprev_rmax = (b_end - 1)->rmax;
do {
*dst = Entry(bprev_rmin + a->rmin, bprev_rmax + a->rmax, a->value);
++dst; ++a;
} while (a != a_end);
}
if (b != b_end) {
RType aprev_rmax = (a_end - 1)->rmax;
do {
*dst = Entry(aprev_rmin + b->rmin, aprev_rmax + b->rmax, b->value);
++dst; ++b;
} while (b != b_end);
}
utils::Assert(dst == data + size, "bug in combine");
}
};
/*!
* \brief template for all quantile sketch algorithm
* that uses merge/prune scheme
* \tparam DType type of data content
* \tparam RType type of rank
* \tparam TSummary actual summary data structure it uses
*/
template<typename DType, typename RType, class TSummary>
class QuantileSketchTemplate {
public:
/*! \brief type of summary type */
typedef TSummary Summary;
/*! \brief the entry type */
typedef typename Summary::Entry Entry;
/*! \brief same as summary, but use STL to backup the space */
struct SummaryContainer : public Summary {
std::vector<Entry> space;
SummaryContainer(const SummaryContainer &src) : Summary(NULL, src.size) {
this->space = src.space;
this->data = BeginPtr(this->space);
}
SummaryContainer(void) : Summary(NULL, 0) {
}
/*! \brief reserve space for summary */
inline void Reserve(size_t size) {
if (size > space.size()) {
space.resize(size);
this->data = BeginPtr(space);
}
}
/*!
* \brief set the space to be merge of all Summary arrays
* \param begin beginning position in the summary array
* \param end ending position in the Summary array
*/
inline void SetMerge(const Summary *begin,
const Summary *end) {
utils::Assert(begin < end, "can not set combine to empty instance");
size_t len = end - begin;
if (len == 1) {
this->Reserve(begin[0].size);
this->CopyFrom(begin[0]);
} else if (len == 2) {
this->Reserve(begin[0].size + begin[1].size);
this->SetMerge(begin[0], begin[1]);
} else {
// recursive merge
SummaryContainer lhs, rhs;
lhs.SetCombine(begin, begin + len / 2);
rhs.SetCombine(begin + len / 2, end);
this->Reserve(lhs.size + rhs.size);
this->SetCombine(lhs, rhs);
}
}
/*!
* \brief do elementwise combination of summary array
* this[i] = combine(this[i], src[i]) for each i
* \param src the source summary
* \param max_nbyte, maximum number of byte allowed in here
*/
inline void Reduce(const Summary &src, size_t max_nbyte) {
this->Reserve((max_nbyte - sizeof(this->size)) / sizeof(Entry));
SummaryContainer temp;
temp.Reserve(this->size + src.size);
temp.SetCombine(*this, src);
this->SetPrune(temp, space.size());
}
/*! \brief return the number of bytes this data structure cost in serialization */
inline static size_t CalcMemCost(size_t nentry) {
return sizeof(size_t) + sizeof(Entry) * nentry;
}
/*! \brief save the data structure into stream */
template<typename TStream>
inline void Save(TStream &fo) const { // NOLINT(*)
fo.Write(&(this->size), sizeof(this->size));
if (this->size != 0) {
fo.Write(this->data, this->size * sizeof(Entry));
}
}
/*! \brief load data structure from input stream */
template<typename TStream>
inline void Load(TStream &fi) { // NOLINT(*)
utils::Check(fi.Read(&this->size, sizeof(this->size)) != 0, "invalid SummaryArray 1");
this->Reserve(this->size);
if (this->size != 0) {
utils::Check(fi.Read(this->data, this->size * sizeof(Entry)) != 0,
"invalid SummaryArray 2");
}
}
};
/*!
* \brief initialize the quantile sketch, given the performance specification
* \param maxn maximum number of data points can be feed into sketch
* \param eps accuracy level of summary
*/
inline void Init(size_t maxn, double eps) {
nlevel = 1;
while (true) {
limit_size = static_cast<size_t>(ceil(nlevel / eps)) + 1;
size_t n = (1UL << nlevel);
if (n * limit_size >= maxn) break;
++nlevel;
}
// check invariant
size_t n = (1UL << nlevel);
utils::Assert(n * limit_size >= maxn, "invalid init parameter");
utils::Assert(nlevel <= limit_size * eps, "invalid init parameter");
// lazy reserve the space, if there is only one value, no need to allocate space
inqueue.queue.resize(1);
inqueue.qtail = 0;
data.clear();
level.clear();
}
/*!
* \brief add an element to a sketch
* \param x the element added to the sketch
*/
inline void Push(DType x, RType w = 1) {
if (w == static_cast<RType>(0)) return;
if (inqueue.qtail == inqueue.queue.size()) {
// jump from lazy one value to limit_size * 2
if (inqueue.queue.size() == 1) {
inqueue.queue.resize(limit_size * 2);
} else {
temp.Reserve(limit_size * 2);
inqueue.MakeSummary(&temp);
// cleanup queue
inqueue.qtail = 0;
this->PushTemp();
}
}
inqueue.Push(x, w);
}
/*! \brief push up temp */
inline void PushTemp(void) {
temp.Reserve(limit_size * 2);
for (size_t l = 1; true; ++l) {
this->InitLevel(l + 1);
// check if level l is empty
if (level[l].size == 0) {
level[l].SetPrune(temp, limit_size);
break;
} else {
// level 0 is actually temp space
level[0].SetPrune(temp, limit_size);
temp.SetCombine(level[0], level[l]);
if (temp.size > limit_size) {
// try next level
level[l].size = 0;
} else {
// if merged record is still smaller, no need to send to next level
level[l].CopyFrom(temp); break;
}
}
}
}
/*! \brief get the summary after finalize */
inline void GetSummary(SummaryContainer *out) {
if (level.size() != 0) {
out->Reserve(limit_size * 2);
} else {
out->Reserve(inqueue.queue.size());
}
inqueue.MakeSummary(out);
if (level.size() != 0) {
level[0].SetPrune(*out, limit_size);
for (size_t l = 1; l < level.size(); ++l) {
if (level[l].size == 0) continue;
if (level[0].size == 0) {
level[0].CopyFrom(level[l]);
} else {
out->SetCombine(level[0], level[l]);
level[0].SetPrune(*out, limit_size);
}
}
out->CopyFrom(level[0]);
} else {
if (out->size > limit_size) {
temp.Reserve(limit_size);
temp.SetPrune(*out, limit_size);
out->CopyFrom(temp);
}
}
}
// used for debug, check if the sketch is valid
inline void CheckValid(RType eps) const {
for (size_t l = 1; l < level.size(); ++l) {
level[l].CheckValid(eps);
}
}
// initialize level space to at least nlevel
inline void InitLevel(size_t nlevel) {
if (level.size() >= nlevel) return;
data.resize(limit_size * nlevel);
level.resize(nlevel, Summary(NULL, 0));
for (size_t l = 0; l < level.size(); ++l) {
level[l].data = BeginPtr(data) + l * limit_size;
}
}
// input data queue
typename Summary::Queue inqueue;
// number of levels
size_t nlevel;
// size of summary in each level
size_t limit_size;
// the level of each summaries
std::vector<Summary> level;
// content of the summary
std::vector<Entry> data;
// temporal summary, used for temp-merge
SummaryContainer temp;
};
/*!
* \brief Quantile sketch use WQSummary
* \tparam DType type of data content
* \tparam RType type of rank
*/
template<typename DType, typename RType = unsigned>
class WQuantileSketch :
public QuantileSketchTemplate<DType, RType, WQSummary<DType, RType> >{
};
/*!
* \brief Quantile sketch use WXQSummary
* \tparam DType type of data content
* \tparam RType type of rank
*/
template<typename DType, typename RType = unsigned>
class WXQuantileSketch :
public QuantileSketchTemplate<DType, RType, WXQSummary<DType, RType> >{
};
/*!
* \brief Quantile sketch use WQSummary
* \tparam DType type of data content
* \tparam RType type of rank
*/
template<typename DType, typename RType = unsigned>
class GKQuantileSketch :
public QuantileSketchTemplate<DType, RType, GKSummary<DType, RType> >{
};
} // namespace utils
} // namespace xgboost
#endif // XGBOOST_UTILS_QUANTILE_H_

108
old_src/utils/random.h
View File

@@ -1,108 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file xgboost_random.h
* \brief PRNG to support random number generation
* \author Tianqi Chen: tianqi.tchen@gmail.com
*
* Use standard PRNG from stdlib
*/
#ifndef XGBOOST_UTILS_RANDOM_H_
#define XGBOOST_UTILS_RANDOM_H_
#include <cmath>
#include <cstdlib>
#include <vector>
#include <algorithm>
#include "./utils.h"
/*! namespace of PRNG */
namespace xgboost {
namespace random {
#ifndef XGBOOST_CUSTOMIZE_PRNG_
/*! \brief seed the PRNG */
inline void Seed(unsigned seed) {
srand(seed);
}
/*! \brief basic function, uniform */
inline double Uniform(void) {
return static_cast<double>(rand()) / (static_cast<double>(RAND_MAX)+1.0); // NOLINT(*)
}
/*! \brief return a real number uniform in (0,1) */
inline double NextDouble2(void) {
return (static_cast<double>(rand()) + 1.0) / (static_cast<double>(RAND_MAX)+2.0); // NOLINT(*)
}
/*! \brief return x~N(0,1) */
inline double Normal(void) {
double x, y, s;
do {
x = 2 * NextDouble2() - 1.0;
y = 2 * NextDouble2() - 1.0;
s = x*x + y*y;
} while (s >= 1.0 || s == 0.0);
return x * sqrt(-2.0 * log(s) / s);
}
#else
// include declarations, to be implemented
void Seed(unsigned seed);
double Uniform(void);
double Normal(void);
#endif
/*! \brief return a real number uniform in [0,1) */
inline double NextDouble(void) {
return Uniform();
}
/*! \brief return a random number in n */
inline uint32_t NextUInt32(uint32_t n) {
return (uint32_t)std::floor(NextDouble() * n);
}
/*! \brief return x~N(mu,sigma^2) */
inline double SampleNormal(double mu, double sigma) {
return Normal() * sigma + mu;
}
/*! \brief return 1 with probability p, coin flip */
inline int SampleBinary(double p) {
return NextDouble() < p;
}
template<typename T>
inline void Shuffle(T *data, size_t sz) {
if (sz == 0) return;
for (uint32_t i = (uint32_t)sz - 1; i > 0; i--) {
std::swap(data[i], data[NextUInt32(i + 1)]);
}
}
// random shuffle the data inside, require PRNG
template<typename T>
inline void Shuffle(std::vector<T> &data) { // NOLINT(*)
Shuffle(&data[0], data.size());
}
/*! \brief random number generator with independent random number seed*/
struct Random{
/*! \brief set random number seed */
inline void Seed(unsigned sd) {
this->rseed = sd;
#if defined(_MSC_VER) || defined(_WIN32)
::xgboost::random::Seed(sd);
#endif
}
/*! \brief return a real number uniform in [0,1) */
inline double RandDouble(void) {
// use rand instead of rand_r in windows, for MSVC it is fine since rand is threadsafe
// For cygwin and mingw, this can slows down parallelism,
// but rand_r is only used in objective-inl.hpp, won't affect speed in general
// todo, replace with another PRNG
#if defined(_MSC_VER) || defined(_WIN32) || defined(XGBOOST_STRICT_CXX98_)
return Uniform();
#else
return static_cast<double>(rand_r(&rseed)) / (static_cast<double>(RAND_MAX) + 1.0);
#endif
}
// random number seed
unsigned rseed;
};
} // namespace random
} // namespace xgboost
#endif // XGBOOST_UTILS_RANDOM_H_

260
old_src/utils/thread.h
View File

@@ -1,260 +0,0 @@
/*!
* Copyright by Contributors
* \file thread.h
* \brief this header include the minimum necessary resource
* for multi-threading that can be compiled in windows, linux, mac
* \author Tianqi Chen
*/
#ifndef XGBOOST_UTILS_THREAD_H_ // NOLINT(*)
#define XGBOOST_UTILS_THREAD_H_ // NOLINT(*)
#ifdef _MSC_VER
#include <windows.h>
#include <process.h>
#include "./utils.h"
namespace xgboost {
namespace utils {
/*! \brief simple semaphore used for synchronization */
class Semaphore {
public :
inline void Init(int init_val) {
sem = CreateSemaphore(NULL, init_val, 10, NULL);
utils::Check(sem != NULL, "create Semaphore error");
}
inline void Destroy(void) {
CloseHandle(sem);
}
inline void Wait(void) {
utils::Check(WaitForSingleObject(sem, INFINITE) == WAIT_OBJECT_0, "WaitForSingleObject error");
}
inline void Post(void) {
utils::Check(ReleaseSemaphore(sem, 1, NULL) != 0, "ReleaseSemaphore error");
}
private:
HANDLE sem;
};
/*! \brief mutex under windows */
class Mutex {
public:
inline void Init(void) {
utils::Check(InitializeCriticalSectionAndSpinCount(&mutex, 0x00000400) != 0,
"Mutex::Init fail");
}
inline void Lock(void) {
EnterCriticalSection(&mutex);
}
inline void Unlock(void) {
LeaveCriticalSection(&mutex);
}
inline void Destroy(void) {
DeleteCriticalSection(&mutex);
}
private:
friend class ConditionVariable;
CRITICAL_SECTION mutex;
};
// conditional variable that uses pthread
class ConditionVariable {
public:
// initialize conditional variable
inline void Init(void) {
InitializeConditionVariable(&cond);
}
// destroy the thread
inline void Destroy(void) {
// DeleteConditionVariable(&cond);
}
// wait on the conditional variable
inline void Wait(Mutex *mutex) {
utils::Check(SleepConditionVariableCS(&cond, &(mutex->mutex), INFINITE) != 0,
"ConditionVariable:Wait fail");
}
inline void Broadcast(void) {
WakeAllConditionVariable(&cond);
}
inline void Signal(void) {
WakeConditionVariable(&cond);
}
private:
CONDITION_VARIABLE cond;
};
/*! \brief simple thread that wraps windows thread */
class Thread {
private:
HANDLE thread_handle;
unsigned thread_id;
public:
inline void Start(unsigned int __stdcall entry(void*p), void *param) {
thread_handle = (HANDLE)_beginthreadex(NULL, 0, entry, param, 0, &thread_id);
}
inline int Join(void) {
WaitForSingleObject(thread_handle, INFINITE);
return 0;
}
};
/*! \brief exit function called from thread */
inline void ThreadExit(void *status) {
_endthreadex(0);
}
#define XGBOOST_THREAD_PREFIX unsigned int __stdcall
} // namespace utils
} // namespace xgboost
#else
// thread interface using g++
#include <semaphore.h>
#include <pthread.h>
#include <errno.h>
namespace xgboost {
namespace utils {
/*!\brief semaphore class */
class Semaphore {
#ifdef __APPLE__
private:
sem_t* semPtr;
char sema_name[20];
private:
inline void GenRandomString(char *s, const int len) {
static const char alphanum[] = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ";
for (int i = 0; i < len; ++i) {
s[i] = alphanum[rand() % (sizeof(alphanum) - 1)];
}
s[len] = 0;
}
public:
inline void Init(int init_val) {
sema_name[0] = '/';
sema_name[1] = 's';
sema_name[2] = 'e';
sema_name[3] = '/';
GenRandomString(&sema_name[4], 16);
if ((semPtr = sem_open(sema_name, O_CREAT, 0644, init_val)) == SEM_FAILED) {
perror("sem_open");
exit(1);
}
utils::Check(semPtr != NULL, "create Semaphore error");
}
inline void Destroy(void) {
if (sem_close(semPtr) == -1) {
perror("sem_close");
exit(EXIT_FAILURE);
}
if (sem_unlink(sema_name) == -1) {
perror("sem_unlink");
exit(EXIT_FAILURE);
}
}
inline void Wait(void) {
sem_wait(semPtr);
}
inline void Post(void) {
sem_post(semPtr);
}
#else
private:
sem_t sem;
public:
inline void Init(int init_val) {
if (sem_init(&sem, 0, init_val) != 0) {
utils::Error("Semaphore.Init:%s", strerror(errno));
}
}
inline void Destroy(void) {
if (sem_destroy(&sem) != 0) {
utils::Error("Semaphore.Destroy:%s", strerror(errno));
}
}
inline void Wait(void) {
if (sem_wait(&sem) != 0) {
utils::Error("Semaphore.Wait:%s", strerror(errno));
}
}
inline void Post(void) {
if (sem_post(&sem) != 0) {
utils::Error("Semaphore.Post:%s", strerror(errno));
}
}
#endif
};
// mutex that works with pthread
class Mutex {
public:
inline void Init(void) {
pthread_mutex_init(&mutex, NULL);
}
inline void Lock(void) {
pthread_mutex_lock(&mutex);
}
inline void Unlock(void) {
pthread_mutex_unlock(&mutex);
}
inline void Destroy(void) {
pthread_mutex_destroy(&mutex);
}
private:
friend class ConditionVariable;
pthread_mutex_t mutex;
};
// conditional variable that uses pthread
class ConditionVariable {
public:
// initialize conditional variable
inline void Init(void) {
pthread_cond_init(&cond, NULL);
}
// destroy the thread
inline void Destroy(void) {
pthread_cond_destroy(&cond);
}
// wait on the conditional variable
inline void Wait(Mutex *mutex) {
pthread_cond_wait(&cond, &(mutex->mutex));
}
inline void Broadcast(void) {
pthread_cond_broadcast(&cond);
}
inline void Signal(void) {
pthread_cond_signal(&cond);
}
private:
pthread_cond_t cond;
};
/*!\brief simple thread class */
class Thread {
private:
pthread_t thread;
public :
inline void Start(void * entry(void*), void *param) { // NOLINT(*)
pthread_attr_t attr;
pthread_attr_init(&attr);
pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_JOINABLE);
pthread_create(&thread, &attr, entry, param);
}
inline int Join(void) {
void *status;
return pthread_join(thread, &status);
}
};
inline void ThreadExit(void *status) {
pthread_exit(status);
}
} // namespace utils
} // namespace xgboost
#define XGBOOST_THREAD_PREFIX void *
#endif // Linux
#endif // XGBOOST_UTILS_THREAD_H_ NOLINT(*)

257
old_src/utils/thread_buffer.h
View File

@@ -1,257 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file thread_buffer.h
* \brief multi-thread buffer, iterator, can be used to create parallel pipeline
* \author Tianqi Chen
*/
#ifndef XGBOOST_UTILS_THREAD_BUFFER_H_
#define XGBOOST_UTILS_THREAD_BUFFER_H_
#include <vector>
#include <cstring>
#include <cstdlib>
#include "./utils.h"
// threading util could not run on solaris
#ifndef XGBOOST_STRICT_CXX98_
#include "./thread.h"
#endif
namespace xgboost {
namespace utils {
#if !defined(XGBOOST_STRICT_CXX98_)
/*!
* \brief buffered loading iterator that uses multithread
* this template method will assume the following parameters
* \tparam Elem element type to be buffered
* \tparam ElemFactory factory type to implement in order to use thread buffer
*/
template<typename Elem, typename ElemFactory>
class ThreadBuffer {
public:
/*!\brief constructor */
ThreadBuffer(void) {
this->init_end = false;
this->buf_size = 30;
}
~ThreadBuffer(void) {
if (init_end) this->Destroy();
}
/*!\brief set parameter, will also pass the parameter to factory */
inline void SetParam(const char *name, const char *val) {
using namespace std;
if (!strcmp( name, "buffer_size")) buf_size = atoi(val);
factory.SetParam(name, val);
}
/*!
* \brief initalize the buffered iterator
* \param param a initialize parameter that will pass to factory, ignore it if not necessary
* \return false if the initialization can't be done, e.g. buffer file hasn't been created
*/
inline bool Init(void) {
if (!factory.Init()) return false;
for (int i = 0; i < buf_size; ++i) {
bufA.push_back(factory.Create());
bufB.push_back(factory.Create());
}
this->init_end = true;
this->StartLoader();
return true;
}
/*!\brief place the iterator before first value */
inline void BeforeFirst(void) {
// wait till last loader end
loading_end.Wait();
// critical zone
current_buf = 1;
factory.BeforeFirst();
// reset terminate limit
endA = endB = buf_size;
// wake up loader for first part
loading_need.Post();
// wait til first part is loaded
loading_end.Wait();
// set current buf to right value
current_buf = 0;
// wake loader for next part
data_loaded = false;
loading_need.Post();
// set buffer value
buf_index = 0;
}
/*! \brief destroy the buffer iterator, will deallocate the buffer */
inline void Destroy(void) {
// wait until the signal is consumed
this->destroy_signal = true;
loading_need.Post();
loader_thread.Join();
loading_need.Destroy();
loading_end.Destroy();
for (size_t i = 0; i < bufA.size(); ++i) {
factory.FreeSpace(bufA[i]);
}
for (size_t i = 0; i < bufB.size(); ++i) {
factory.FreeSpace(bufB[i]);
}
bufA.clear(); bufB.clear();
factory.Destroy();
this->init_end = false;
}
/*!
* \brief get the next element needed in buffer
* \param elem element to store into
* \return whether reaches end of data
*/
inline bool Next(Elem &elem) { // NOLINT(*)
// end of buffer try to switch
if (buf_index == buf_size) {
this->SwitchBuffer();
buf_index = 0;
}
if (buf_index >= (current_buf ? endA : endB)) {
return false;
}
std::vector<Elem> &buf = current_buf ? bufA : bufB;
elem = buf[buf_index];
++buf_index;
return true;
}
/*!
* \brief get the factory object
*/
inline ElemFactory &get_factory(void) {
return factory;
}
inline const ElemFactory &get_factory(void) const {
return factory;
}
// size of buffer
int buf_size;
private:
// factory object used to load configures
ElemFactory factory;
// index in current buffer
int buf_index;
// indicate which one is current buffer
int current_buf;
// max limit of visit, also marks termination
int endA, endB;
// double buffer, one is accessed by loader
// the other is accessed by consumer
// buffer of the data
std::vector<Elem> bufA, bufB;
// initialization end
bool init_end;
// singal whether the data is loaded
bool data_loaded;
// signal to kill the thread
bool destroy_signal;
// thread object
Thread loader_thread;
// signal of the buffer
Semaphore loading_end, loading_need;
/*!
* \brief slave thread
* this implementation is like producer-consumer style
*/
inline void RunLoader(void) {
while (!destroy_signal) {
// sleep until loading is needed
loading_need.Wait();
std::vector<Elem> &buf = current_buf ? bufB : bufA;
int i;
for (i = 0; i < buf_size ; ++i) {
if (!factory.LoadNext(buf[i])) {
int &end = current_buf ? endB : endA;
end = i; // marks the termination
break;
}
}
// signal that loading is done
data_loaded = true;
loading_end.Post();
}
}
/*!\brief entry point of loader thread */
inline static XGBOOST_THREAD_PREFIX LoaderEntry(void *pthread) {
static_cast< ThreadBuffer<Elem, ElemFactory>* >(pthread)->RunLoader();
return NULL;
}
/*!\brief start loader thread */
inline void StartLoader(void) {
destroy_signal = false;
// set param
current_buf = 1;
loading_need.Init(1);
loading_end .Init(0);
// reset terminate limit
endA = endB = buf_size;
loader_thread.Start(LoaderEntry, this);
// wait until first part of data is loaded
loading_end.Wait();
// set current buf to right value
current_buf = 0;
// wake loader for next part
data_loaded = false;
loading_need.Post();
buf_index = 0;
}
/*!\brief switch double buffer */
inline void SwitchBuffer(void) {
loading_end.Wait();
// loader shall be sleep now, critcal zone!
current_buf = !current_buf;
// wake up loader
data_loaded = false;
loading_need.Post();
}
};
#else
// a dummy single threaded ThreadBuffer
// use this to resolve R's solaris compatibility for now
template<typename Elem, typename ElemFactory>
class ThreadBuffer {
public:
ThreadBuffer() : init_end_(false) {}
~ThreadBuffer() {
if (init_end_) {
factory_.FreeSpace(data_);
factory_.Destroy();
}
}
inline void SetParam(const char *name, const char *val) {
}
inline bool Init(void) {
if (!factory_.Init()) return false;
data_ = factory_.Create();
return (init_end_ = true);
}
inline void BeforeFirst(void) {
factory_.BeforeFirst();
}
inline bool Next(Elem &elem) { // NOLINT(*)
if (factory_.LoadNext(data_)) {
elem = data_; return true;
} else {
return false;
}
}
inline ElemFactory &get_factory() {
return factory_;
}
inline const ElemFactory &get_factory() const {
return factory_;
}
private:
// initialized
bool init_end_;
// current data
Elem data_;
// factory object used to load configures
ElemFactory factory_;
};
#endif // !defined(XGBOOST_STRICT_CXX98_)
} // namespace utils
} // namespace xgboost
#endif // XGBOOST_UTILS_THREAD_BUFFER_H_

188
old_src/utils/utils.h
View File

@@ -1,188 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file utils.h
* \brief simple utils to support the code
* \author Tianqi Chen
*/
#ifndef XGBOOST_UTILS_UTILS_H_
#define XGBOOST_UTILS_UTILS_H_
#define _CRT_SECURE_NO_WARNINGS
#include <cstdio>
#include <string>
#include <cstdlib>
#include <vector>
#include <stdexcept>
#ifndef XGBOOST_STRICT_CXX98_
#include <cstdarg>
#endif
#if !defined(__GNUC__)
#define fopen64 std::fopen
#endif
#ifdef _MSC_VER
// NOTE: sprintf_s is not equivalent to snprintf,
// they are equivalent when success, which is sufficient for our case
#define snprintf sprintf_s
#define vsnprintf vsprintf_s
#else
#ifdef _FILE_OFFSET_BITS
#if _FILE_OFFSET_BITS == 32
#pragma message("Warning: FILE OFFSET BITS defined to be 32 bit")
#endif
#endif
#ifdef __APPLE__
#define off64_t off_t
#define fopen64 std::fopen
#endif
extern "C" {
#include <sys/types.h>
}
#endif
#ifdef _MSC_VER
typedef unsigned char uint8_t;
typedef unsigned __int16 uint16_t;
typedef unsigned __int32 uint32_t;
typedef unsigned __int64 uint64_t;
typedef __int64 int64_t;
#else
#include <inttypes.h>
#endif
namespace xgboost {
/*! \brief namespace for helper utils of the project */
namespace utils {
/*! \brief error message buffer length */
const int kPrintBuffer = 1 << 12;
#ifndef XGBOOST_CUSTOMIZE_MSG_
/*!
* \brief handling of Assert error, caused by inappropriate input
* \param msg error message
*/
inline void HandleAssertError(const char *msg) {
fprintf(stderr, "AssertError:%s\n", msg);
exit(-1);
}
/*!
* \brief handling of Check error, caused by inappropriate input
* \param msg error message
*/
inline void HandleCheckError(const char *msg) {
throw std::runtime_error(msg);
}
inline void HandlePrint(const char *msg) {
printf("%s", msg);
}
#else
#ifndef XGBOOST_STRICT_CXX98_
// include declarations, some one must implement this
void HandleAssertError(const char *msg);
void HandleCheckError(const char *msg);
void HandlePrint(const char *msg);
#endif
#endif
#ifdef XGBOOST_STRICT_CXX98_
// these function pointers are to be assigned
extern "C" void (*Printf)(const char *fmt, ...);
extern "C" int (*SPrintf)(char *buf, size_t size, const char *fmt, ...);
extern "C" void (*Assert)(int exp, const char *fmt, ...);
extern "C" void (*Check)(int exp, const char *fmt, ...);
extern "C" void (*Error)(const char *fmt, ...);
#else
/*! \brief printf, print message to the console */
inline void Printf(const char *fmt, ...) {
std::string msg(kPrintBuffer, '\0');
va_list args;
va_start(args, fmt);
vsnprintf(&msg[0], kPrintBuffer, fmt, args);
va_end(args);
HandlePrint(msg.c_str());
}
/*! \brief portable version of snprintf */
inline int SPrintf(char *buf, size_t size, const char *fmt, ...) {
va_list args;
va_start(args, fmt);
int ret = vsnprintf(buf, size, fmt, args);
va_end(args);
return ret;
}
/*! \brief assert an condition is true, use this to handle debug information */
inline void Assert(bool exp, const char *fmt, ...) {
if (!exp) {
std::string msg(kPrintBuffer, '\0');
va_list args;
va_start(args, fmt);
vsnprintf(&msg[0], kPrintBuffer, fmt, args);
va_end(args);
HandleAssertError(msg.c_str());
}
}
/*!\brief same as assert, but this is intended to be used as message for user*/
inline void Check(bool exp, const char *fmt, ...) {
if (!exp) {
std::string msg(kPrintBuffer, '\0');
va_list args;
va_start(args, fmt);
vsnprintf(&msg[0], kPrintBuffer, fmt, args);
va_end(args);
HandleCheckError(msg.c_str());
}
}
/*! \brief report error message, same as check */
inline void Error(const char *fmt, ...) {
{
std::string msg(kPrintBuffer, '\0');
va_list args;
va_start(args, fmt);
vsnprintf(&msg[0], kPrintBuffer, fmt, args);
va_end(args);
HandleCheckError(msg.c_str());
}
}
#endif
/*! \brief replace fopen, report error when the file open fails */
inline std::FILE *FopenCheck(const char *fname, const char *flag) {
std::FILE *fp = fopen64(fname, flag);
Check(fp != NULL, "can not open file \"%s\"\n", fname);
return fp;
}
} // namespace utils
// easy utils that can be directly accessed in xgboost
/*! \brief get the beginning address of a vector */
template<typename T>
inline T *BeginPtr(std::vector<T> &vec) { // NOLINT(*)
if (vec.size() == 0) {
return NULL;
} else {
return &vec[0];
}
}
/*! \brief get the beginning address of a vector */
template<typename T>
inline const T *BeginPtr(const std::vector<T> &vec) {
if (vec.size() == 0) {
return NULL;
} else {
return &vec[0];
}
}
inline char* BeginPtr(std::string &str) { // NOLINT(*)
if (str.length() == 0) return NULL;
return &str[0];
}
inline const char* BeginPtr(const std::string &str) {
if (str.length() == 0) return NULL;
return &str[0];
}
} // namespace xgboost
#endif // XGBOOST_UTILS_UTILS_H_