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

[REFACTOR] cleanup structure

Browse Source
This commit is contained in:
tqchen
2015-11-24 14:25:56 -08:00
parent 5ed4dc4f60
commit d530e0c14f
60 changed files with 42 additions and 51 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
src/README.md
View File

@@ -1,26 +0,0 @@
Coding Guide
======
This file is intended to be notes about code structure in xgboost
Project Logical Layout
=======
* Dependency order: io->learner->gbm->tree
- All module depends on data.h
* tree are implementations of tree construction algorithms.
* gbm is gradient boosting interface, that takes trees and other base learner to do boosting.
- gbm only takes gradient as sufficient statistics, it does not compute the gradient.
* learner is learning module that computes gradient for specific object, and pass it to GBM
File Naming Convention
=======
* .h files are data structures and interface, which are needed to use functions in that layer.
* -inl.hpp files are implementations of interface, like cpp file in most project.
- You only need to understand the interface file to understand the usage of that layer
* In each folder, there can be a .cpp file, that compiles the module of that layer
How to Hack the Code
======
* Add objective function: add to learner/objective-inl.hpp and register it in learner/objective.h ```CreateObjFunction```
- You can also directly do it in python
* Add new evaluation metric: add to learner/evaluation-inl.hpp and register it in learner/evaluation.h ```CreateEvaluator```
* Add wrapper for a new language, most likely you can do it by taking the functions in python/xgboost_wrapper.h, which is purely C based, and call these C functions to use xgboost

2
src/c_api.cc Normal file
View File

@@ -0,0 +1,2 @@
#include <xgboost/c_api.h>

166
src/data.h
View File

@@ -1,166 +0,0 @@
/*!
* Copyright (c) 2014 by Contributors
* \file data.h
* \brief the input data structure for gradient boosting
* \author Tianqi Chen
*/
#ifndef XGBOOST_DATA_H_
#define XGBOOST_DATA_H_
#include <cstdio>
#include <vector>
#include "utils/utils.h"
#include "utils/iterator.h"
namespace xgboost {
/*!
* \brief unsigned integer type used in boost,
* used for feature index and row index
*/
typedef unsigned bst_uint;
/*! \brief float type, used for storing statistics */
typedef float bst_float;
const float rt_eps = 1e-5f;
// min gap between feature values to allow a split happen
const float rt_2eps = rt_eps * 2.0f;
/*! \brief gradient statistics pair usually needed in gradient boosting */
struct bst_gpair {
/*! \brief gradient statistics */
bst_float grad;
/*! \brief second order gradient statistics */
bst_float hess;
bst_gpair(void) {}
bst_gpair(bst_float grad, bst_float hess) : grad(grad), hess(hess) {}
};
/*!
* \brief extra information that might be needed by gbm and tree module
* this information is not necessarily present, and can be empty
*/
struct BoosterInfo {
/*! \brief number of rows in the data */
size_t num_row;
/*! \brief number of columns in the data */
size_t num_col;
/*!
* \brief specified root index of each instance,
* can be used for multi task setting
*/
std::vector<unsigned> root_index;
/*! \brief set fold indicator */
std::vector<unsigned> fold_index;
/*! \brief number of rows, number of columns */
BoosterInfo(void) : num_row(0), num_col(0) {
}
/*! \brief get root of i-th instance */
inline unsigned GetRoot(size_t i) const {
return root_index.size() == 0 ? 0 : root_index[i];
}
};
/*! \brief read-only sparse instance batch in CSR format */
struct SparseBatch {
/*! \brief an entry of sparse vector */
struct Entry {
/*! \brief feature index */
bst_uint index;
/*! \brief feature value */
bst_float fvalue;
// default constructor
Entry(void) {}
Entry(bst_uint index, bst_float fvalue) : index(index), fvalue(fvalue) {}
/*! \brief reversely compare feature values */
inline static bool CmpValue(const Entry &a, const Entry &b) {
return a.fvalue < b.fvalue;
}
};
/*! \brief an instance of sparse vector in the batch */
struct Inst {
/*! \brief pointer to the elements*/
const Entry *data;
/*! \brief length of the instance */
bst_uint length;
/*! \brief constructor */
Inst(const Entry *data, bst_uint length) : data(data), length(length) {}
/*! \brief get i-th pair in the sparse vector*/
inline const Entry& operator[](size_t i) const {
return data[i];
}
};
/*! \brief batch size */
size_t size;
};
/*! \brief read-only row batch, used to access row continuously */
struct RowBatch : public SparseBatch {
/*! \brief the offset of rowid of this batch */
size_t base_rowid;
/*! \brief array[size+1], row pointer of each of the elements */
const size_t *ind_ptr;
/*! \brief array[ind_ptr.back()], content of the sparse element */
const Entry *data_ptr;
/*! \brief get i-th row from the batch */
inline Inst operator[](size_t i) const {
return Inst(data_ptr + ind_ptr[i], static_cast<bst_uint>(ind_ptr[i+1] - ind_ptr[i]));
}
};
/*!
* \brief read-only column batch, used to access columns,
* the columns are not required to be continuous
*/
struct ColBatch : public SparseBatch {
/*! \brief column index of each columns in the data */
const bst_uint *col_index;
/*! \brief pointer to the column data */
const Inst *col_data;
/*! \brief get i-th column from the batch */
inline Inst operator[](size_t i) const {
return col_data[i];
}
};
/**
* \brief interface of feature matrix, needed for tree construction
* this interface defines two ways to access features:
* row access is defined by iterator of RowBatch
* col access is optional, checked by HaveColAccess, and defined by iterator of ColBatch
*/
class IFMatrix {
public:
// the interface only need to guarantee row iter
// column iter is active, when ColIterator is called, row_iter can be disabled
/*! \brief get the row iterator associated with FMatrix */
virtual utils::IIterator<RowBatch> *RowIterator(void) = 0;
/*!\brief get column iterator */
virtual utils::IIterator<ColBatch> *ColIterator(void) = 0;
/*!
* \brief get the column iterator associated with FMatrix with subset of column features
* \param fset is the list of column index set that must be contained in the returning Column iterator
* \return the column iterator, initialized so that it reads the elements in fset
*/
virtual utils::IIterator<ColBatch> *ColIterator(const std::vector<bst_uint> &fset) = 0;
/*!
* \brief check if column access is supported, if not, initialize column access
* \param enabled whether certain feature should be included in column access
* \param subsample subsample ratio when generating column access
* \param max_row_perbatch auxiliary information, maximum row used in each column batch
* this is a hint information that can be ignored by the implementation
*/
virtual void InitColAccess(const std::vector<bool> &enabled,
float subsample,
size_t max_row_perbatch) = 0;
// the following are column meta data, should be able to answer them fast
/*! \return whether column access is enabled */
virtual bool HaveColAccess(void) const = 0;
/*! \return number of columns in the FMatrix */
virtual size_t NumCol(void) const = 0;
/*! \brief get number of non-missing entries in column */
virtual size_t GetColSize(size_t cidx) const = 0;
/*! \brief get column density */
virtual float GetColDensity(size_t cidx) const = 0;
/*! \brief reference of buffered rowset */
virtual const std::vector<bst_uint> &buffered_rowset(void) const = 0;
// virtual destructor
virtual ~IFMatrix(void){}
};
} // namespace xgboost
#endif // XGBOOST_DATA_H_

292
src/gbm/gblinear-inl.hpp
View File

@@ -1,292 +0,0 @@
/*!
* Copyright by Contributors
* \file gblinear-inl.hpp
* \brief Implementation of Linear booster, with L1/L2 regularization: Elastic Net
* the update rule is parallel coordinate descent (shotgun)
* \author Tianqi Chen
*/
#ifndef XGBOOST_GBM_GBLINEAR_INL_HPP_
#define XGBOOST_GBM_GBLINEAR_INL_HPP_
#include <vector>
#include <string>
#include <sstream>
#include <algorithm>
#include "./gbm.h"
#include "../tree/updater.h"
namespace xgboost {
namespace gbm {
/*!
* \brief gradient boosted linear model
* \tparam FMatrix the data type updater taking
*/
class GBLinear : public IGradBooster {
public:
virtual ~GBLinear(void) {
}
// set model parameters
virtual void SetParam(const char *name, const char *val) {
using namespace std;
if (!strncmp(name, "bst:", 4)) {
param.SetParam(name + 4, val);
}
if (model.weight.size() == 0) {
model.param.SetParam(name, val);
}
}
virtual void LoadModel(utils::IStream &fi, bool with_pbuffer) { // NOLINT(*)
model.LoadModel(fi);
}
virtual void SaveModel(utils::IStream &fo, bool with_pbuffer) const { // NOLINT(*)
model.SaveModel(fo);
}
virtual void InitModel(void) {
model.InitModel();
}
virtual void DoBoost(IFMatrix *p_fmat,
int64_t buffer_offset,
const BoosterInfo &info,
std::vector<bst_gpair> *in_gpair) {
std::vector<bst_gpair> &gpair = *in_gpair;
const int ngroup = model.param.num_output_group;
const std::vector<bst_uint> &rowset = p_fmat->buffered_rowset();
// for all the output group
for (int gid = 0; gid < ngroup; ++gid) {
double sum_grad = 0.0, sum_hess = 0.0;
const bst_omp_uint ndata = static_cast<bst_omp_uint>(rowset.size());
#pragma omp parallel for schedule(static) reduction(+: sum_grad, sum_hess)
for (bst_omp_uint i = 0; i < ndata; ++i) {
bst_gpair &p = gpair[rowset[i] * ngroup + gid];
if (p.hess >= 0.0f) {
sum_grad += p.grad; sum_hess += p.hess;
}
}
// remove bias effect
bst_float dw = static_cast<bst_float>(
param.learning_rate * param.CalcDeltaBias(sum_grad, sum_hess, model.bias()[gid]));
model.bias()[gid] += dw;
// update grad value
#pragma omp parallel for schedule(static)
for (bst_omp_uint i = 0; i < ndata; ++i) {
bst_gpair &p = gpair[rowset[i] * ngroup + gid];
if (p.hess >= 0.0f) {
p.grad += p.hess * dw;
}
}
}
utils::IIterator<ColBatch> *iter = p_fmat->ColIterator();
while (iter->Next()) {
// number of features
const ColBatch &batch = iter->Value();
const bst_omp_uint nfeat = static_cast<bst_omp_uint>(batch.size);
#pragma omp parallel for schedule(static)
for (bst_omp_uint i = 0; i < nfeat; ++i) {
const bst_uint fid = batch.col_index[i];
ColBatch::Inst col = batch[i];
for (int gid = 0; gid < ngroup; ++gid) {
double sum_grad = 0.0, sum_hess = 0.0;
for (bst_uint j = 0; j < col.length; ++j) {
const float v = col[j].fvalue;
bst_gpair &p = gpair[col[j].index * ngroup + gid];
if (p.hess < 0.0f) continue;
sum_grad += p.grad * v;
sum_hess += p.hess * v * v;
}
float &w = model[fid][gid];
bst_float dw = static_cast<bst_float>(param.learning_rate *
param.CalcDelta(sum_grad, sum_hess, w));
w += dw;
// update grad value
for (bst_uint j = 0; j < col.length; ++j) {
bst_gpair &p = gpair[col[j].index * ngroup + gid];
if (p.hess < 0.0f) continue;
p.grad += p.hess * col[j].fvalue * dw;
}
}
}
}
}
virtual void Predict(IFMatrix *p_fmat,
int64_t buffer_offset,
const BoosterInfo &info,
std::vector<float> *out_preds,
unsigned ntree_limit = 0) {
utils::Check(ntree_limit == 0,
"GBLinear::Predict ntrees is only valid for gbtree predictor");
std::vector<float> &preds = *out_preds;
preds.resize(0);
// start collecting the prediction
utils::IIterator<RowBatch> *iter = p_fmat->RowIterator();
const int ngroup = model.param.num_output_group;
while (iter->Next()) {
const RowBatch &batch = iter->Value();
utils::Assert(batch.base_rowid * ngroup == preds.size(),
"base_rowid is not set correctly");
// output convention: nrow * k, where nrow is number of rows
// k is number of group
preds.resize(preds.size() + batch.size * ngroup);
// parallel over local batch
const bst_omp_uint nsize = static_cast<bst_omp_uint>(batch.size);
#pragma omp parallel for schedule(static)
for (bst_omp_uint i = 0; i < nsize; ++i) {
const size_t ridx = batch.base_rowid + i;
// loop over output groups
for (int gid = 0; gid < ngroup; ++gid) {
this->Pred(batch[i], &preds[ridx * ngroup]);
}
}
}
}
virtual void Predict(const SparseBatch::Inst &inst,
std::vector<float> *out_preds,
unsigned ntree_limit,
unsigned root_index) {
const int ngroup = model.param.num_output_group;
for (int gid = 0; gid < ngroup; ++gid) {
this->Pred(inst, BeginPtr(*out_preds));
}
}
virtual void PredictLeaf(IFMatrix *p_fmat,
const BoosterInfo &info,
std::vector<float> *out_preds,
unsigned ntree_limit = 0) {
utils::Error("gblinear does not support predict leaf index");
}
virtual std::vector<std::string> DumpModel(const utils::FeatMap& fmap, int option) {
std::stringstream fo("");
fo << "bias:\n";
for (int i = 0; i < model.param.num_output_group; ++i) {
fo << model.bias()[i] << std::endl;
}
fo << "weight:\n";
for (int i = 0; i < model.param.num_output_group; ++i) {
for (unsigned j = 0; j <model.param.num_feature; ++j) {
fo << model[i][j] << std::endl;
}
}
std::vector<std::string> v;
v.push_back(fo.str());
return v;
}
protected:
inline void Pred(const RowBatch::Inst &inst, float *preds) {
for (int gid = 0; gid < model.param.num_output_group; ++gid) {
float psum = model.bias()[gid];
for (bst_uint i = 0; i < inst.length; ++i) {
if (inst[i].index >= model.param.num_feature) continue;
psum += inst[i].fvalue * model[inst[i].index][gid];
}
preds[gid] = psum;
}
}
// training parameter
struct ParamTrain {
/*! \brief learning_rate */
float learning_rate;
/*! \brief regularization weight for L2 norm */
float reg_lambda;
/*! \brief regularization weight for L1 norm */
float reg_alpha;
/*! \brief regularization weight for L2 norm in bias */
float reg_lambda_bias;
// parameter
ParamTrain(void) {
reg_alpha = 0.0f;
reg_lambda = 0.0f;
reg_lambda_bias = 0.0f;
learning_rate = 1.0f;
}
inline void SetParam(const char *name, const char *val) {
using namespace std;
// sync-names
if (!strcmp("eta", name)) learning_rate = static_cast<float>(atof(val));
if (!strcmp("lambda", name)) reg_lambda = static_cast<float>(atof(val));
if (!strcmp( "alpha", name)) reg_alpha = static_cast<float>(atof(val));
if (!strcmp( "lambda_bias", name)) reg_lambda_bias = static_cast<float>(atof(val));
// real names
if (!strcmp( "learning_rate", name)) learning_rate = static_cast<float>(atof(val));
if (!strcmp( "reg_lambda", name)) reg_lambda = static_cast<float>(atof(val));
if (!strcmp( "reg_alpha", name)) reg_alpha = static_cast<float>(atof(val));
if (!strcmp( "reg_lambda_bias", name)) reg_lambda_bias = static_cast<float>(atof(val));
}
// given original weight calculate delta
inline double CalcDelta(double sum_grad, double sum_hess, double w) {
if (sum_hess < 1e-5f) return 0.0f;
double tmp = w - (sum_grad + reg_lambda * w) / (sum_hess + reg_lambda);
if (tmp >=0) {
return std::max(-(sum_grad + reg_lambda * w + reg_alpha) / (sum_hess + reg_lambda), -w);
} else {
return std::min(-(sum_grad + reg_lambda * w - reg_alpha) / (sum_hess + reg_lambda), -w);
}
}
// given original weight calculate delta bias
inline double CalcDeltaBias(double sum_grad, double sum_hess, double w) {
return - (sum_grad + reg_lambda_bias * w) / (sum_hess + reg_lambda_bias);
}
};
// model for linear booster
class Model {
public:
// model parameter
struct Param {
// number of feature dimension
unsigned num_feature;
// number of output group
int num_output_group;
// reserved field
int reserved[32];
// constructor
Param(void) {
num_feature = 0;
num_output_group = 1;
std::memset(reserved, 0, sizeof(reserved));
}
inline void SetParam(const char *name, const char *val) {
using namespace std;
if (!strcmp(name, "bst:num_feature")) num_feature = static_cast<unsigned>(atoi(val));
if (!strcmp(name, "num_output_group")) num_output_group = atoi(val);
}
};
// parameter
Param param;
// weight for each of feature, bias is the last one
std::vector<float> weight;
// initialize the model parameter
inline void InitModel(void) {
// bias is the last weight
weight.resize((param.num_feature + 1) * param.num_output_group);
std::fill(weight.begin(), weight.end(), 0.0f);
}
// save the model to file
inline void SaveModel(utils::IStream &fo) const { // NOLINT(*)
fo.Write(&param, sizeof(Param));
fo.Write(weight);
}
// load model from file
inline void LoadModel(utils::IStream &fi) { // NOLINT(*)
utils::Assert(fi.Read(&param, sizeof(Param)) != 0, "Load LinearBooster");
fi.Read(&weight);
}
// model bias
inline float* bias(void) {
return &weight[param.num_feature * param.num_output_group];
}
// get i-th weight
inline float* operator[](size_t i) {
return &weight[i * param.num_output_group];
}
};
// model field
Model model;
// training parameter
ParamTrain param;
// Per feature: shuffle index of each feature index
std::vector<bst_uint> feat_index;
};
} // namespace gbm
} // namespace xgboost
#endif // XGBOOST_GBM_GBLINEAR_INL_HPP_

21
src/gbm/gbm.cpp
View File

@@ -1,21 +0,0 @@
// Copyright by Contributors
#define _CRT_SECURE_NO_WARNINGS
#define _CRT_SECURE_NO_DEPRECATE
#define NOMINMAX
#include <cstring>
#include "./gbm.h"
#include "./gbtree-inl.hpp"
#include "./gblinear-inl.hpp"
namespace xgboost {
namespace gbm {
IGradBooster* CreateGradBooster(const char *name) {
using namespace std;
if (!strcmp("gbtree", name)) return new GBTree();
if (!strcmp("gblinear", name)) return new GBLinear();
utils::Error("unknown booster type: %s", name);
return NULL;
}
} // namespace gbm
} // namespace xgboost

136
src/gbm/gbm.h
View File

@@ -1,136 +0,0 @@
/*!
* Copyright by Contributors
* \file gbm.h
* \brief interface of gradient booster, that learns through gradient statistics
* \author Tianqi Chen
*/
#ifndef XGBOOST_GBM_GBM_H_
#define XGBOOST_GBM_GBM_H_
#include <vector>
#include <string>
#include "../data.h"
#include "../utils/io.h"
#include "../utils/fmap.h"
namespace xgboost {
/*! \brief namespace for gradient booster */
namespace gbm {
/*!
* \brief interface of gradient boosting model
*/
class IGradBooster {
public:
/*!
* \brief set parameters from outside
* \param name name of the parameter
* \param val value of the parameter
*/
virtual void SetParam(const char *name, const char *val) = 0;
/*!
* \brief load model from stream
* \param fi input stream
* \param with_pbuffer whether the incoming data contains pbuffer
*/
virtual void LoadModel(utils::IStream &fi, bool with_pbuffer) = 0; // NOLINT(*)
/*!
* \brief save model to stream
* \param fo output stream
* \param with_pbuffer whether save out pbuffer
*/
virtual void SaveModel(utils::IStream &fo, bool with_pbuffer) const = 0; // NOLINT(*)
/*!
* \brief initialize the model
*/
virtual void InitModel(void) = 0;
/*!
* \brief reset the predict buffer
* this will invalidate all the previous cached results
* and recalculate from scratch
*/
virtual void ResetPredBuffer(size_t num_pbuffer) {}
/*!
* \brief whether the model allow lazy checkpoint
* return true if model is only updated in DoBoost
* after all Allreduce calls
*/
virtual bool AllowLazyCheckPoint(void) const {
return false;
}
/*!
* \brief perform update to the model(boosting)
* \param p_fmat feature matrix that provide access to features
* \param buffer_offset buffer index offset of these instances, if equals -1
* this means we do not have buffer index allocated to the gbm
* \param info meta information about training
* \param in_gpair address of the gradient pair statistics of the data
* the booster may change content of gpair
*/
virtual void DoBoost(IFMatrix *p_fmat,
int64_t buffer_offset,
const BoosterInfo &info,
std::vector<bst_gpair> *in_gpair) = 0;
/*!
* \brief generate predictions for given feature matrix
* \param p_fmat feature matrix
* \param buffer_offset buffer index offset of these instances, if equals -1
* this means we do not have buffer index allocated to the gbm
* a buffer index is assigned to each instance that requires repeative prediction
* the size of buffer is set by convention using IGradBooster.SetParam("num_pbuffer","size")
* \param info extra side information that may be needed for prediction
* \param out_preds output vector to hold the predictions
* \param ntree_limit limit the number of trees used in prediction, when it equals 0, this means
* we do not limit number of trees, this parameter is only valid for gbtree, but not for gblinear
*/
virtual void Predict(IFMatrix *p_fmat,
int64_t buffer_offset,
const BoosterInfo &info,
std::vector<float> *out_preds,
unsigned ntree_limit = 0) = 0;
/*!
* \brief online prediction function, predict score for one instance at a time
* NOTE: use the batch prediction interface if possible, batch prediction is usually
* more efficient than online prediction
* This function is NOT threadsafe, make sure you only call from one thread
*
* \param inst the instance you want to predict
* \param out_preds output vector to hold the predictions
* \param ntree_limit limit the number of trees used in prediction
* \param root_index the root index
* \sa Predict
*/
virtual void Predict(const SparseBatch::Inst &inst,
std::vector<float> *out_preds,
unsigned ntree_limit = 0,
unsigned root_index = 0) = 0;
/*!
* \brief predict the leaf index of each tree, the output will be nsample * ntree vector
* this is only valid in gbtree predictor
* \param p_fmat feature matrix
* \param info extra side information that may be needed for prediction
* \param out_preds output vector to hold the predictions
* \param ntree_limit limit the number of trees used in prediction, when it equals 0, this means
* we do not limit number of trees, this parameter is only valid for gbtree, but not for gblinear
*/
virtual void PredictLeaf(IFMatrix *p_fmat,
const BoosterInfo &info,
std::vector<float> *out_preds,
unsigned ntree_limit = 0) = 0;
/*!
* \brief dump the model in text format
* \param fmap feature map that may help give interpretations of feature
* \param option extra option of the dump model
* \return a vector of dump for boosters
*/
virtual std::vector<std::string> DumpModel(const utils::FeatMap& fmap, int option) = 0;
// destrcutor
virtual ~IGradBooster(void){}
};
/*!
* \breif create a gradient booster from given name
* \param name name of gradient booster
*/
IGradBooster* CreateGradBooster(const char *name);
} // namespace gbm
} // namespace xgboost
#endif // XGBOOST_GBM_GBM_H_

520
src/gbm/gbtree-inl.hpp
View File

@@ -1,520 +0,0 @@
/*!
* Copyright by Contributors
* \file gbtree-inl.hpp
* \brief gradient boosted tree implementation
* \author Tianqi Chen
*/
#ifndef XGBOOST_GBM_GBTREE_INL_HPP_
#define XGBOOST_GBM_GBTREE_INL_HPP_
#include <vector>
#include <utility>
#include <string>
#include <limits>
#include "./gbm.h"
#include "../utils/omp.h"
#include "../tree/updater.h"
namespace xgboost {
namespace gbm {
/*!
* \brief gradient boosted tree
*/
class GBTree : public IGradBooster {
public:
GBTree(void) {
}
virtual ~GBTree(void) {
this->Clear();
}
virtual void SetParam(const char *name, const char *val) {
using namespace std;
if (!strncmp(name, "bst:", 4)) {
cfg.push_back(std::make_pair(std::string(name+4), std::string(val)));
// set into updaters, if already initialized
for (size_t i = 0; i < updaters.size(); ++i) {
updaters[i]->SetParam(name+4, val);
}
}
if (!strcmp(name, "silent")) {
this->SetParam("bst:silent", val);
}
tparam.SetParam(name, val);
if (trees.size() == 0) mparam.SetParam(name, val);
}
virtual void LoadModel(utils::IStream &fi, bool with_pbuffer) { // NOLINT(*)
this->Clear();
utils::Check(fi.Read(&mparam, sizeof(ModelParam)) != 0,
"GBTree: invalid model file");
trees.resize(mparam.num_trees);
for (size_t i = 0; i < trees.size(); ++i) {
trees[i] = new tree::RegTree();
trees[i]->LoadModel(fi);
}
tree_info.resize(mparam.num_trees);
if (mparam.num_trees != 0) {
utils::Check(fi.Read(&tree_info[0], sizeof(int) * mparam.num_trees) != 0,
"GBTree: invalid model file");
}
if (mparam.num_pbuffer != 0 && with_pbuffer) {
pred_buffer.resize(mparam.PredBufferSize());
pred_counter.resize(mparam.PredBufferSize());
utils::Check(fi.Read(&pred_buffer[0], pred_buffer.size() * sizeof(float)) != 0,
"GBTree: invalid model file");
utils::Check(fi.Read(&pred_counter[0], pred_counter.size() * sizeof(unsigned)) != 0,
"GBTree: invalid model file");
}
}
virtual void SaveModel(utils::IStream &fo, bool with_pbuffer) const { // NOLINT(*)
utils::Assert(mparam.num_trees == static_cast<int>(trees.size()), "GBTree");
if (with_pbuffer) {
fo.Write(&mparam, sizeof(ModelParam));
} else {
ModelParam p = mparam;
p.num_pbuffer = 0;
fo.Write(&p, sizeof(ModelParam));
}
for (size_t i = 0; i < trees.size(); ++i) {
trees[i]->SaveModel(fo);
}
if (tree_info.size() != 0) {
fo.Write(BeginPtr(tree_info), sizeof(int) * tree_info.size());
}
if (mparam.num_pbuffer != 0 && with_pbuffer) {
fo.Write(BeginPtr(pred_buffer), pred_buffer.size() * sizeof(float));
fo.Write(BeginPtr(pred_counter), pred_counter.size() * sizeof(unsigned));
}
}
// initialize the predict buffer
virtual void InitModel(void) {
pred_buffer.clear(); pred_counter.clear();
pred_buffer.resize(mparam.PredBufferSize(), 0.0f);
pred_counter.resize(mparam.PredBufferSize(), 0);
utils::Assert(mparam.num_trees == 0, "GBTree: model already initialized");
utils::Assert(trees.size() == 0, "GBTree: model already initialized");
}
virtual void ResetPredBuffer(size_t num_pbuffer) {
mparam.num_pbuffer = static_cast<int64_t>(num_pbuffer);
pred_buffer.clear(); pred_counter.clear();
pred_buffer.resize(mparam.PredBufferSize(), 0.0f);
pred_counter.resize(mparam.PredBufferSize(), 0);
}
virtual bool AllowLazyCheckPoint(void) const {
return !(tparam.distcol_mode != 0 && mparam.num_output_group != 1);
}
virtual void DoBoost(IFMatrix *p_fmat,
int64_t buffer_offset,
const BoosterInfo &info,
std::vector<bst_gpair> *in_gpair) {
const std::vector<bst_gpair> &gpair = *in_gpair;
std::vector<std::vector<tree::RegTree*> > new_trees;
if (mparam.num_output_group == 1) {
new_trees.push_back(BoostNewTrees(gpair, p_fmat, buffer_offset, info, 0));
} else {
const int ngroup = mparam.num_output_group;
utils::Check(gpair.size() % ngroup == 0,
"must have exactly ngroup*nrow gpairs");
std::vector<bst_gpair> tmp(gpair.size()/ngroup);
for (int gid = 0; gid < ngroup; ++gid) {
bst_omp_uint nsize = static_cast<bst_omp_uint>(tmp.size());
#pragma omp parallel for schedule(static)
for (bst_omp_uint i = 0; i < nsize; ++i) {
tmp[i] = gpair[i * ngroup + gid];
}
new_trees.push_back(BoostNewTrees(tmp, p_fmat, buffer_offset, info, gid));
}
}
for (int gid = 0; gid < mparam.num_output_group; ++gid) {
this->CommitModel(new_trees[gid], gid);
}
}
virtual void Predict(IFMatrix *p_fmat,
int64_t buffer_offset,
const BoosterInfo &info,
std::vector<float> *out_preds,
unsigned ntree_limit = 0) {
int nthread;
#pragma omp parallel
{
nthread = omp_get_num_threads();
}
InitThreadTemp(nthread);
std::vector<float> &preds = *out_preds;
const size_t stride = info.num_row * mparam.num_output_group;
preds.resize(stride * (mparam.size_leaf_vector+1));
// start collecting the prediction
utils::IIterator<RowBatch> *iter = p_fmat->RowIterator();
iter->BeforeFirst();
while (iter->Next()) {
const RowBatch &batch = iter->Value();
// parallel over local batch
const bst_omp_uint nsize = static_cast<bst_omp_uint>(batch.size);
#pragma omp parallel for schedule(static)
for (bst_omp_uint i = 0; i < nsize; ++i) {
const int tid = omp_get_thread_num();
tree::RegTree::FVec &feats = thread_temp[tid];
int64_t ridx = static_cast<int64_t>(batch.base_rowid + i);
utils::Assert(static_cast<size_t>(ridx) < info.num_row, "data row index exceed bound");
// loop over output groups
for (int gid = 0; gid < mparam.num_output_group; ++gid) {
this->Pred(batch[i],
buffer_offset < 0 ? -1 : buffer_offset + ridx,
gid, info.GetRoot(ridx), &feats,
&preds[ridx * mparam.num_output_group + gid], stride,
ntree_limit);
}
}
}
}
virtual void Predict(const SparseBatch::Inst &inst,
std::vector<float> *out_preds,
unsigned ntree_limit,
unsigned root_index) {
if (thread_temp.size() == 0) {
thread_temp.resize(1, tree::RegTree::FVec());
thread_temp[0].Init(mparam.num_feature);
}
out_preds->resize(mparam.num_output_group * (mparam.size_leaf_vector+1));
// loop over output groups
for (int gid = 0; gid < mparam.num_output_group; ++gid) {
this->Pred(inst, -1, gid, root_index, &thread_temp[0],
&(*out_preds)[gid], mparam.num_output_group,
ntree_limit);
}
}
virtual void PredictLeaf(IFMatrix *p_fmat,
const BoosterInfo &info,
std::vector<float> *out_preds,
unsigned ntree_limit) {
int nthread;
#pragma omp parallel
{
nthread = omp_get_num_threads();
}
InitThreadTemp(nthread);
this->PredPath(p_fmat, info, out_preds, ntree_limit);
}
virtual std::vector<std::string> DumpModel(const utils::FeatMap& fmap, int option) {
std::vector<std::string> dump;
for (size_t i = 0; i < trees.size(); i++) {
dump.push_back(trees[i]->DumpModel(fmap, option&1));
}
return dump;
}
protected:
// clear the model
inline void Clear(void) {
for (size_t i = 0; i < trees.size(); ++i) {
delete trees[i];
}
for (size_t i = 0; i < updaters.size(); ++i) {
delete updaters[i];
}
updaters.clear();
trees.clear();
pred_buffer.clear();
pred_counter.clear();
}
// initialize updater before using them
inline void InitUpdater(void) {
if (tparam.updater_initialized != 0) return;
for (size_t i = 0; i < updaters.size(); ++i) {
delete updaters[i];
}
updaters.clear();
std::string tval = tparam.updater_seq;
char *pstr;
pstr = std::strtok(&tval[0], ",");
while (pstr != NULL) {
updaters.push_back(tree::CreateUpdater(pstr));
for (size_t j = 0; j < cfg.size(); ++j) {
// set parameters
updaters.back()->SetParam(cfg[j].first.c_str(), cfg[j].second.c_str());
}
pstr = std::strtok(NULL, ",");
}
tparam.updater_initialized = 1;
}
// do group specific group
inline std::vector<tree::RegTree*>
BoostNewTrees(const std::vector<bst_gpair> &gpair,
IFMatrix *p_fmat,
int64_t buffer_offset,
const BoosterInfo &info,
int bst_group) {
std::vector<tree::RegTree *> new_trees;
this->InitUpdater();
// create the trees
for (int i = 0; i < tparam.num_parallel_tree; ++i) {
new_trees.push_back(new tree::RegTree());
for (size_t j = 0; j < cfg.size(); ++j) {
new_trees.back()->param.SetParam(cfg[j].first.c_str(), cfg[j].second.c_str());
}
new_trees.back()->InitModel();
}
// update the trees
for (size_t i = 0; i < updaters.size(); ++i) {
updaters[i]->Update(gpair, p_fmat, info, new_trees);
}
// optimization, update buffer, if possible
// this is only under distributed column mode
// for safety check of lazy checkpoint
if (
buffer_offset >= 0 &&
new_trees.size() == 1 && updaters.size() > 0 &&
updaters.back()->GetLeafPosition() != NULL) {
utils::Check(info.num_row == p_fmat->buffered_rowset().size(),
"distributed mode is not compatible with prob_buffer_row");
this->UpdateBufferByPosition(p_fmat,
buffer_offset, bst_group,
*new_trees[0],
updaters.back()->GetLeafPosition());
}
return new_trees;
}
// commit new trees all at once
inline void CommitModel(const std::vector<tree::RegTree*> &new_trees, int bst_group) {
for (size_t i = 0; i < new_trees.size(); ++i) {
trees.push_back(new_trees[i]);
tree_info.push_back(bst_group);
}
mparam.num_trees += static_cast<int>(new_trees.size());
}
// update buffer by pre-cached position
inline void UpdateBufferByPosition(IFMatrix *p_fmat,
int64_t buffer_offset,
int bst_group,
const tree::RegTree &new_tree,
const int* leaf_position) {
const std::vector<bst_uint> &rowset = p_fmat->buffered_rowset();
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 int64_t bid = mparam.BufferOffset(buffer_offset + ridx, bst_group);
const int tid = leaf_position[ridx];
utils::Assert(pred_counter[bid] == trees.size(), "cached buffer not up to date");
utils::Assert(tid >= 0, "invalid leaf position");
pred_buffer[bid] += new_tree[tid].leaf_value();
for (int i = 0; i < mparam.size_leaf_vector; ++i) {
pred_buffer[bid + i + 1] += new_tree.leafvec(tid)[i];
}
pred_counter[bid] += tparam.num_parallel_tree;
}
}
// make a prediction for a single instance
inline void Pred(const RowBatch::Inst &inst,
int64_t buffer_index,
int bst_group,
unsigned root_index,
tree::RegTree::FVec *p_feats,
float *out_pred, size_t stride,
unsigned ntree_limit) {
size_t itop = 0;
float psum = 0.0f;
// sum of leaf vector
std::vector<float> vec_psum(mparam.size_leaf_vector, 0.0f);
const int64_t bid = mparam.BufferOffset(buffer_index, bst_group);
// number of valid trees
unsigned treeleft = ntree_limit == 0 ? std::numeric_limits<unsigned>::max() : ntree_limit;
// load buffered results if any
if (bid >= 0 && ntree_limit == 0) {
itop = pred_counter[bid];
psum = pred_buffer[bid];
for (int i = 0; i < mparam.size_leaf_vector; ++i) {
vec_psum[i] = pred_buffer[bid + i + 1];
}
}
if (itop != trees.size()) {
p_feats->Fill(inst);
for (size_t i = itop; i < trees.size(); ++i) {
if (tree_info[i] == bst_group) {
int tid = trees[i]->GetLeafIndex(*p_feats, root_index);
psum += (*trees[i])[tid].leaf_value();
for (int j = 0; j < mparam.size_leaf_vector; ++j) {
vec_psum[j] += trees[i]->leafvec(tid)[j];
}
if (--treeleft == 0) break;
}
}
p_feats->Drop(inst);
}
// updated the buffered results
if (bid >= 0 && ntree_limit == 0) {
pred_counter[bid] = static_cast<unsigned>(trees.size());
pred_buffer[bid] = psum;
for (int i = 0; i < mparam.size_leaf_vector; ++i) {
pred_buffer[bid + i + 1] = vec_psum[i];
}
}
out_pred[0] = psum;
for (int i = 0; i < mparam.size_leaf_vector; ++i) {
out_pred[stride * (i + 1)] = vec_psum[i];
}
}
// predict independent leaf index
inline void PredPath(IFMatrix *p_fmat,
const BoosterInfo &info,
std::vector<float> *out_preds,
unsigned ntree_limit) {
// number of valid trees
if (ntree_limit == 0 || ntree_limit > trees.size()) {
ntree_limit = static_cast<unsigned>(trees.size());
}
std::vector<float> &preds = *out_preds;
preds.resize(info.num_row * ntree_limit);
// start collecting the prediction
utils::IIterator<RowBatch> *iter = p_fmat->RowIterator();
iter->BeforeFirst();
while (iter->Next()) {
const RowBatch &batch = iter->Value();
// parallel over local batch
const bst_omp_uint nsize = static_cast<bst_omp_uint>(batch.size);
#pragma omp parallel for schedule(static)
for (bst_omp_uint i = 0; i < nsize; ++i) {
const int tid = omp_get_thread_num();
size_t ridx = static_cast<size_t>(batch.base_rowid + i);
tree::RegTree::FVec &feats = thread_temp[tid];
feats.Fill(batch[i]);
for (unsigned j = 0; j < ntree_limit; ++j) {
int tid = trees[j]->GetLeafIndex(feats, info.GetRoot(ridx));
preds[ridx * ntree_limit + j] = static_cast<float>(tid);
}
feats.Drop(batch[i]);
}
}
}
// init thread buffers
inline void InitThreadTemp(int nthread) {
int prev_thread_temp_size = thread_temp.size();
if (prev_thread_temp_size < nthread) {
thread_temp.resize(nthread, tree::RegTree::FVec());
for (int i = prev_thread_temp_size; i < nthread; ++i) {
thread_temp[i].Init(mparam.num_feature);
}
}
}
// --- data structure ---
/*! \brief training parameters */
struct TrainParam {
/*! \brief number of threads */
int nthread;
/*!
* \brief number of parallel trees constructed each iteration
* use this option to support boosted random forest
*/
int num_parallel_tree;
/*! \brief whether updater is already initialized */
int updater_initialized;
/*! \brief distributed column mode */
int distcol_mode;
/*! \brief tree updater sequence */
std::string updater_seq;
// construction
TrainParam(void) {
nthread = 0;
updater_seq = "grow_colmaker,prune";
num_parallel_tree = 1;
updater_initialized = 0;
distcol_mode = 0;
}
inline void SetParam(const char *name, const char *val){
using namespace std;
if (!strcmp(name, "updater") &&
strcmp(updater_seq.c_str(), val) != 0) {
updater_seq = val;
updater_initialized = 0;
}
if (!strcmp(name, "dsplit") && !strcmp(val, "col")) {
distcol_mode = 1;
}
if (!strcmp(name, "nthread")) {
omp_set_num_threads(nthread = atoi(val));
}
if (!strcmp(name, "num_parallel_tree")) {
num_parallel_tree = atoi(val);
}
}
};
/*! \brief model parameters */
struct ModelParam {
/*! \brief number of trees */
int num_trees;
/*! \brief number of root: default 0, means single tree */
int num_roots;
/*! \brief number of features to be used by trees */
int num_feature;
/*! \brief size of prediction buffer allocated used for buffering */
int64_t num_pbuffer;
/*!
* \brief how many output group a single instance can produce
* this affects the behavior of number of output we have:
* suppose we have n instance and k group, output will be k*n
*/
int num_output_group;
/*! \brief size of leaf vector needed in tree */
int size_leaf_vector;
/*! \brief reserved parameters */
int reserved[31];
/*! \brief constructor */
ModelParam(void) {
std::memset(this, 0, sizeof(ModelParam));
num_trees = 0;
num_roots = num_feature = 0;
num_pbuffer = 0;
num_output_group = 1;
size_leaf_vector = 0;
}
/*!
* \brief set parameters from outside
* \param name name of the parameter
* \param val value of the parameter
*/
inline void SetParam(const char *name, const char *val) {
using namespace std;
if (!strcmp("num_pbuffer", name)) num_pbuffer = atol(val);
if (!strcmp("num_output_group", name)) num_output_group = atol(val);
if (!strcmp("bst:num_roots", name)) num_roots = atoi(val);
if (!strcmp("bst:num_feature", name)) num_feature = atoi(val);
if (!strcmp("bst:size_leaf_vector", name)) size_leaf_vector = atoi(val);
}
/*! \return size of prediction buffer actually needed */
inline size_t PredBufferSize(void) const {
return num_output_group * num_pbuffer * (size_leaf_vector + 1);
}
/*!
* \brief get the buffer offset given a buffer index and group id
* \return calculated buffer offset
*/
inline int64_t BufferOffset(int64_t buffer_index, int bst_group) const {
if (buffer_index < 0) return -1;
utils::Check(buffer_index < num_pbuffer, "buffer_index exceed num_pbuffer");
return (buffer_index + num_pbuffer * bst_group) * (size_leaf_vector + 1);
}
};
// training parameter
TrainParam tparam;
// model parameter
ModelParam mparam;
/*! \brief vector of trees stored in the model */
std::vector<tree::RegTree*> trees;
/*! \brief some information indicator of the tree, reserved */
std::vector<int> tree_info;
/*! \brief prediction buffer */
std::vector<float> pred_buffer;
/*! \brief prediction buffer counter, remember the prediction */
std::vector<unsigned> pred_counter;
// ----training fields----
// configurations for tree
std::vector< std::pair<std::string, std::string> > cfg;
// temporal storage for per thread
std::vector<tree::RegTree::FVec> thread_temp;
// the updaters that can be applied to each of tree
std::vector<tree::IUpdater*> updaters;
};
} // namespace gbm
} // namespace xgboost
#endif // XGBOOST_GBM_GBTREE_INL_HPP_

229
src/io/dmlc_simple.cpp
View File

@@ -1,229 +0,0 @@
// Copyright by Contributors
#define _CRT_SECURE_NO_WARNINGS
#define _CRT_SECURE_NO_DEPRECATE
#define NOMINMAX
#include <string>
#include "../utils/io.h"
// implements a single no split version of DMLC
// in case we want to avoid dependency on dmlc-core
namespace xgboost {
namespace utils {
/*!
* \brief line split implementation from single FILE
* simply returns lines of files, used for stdin
*/
class SingleFileSplit : public dmlc::InputSplit {
public:
explicit SingleFileSplit(const char *fname)
: use_stdin_(false),
chunk_begin_(NULL), chunk_end_(NULL) {
if (!std::strcmp(fname, "stdin")) {
#ifndef XGBOOST_STRICT_CXX98_
use_stdin_ = true; fp_ = stdin;
#endif
}
if (!use_stdin_) {
fp_ = utils::FopenCheck(fname, "rb");
}
buffer_.resize(kBufferSize);
}
virtual ~SingleFileSplit(void) {
if (!use_stdin_) std::fclose(fp_);
}
virtual size_t Read(void *ptr, size_t size) {
return std::fread(ptr, 1, size, fp_);
}
virtual void Write(const void *ptr, size_t size) {
utils::Error("cannot do write in inputsplit");
}
virtual void BeforeFirst(void) {
std::fseek(fp_, 0, SEEK_SET);
}
virtual bool NextRecord(Blob *out_rec) {
if (chunk_begin_ == chunk_end_) {
if (!LoadChunk()) return false;
}
char *next = FindNextRecord(chunk_begin_,
chunk_end_);
out_rec->dptr = chunk_begin_;
out_rec->size = next - chunk_begin_;
chunk_begin_ = next;
return true;
}
virtual bool NextChunk(Blob *out_chunk) {
if (chunk_begin_ == chunk_end_) {
if (!LoadChunk()) return false;
}
out_chunk->dptr = chunk_begin_;
out_chunk->size = chunk_end_ - chunk_begin_;
chunk_begin_ = chunk_end_;
return true;
}
inline bool ReadChunk(void *buf, size_t *size) {
size_t max_size = *size;
if (max_size <= overflow_.length()) {
*size = 0; return true;
}
if (overflow_.length() != 0) {
std::memcpy(buf, BeginPtr(overflow_), overflow_.length());
}
size_t olen = overflow_.length();
overflow_.resize(0);
size_t nread = this->Read(reinterpret_cast<char*>(buf) + olen,
max_size - olen);
nread += olen;
if (nread == 0) return false;
if (nread != max_size) {
*size = nread;
return true;
} else {
const char *bptr = reinterpret_cast<const char*>(buf);
// return the last position where a record starts
const char *bend = this->FindLastRecordBegin(bptr, bptr + max_size);
*size = bend - bptr;
overflow_.resize(max_size - *size);
if (overflow_.length() != 0) {
std::memcpy(BeginPtr(overflow_), bend, overflow_.length());
}
return true;
}
}
protected:
inline const char* FindLastRecordBegin(const char *begin,
const char *end) {
if (begin == end) return begin;
for (const char *p = end - 1; p != begin; --p) {
if (*p == '\n' || *p == '\r') return p + 1;
}
return begin;
}
inline char* FindNextRecord(char *begin, char *end) {
char *p;
for (p = begin; p != end; ++p) {
if (*p == '\n' || *p == '\r') break;
}
for (; p != end; ++p) {
if (*p != '\n' && *p != '\r') return p;
}
return end;
}
inline bool LoadChunk(void) {
while (true) {
size_t size = buffer_.length();
if (!ReadChunk(BeginPtr(buffer_), &size)) return false;
if (size == 0) {
buffer_.resize(buffer_.length() * 2);
} else {
chunk_begin_ = reinterpret_cast<char *>(BeginPtr(buffer_));
chunk_end_ = chunk_begin_ + size;
break;
}
}
return true;
}
private:
// buffer size
static const size_t kBufferSize = 1 << 18UL;
// file
std::FILE *fp_;
bool use_stdin_;
// internal overflow
std::string overflow_;
// internal buffer
std::string buffer_;
// beginning of chunk
char *chunk_begin_;
// end of chunk
char *chunk_end_;
};
class StdFile : public dmlc::Stream {
public:
explicit StdFile(std::FILE *fp, bool use_stdio)
: fp(fp), use_stdio(use_stdio) {
}
virtual ~StdFile(void) {
this->Close();
}
virtual size_t Read(void *ptr, size_t size) {
return std::fread(ptr, 1, size, fp);
}
virtual void Write(const void *ptr, size_t size) {
Check(std::fwrite(ptr, size, 1, fp) == 1, "StdFile::Write: fwrite error!");
}
virtual void Seek(size_t pos) {
std::fseek(fp, static_cast<long>(pos), SEEK_SET); // NOLINT(*)
}
virtual size_t Tell(void) {
return std::ftell(fp);
}
virtual bool AtEnd(void) const {
return std::feof(fp) != 0;
}
inline void Close(void) {
if (fp != NULL && !use_stdio) {
std::fclose(fp); fp = NULL;
}
}
private:
std::FILE *fp;
bool use_stdio;
};
} // namespace utils
} // namespace xgboost
namespace dmlc {
InputSplit* InputSplit::Create(const char *uri,
unsigned part,
unsigned nsplit,
const char *type) {
using namespace std;
using namespace xgboost;
const char *msg = "xgboost is compiled in local mode\n"\
"to use hdfs, s3 or distributed version, compile with make dmlc=1";
utils::Check(strncmp(uri, "s3://", 5) != 0, msg);
utils::Check(strncmp(uri, "hdfs://", 7) != 0, msg);
utils::Check(nsplit == 1, msg);
return new utils::SingleFileSplit(uri);
}
Stream *Stream::Create(const char *fname, const char * const mode, bool allow_null) {
using namespace std;
using namespace xgboost;
const char *msg = "xgboost is compiled in local mode\n"\
"to use hdfs, s3 or distributed version, compile with make dmlc=1";
utils::Check(strncmp(fname, "s3://", 5) != 0, msg);
utils::Check(strncmp(fname, "hdfs://", 7) != 0, msg);
std::FILE *fp = NULL;
bool use_stdio = false;
using namespace std;
#ifndef XGBOOST_STRICT_CXX98_
if (!strcmp(fname, "stdin")) {
use_stdio = true; fp = stdin;
}
if (!strcmp(fname, "stdout")) {
use_stdio = true; fp = stdout;
}
#endif
if (!strncmp(fname, "file://", 7)) fname += 7;
if (!use_stdio) {
std::string flag = mode;
if (flag == "w") flag = "wb";
if (flag == "r") flag = "rb";
fp = fopen64(fname, flag.c_str());
}
if (fp != NULL) {
return new utils::StdFile(fp, use_stdio);
} else {
utils::Check(allow_null, "fail to open file %s", fname);
return NULL;
}
}
} // namespace dmlc

97
src/io/io.cpp
View File

@@ -1,97 +0,0 @@
// Copyright 2014 by Contributors
#define _CRT_SECURE_NO_WARNINGS
#define _CRT_SECURE_NO_DEPRECATE
#define NOMINMAX
#include <string>
#include "./io.h"
#include "../utils/io.h"
#include "../utils/utils.h"
#include "simple_dmatrix-inl.hpp"
#include "page_dmatrix-inl.hpp"
namespace xgboost {
namespace io {
DataMatrix* LoadDataMatrix(const char *fname,
bool silent,
bool savebuffer,
bool loadsplit,
const char *cache_file) {
using namespace std;
std::string fname_ = fname;
const char *dlm = strchr(fname, '#');
if (dlm != NULL) {
utils::Check(strchr(dlm + 1, '#') == NULL,
"only one `#` is allowed in file path for cachefile specification");
utils::Check(cache_file == NULL,
"can only specify the cachefile with `#` or argument, not both");
fname_ = std::string(fname, dlm - fname);
fname = fname_.c_str();
cache_file = dlm +1;
}
if (cache_file == NULL) {
if (!std::strcmp(fname, "stdin") ||
!std::strncmp(fname, "s3://", 5) ||
!std::strncmp(fname, "hdfs://", 7) ||
loadsplit) {
DMatrixSimple *dmat = new DMatrixSimple();
dmat->LoadText(fname, silent, loadsplit);
return dmat;
}
int magic;
utils::FileStream fs(utils::FopenCheck(fname, "rb"));
utils::Check(fs.Read(&magic, sizeof(magic)) != 0, "invalid input file format");
fs.Seek(0);
if (magic == DMatrixSimple::kMagic) {
DMatrixSimple *dmat = new DMatrixSimple();
dmat->LoadBinary(fs, silent, fname);
fs.Close();
return dmat;
}
fs.Close();
DMatrixSimple *dmat = new DMatrixSimple();
dmat->CacheLoad(fname, silent, savebuffer);
return dmat;
} else {
std::string cache_fname = cache_file;
if (loadsplit) {
std::ostringstream os;
os << cache_file << ".r" << rabit::GetRank();
cache_fname = os.str();
cache_file = cache_fname.c_str();
}
FILE *fi = fopen64(cache_file, "rb");
if (fi != NULL) {
DMatrixPage *dmat = new DMatrixPage();
utils::FileStream fs(fi);
dmat->LoadBinary(fs, silent, cache_file);
fs.Close();
return dmat;
} else {
if (fname[0] == '!') {
DMatrixHalfRAM *dmat = new DMatrixHalfRAM();
dmat->LoadText(fname + 1, cache_file, false, loadsplit);
return dmat;
} else {
DMatrixPage *dmat = new DMatrixPage();
dmat->LoadText(fname, cache_file, false, loadsplit);
return dmat;
}
}
}
}
void SaveDataMatrix(const DataMatrix &dmat, const char *fname, bool silent) {
if (dmat.magic == DMatrixSimple::kMagic) {
const DMatrixSimple *p_dmat = static_cast<const DMatrixSimple*>(&dmat);
p_dmat->SaveBinary(fname, silent);
} else {
DMatrixSimple smat;
smat.CopyFrom(dmat);
smat.SaveBinary(fname, silent);
}
}
} // namespace io
} // namespace xgboost

47
src/io/io.h
View File

@@ -1,47 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file io.h
* \brief handles input data format of xgboost
* I/O module handles a specific DMatrix format
* \author Tianqi Chen
*/
#ifndef XGBOOST_IO_IO_H_
#define XGBOOST_IO_IO_H_
#include "../data.h"
#include "../learner/dmatrix.h"
namespace xgboost {
/*! \brief namespace related to data format */
namespace io {
/*! \brief DMatrix object that I/O module support save/load */
typedef learner::DMatrix DataMatrix;
/*!
* \brief load DataMatrix from stream
* \param fname file name to be loaded
* \param silent whether print message during loading
* \param savebuffer whether temporal buffer the file if the file is in text format
* \param loadsplit whether we only load a split of input files
* such that each worker node get a split of the data
* \param cache_file name of cache_file, used by external memory version
* can be NULL, if cache_file is specified, this will be the temporal
* space that can be re-used to store intermediate data
* \return a loaded DMatrix
*/
DataMatrix* LoadDataMatrix(const char *fname,
bool silent,
bool savebuffer,
bool loadsplit,
const char *cache_file = NULL);
/*!
* \brief save DataMatrix into stream,
* note: the saved dmatrix format may not be in exactly same as input
* SaveDMatrix will choose the best way to materialize the dmatrix.
* \param dmat the dmatrix to be saved
* \param fname file name to be saved
* \param silent whether print message during saving
*/
void SaveDataMatrix(const DataMatrix &dmat, const char *fname, bool silent = false);
} // namespace io
} // namespace xgboost
#endif // XGBOOST_IO_IO_H_

212
src/io/libsvm_parser.h
View File

@@ -1,212 +0,0 @@
/*!
* Copyright (c) 2015 by Contributors
* \file libsvm_parser.h
* \brief iterator parser to parse libsvm format
* \author Tianqi Chen
*/
#ifndef XGBOOST_IO_LIBSVM_PARSER_H_
#define XGBOOST_IO_LIBSVM_PARSER_H_
#define NOMINMAX
#include <vector>
#include <cstring>
#include <cctype>
#include <algorithm>
#include "../utils/omp.h"
#include "../utils/utils.h"
#include "../sync/sync.h"
#include "../utils/thread_buffer.h"
#include "./sparse_batch_page.h"
namespace xgboost {
namespace io {
/*! \brief page returned by libsvm parser */
struct LibSVMPage : public SparsePage {
std::vector<float> label;
// overload clear
inline void Clear() {
SparsePage::Clear();
label.clear();
}
};
/*!
* \brief libsvm parser that parses the input lines
* and returns rows in input data
* factory that was used by threadbuffer template
*/
class LibSVMPageFactory {
public:
LibSVMPageFactory()
: bytes_read_(0), at_head_(true) {
}
inline bool Init(void) {
return true;
}
inline void Setup(dmlc::InputSplit *source,
int nthread) {
source_ = source;
int maxthread;
#pragma omp parallel
{
maxthread = omp_get_num_procs();
}
maxthread = std::max(maxthread / 2, 1);
nthread_ = std::min(maxthread, nthread);
}
inline void SetParam(const char *name, const char *val) {}
inline bool LoadNext(std::vector<LibSVMPage> *data) {
return FillData(data);
}
inline void FreeSpace(std::vector<LibSVMPage> *a) {
delete a;
}
inline std::vector<LibSVMPage> *Create(void) {
return new std::vector<LibSVMPage>();
}
inline void BeforeFirst(void) {
utils::Assert(at_head_, "cannot call beforefirst");
}
inline void Destroy(void) {
delete source_;
}
inline size_t bytes_read(void) const {
return bytes_read_;
}
protected:
inline bool FillData(std::vector<LibSVMPage> *data) {
dmlc::InputSplit::Blob chunk;
if (!source_->NextChunk(&chunk)) return false;
int nthread;
#pragma omp parallel num_threads(nthread_)
{
nthread = omp_get_num_threads();
}
// reserve space for data
data->resize(nthread);
bytes_read_ += chunk.size;
utils::Assert(chunk.size != 0, "LibSVMParser.FileData");
char *head = reinterpret_cast<char*>(chunk.dptr);
#pragma omp parallel num_threads(nthread_)
{
// threadid
int tid = omp_get_thread_num();
size_t nstep = (chunk.size + nthread - 1) / nthread;
size_t sbegin = std::min(tid * nstep, chunk.size);
size_t send = std::min((tid + 1) * nstep, chunk.size);
char *pbegin = BackFindEndLine(head + sbegin, head);
char *pend;
if (tid + 1 == nthread) {
pend = head + send;
} else {
pend = BackFindEndLine(head + send, head);
}
ParseBlock(pbegin, pend, &(*data)[tid]);
}
return true;
}
/*!
* \brief parse data into out
* \param begin beginning of buffer
* \param end end of buffer
*/
inline void ParseBlock(char *begin,
char *end,
LibSVMPage *out) {
using namespace std;
out->Clear();
char *p = begin;
while (p != end) {
while (isspace(*p) && p != end) ++p;
if (p == end) break;
char *head = p;
while (isdigit(*p) && p != end) ++p;
if (*p == ':') {
out->data.push_back(SparseBatch::Entry(atol(head),
static_cast<bst_float>(atof(p + 1))));
} else {
if (out->label.size() != 0) {
out->offset.push_back(out->data.size());
}
out->label.push_back(static_cast<float>(atof(head)));
}
while (!isspace(*p) && p != end) ++p;
}
if (out->label.size() != 0) {
out->offset.push_back(out->data.size());
}
utils::Check(out->label.size() + 1 == out->offset.size(),
"LibSVMParser inconsistent");
}
/*!
* \brief start from bptr, go backward and find first endof line
* \param bptr end position to go backward
* \param begin the beginning position of buffer
* \return position of first endof line going backward
*/
inline char* BackFindEndLine(char *bptr,
char *begin) {
for (; bptr != begin; --bptr) {
if (*bptr == '\n' || *bptr == '\r') return bptr;
}
return begin;
}
private:
// nthread
int nthread_;
// number of bytes readed
size_t bytes_read_;
// at beginning, at end of stream
bool at_head_;
// source split that provides the data
dmlc::InputSplit *source_;
};
class LibSVMParser : public utils::IIterator<LibSVMPage> {
public:
explicit LibSVMParser(dmlc::InputSplit *source,
int nthread)
: at_end_(false), data_ptr_(0), data_(NULL) {
itr.SetParam("buffer_size", "2");
itr.get_factory().Setup(source, nthread);
itr.Init();
}
virtual void BeforeFirst(void) {
itr.BeforeFirst();
}
virtual bool Next(void) {
if (at_end_) return false;
while (true) {
if (data_ == NULL || data_ptr_ >= data_->size()) {
if (!itr.Next(data_)) {
at_end_ = true; return false;
} else {
data_ptr_ = 0;
}
}
while (data_ptr_ < data_->size()) {
data_ptr_ += 1;
if ((*data_)[data_ptr_ - 1].Size() != 0) {
return true;
}
}
}
return true;
}
virtual const LibSVMPage &Value(void) const {
return (*data_)[data_ptr_ - 1];
}
inline size_t bytes_read(void) const {
return itr.get_factory().bytes_read();
}
private:
bool at_end_;
size_t data_ptr_;
std::vector<LibSVMPage> *data_;
utils::ThreadBuffer<std::vector<LibSVMPage>*, LibSVMPageFactory> itr;
};
} // namespace io
} // namespace xgboost
#endif // XGBOOST_IO_LIBSVM_PARSER_H_

260
src/io/page_dmatrix-inl.hpp
View File

@@ -1,260 +0,0 @@
/*!
* Copyright (c) 2014 by Contributors
* \file page_dmatrix-inl.hpp
* row iterator based on sparse page
* \author Tianqi Chen
*/
#ifndef XGBOOST_IO_PAGE_DMATRIX_INL_HPP_
#define XGBOOST_IO_PAGE_DMATRIX_INL_HPP_
#include <vector>
#include <string>
#include <algorithm>
#include "../data.h"
#include "../utils/iterator.h"
#include "../utils/thread_buffer.h"
#include "./simple_fmatrix-inl.hpp"
#include "./sparse_batch_page.h"
#include "./page_fmatrix-inl.hpp"
#include "./libsvm_parser.h"
namespace xgboost {
namespace io {
/*! \brief thread buffer iterator */
class ThreadRowPageIterator: public utils::IIterator<RowBatch> {
public:
ThreadRowPageIterator(void) {
itr.SetParam("buffer_size", "4");
page_ = NULL;
base_rowid_ = 0;
}
virtual ~ThreadRowPageIterator(void) {}
virtual void Init(void) {
}
virtual void BeforeFirst(void) {
itr.BeforeFirst();
base_rowid_ = 0;
}
virtual bool Next(void) {
if (!itr.Next(page_)) return false;
out_ = page_->GetRowBatch(base_rowid_);
base_rowid_ += out_.size;
return true;
}
virtual const RowBatch &Value(void) const {
return out_;
}
/*! \brief load and initialize the iterator with fi */
inline void Load(const utils::FileStream &fi) {
itr.get_factory().SetFile(fi, 0);
itr.Init();
this->BeforeFirst();
}
private:
// base row id
size_t base_rowid_;
// output data
RowBatch out_;
SparsePage *page_;
utils::ThreadBuffer<SparsePage*, SparsePageFactory> itr;
};
/*! \brief data matrix using page */
template<int TKMagic>
class DMatrixPageBase : public DataMatrix {
public:
DMatrixPageBase(void) : DataMatrix(kMagic) {
iter_ = new ThreadRowPageIterator();
}
// virtual destructor
virtual ~DMatrixPageBase(void) {
// do not delete row iterator, since it is owned by fmat
// to be cleaned up in a more clear way
}
/*! \brief save a DataMatrix as DMatrixPage */
inline static void Save(const char *fname_, const DataMatrix &mat, bool silent) {
std::string fname = fname_;
utils::FileStream fs(utils::FopenCheck(fname.c_str(), "wb"));
int magic = kMagic;
fs.Write(&magic, sizeof(magic));
mat.info.SaveBinary(fs);
fs.Close();
fname += ".row.blob";
utils::IIterator<RowBatch> *iter = mat.fmat()->RowIterator();
utils::FileStream fbin(utils::FopenCheck(fname.c_str(), "wb"));
SparsePage page;
iter->BeforeFirst();
while (iter->Next()) {
const RowBatch &batch = iter->Value();
for (size_t i = 0; i < batch.size; ++i) {
page.Push(batch[i]);
if (page.MemCostBytes() >= kPageSize) {
page.Save(&fbin); page.Clear();
}
}
}
if (page.data.size() != 0) page.Save(&fbin);
fbin.Close();
if (!silent) {
utils::Printf("DMatrixPage: %lux%lu is saved to %s\n",
static_cast<unsigned long>(mat.info.num_row()), // NOLINT(*)
static_cast<unsigned long>(mat.info.num_col()), fname_); // NOLINT(*)
}
}
/*! \brief load and initialize the iterator with fi */
inline void LoadBinary(utils::FileStream &fi, // NOLINT(*)
bool silent,
const char *fname_) {
this->set_cache_file(fname_);
std::string fname = fname_;
int tmagic;
utils::Check(fi.Read(&tmagic, sizeof(tmagic)) != 0, "invalid input file format");
this->CheckMagic(tmagic);
this->info.LoadBinary(fi);
// load in the row data file
fname += ".row.blob";
utils::FileStream fs(utils::FopenCheck(fname.c_str(), "rb"));
iter_->Load(fs);
if (!silent) {
utils::Printf("DMatrixPage: %lux%lu matrix is loaded",
static_cast<unsigned long>(info.num_row()), // NOLINT(*)
static_cast<unsigned long>(info.num_col())); // NOLINT(*)
if (fname_ != NULL) {
utils::Printf(" from %s\n", fname_);
} else {
utils::Printf("\n");
}
if (info.group_ptr.size() != 0) {
utils::Printf("data contains %u groups\n", (unsigned)info.group_ptr.size() - 1);
}
}
}
/*! \brief save a LibSVM format file as DMatrixPage */
inline void LoadText(const char *uri,
const char* cache_file,
bool silent,
bool loadsplit) {
if (!silent) {
utils::Printf("start generate text file from %s\n", uri);
}
int rank = 0, npart = 1;
if (loadsplit) {
rank = rabit::GetRank();
npart = rabit::GetWorldSize();
}
this->set_cache_file(cache_file);
std::string fname_row = std::string(cache_file) + ".row.blob";
utils::FileStream fo(utils::FopenCheck(fname_row.c_str(), "wb"));
SparsePage page;
size_t bytes_write = 0;
double tstart = rabit::utils::GetTime();
LibSVMParser parser(
dmlc::InputSplit::Create(uri, rank, npart, "text"), 16);
info.Clear();
while (parser.Next()) {
const LibSVMPage &batch = parser.Value();
size_t nlabel = info.labels.size();
info.labels.resize(nlabel + batch.label.size());
if (batch.label.size() != 0) {
std::memcpy(BeginPtr(info.labels) + nlabel,
BeginPtr(batch.label),
batch.label.size() * sizeof(float));
}
page.Push(batch);
for (size_t i = 0; i < batch.data.size(); ++i) {
info.info.num_col = std::max(info.info.num_col,
static_cast<size_t>(batch.data[i].index+1));
}
if (page.MemCostBytes() >= kPageSize) {
bytes_write += page.MemCostBytes();
page.Save(&fo);
page.Clear();
double tdiff = rabit::utils::GetTime() - tstart;
if (!silent) {
utils::Printf("Writting to %s in %g MB/s, %lu MB written\n",
cache_file, (bytes_write >> 20UL) / tdiff,
(bytes_write >> 20UL));
}
}
info.info.num_row += batch.label.size();
}
if (page.data.size() != 0) {
page.Save(&fo);
}
fo.Close();
iter_->Load(utils::FileStream(utils::FopenCheck(fname_row.c_str(), "rb")));
// save data matrix
utils::FileStream fs(utils::FopenCheck(cache_file, "wb"));
int tmagic = kMagic;
fs.Write(&tmagic, sizeof(tmagic));
this->info.SaveBinary(fs);
fs.Close();
if (!silent) {
utils::Printf("DMatrixPage: %lux%lu is parsed from %s\n",
static_cast<unsigned long>(info.num_row()), // NOLINT(*)
static_cast<unsigned long>(info.num_col()), // NOLINT(*)
uri);
}
}
/*! \brief magic number used to identify DMatrix */
static const int kMagic = TKMagic;
/*! \brief page size 32 MB */
static const size_t kPageSize = 32UL << 20UL;
protected:
virtual void set_cache_file(const std::string &cache_file) = 0;
virtual void CheckMagic(int tmagic) = 0;
/*! \brief row iterator */
ThreadRowPageIterator *iter_;
};
class DMatrixPage : public DMatrixPageBase<0xffffab02> {
public:
DMatrixPage(void) {
fmat_ = new FMatrixPage(iter_, this->info);
}
virtual ~DMatrixPage(void) {
delete fmat_;
}
virtual IFMatrix *fmat(void) const {
return fmat_;
}
virtual void set_cache_file(const std::string &cache_file) {
fmat_->set_cache_file(cache_file);
}
virtual void CheckMagic(int tmagic) {
utils::Check(tmagic == DMatrixPageBase<0xffffab02>::kMagic ||
tmagic == DMatrixPageBase<0xffffab03>::kMagic,
"invalid format,magic number mismatch");
}
/*! \brief the real fmatrix */
FMatrixPage *fmat_;
};
// mix of FMatrix S and DMatrix
// cost half of ram usually as DMatrixSimple
class DMatrixHalfRAM : public DMatrixPageBase<0xffffab03> {
public:
DMatrixHalfRAM(void) {
fmat_ = new FMatrixS(iter_, this->info);
}
virtual ~DMatrixHalfRAM(void) {
delete fmat_;
}
virtual IFMatrix *fmat(void) const {
return fmat_;
}
virtual void set_cache_file(const std::string &cache_file) {
}
virtual void CheckMagic(int tmagic) {
utils::Check(tmagic == DMatrixPageBase<0xffffab02>::kMagic ||
tmagic == DMatrixPageBase<0xffffab03>::kMagic,
"invalid format,magic number mismatch");
}
/*! \brief the real fmatrix */
IFMatrix *fmat_;
};
} // namespace io
} // namespace xgboost
#endif // XGBOOST_IO_PAGE_ROW_ITER_INL_HPP_

360
src/io/page_fmatrix-inl.hpp
View File

@@ -1,360 +0,0 @@
/*!
* Copyright (c) 2014 by Contributors
* \file page_fmatrix-inl.hpp
* col iterator based on sparse page
* \author Tianqi Chen
*/
#ifndef XGBOOST_IO_PAGE_FMATRIX_INL_HPP_
#define XGBOOST_IO_PAGE_FMATRIX_INL_HPP_
#include <vector>
#include <string>
#include <algorithm>
namespace xgboost {
namespace io {
/*! \brief thread buffer iterator */
class ThreadColPageIterator: public utils::IIterator<ColBatch> {
public:
ThreadColPageIterator(void) {
itr.SetParam("buffer_size", "2");
page_ = NULL;
}
virtual ~ThreadColPageIterator(void) {}
virtual void Init(void) {}
virtual void BeforeFirst(void) {
itr.BeforeFirst();
}
virtual bool Next(void) {
if (!itr.Next(page_)) return false;
out_.col_index = BeginPtr(itr.get_factory().index_set());
col_data_.resize(page_->offset.size() - 1, SparseBatch::Inst(NULL, 0));
for (size_t i = 0; i < col_data_.size(); ++i) {
col_data_[i] = SparseBatch::Inst
(BeginPtr(page_->data) + page_->offset[i],
static_cast<bst_uint>(page_->offset[i + 1] - page_->offset[i]));
}
out_.col_data = BeginPtr(col_data_);
out_.size = col_data_.size();
return true;
}
virtual const ColBatch &Value(void) const {
return out_;
}
/*! \brief load and initialize the iterator with fi */
inline void SetFile(const utils::FileStream &fi) {
itr.get_factory().SetFile(fi);
itr.Init();
}
// set index set
inline void SetIndexSet(const std::vector<bst_uint> &fset, bool load_all) {
itr.get_factory().SetIndexSet(fset, load_all);
}
private:
// output data
ColBatch out_;
SparsePage *page_;
std::vector<SparseBatch::Inst> col_data_;
utils::ThreadBuffer<SparsePage*, SparsePageFactory> itr;
};
struct ColConvertFactory {
inline bool Init(void) {
return true;
}
inline void Setup(float pkeep,
size_t max_row_perbatch,
size_t num_col,
utils::IIterator<RowBatch> *iter,
std::vector<bst_uint> *buffered_rowset,
const std::vector<bool> *enabled) {
pkeep_ = pkeep;
max_row_perbatch_ = max_row_perbatch;
num_col_ = num_col;
iter_ = iter;
buffered_rowset_ = buffered_rowset;
enabled_ = enabled;
}
inline SparsePage *Create(void) {
return new SparsePage();
}
inline void FreeSpace(SparsePage *a) {
delete a;
}
inline void SetParam(const char *name, const char *val) {}
inline bool LoadNext(SparsePage *val) {
tmp_.Clear();
size_t btop = buffered_rowset_->size();
while (iter_->Next()) {
const RowBatch &batch = iter_->Value();
for (size_t i = 0; i < batch.size; ++i) {
bst_uint ridx = static_cast<bst_uint>(batch.base_rowid + i);
if (pkeep_ == 1.0f || random::SampleBinary(pkeep_)) {
buffered_rowset_->push_back(ridx);
tmp_.Push(batch[i]);
}
}
if (tmp_.MemCostBytes() >= kPageSize ||
tmp_.Size() >= max_row_perbatch_) {
this->MakeColPage(tmp_, BeginPtr(*buffered_rowset_) + btop,
*enabled_, val);
return true;
}
}
if (tmp_.Size() != 0) {
this->MakeColPage(tmp_, BeginPtr(*buffered_rowset_) + btop,
*enabled_, val);
return true;
} else {
return false;
}
}
inline void Destroy(void) {}
inline void BeforeFirst(void) {}
inline void MakeColPage(const SparsePage &prow,
const bst_uint *ridx,
const std::vector<bool> &enabled,
SparsePage *pcol) {
pcol->Clear();
int nthread;
#pragma omp parallel
{
nthread = omp_get_num_threads();
int max_nthread = std::max(omp_get_num_procs() / 2 - 4, 1);
if (nthread > max_nthread) {
nthread = max_nthread;
}
}
pcol->Clear();
utils::ParallelGroupBuilder<SparseBatch::Entry>
builder(&pcol->offset, &pcol->data);
builder.InitBudget(num_col_, nthread);
bst_omp_uint ndata = static_cast<bst_uint>(prow.Size());
#pragma omp parallel for schedule(static) num_threads(nthread)
for (bst_omp_uint i = 0; i < ndata; ++i) {
int tid = omp_get_thread_num();
for (size_t j = prow.offset[i]; j < prow.offset[i+1]; ++j) {
const SparseBatch::Entry &e = prow.data[j];
if (enabled[e.index]) {
builder.AddBudget(e.index, tid);
}
}
}
builder.InitStorage();
#pragma omp parallel for schedule(static) num_threads(nthread)
for (bst_omp_uint i = 0; i < ndata; ++i) {
int tid = omp_get_thread_num();
for (size_t j = prow.offset[i]; j < prow.offset[i+1]; ++j) {
const SparseBatch::Entry &e = prow.data[j];
builder.Push(e.index,
SparseBatch::Entry(ridx[i], e.fvalue),
tid);
}
}
utils::Assert(pcol->Size() == num_col_, "inconsistent col data");
// sort columns
bst_omp_uint ncol = static_cast<bst_omp_uint>(pcol->Size());
#pragma omp parallel for schedule(dynamic, 1) num_threads(nthread)
for (bst_omp_uint i = 0; i < ncol; ++i) {
if (pcol->offset[i] < pcol->offset[i + 1]) {
std::sort(BeginPtr(pcol->data) + pcol->offset[i],
BeginPtr(pcol->data) + pcol->offset[i + 1],
SparseBatch::Entry::CmpValue);
}
}
}
// probability of keep
float pkeep_;
// maximum number of rows per batch
size_t max_row_perbatch_;
// number of columns
size_t num_col_;
// row batch iterator
utils::IIterator<RowBatch> *iter_;
// buffered rowset
std::vector<bst_uint> *buffered_rowset_;
// enabled marks
const std::vector<bool> *enabled_;
// internal temp cache
SparsePage tmp_;
/*! \brief page size 256 M */
static const size_t kPageSize = 256 << 20UL;
};
/*!
* \brief sparse matrix that support column access, CSC
*/
class FMatrixPage : public IFMatrix {
public:
typedef SparseBatch::Entry Entry;
/*! \brief constructor */
FMatrixPage(utils::IIterator<RowBatch> *iter,
const learner::MetaInfo &info) : info(info) {
this->iter_ = iter;
}
// destructor
virtual ~FMatrixPage(void) {
if (iter_ != NULL) delete iter_;
}
/*! \return whether column access is enabled */
virtual bool HaveColAccess(void) const {
return col_size_.size() != 0;
}
/*! \brief get number of columns */
virtual size_t NumCol(void) const {
utils::Check(this->HaveColAccess(), "NumCol:need column access");
return col_size_.size();
}
/*! \brief get number of buffered rows */
virtual const std::vector<bst_uint> &buffered_rowset(void) const {
return buffered_rowset_;
}
/*! \brief get column size */
virtual size_t GetColSize(size_t cidx) const {
return col_size_[cidx];
}
/*! \brief get column density */
virtual float GetColDensity(size_t cidx) const {
size_t nmiss = num_buffered_row_ - (col_size_[cidx]);
return 1.0f - (static_cast<float>(nmiss)) / num_buffered_row_;
}
virtual void InitColAccess(const std::vector<bool> &enabled,
float pkeep, size_t max_row_perbatch) {
if (this->HaveColAccess()) return;
if (TryLoadColData()) return;
this->InitColData(enabled, pkeep, max_row_perbatch);
utils::Check(TryLoadColData(), "failed on creating col.blob");
}
/*!
* \brief get the row iterator associated with FMatrix
*/
virtual utils::IIterator<RowBatch>* RowIterator(void) {
iter_->BeforeFirst();
return iter_;
}
/*!
* \brief get the column based iterator
*/
virtual utils::IIterator<ColBatch>* ColIterator(void) {
size_t ncol = this->NumCol();
col_index_.resize(0);
for (size_t i = 0; i < ncol; ++i) {
col_index_.push_back(static_cast<bst_uint>(i));
}
col_iter_.SetIndexSet(col_index_, false);
col_iter_.BeforeFirst();
return &col_iter_;
}
/*!
* \brief column based iterator
*/
virtual utils::IIterator<ColBatch> *ColIterator(const std::vector<bst_uint> &fset) {
size_t ncol = this->NumCol();
col_index_.resize(0);
for (size_t i = 0; i < fset.size(); ++i) {
if (fset[i] < ncol) col_index_.push_back(fset[i]);
}
col_iter_.SetIndexSet(col_index_, false);
col_iter_.BeforeFirst();
return &col_iter_;
}
// set the cache file name
inline void set_cache_file(const std::string &cache_file) {
col_data_name_ = std::string(cache_file) + ".col.blob";
col_meta_name_ = std::string(cache_file) + ".col.meta";
}
protected:
inline bool TryLoadColData(void) {
std::FILE *fi = fopen64(col_meta_name_.c_str(), "rb");
if (fi == NULL) return false;
utils::FileStream fs(fi);
LoadMeta(&fs);
fs.Close();
fi = utils::FopenCheck(col_data_name_.c_str(), "rb");
if (fi == NULL) return false;
col_iter_.SetFile(utils::FileStream(fi));
return true;
}
inline void LoadMeta(utils::IStream *fi) {
utils::Check(fi->Read(&num_buffered_row_, sizeof(num_buffered_row_)) != 0,
"invalid col.blob file");
utils::Check(fi->Read(&buffered_rowset_),
"invalid col.blob file");
utils::Check(fi->Read(&col_size_),
"invalid col.blob file");
}
inline void SaveMeta(utils::IStream *fo) {
fo->Write(&num_buffered_row_, sizeof(num_buffered_row_));
fo->Write(buffered_rowset_);
fo->Write(col_size_);
}
/*!
* \brief initialize column data
* \param enabled the list of enabled columns
* \param pkeep probability to keep a row
* \param max_row_perbatch maximum row per batch
*/
inline void InitColData(const std::vector<bool> &enabled,
float pkeep, size_t max_row_perbatch) {
// clear rowset
buffered_rowset_.clear();
col_size_.resize(info.num_col());
std::fill(col_size_.begin(), col_size_.end(), 0);
utils::FileStream fo;
fo = utils::FileStream(utils::FopenCheck(col_data_name_.c_str(), "wb"));
iter_->BeforeFirst();
double tstart = rabit::utils::GetTime();
size_t bytes_write = 0;
utils::ThreadBuffer<SparsePage*, ColConvertFactory> citer;
citer.SetParam("buffer_size", "2");
citer.get_factory().Setup(pkeep, max_row_perbatch, info.num_col(),
iter_, &buffered_rowset_, &enabled);
citer.Init();
SparsePage *pcol;
while (citer.Next(pcol)) {
for (size_t i = 0; i < pcol->Size(); ++i) {
col_size_[i] += pcol->offset[i + 1] - pcol->offset[i];
}
pcol->Save(&fo);
size_t spage = pcol->MemCostBytes();
bytes_write += spage;
double tnow = rabit::utils::GetTime();
double tdiff = tnow - tstart;
utils::Printf("Writing to %s in %g MB/s, %lu MB written\n",
col_data_name_.c_str(),
(bytes_write >> 20UL) / tdiff,
(bytes_write >> 20UL));
}
fo.Close();
num_buffered_row_ = buffered_rowset_.size();
fo = utils::FileStream(utils::FopenCheck(col_meta_name_.c_str(), "wb"));
this->SaveMeta(&fo);
fo.Close();
}
private:
/*! \brief page size 256 M */
static const size_t kPageSize = 256 << 20UL;
// shared meta info with DMatrix
const learner::MetaInfo &info;
// row iterator
utils::IIterator<RowBatch> *iter_;
/*! \brief column based data file name */
std::string col_data_name_;
/*! \brief column based data file name */
std::string col_meta_name_;
/*! \brief list of row index that are buffered */
std::vector<bst_uint> buffered_rowset_;
// number of buffered rows
size_t num_buffered_row_;
// count for column data
std::vector<size_t> col_size_;
// internal column index for output
std::vector<bst_uint> col_index_;
// internal thread backed col iterator
ThreadColPageIterator col_iter_;
};
} // namespace io
} // namespace xgboost
#endif // XGBOOST_IO_PAGE_FMATRIX_INL_HPP_

324
src/io/simple_dmatrix-inl.hpp
View File

@@ -1,324 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file simple_dmatrix-inl.hpp
* \brief simple implementation of DMatrixS that can be used
* the data format of xgboost is templatized, which means it can accept
* any data structure that implements the function defined by FMatrix
* this file is a specific implementation of input data structure that can be used by BoostLearner
* \author Tianqi Chen
*/
#ifndef XGBOOST_IO_SIMPLE_DMATRIX_INL_HPP_
#define XGBOOST_IO_SIMPLE_DMATRIX_INL_HPP_
#include <string>
#include <cstring>
#include <vector>
#include <sstream>
#include <algorithm>
#include "../data.h"
#include "../utils/utils.h"
#include "../learner/dmatrix.h"
#include "./io.h"
#include "./simple_fmatrix-inl.hpp"
#include "../sync/sync.h"
#include "./libsvm_parser.h"
namespace xgboost {
namespace io {
/*! \brief implementation of DataMatrix, in CSR format */
class DMatrixSimple : public DataMatrix {
public:
// constructor
DMatrixSimple(void) : DataMatrix(kMagic) {
fmat_ = new FMatrixS(new OneBatchIter(this), this->info);
this->Clear();
}
// virtual destructor
virtual ~DMatrixSimple(void) {
delete fmat_;
}
virtual IFMatrix *fmat(void) const {
return fmat_;
}
/*! \brief clear the storage */
inline void Clear(void) {
row_ptr_.clear();
row_ptr_.push_back(0);
row_data_.clear();
info.Clear();
}
/*! \brief copy content data from source matrix */
inline void CopyFrom(const DataMatrix &src) {
this->Clear();
this->info = src.info;
// clone data contents from src matrix
utils::IIterator<RowBatch> *iter = src.fmat()->RowIterator();
iter->BeforeFirst();
while (iter->Next()) {
const RowBatch &batch = iter->Value();
for (size_t i = 0; i < batch.size; ++i) {
RowBatch::Inst inst = batch[i];
row_data_.resize(row_data_.size() + inst.length);
if (inst.length != 0) {
std::memcpy(&row_data_[row_ptr_.back()], inst.data,
sizeof(RowBatch::Entry) * inst.length);
}
row_ptr_.push_back(row_ptr_.back() + inst.length);
}
}
}
/*!
* \brief add a row to the matrix
* \param feats features
* \return the index of added row
*/
inline size_t AddRow(const std::vector<RowBatch::Entry> &feats) {
for (size_t i = 0; i < feats.size(); ++i) {
row_data_.push_back(feats[i]);
info.info.num_col = std::max(info.info.num_col,
static_cast<size_t>(feats[i].index+1));
}
row_ptr_.push_back(row_ptr_.back() + feats.size());
info.info.num_row += 1;
return row_ptr_.size() - 2;
}
/*!
* \brief load split of input, used in distributed mode
* \param uri the uri of input
* \param loadsplit whether loadsplit of data or all the data
* \param silent whether print information or not
*/
inline void LoadText(const char *uri, bool silent = false, bool loadsplit = false) {
int rank = 0, npart = 1;
if (loadsplit) {
rank = rabit::GetRank();
npart = rabit::GetWorldSize();
}
LibSVMParser parser(
dmlc::InputSplit::Create(uri, rank, npart, "text"), 16);
this->Clear();
while (parser.Next()) {
const LibSVMPage &batch = parser.Value();
size_t nlabel = info.labels.size();
info.labels.resize(nlabel + batch.label.size());
if (batch.label.size() != 0) {
std::memcpy(BeginPtr(info.labels) + nlabel,
BeginPtr(batch.label),
batch.label.size() * sizeof(float));
}
size_t ndata = row_data_.size();
row_data_.resize(ndata + batch.data.size());
if (batch.data.size() != 0) {
std::memcpy(BeginPtr(row_data_) + ndata,
BeginPtr(batch.data),
batch.data.size() * sizeof(RowBatch::Entry));
}
row_ptr_.resize(row_ptr_.size() + batch.label.size());
for (size_t i = 0; i < batch.label.size(); ++i) {
row_ptr_[nlabel + i + 1] = row_ptr_[nlabel] + batch.offset[i + 1];
}
info.info.num_row += batch.Size();
for (size_t i = 0; i < batch.data.size(); ++i) {
info.info.num_col = std::max(info.info.num_col,
static_cast<size_t>(batch.data[i].index+1));
}
}
if (!silent) {
utils::Printf("%lux%lu matrix with %lu entries is loaded from %s\n",
static_cast<unsigned long>(info.num_row()), // NOLINT(*)
static_cast<unsigned long>(info.num_col()), // NOLINT(*)
static_cast<unsigned long>(row_data_.size()), uri); // NOLINT(*)
}
// try to load in additional file
if (!loadsplit) {
std::string name = uri;
std::string gname = name + ".group";
if (info.TryLoadGroup(gname.c_str(), silent)) {
utils::Check(info.group_ptr.back() == info.num_row(),
"DMatrix: group data does not match the number of rows in features");
}
std::string wname = name + ".weight";
if (info.TryLoadFloatInfo("weight", wname.c_str(), silent)) {
utils::Check(info.weights.size() == info.num_row(),
"DMatrix: weight data does not match the number of rows in features");
}
std::string mname = name + ".base_margin";
if (info.TryLoadFloatInfo("base_margin", mname.c_str(), silent)) {
}
}
}
/*!
* \brief load from binary file
* \param fname name of binary data
* \param silent whether print information or not
* \return whether loading is success
*/
inline bool LoadBinary(const char* fname, bool silent = false) {
std::FILE *fp = fopen64(fname, "rb");
if (fp == NULL) return false;
utils::FileStream fs(fp);
this->LoadBinary(fs, silent, fname);
fs.Close();
return true;
}
/*!
* \brief load from binary stream
* \param fs input file stream
* \param silent whether print information during loading
* \param fname file name, used to print message
*/
inline void LoadBinary(utils::IStream &fs, bool silent = false, const char *fname = NULL) { // NOLINT(*)
int tmagic;
utils::Check(fs.Read(&tmagic, sizeof(tmagic)) != 0, "invalid input file format");
utils::Check(tmagic == kMagic, "\"%s\" invalid format, magic number mismatch",
fname == NULL ? "" : fname);
info.LoadBinary(fs);
LoadBinary(fs, &row_ptr_, &row_data_);
fmat_->LoadColAccess(fs);
if (!silent) {
utils::Printf("%lux%lu matrix with %lu entries is loaded",
static_cast<unsigned long>(info.num_row()), // NOLINT(*)
static_cast<unsigned long>(info.num_col()), // NOLINT(*)
static_cast<unsigned long>(row_data_.size())); // NOLINT(*)
if (fname != NULL) {
utils::Printf(" from %s\n", fname);
} else {
utils::Printf("\n");
}
if (info.group_ptr.size() != 0) {
utils::Printf("data contains %u groups\n", (unsigned)info.group_ptr.size()-1);
}
}
}
/*!
* \brief save to binary file
* \param fname name of binary data
* \param silent whether print information or not
*/
inline void SaveBinary(const char* fname, bool silent = false) const {
utils::FileStream fs(utils::FopenCheck(fname, "wb"));
int tmagic = kMagic;
fs.Write(&tmagic, sizeof(tmagic));
info.SaveBinary(fs);
SaveBinary(fs, row_ptr_, row_data_);
fmat_->SaveColAccess(fs);
fs.Close();
if (!silent) {
utils::Printf("%lux%lu matrix with %lu entries is saved to %s\n",
static_cast<unsigned long>(info.num_row()), // NOLINT(*)
static_cast<unsigned long>(info.num_col()), // NOLINT(*)
static_cast<unsigned long>(row_data_.size()), fname); // NOLINT(*)
if (info.group_ptr.size() != 0) {
utils::Printf("data contains %u groups\n",
static_cast<unsigned>(info.group_ptr.size()-1));
}
}
}
/*!
* \brief cache load data given a file name, if filename ends with .buffer, direct load binary
* otherwise the function will first check if fname + '.buffer' exists,
* if binary buffer exists, it will reads from binary buffer, otherwise, it will load from text file,
* and try to create a buffer file
* \param fname name of binary data
* \param silent whether print information or not
* \param savebuffer whether do save binary buffer if it is text
*/
inline void CacheLoad(const char *fname, bool silent = false, bool savebuffer = true) {
using namespace std;
size_t len = strlen(fname);
if (len > 8 && !strcmp(fname + len - 7, ".buffer")) {
if (!this->LoadBinary(fname, silent)) {
utils::Error("can not open file \"%s\"", fname);
}
return;
}
char bname[1024];
utils::SPrintf(bname, sizeof(bname), "%s.buffer", fname);
if (!this->LoadBinary(bname, silent)) {
this->LoadText(fname, silent);
if (savebuffer) this->SaveBinary(bname, silent);
}
}
// data fields
/*! \brief row pointer of CSR sparse storage */
std::vector<size_t> row_ptr_;
/*! \brief data in the row */
std::vector<RowBatch::Entry> row_data_;
/*! \brief the real fmatrix */
FMatrixS *fmat_;
/*! \brief magic number used to identify DMatrix */
static const int kMagic = 0xffffab01;
protected:
/*!
* \brief save data to binary stream
* \param fo output stream
* \param ptr pointer data
* \param data data content
*/
inline static void SaveBinary(utils::IStream &fo, // NOLINT(*)
const std::vector<size_t> &ptr,
const std::vector<RowBatch::Entry> &data) {
size_t nrow = ptr.size() - 1;
fo.Write(&nrow, sizeof(size_t));
fo.Write(BeginPtr(ptr), ptr.size() * sizeof(size_t));
if (data.size() != 0) {
fo.Write(BeginPtr(data), data.size() * sizeof(RowBatch::Entry));
}
}
/*!
* \brief load data from binary stream
* \param fi input stream
* \param out_ptr pointer data
* \param out_data data content
*/
inline static void LoadBinary(utils::IStream &fi, // NOLINT(*)
std::vector<size_t> *out_ptr,
std::vector<RowBatch::Entry> *out_data) {
size_t nrow;
utils::Check(fi.Read(&nrow, sizeof(size_t)) != 0, "invalid input file format");
out_ptr->resize(nrow + 1);
utils::Check(fi.Read(BeginPtr(*out_ptr), out_ptr->size() * sizeof(size_t)) != 0,
"invalid input file format");
out_data->resize(out_ptr->back());
if (out_data->size() != 0) {
utils::Assert(fi.Read(BeginPtr(*out_data), out_data->size() * sizeof(RowBatch::Entry)) != 0,
"invalid input file format");
}
}
// one batch iterator that return content in the matrix
struct OneBatchIter: utils::IIterator<RowBatch> {
explicit OneBatchIter(DMatrixSimple *parent)
: at_first_(true), parent_(parent) {}
virtual ~OneBatchIter(void) {}
virtual void BeforeFirst(void) {
at_first_ = true;
}
virtual bool Next(void) {
if (!at_first_) return false;
at_first_ = false;
batch_.size = parent_->row_ptr_.size() - 1;
batch_.base_rowid = 0;
batch_.ind_ptr = BeginPtr(parent_->row_ptr_);
batch_.data_ptr = BeginPtr(parent_->row_data_);
return true;
}
virtual const RowBatch &Value(void) const {
return batch_;
}
private:
// whether is at first
bool at_first_;
// pointer to parent
DMatrixSimple *parent_;
// temporal space for batch
RowBatch batch_;
};
};
} // namespace io
} // namespace xgboost
#endif // namespace XGBOOST_IO_SIMPLE_DMATRIX_INL_HPP_

374
src/io/simple_fmatrix-inl.hpp
View File

@@ -1,374 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file simple_fmatrix-inl.hpp
* \brief the input data structure for gradient boosting
* \author Tianqi Chen
*/
#ifndef XGBOOST_IO_SIMPLE_FMATRIX_INL_HPP_
#define XGBOOST_IO_SIMPLE_FMATRIX_INL_HPP_
#include <limits>
#include <algorithm>
#include <vector>
#include "../data.h"
#include "../utils/utils.h"
#include "../utils/random.h"
#include "../utils/omp.h"
#include "../learner/dmatrix.h"
#include "../utils/group_data.h"
#include "./sparse_batch_page.h"
namespace xgboost {
namespace io {
/*!
* \brief sparse matrix that support column access, CSC
*/
class FMatrixS : public IFMatrix {
public:
typedef SparseBatch::Entry Entry;
/*! \brief constructor */
FMatrixS(utils::IIterator<RowBatch> *iter,
const learner::MetaInfo &info)
: info_(info) {
this->iter_ = iter;
}
// destructor
virtual ~FMatrixS(void) {
if (iter_ != NULL) delete iter_;
}
/*! \return whether column access is enabled */
virtual bool HaveColAccess(void) const {
return col_size_.size() != 0;
}
/*! \brief get number of columns */
virtual size_t NumCol(void) const {
utils::Check(this->HaveColAccess(), "NumCol:need column access");
return col_size_.size();
}
/*! \brief get number of buffered rows */
virtual const std::vector<bst_uint> &buffered_rowset(void) const {
return buffered_rowset_;
}
/*! \brief get column size */
virtual size_t GetColSize(size_t cidx) const {
return col_size_[cidx];
}
/*! \brief get column density */
virtual float GetColDensity(size_t cidx) const {
size_t nmiss = buffered_rowset_.size() - col_size_[cidx];
return 1.0f - (static_cast<float>(nmiss)) / buffered_rowset_.size();
}
virtual void InitColAccess(const std::vector<bool> &enabled,
float pkeep, size_t max_row_perbatch) {
if (this->HaveColAccess()) return;
this->InitColData(enabled, pkeep, max_row_perbatch);
}
/*!
* \brief get the row iterator associated with FMatrix
*/
virtual utils::IIterator<RowBatch>* RowIterator(void) {
iter_->BeforeFirst();
return iter_;
}
/*!
* \brief get the column based iterator
*/
virtual utils::IIterator<ColBatch>* ColIterator(void) {
size_t ncol = this->NumCol();
col_iter_.col_index_.resize(ncol);
for (size_t i = 0; i < ncol; ++i) {
col_iter_.col_index_[i] = static_cast<bst_uint>(i);
}
col_iter_.BeforeFirst();
return &col_iter_;
}
/*!
* \brief column based iterator
*/
virtual utils::IIterator<ColBatch> *ColIterator(const std::vector<bst_uint> &fset) {
size_t ncol = this->NumCol();
col_iter_.col_index_.resize(0);
for (size_t i = 0; i < fset.size(); ++i) {
if (fset[i] < ncol) col_iter_.col_index_.push_back(fset[i]);
}
col_iter_.BeforeFirst();
return &col_iter_;
}
/*!
* \brief save column access data into stream
* \param fo output stream to save to
*/
inline void SaveColAccess(utils::IStream &fo) const { // NOLINT(*)
size_t n = 0;
fo.Write(&n, sizeof(n));
}
/*!
* \brief load column access data from stream
* \param fo output stream to load from
*/
inline void LoadColAccess(utils::IStream &fi) { // NOLINT(*)
// do nothing in load col access
}
protected:
/*!
* \brief initialize column data
* \param enabled the list of enabled columns
* \param pkeep probability to keep a row
* \param max_row_perbatch maximum row per batch
*/
inline void InitColData(const std::vector<bool> &enabled,
float pkeep, size_t max_row_perbatch) {
col_iter_.Clear();
if (info_.num_row() < max_row_perbatch) {
SparsePage *page = new SparsePage();
this->MakeOneBatch(enabled, pkeep, page);
col_iter_.cpages_.push_back(page);
} else {
this->MakeManyBatch(enabled, pkeep, max_row_perbatch);
}
// setup col-size
col_size_.resize(info_.num_col());
std::fill(col_size_.begin(), col_size_.end(), 0);
for (size_t i = 0; i < col_iter_.cpages_.size(); ++i) {
SparsePage *pcol = col_iter_.cpages_[i];
for (size_t j = 0; j < pcol->Size(); ++j) {
col_size_[j] += pcol->offset[j + 1] - pcol->offset[j];
}
}
}
/*!
* \brief make column page from iterator
* \param pkeep probability to keep a row
* \param pcol the target column
*/
inline void MakeOneBatch(const std::vector<bool> &enabled,
float pkeep,
SparsePage *pcol) {
// clear rowset
buffered_rowset_.clear();
// bit map
int nthread;
std::vector<bool> bmap;
#pragma omp parallel
{
nthread = omp_get_num_threads();
}
pcol->Clear();
utils::ParallelGroupBuilder<SparseBatch::Entry>
builder(&pcol->offset, &pcol->data);
builder.InitBudget(info_.num_col(), nthread);
// start working
iter_->BeforeFirst();
while (iter_->Next()) {
const RowBatch &batch = iter_->Value();
bmap.resize(bmap.size() + batch.size, true);
long batch_size = static_cast<long>(batch.size); // NOLINT(*)
for (long i = 0; i < batch_size; ++i) { // NOLINT(*)
bst_uint ridx = static_cast<bst_uint>(batch.base_rowid + i);
if (pkeep == 1.0f || random::SampleBinary(pkeep)) {
buffered_rowset_.push_back(ridx);
} else {
bmap[i] = false;
}
}
#pragma omp parallel for schedule(static)
for (long i = 0; i < batch_size; ++i) { // NOLINT(*)
int tid = omp_get_thread_num();
bst_uint ridx = static_cast<bst_uint>(batch.base_rowid + i);
if (bmap[ridx]) {
RowBatch::Inst inst = batch[i];
for (bst_uint j = 0; j < inst.length; ++j) {
if (enabled[inst[j].index]) {
builder.AddBudget(inst[j].index, tid);
}
}
}
}
}
builder.InitStorage();
iter_->BeforeFirst();
while (iter_->Next()) {
const RowBatch &batch = iter_->Value();
#pragma omp parallel for schedule(static)
for (long i = 0; i < static_cast<long>(batch.size); ++i) { // NOLINT(*)
int tid = omp_get_thread_num();
bst_uint ridx = static_cast<bst_uint>(batch.base_rowid + i);
if (bmap[ridx]) {
RowBatch::Inst inst = batch[i];
for (bst_uint j = 0; j < inst.length; ++j) {
if (enabled[inst[j].index]) {
builder.Push(inst[j].index,
Entry((bst_uint)(batch.base_rowid+i),
inst[j].fvalue), tid);
}
}
}
}
}
utils::Assert(pcol->Size() == info_.num_col(),
"inconsistent col data");
// sort columns
bst_omp_uint ncol = static_cast<bst_omp_uint>(pcol->Size());
#pragma omp parallel for schedule(dynamic, 1) num_threads(nthread)
for (bst_omp_uint i = 0; i < ncol; ++i) {
if (pcol->offset[i] < pcol->offset[i + 1]) {
std::sort(BeginPtr(pcol->data) + pcol->offset[i],
BeginPtr(pcol->data) + pcol->offset[i + 1],
SparseBatch::Entry::CmpValue);
}
}
}
inline void MakeManyBatch(const std::vector<bool> &enabled,
float pkeep, size_t max_row_perbatch) {
size_t btop = 0;
buffered_rowset_.clear();
// internal temp cache
SparsePage tmp; tmp.Clear();
iter_->BeforeFirst();
while (iter_->Next()) {
const RowBatch &batch = iter_->Value();
for (size_t i = 0; i < batch.size; ++i) {
bst_uint ridx = static_cast<bst_uint>(batch.base_rowid + i);
if (pkeep == 1.0f || random::SampleBinary(pkeep)) {
buffered_rowset_.push_back(ridx);
tmp.Push(batch[i]);
}
if (tmp.Size() >= max_row_perbatch) {
SparsePage *page = new SparsePage();
this->MakeColPage(tmp.GetRowBatch(0),
BeginPtr(buffered_rowset_) + btop,
enabled, page);
col_iter_.cpages_.push_back(page);
btop = buffered_rowset_.size();
tmp.Clear();
}
}
}
if (tmp.Size() != 0) {
SparsePage *page = new SparsePage();
this->MakeColPage(tmp.GetRowBatch(0),
BeginPtr(buffered_rowset_) + btop,
enabled, page);
col_iter_.cpages_.push_back(page);
}
}
// make column page from subset of rowbatchs
inline void MakeColPage(const RowBatch &batch,
const bst_uint *ridx,
const std::vector<bool> &enabled,
SparsePage *pcol) {
int nthread;
#pragma omp parallel
{
nthread = omp_get_num_threads();
int max_nthread = std::max(omp_get_num_procs() / 2 - 2, 1);
if (nthread > max_nthread) {
nthread = max_nthread;
}
}
pcol->Clear();
utils::ParallelGroupBuilder<SparseBatch::Entry>
builder(&pcol->offset, &pcol->data);
builder.InitBudget(info_.num_col(), nthread);
bst_omp_uint ndata = static_cast<bst_uint>(batch.size);
#pragma omp parallel for schedule(static) num_threads(nthread)
for (bst_omp_uint i = 0; i < ndata; ++i) {
int tid = omp_get_thread_num();
RowBatch::Inst inst = batch[i];
for (bst_uint j = 0; j < inst.length; ++j) {
const SparseBatch::Entry &e = inst[j];
if (enabled[e.index]) {
builder.AddBudget(e.index, tid);
}
}
}
builder.InitStorage();
#pragma omp parallel for schedule(static) num_threads(nthread)
for (bst_omp_uint i = 0; i < ndata; ++i) {
int tid = omp_get_thread_num();
RowBatch::Inst inst = batch[i];
for (bst_uint j = 0; j < inst.length; ++j) {
const SparseBatch::Entry &e = inst[j];
builder.Push(e.index,
SparseBatch::Entry(ridx[i], e.fvalue),
tid);
}
}
utils::Assert(pcol->Size() == info_.num_col(), "inconsistent col data");
// sort columns
bst_omp_uint ncol = static_cast<bst_omp_uint>(pcol->Size());
#pragma omp parallel for schedule(dynamic, 1) num_threads(nthread)
for (bst_omp_uint i = 0; i < ncol; ++i) {
if (pcol->offset[i] < pcol->offset[i + 1]) {
std::sort(BeginPtr(pcol->data) + pcol->offset[i],
BeginPtr(pcol->data) + pcol->offset[i + 1],
SparseBatch::Entry::CmpValue);
}
}
}
private:
// one batch iterator that return content in the matrix
struct ColBatchIter: utils::IIterator<ColBatch> {
ColBatchIter(void) : data_ptr_(0) {}
virtual ~ColBatchIter(void) {
this->Clear();
}
virtual void BeforeFirst(void) {
data_ptr_ = 0;
}
virtual bool Next(void) {
if (data_ptr_ >= cpages_.size()) return false;
data_ptr_ += 1;
SparsePage *pcol = cpages_[data_ptr_ - 1];
batch_.size = col_index_.size();
col_data_.resize(col_index_.size(), SparseBatch::Inst(NULL, 0));
for (size_t i = 0; i < col_data_.size(); ++i) {
const bst_uint ridx = col_index_[i];
col_data_[i] = SparseBatch::Inst
(BeginPtr(pcol->data) + pcol->offset[ridx],
static_cast<bst_uint>(pcol->offset[ridx + 1] - pcol->offset[ridx]));
}
batch_.col_index = BeginPtr(col_index_);
batch_.col_data = BeginPtr(col_data_);
return true;
}
virtual const ColBatch &Value(void) const {
return batch_;
}
inline void Clear(void) {
for (size_t i = 0; i < cpages_.size(); ++i) {
delete cpages_[i];
}
cpages_.clear();
}
// data content
std::vector<bst_uint> col_index_;
// column content
std::vector<ColBatch::Inst> col_data_;
// column sparse pages
std::vector<SparsePage*> cpages_;
// data pointer
size_t data_ptr_;
// temporal space for batch
ColBatch batch_;
};
// --- data structure used to support InitColAccess --
// column iterator
ColBatchIter col_iter_;
// shared meta info with DMatrix
const learner::MetaInfo &info_;
// row iterator
utils::IIterator<RowBatch> *iter_;
/*! \brief list of row index that are buffered */
std::vector<bst_uint> buffered_rowset_;
// count for column data
std::vector<size_t> col_size_;
};
} // namespace io
} // namespace xgboost
#endif // XGBOOST_IO_SLICE_FMATRIX_INL_HPP_

272
src/io/sparse_batch_page.h
View File

@@ -1,272 +0,0 @@
/*!
* Copyright (c) 2014 by Contributors
* \file sparse_batch_page.h
* content holder of sparse batch that can be saved to disk
* the representation can be effectively
* use in external memory computation
* \author Tianqi Chen
*/
#ifndef XGBOOST_IO_SPARSE_BATCH_PAGE_H_
#define XGBOOST_IO_SPARSE_BATCH_PAGE_H_
#include <vector>
#include <algorithm>
#include "../data.h"
namespace xgboost {
namespace io {
/*!
* \brief storage unit of sparse batch
*/
class SparsePage {
public:
/*! \brief offset of the segments */
std::vector<size_t> offset;
/*! \brief the data of the segments */
std::vector<SparseBatch::Entry> data;
/*! \brief constructor */
SparsePage() {
this->Clear();
}
/*! \return number of instance in the page */
inline size_t Size() const {
return offset.size() - 1;
}
/*!
* \brief load only the segments we are interested in
* \param fi the input stream of the file
* \param sorted_index_set sorted index of segments we are interested in
* \return true of the loading as successful, false if end of file was reached
*/
inline bool Load(utils::ISeekStream *fi,
const std::vector<bst_uint> &sorted_index_set) {
if (!fi->Read(&disk_offset_)) return false;
// setup the offset
offset.clear(); offset.push_back(0);
for (size_t i = 0; i < sorted_index_set.size(); ++i) {
bst_uint fid = sorted_index_set[i];
utils::Check(fid + 1 < disk_offset_.size(), "bad col.blob format");
size_t size = disk_offset_[fid + 1] - disk_offset_[fid];
offset.push_back(offset.back() + size);
}
data.resize(offset.back());
// read in the data
size_t begin = fi->Tell();
size_t curr_offset = 0;
for (size_t i = 0; i < sorted_index_set.size();) {
bst_uint fid = sorted_index_set[i];
if (disk_offset_[fid] != curr_offset) {
utils::Assert(disk_offset_[fid] > curr_offset, "fset index was not sorted");
fi->Seek(begin + disk_offset_[fid] * sizeof(SparseBatch::Entry));
curr_offset = disk_offset_[fid];
}
size_t j, size_to_read = 0;
for (j = i; j < sorted_index_set.size(); ++j) {
if (disk_offset_[sorted_index_set[j]] == disk_offset_[fid] + size_to_read) {
size_to_read += offset[j + 1] - offset[j];
} else {
break;
}
}
if (size_to_read != 0) {
utils::Check(fi->Read(BeginPtr(data) + offset[i],
size_to_read * sizeof(SparseBatch::Entry)) != 0,
"Invalid SparsePage file");
curr_offset += size_to_read;
}
i = j;
}
// seek to end of record
if (curr_offset != disk_offset_.back()) {
fi->Seek(begin + disk_offset_.back() * sizeof(SparseBatch::Entry));
}
return true;
}
/*!
* \brief load all the segments
* \param fi the input stream of the file
* \return true of the loading as successful, false if end of file was reached
*/
inline bool Load(utils::IStream *fi) {
if (!fi->Read(&offset)) return false;
utils::Check(offset.size() != 0, "Invalid SparsePage file");
data.resize(offset.back());
if (data.size() != 0) {
utils::Check(fi->Read(BeginPtr(data), data.size() * sizeof(SparseBatch::Entry)) != 0,
"Invalid SparsePage file");
}
return true;
}
/*!
* \brief save the data to fo, when a page was written
* to disk it must contain all the elements in the
* \param fo output stream
*/
inline void Save(utils::IStream *fo) const {
utils::Assert(offset.size() != 0 && offset[0] == 0, "bad offset");
utils::Assert(offset.back() == data.size(), "in consistent SparsePage");
fo->Write(offset);
if (data.size() != 0) {
fo->Write(BeginPtr(data), data.size() * sizeof(SparseBatch::Entry));
}
}
/*! \return estimation of memory cost of this page */
inline size_t MemCostBytes(void) const {
return offset.size() * sizeof(size_t) + data.size() * sizeof(SparseBatch::Entry);
}
/*! \brief clear the page */
inline void Clear(void) {
offset.clear();
offset.push_back(0);
data.clear();
}
/*!
* \brief load all the segments and add it to existing batch
* \param fi the input stream of the file
* \return true of the loading as successful, false if end of file was reached
*/
inline bool PushLoad(utils::IStream *fi) {
if (!fi->Read(&disk_offset_)) return false;
data.resize(offset.back() + disk_offset_.back());
if (disk_offset_.back() != 0) {
utils::Check(fi->Read(BeginPtr(data) + offset.back(),
disk_offset_.back() * sizeof(SparseBatch::Entry)) != 0,
"Invalid SparsePage file");
}
size_t top = offset.back();
size_t begin = offset.size();
offset.resize(offset.size() + disk_offset_.size());
for (size_t i = 0; i < disk_offset_.size(); ++i) {
offset[i + begin] = top + disk_offset_[i];
}
return true;
}
/*!
* \brief Push row batch into the page
* \param batch the row batch
*/
inline void Push(const RowBatch &batch) {
data.resize(offset.back() + batch.ind_ptr[batch.size]);
std::memcpy(BeginPtr(data) + offset.back(),
batch.data_ptr + batch.ind_ptr[0],
sizeof(SparseBatch::Entry) * batch.ind_ptr[batch.size]);
size_t top = offset.back();
size_t begin = offset.size();
offset.resize(offset.size() + batch.size);
for (size_t i = 0; i < batch.size; ++i) {
offset[i + begin] = top + batch.ind_ptr[i + 1] - batch.ind_ptr[0];
}
}
/*!
* \brief Push a sparse page
* \param batch the row page
*/
inline void Push(const SparsePage &batch) {
size_t top = offset.back();
data.resize(top + batch.data.size());
std::memcpy(BeginPtr(data) + top,
BeginPtr(batch.data),
sizeof(SparseBatch::Entry) * batch.data.size());
size_t begin = offset.size();
offset.resize(begin + batch.Size());
for (size_t i = 0; i < batch.Size(); ++i) {
offset[i + begin] = top + batch.offset[i + 1];
}
}
/*!
* \brief Push one instance into page
* \param row an instance row
*/
inline void Push(const SparseBatch::Inst &inst) {
offset.push_back(offset.back() + inst.length);
size_t begin = data.size();
data.resize(begin + inst.length);
if (inst.length != 0) {
std::memcpy(BeginPtr(data) + begin, inst.data,
sizeof(SparseBatch::Entry) * inst.length);
}
}
/*!
* \param base_rowid base_rowid of the data
* \return row batch representation of the page
*/
inline RowBatch GetRowBatch(size_t base_rowid) const {
RowBatch out;
out.base_rowid = base_rowid;
out.ind_ptr = BeginPtr(offset);
out.data_ptr = BeginPtr(data);
out.size = offset.size() - 1;
return out;
}
private:
/*! \brief external memory column offset */
std::vector<size_t> disk_offset_;
};
/*!
* \brief factory class for SparsePage,
* used in threadbuffer template
*/
class SparsePageFactory {
public:
SparsePageFactory(void)
: action_load_all_(true), set_load_all_(true) {}
inline void SetFile(const utils::FileStream &fi,
size_t file_begin = 0) {
fi_ = fi;
file_begin_ = file_begin;
}
inline const std::vector<bst_uint> &index_set(void) const {
return action_index_set_;
}
// set index set, will be used after next before first
inline void SetIndexSet(const std::vector<bst_uint> &index_set,
bool load_all) {
set_load_all_ = load_all;
if (!set_load_all_) {
set_index_set_ = index_set;
std::sort(set_index_set_.begin(), set_index_set_.end());
}
}
inline bool Init(void) {
return true;
}
inline void SetParam(const char *name, const char *val) {}
inline bool LoadNext(SparsePage *val) {
if (!action_load_all_) {
if (action_index_set_.size() == 0) {
return false;
} else {
return val->Load(&fi_, action_index_set_);
}
} else {
return val->Load(&fi_);
}
}
inline SparsePage *Create(void) {
return new SparsePage();
}
inline void FreeSpace(SparsePage *a) {
delete a;
}
inline void Destroy(void) {
fi_.Close();
}
inline void BeforeFirst(void) {
fi_.Seek(file_begin_);
action_load_all_ = set_load_all_;
if (!set_load_all_) {
action_index_set_ = set_index_set_;
}
}
private:
bool action_load_all_, set_load_all_;
size_t file_begin_;
utils::FileStream fi_;
std::vector<bst_uint> action_index_set_;
std::vector<bst_uint> set_index_set_;
};
} // namespace io
} // namespace xgboost
#endif // XGBOOST_IO_SPARSE_BATCH_PAGE_H_

176
src/learner/dmatrix.h
View File

@@ -1,176 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file dmatrix.h
* \brief meta data and template data structure
* used for regression/classification/ranking
* \author Tianqi Chen
*/
#ifndef XGBOOST_LEARNER_DMATRIX_H_
#define XGBOOST_LEARNER_DMATRIX_H_
#include <vector>
#include <cstring>
#include "../data.h"
#include "../utils/io.h"
namespace xgboost {
namespace learner {
/*!
* \brief meta information needed in training, including label, weight
*/
struct MetaInfo {
/*!
* \brief information needed by booster
* BoosterInfo does not implement save and load,
* all serialization is done in MetaInfo
*/
BoosterInfo info;
/*! \brief label of each instance */
std::vector<float> labels;
/*!
* \brief the index of begin and end of a group
* needed when the learning task is ranking
*/
std::vector<bst_uint> group_ptr;
/*! \brief weights of each instance, optional */
std::vector<float> weights;
/*!
* \brief initialized margins,
* if specified, xgboost will start from this initial margin
* can be used to specify initial prediction to boost from
*/
std::vector<float> base_margin;
/*! \brief version flag, used to check version of this info */
static const int kVersion = 0;
// constructor
MetaInfo(void) {}
/*! \return number of rows in dataset */
inline size_t num_row(void) const {
return info.num_row;
}
/*! \return number of columns in dataset */
inline size_t num_col(void) const {
return info.num_col;
}
/*! \brief clear all the information */
inline void Clear(void) {
labels.clear();
group_ptr.clear();
weights.clear();
info.root_index.clear();
base_margin.clear();
info.num_row = info.num_col = 0;
}
/*! \brief get weight of each instances */
inline float GetWeight(size_t i) const {
if (weights.size() != 0) {
return weights[i];
} else {
return 1.0f;
}
}
inline void SaveBinary(utils::IStream &fo) const { // NOLINT(*)
int version = kVersion;
fo.Write(&version, sizeof(version));
fo.Write(&info.num_row, sizeof(info.num_row));
fo.Write(&info.num_col, sizeof(info.num_col));
fo.Write(labels);
fo.Write(group_ptr);
fo.Write(weights);
fo.Write(info.root_index);
fo.Write(base_margin);
}
inline void LoadBinary(utils::IStream &fi) { // NOLINT(*)
int version;
utils::Check(fi.Read(&version, sizeof(version)) != 0, "MetaInfo: invalid format");
utils::Check(fi.Read(&info.num_row, sizeof(info.num_row)) != 0, "MetaInfo: invalid format");
utils::Check(fi.Read(&info.num_col, sizeof(info.num_col)) != 0, "MetaInfo: invalid format");
utils::Check(fi.Read(&labels), "MetaInfo: invalid format");
utils::Check(fi.Read(&group_ptr), "MetaInfo: invalid format");
utils::Check(fi.Read(&weights), "MetaInfo: invalid format");
utils::Check(fi.Read(&info.root_index), "MetaInfo: invalid format");
utils::Check(fi.Read(&base_margin), "MetaInfo: invalid format");
}
// try to load group information from file, if exists
inline bool TryLoadGroup(const char* fname, bool silent = false) {
using namespace std;
FILE *fi = fopen64(fname, "r");
if (fi == NULL) return false;
group_ptr.push_back(0);
unsigned nline;
while (fscanf(fi, "%u", &nline) == 1) {
group_ptr.push_back(group_ptr.back()+nline);
}
if (!silent) {
utils::Printf("%u groups are loaded from %s\n",
static_cast<unsigned>(group_ptr.size()-1), fname);
}
fclose(fi);
return true;
}
inline std::vector<float>& GetFloatInfo(const char *field) {
using namespace std;
if (!strcmp(field, "label")) return labels;
if (!strcmp(field, "weight")) return weights;
if (!strcmp(field, "base_margin")) return base_margin;
utils::Error("unknown field %s", field);
return labels;
}
inline const std::vector<float>& GetFloatInfo(const char *field) const {
return ((MetaInfo*)this)->GetFloatInfo(field); // NOLINT(*)
}
inline std::vector<unsigned> &GetUIntInfo(const char *field) {
using namespace std;
if (!strcmp(field, "root_index")) return info.root_index;
if (!strcmp(field, "fold_index")) return info.fold_index;
utils::Error("unknown field %s", field);
return info.root_index;
}
inline const std::vector<unsigned> &GetUIntInfo(const char *field) const {
return ((MetaInfo*)this)->GetUIntInfo(field); // NOLINT(*)
}
// try to load weight information from file, if exists
inline bool TryLoadFloatInfo(const char *field, const char* fname, bool silent = false) {
using namespace std;
std::vector<float> &data = this->GetFloatInfo(field);
FILE *fi = fopen64(fname, "r");
if (fi == NULL) return false;
float wt;
while (fscanf(fi, "%f", &wt) == 1) {
data.push_back(wt);
}
if (!silent) {
utils::Printf("loading %s from %s\n", field, fname);
}
fclose(fi);
return true;
}
};
/*!
* \brief data object used for learning,
* \tparam FMatrix type of feature data source
*/
struct DMatrix {
/*!
* \brief magic number associated with this object
* used to check if it is specific instance
*/
const int magic;
/*! \brief meta information about the dataset */
MetaInfo info;
/*!
* \brief cache pointer to verify if the data structure is cached in some learner
* used to verify if DMatrix is cached
*/
void *cache_learner_ptr_;
/*! \brief default constructor */
explicit DMatrix(int magic) : magic(magic), cache_learner_ptr_(NULL) {}
/*! \brief get feature matrix about data content */
virtual IFMatrix *fmat(void) const = 0;
// virtual destructor
virtual ~DMatrix(void){}
};
} // namespace learner
} // namespace xgboost
#endif // XGBOOST_LEARNER_DMATRIX_H_

589
src/learner/evaluation-inl.hpp
View File

@@ -1,589 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file xgboost_evaluation-inl.hpp
* \brief evaluation metrics for regression and classification and rank
* \author Kailong Chen, Tianqi Chen
*/
#ifndef XGBOOST_LEARNER_EVALUATION_INL_HPP_
#define XGBOOST_LEARNER_EVALUATION_INL_HPP_
#include <vector>
#include <utility>
#include <string>
#include <cmath>
#include <climits>
#include <algorithm>
#include "../sync/sync.h"
#include "../utils/math.h"
#include "./evaluation.h"
#include "./helper_utils.h"
namespace xgboost {
namespace learner {
/*!
* \brief base class of element-wise evaluation
* \tparam Derived the name of subclass
*/
template<typename Derived>
struct EvalEWiseBase : public IEvaluator {
virtual float Eval(const std::vector<float> &preds,
const MetaInfo &info,
bool distributed) const {
utils::Check(info.labels.size() != 0, "label set cannot be empty");
utils::Check(preds.size() == info.labels.size(),
"label and prediction size not match"\
"hint: use merror or mlogloss for multi-class classification");
const bst_omp_uint ndata = static_cast<bst_omp_uint>(info.labels.size());
float sum = 0.0, wsum = 0.0;
#pragma omp parallel for reduction(+: sum, wsum) schedule(static)
for (bst_omp_uint i = 0; i < ndata; ++i) {
const float wt = info.GetWeight(i);
sum += Derived::EvalRow(info.labels[i], preds[i]) * wt;
wsum += wt;
}
float dat[2]; dat[0] = sum, dat[1] = wsum;
if (distributed) {
rabit::Allreduce<rabit::op::Sum>(dat, 2);
}
return Derived::GetFinal(dat[0], dat[1]);
}
/*!
* \brief to be implemented by subclass,
* get evaluation result from one row
* \param label label of current instance
* \param pred prediction value of current instance
*/
inline static float EvalRow(float label, float pred);
/*!
* \brief to be overridden by subclass, final transformation
* \param esum the sum statistics returned by EvalRow
* \param wsum sum of weight
*/
inline static float GetFinal(float esum, float wsum) {
return esum / wsum;
}
};
/*! \brief RMSE */
struct EvalRMSE : public EvalEWiseBase<EvalRMSE> {
virtual const char *Name(void) const {
return "rmse";
}
inline static float EvalRow(float label, float pred) {
float diff = label - pred;
return diff * diff;
}
inline static float GetFinal(float esum, float wsum) {
return std::sqrt(esum / wsum);
}
};
/*! \brief logloss */
struct EvalLogLoss : public EvalEWiseBase<EvalLogLoss> {
virtual const char *Name(void) const {
return "logloss";
}
inline static float EvalRow(float y, float py) {
const float eps = 1e-16f;
const float pneg = 1.0f - py;
if (py < eps) {
return -y * std::log(eps) - (1.0f - y) * std::log(1.0f - eps);
} else if (pneg < eps) {
return -y * std::log(1.0f - eps) - (1.0f - y) * std::log(eps);
} else {
return -y * std::log(py) - (1.0f - y) * std::log(pneg);
}
}
};
/*! \brief error */
struct EvalError : public EvalEWiseBase<EvalError> {
virtual const char *Name(void) const {
return "error";
}
inline static float EvalRow(float label, float pred) {
// assume label is in [0,1]
return pred > 0.5f ? 1.0f - label : label;
}
};
/*! \brief log-likelihood of Poission distribution */
struct EvalPoissionNegLogLik : public EvalEWiseBase<EvalPoissionNegLogLik> {
virtual const char *Name(void) const {
return "poisson-nloglik";
}
inline static float EvalRow(float y, float py) {
const float eps = 1e-16f;
if (py < eps) py = eps;
return utils::LogGamma(y + 1.0f) + py - std::log(py) * y;
}
};
/*!
* \brief base class of multi-class evaluation
* \tparam Derived the name of subclass
*/
template<typename Derived>
struct EvalMClassBase : public IEvaluator {
virtual float Eval(const std::vector<float> &preds,
const MetaInfo &info,
bool distributed) const {
utils::Check(info.labels.size() != 0, "label set cannot be empty");
utils::Check(preds.size() % info.labels.size() == 0,
"label and prediction size not match");
const size_t nclass = preds.size() / info.labels.size();
utils::Check(nclass > 1,
"mlogloss and merror are only used for multi-class classification,"\
" use logloss for binary classification");
const bst_omp_uint ndata = static_cast<bst_omp_uint>(info.labels.size());
float sum = 0.0, wsum = 0.0;
int label_error = 0;
#pragma omp parallel for reduction(+: sum, wsum) schedule(static)
for (bst_omp_uint i = 0; i < ndata; ++i) {
const float wt = info.GetWeight(i);
int label = static_cast<int>(info.labels[i]);
if (label >= 0 && label < static_cast<int>(nclass)) {
sum += Derived::EvalRow(label,
BeginPtr(preds) + i * nclass,
nclass) * wt;
wsum += wt;
} else {
label_error = label;
}
}
utils::Check(label_error >= 0 && label_error < static_cast<int>(nclass),
"MultiClassEvaluation: label must be in [0, num_class)," \
" num_class=%d but found %d in label",
static_cast<int>(nclass), label_error);
float dat[2]; dat[0] = sum, dat[1] = wsum;
if (distributed) {
rabit::Allreduce<rabit::op::Sum>(dat, 2);
}
return Derived::GetFinal(dat[0], dat[1]);
}
/*!
* \brief to be implemented by subclass,
* get evaluation result from one row
* \param label label of current instance
* \param pred prediction value of current instance
* \param nclass number of class in the prediction
*/
inline static float EvalRow(int label,
const float *pred,
size_t nclass);
/*!
* \brief to be overridden by subclass, final transformation
* \param esum the sum statistics returned by EvalRow
* \param wsum sum of weight
*/
inline static float GetFinal(float esum, float wsum) {
return esum / wsum;
}
// used to store error message
const char *error_msg_;
};
/*! \brief match error */
struct EvalMatchError : public EvalMClassBase<EvalMatchError> {
virtual const char *Name(void) const {
return "merror";
}
inline static float EvalRow(int label,
const float *pred,
size_t nclass) {
return FindMaxIndex(pred, nclass) != static_cast<int>(label);
}
};
/*! \brief match error */
struct EvalMultiLogLoss : public EvalMClassBase<EvalMultiLogLoss> {
virtual const char *Name(void) const {
return "mlogloss";
}
inline static float EvalRow(int label,
const float *pred,
size_t nclass) {
const float eps = 1e-16f;
size_t k = static_cast<size_t>(label);
if (pred[k] > eps) {
return -std::log(pred[k]);
} else {
return -std::log(eps);
}
}
};
/*! \brief ctest */
struct EvalCTest: public IEvaluator {
EvalCTest(IEvaluator *base, const char *name)
: base_(base), name_(name) {}
virtual ~EvalCTest(void) {
delete base_;
}
virtual const char *Name(void) const {
return name_.c_str();
}
virtual float Eval(const std::vector<float> &preds,
const MetaInfo &info,
bool distributed) const {
utils::Check(!distributed, "metric %s do not support distributed evaluation", name_.c_str());
utils::Check(preds.size() % info.labels.size() == 0,
"label and prediction size not match");
size_t ngroup = preds.size() / info.labels.size() - 1;
const unsigned ndata = static_cast<unsigned>(info.labels.size());
utils::Check(ngroup > 1, "pred size does not meet requirement");
utils::Check(ndata == info.info.fold_index.size(), "need fold index");
double wsum = 0.0;
for (size_t k = 0; k < ngroup; ++k) {
std::vector<float> tpred;
MetaInfo tinfo;
for (unsigned i = 0; i < ndata; ++i) {
if (info.info.fold_index[i] == k) {
tpred.push_back(preds[i + (k + 1) * ndata]);
tinfo.labels.push_back(info.labels[i]);
tinfo.weights.push_back(info.GetWeight(i));
}
}
wsum += base_->Eval(tpred, tinfo);
}
return static_cast<float>(wsum / ngroup);
}
private:
IEvaluator *base_;
std::string name_;
};
/*! \brief AMS: also records best threshold */
struct EvalAMS : public IEvaluator {
public:
explicit EvalAMS(const char *name) {
name_ = name;
// note: ams@0 will automatically select which ratio to go
utils::Check(std::sscanf(name, "ams@%f", &ratio_) == 1, "invalid ams format");
}
virtual float Eval(const std::vector<float> &preds,
const MetaInfo &info,
bool distributed) const {
utils::Check(!distributed, "metric AMS do not support distributed evaluation");
using namespace std;
const bst_omp_uint ndata = static_cast<bst_omp_uint>(info.labels.size());
utils::Check(info.weights.size() == ndata, "we need weight to evaluate ams");
std::vector< std::pair<float, unsigned> > rec(ndata);
#pragma omp parallel for schedule(static)
for (bst_omp_uint i = 0; i < ndata; ++i) {
rec[i] = std::make_pair(preds[i], i);
}
std::sort(rec.begin(), rec.end(), CmpFirst);
unsigned ntop = static_cast<unsigned>(ratio_ * ndata);
if (ntop == 0) ntop = ndata;
const double br = 10.0;
unsigned thresindex = 0;
double s_tp = 0.0, b_fp = 0.0, tams = 0.0;
for (unsigned i = 0; i < static_cast<unsigned>(ndata-1) && i < ntop; ++i) {
const unsigned ridx = rec[i].second;
const float wt = info.weights[ridx];
if (info.labels[ridx] > 0.5f) {
s_tp += wt;
} else {
b_fp += wt;
}
if (rec[i].first != rec[i+1].first) {
double ams = sqrt(2*((s_tp+b_fp+br) * log(1.0 + s_tp/(b_fp+br)) - s_tp));
if (tams < ams) {
thresindex = i;
tams = ams;
}
}
}
if (ntop == ndata) {
utils::Printf("\tams-ratio=%g", static_cast<float>(thresindex) / ndata);
return static_cast<float>(tams);
} else {
return static_cast<float>(sqrt(2*((s_tp+b_fp+br) * log(1.0 + s_tp/(b_fp+br)) - s_tp)));
}
}
virtual const char *Name(void) const {
return name_.c_str();
}
private:
std::string name_;
float ratio_;
};
/*! \brief precision with cut off at top percentile */
struct EvalPrecisionRatio : public IEvaluator{
public:
explicit EvalPrecisionRatio(const char *name) : name_(name) {
using namespace std;
if (sscanf(name, "apratio@%f", &ratio_) == 1) {
use_ap = 1;
} else {
utils::Assert(sscanf(name, "pratio@%f", &ratio_) == 1, "BUG");
use_ap = 0;
}
}
virtual float Eval(const std::vector<float> &preds,
const MetaInfo &info,
bool distributed) const {
utils::Check(!distributed, "metric %s do not support distributed evaluation", Name());
utils::Check(info.labels.size() != 0, "label set cannot be empty");
utils::Assert(preds.size() % info.labels.size() == 0,
"label size predict size not match");
std::vector< std::pair<float, unsigned> > rec;
for (size_t j = 0; j < info.labels.size(); ++j) {
rec.push_back(std::make_pair(preds[j], static_cast<unsigned>(j)));
}
std::sort(rec.begin(), rec.end(), CmpFirst);
double pratio = CalcPRatio(rec, info);
return static_cast<float>(pratio);
}
virtual const char *Name(void) const {
return name_.c_str();
}
protected:
inline double CalcPRatio(const std::vector< std::pair<float, unsigned> >& rec,
const MetaInfo &info) const {
size_t cutoff = static_cast<size_t>(ratio_ * rec.size());
double wt_hit = 0.0, wsum = 0.0, wt_sum = 0.0;
for (size_t j = 0; j < cutoff; ++j) {
const float wt = info.GetWeight(j);
wt_hit += info.labels[rec[j].second] * wt;
wt_sum += wt;
wsum += wt_hit / wt_sum;
}
if (use_ap != 0) {
return wsum / cutoff;
} else {
return wt_hit / wt_sum;
}
}
int use_ap;
float ratio_;
std::string name_;
};
/*! \brief Area Under Curve, for both classification and rank */
struct EvalAuc : public IEvaluator {
virtual float Eval(const std::vector<float> &preds,
const MetaInfo &info,
bool distributed) const {
utils::Check(info.labels.size() != 0, "label set cannot be empty");
utils::Check(preds.size() % info.labels.size() == 0,
"label size predict size not match");
std::vector<unsigned> tgptr(2, 0);
tgptr[1] = static_cast<unsigned>(info.labels.size());
const std::vector<unsigned> &gptr = info.group_ptr.size() == 0 ? tgptr : info.group_ptr;
utils::Check(gptr.back() == info.labels.size(),
"EvalAuc: group structure must match number of prediction");
const bst_omp_uint ngroup = static_cast<bst_omp_uint>(gptr.size() - 1);
// sum statistics
double sum_auc = 0.0f;
#pragma omp parallel reduction(+:sum_auc)
{
// each thread takes a local rec
std::vector< std::pair<float, unsigned> > rec;
#pragma omp for schedule(static)
for (bst_omp_uint k = 0; k < ngroup; ++k) {
rec.clear();
for (unsigned j = gptr[k]; j < gptr[k + 1]; ++j) {
rec.push_back(std::make_pair(preds[j], j));
}
std::sort(rec.begin(), rec.end(), CmpFirst);
// calculate AUC
double sum_pospair = 0.0;
double sum_npos = 0.0, sum_nneg = 0.0, buf_pos = 0.0, buf_neg = 0.0;
for (size_t j = 0; j < rec.size(); ++j) {
const float wt = info.GetWeight(rec[j].second);
const float ctr = info.labels[rec[j].second];
// keep bucketing predictions in same bucket
if (j != 0 && rec[j].first != rec[j - 1].first) {
sum_pospair += buf_neg * (sum_npos + buf_pos *0.5);
sum_npos += buf_pos;
sum_nneg += buf_neg;
buf_neg = buf_pos = 0.0f;
}
buf_pos += ctr * wt;
buf_neg += (1.0f - ctr) * wt;
}
sum_pospair += buf_neg * (sum_npos + buf_pos *0.5);
sum_npos += buf_pos;
sum_nneg += buf_neg;
// check weird conditions
utils::Check(sum_npos > 0.0 && sum_nneg > 0.0,
"AUC: the dataset only contains pos or neg samples");
// this is the AUC
sum_auc += sum_pospair / (sum_npos*sum_nneg);
}
}
if (distributed) {
float dat[2];
dat[0] = static_cast<float>(sum_auc);
dat[1] = static_cast<float>(ngroup);
// approximately estimate auc using mean
rabit::Allreduce<rabit::op::Sum>(dat, 2);
return dat[0] / dat[1];
} else {
return static_cast<float>(sum_auc) / ngroup;
}
}
virtual const char *Name(void) const {
return "auc";
}
};
/*! \brief Evaluate rank list */
struct EvalRankList : public IEvaluator {
public:
virtual float Eval(const std::vector<float> &preds,
const MetaInfo &info,
bool distributed) const {
utils::Check(preds.size() == info.labels.size(),
"label size predict size not match");
// quick consistency when group is not available
std::vector<unsigned> tgptr(2, 0);
tgptr[1] = static_cast<unsigned>(preds.size());
const std::vector<unsigned> &gptr = info.group_ptr.size() == 0 ? tgptr : info.group_ptr;
utils::Assert(gptr.size() != 0, "must specify group when constructing rank file");
utils::Assert(gptr.back() == preds.size(),
"EvalRanklist: group structure must match number of prediction");
const bst_omp_uint ngroup = static_cast<bst_omp_uint>(gptr.size() - 1);
// sum statistics
double sum_metric = 0.0f;
#pragma omp parallel reduction(+:sum_metric)
{
// each thread takes a local rec
std::vector< std::pair<float, unsigned> > rec;
#pragma omp for schedule(static)
for (bst_omp_uint k = 0; k < ngroup; ++k) {
rec.clear();
for (unsigned j = gptr[k]; j < gptr[k + 1]; ++j) {
rec.push_back(std::make_pair(preds[j], static_cast<int>(info.labels[j])));
}
sum_metric += this->EvalMetric(rec);
}
}
if (distributed) {
float dat[2];
dat[0] = static_cast<float>(sum_metric);
dat[1] = static_cast<float>(ngroup);
// approximately estimate the metric using mean
rabit::Allreduce<rabit::op::Sum>(dat, 2);
return dat[0] / dat[1];
} else {
return static_cast<float>(sum_metric) / ngroup;
}
}
virtual const char *Name(void) const {
return name_.c_str();
}
protected:
explicit EvalRankList(const char *name) {
using namespace std;
name_ = name;
minus_ = false;
if (sscanf(name, "%*[^@]@%u[-]?", &topn_) != 1) {
topn_ = UINT_MAX;
}
if (name[strlen(name) - 1] == '-') {
minus_ = true;
}
}
/*! \return evaluation metric, given the pair_sort record, (pred,label) */
virtual float EvalMetric(std::vector< std::pair<float, unsigned> > &pair_sort) const = 0; // NOLINT(*)
protected:
unsigned topn_;
std::string name_;
bool minus_;
};
/*! \brief Precision at N, for both classification and rank */
struct EvalPrecision : public EvalRankList{
public:
explicit EvalPrecision(const char *name) : EvalRankList(name) {}
protected:
virtual float EvalMetric(std::vector< std::pair<float, unsigned> > &rec) const {
// calculate Precision
std::sort(rec.begin(), rec.end(), CmpFirst);
unsigned nhit = 0;
for (size_t j = 0; j < rec.size() && j < this->topn_; ++j) {
nhit += (rec[j].second != 0);
}
return static_cast<float>(nhit) / topn_;
}
};
/*! \brief NDCG: Normalized Discounted Cumulative Gain at N */
struct EvalNDCG : public EvalRankList{
public:
explicit EvalNDCG(const char *name) : EvalRankList(name) {}
protected:
inline float CalcDCG(const std::vector< std::pair<float, unsigned> > &rec) const {
double sumdcg = 0.0;
for (size_t i = 0; i < rec.size() && i < this->topn_; ++i) {
const unsigned rel = rec[i].second;
if (rel != 0) {
sumdcg += ((1 << rel) - 1) / std::log(i + 2.0);
}
}
return static_cast<float>(sumdcg);
}
virtual float EvalMetric(std::vector< std::pair<float, unsigned> > &rec) const { // NOLINT(*)
std::stable_sort(rec.begin(), rec.end(), CmpFirst);
float dcg = this->CalcDCG(rec);
std::stable_sort(rec.begin(), rec.end(), CmpSecond);
float idcg = this->CalcDCG(rec);
if (idcg == 0.0f) {
if (minus_) {
return 0.0f;
} else {
return 1.0f;
}
}
return dcg/idcg;
}
};
/*! \brief Mean Average Precision at N, for both classification and rank */
struct EvalMAP : public EvalRankList {
public:
explicit EvalMAP(const char *name) : EvalRankList(name) {}
protected:
virtual float EvalMetric(std::vector< std::pair<float, unsigned> > &rec) const {
std::sort(rec.begin(), rec.end(), CmpFirst);
unsigned nhits = 0;
double sumap = 0.0;
for (size_t i = 0; i < rec.size(); ++i) {
if (rec[i].second != 0) {
nhits += 1;
if (i < this->topn_) {
sumap += static_cast<float>(nhits) / (i+1);
}
}
}
if (nhits != 0) {
sumap /= nhits;
return static_cast<float>(sumap);
} else {
if (minus_) {
return 0.0f;
} else {
return 1.0f;
}
}
}
};
} // namespace learner
} // namespace xgboost
#endif // XGBOOST_LEARNER_EVALUATION_INL_HPP_

101
src/learner/evaluation.h
View File

@@ -1,101 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file evaluation.h
* \brief interface of evaluation function supported in xgboost
* \author Tianqi Chen, Kailong Chen
*/
#ifndef XGBOOST_LEARNER_EVALUATION_H_
#define XGBOOST_LEARNER_EVALUATION_H_
#include <string>
#include <vector>
#include <cstdio>
#include "../utils/utils.h"
#include "./dmatrix.h"
namespace xgboost {
namespace learner {
/*! \brief evaluator that evaluates the loss metrics */
struct IEvaluator{
/*!
* \brief evaluate a specific metric
* \param preds prediction
* \param info information, including label etc.
* \param distributed whether a call to Allreduce is needed to gather
* the average statistics across all the node,
* this is only supported by some metrics
*/
virtual float Eval(const std::vector<float> &preds,
const MetaInfo &info,
bool distributed = false) const = 0;
/*! \return name of metric */
virtual const char *Name(void) const = 0;
/*! \brief virtual destructor */
virtual ~IEvaluator(void) {}
};
} // namespace learner
} // namespace xgboost
// include implementations of evaluation functions
#include "evaluation-inl.hpp"
// factory function
namespace xgboost {
namespace learner {
inline IEvaluator* CreateEvaluator(const char *name) {
using namespace std;
if (!strcmp(name, "rmse")) return new EvalRMSE();
if (!strcmp(name, "error")) return new EvalError();
if (!strcmp(name, "merror")) return new EvalMatchError();
if (!strcmp(name, "logloss")) return new EvalLogLoss();
if (!strcmp(name, "mlogloss")) return new EvalMultiLogLoss();
if (!strcmp(name, "poisson-nloglik")) return new EvalPoissionNegLogLik();
if (!strcmp(name, "auc")) return new EvalAuc();
if (!strncmp(name, "ams@", 4)) return new EvalAMS(name);
if (!strncmp(name, "pre@", 4)) return new EvalPrecision(name);
if (!strncmp(name, "pratio@", 7)) return new EvalPrecisionRatio(name);
if (!strncmp(name, "map", 3)) return new EvalMAP(name);
if (!strncmp(name, "ndcg", 4)) return new EvalNDCG(name);
if (!strncmp(name, "ct-", 3)) return new EvalCTest(CreateEvaluator(name+3), name);
utils::Error("unknown evaluation metric type: %s", name);
return NULL;
}
/*! \brief a set of evaluators */
class EvalSet{
public:
inline void AddEval(const char *name) {
using namespace std;
for (size_t i = 0; i < evals_.size(); ++i) {
if (!strcmp(name, evals_[i]->Name())) return;
}
evals_.push_back(CreateEvaluator(name));
}
~EvalSet(void) {
for (size_t i = 0; i < evals_.size(); ++i) {
delete evals_[i];
}
}
inline std::string Eval(const char *evname,
const std::vector<float> &preds,
const MetaInfo &info,
bool distributed = false) {
std::string result = "";
for (size_t i = 0; i < evals_.size(); ++i) {
float res = evals_[i]->Eval(preds, info, distributed);
char tmp[1024];
utils::SPrintf(tmp, sizeof(tmp), "\t%s-%s:%f", evname, evals_[i]->Name(), res);
result += tmp;
}
return result;
}
inline size_t Size(void) const {
return evals_.size();
}
private:
std::vector<const IEvaluator*> evals_;
};
} // namespace learner
} // namespace xgboost
#endif // XGBOOST_LEARNER_EVALUATION_H_

80
src/learner/helper_utils.h
View File

@@ -1,80 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file helper_utils.h
* \brief useful helper functions
* \author Tianqi Chen, Kailong Chen
*/
#ifndef XGBOOST_LEARNER_HELPER_UTILS_H_
#define XGBOOST_LEARNER_HELPER_UTILS_H_
#include <utility>
#include <vector>
#include <cmath>
#include <algorithm>
namespace xgboost {
namespace learner {
// simple helper function to do softmax
inline static void Softmax(std::vector<float>* p_rec) {
std::vector<float> &rec = *p_rec;
float wmax = rec[0];
for (size_t i = 1; i < rec.size(); ++i) {
wmax = std::max(rec[i], wmax);
}
double wsum = 0.0f;
for (size_t i = 0; i < rec.size(); ++i) {
rec[i] = std::exp(rec[i]-wmax);
wsum += rec[i];
}
for (size_t i = 0; i < rec.size(); ++i) {
rec[i] /= static_cast<float>(wsum);
}
}
inline static int FindMaxIndex(const float *rec, size_t size) {
size_t mxid = 0;
for (size_t i = 1; i < size; ++i) {
if (rec[i] > rec[mxid]) {
mxid = i;
}
}
return static_cast<int>(mxid);
}
// simple helper function to do softmax
inline static int FindMaxIndex(const std::vector<float>& rec) {
return FindMaxIndex(BeginPtr(rec), rec.size());
}
// perform numerically safe logsum
inline float LogSum(float x, float y) {
if (x < y) {
return y + std::log(std::exp(x - y) + 1.0f);
} else {
return x + std::log(std::exp(y - x) + 1.0f);
}
}
// numerically safe logsum
inline float LogSum(const float *rec, size_t size) {
float mx = rec[0];
for (size_t i = 1; i < size; ++i) {
mx = std::max(mx, rec[i]);
}
float sum = 0.0f;
for (size_t i = 0; i < size; ++i) {
sum += std::exp(rec[i] - mx);
}
return mx + std::log(sum);
}
// comparator functions for sorting pairs in descending order
inline static bool CmpFirst(const std::pair<float, unsigned> &a,
const std::pair<float, unsigned> &b) {
return a.first > b.first;
}
inline static bool CmpSecond(const std::pair<float, unsigned> &a,
const std::pair<float, unsigned> &b) {
return a.second > b.second;
}
} // namespace learner
} // namespace xgboost
#endif // XGBOOST_LEARNER_HELPER_UTILS_H_

547
src/learner/learner-inl.hpp
View File

@@ -1,547 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file learner-inl.hpp
* \brief learning algorithm
* \author Tianqi Chen
*/
#ifndef XGBOOST_LEARNER_LEARNER_INL_HPP_
#define XGBOOST_LEARNER_LEARNER_INL_HPP_
#include <algorithm>
#include <vector>
#include <utility>
#include <string>
#include <limits>
#include "../sync/sync.h"
#include "../utils/io.h"
#include "./objective.h"
#include "./evaluation.h"
#include "../gbm/gbm.h"
namespace xgboost {
/*! \brief namespace for learning algorithm */
namespace learner {
/*!
* \brief learner that performs gradient boosting for a specific objective function.
* It does training and prediction.
*/
class BoostLearner : public rabit::Serializable {
public:
BoostLearner(void) {
obj_ = NULL;
gbm_ = NULL;
name_obj_ = "reg:linear";
name_gbm_ = "gbtree";
silent = 0;
prob_buffer_row = 1.0f;
distributed_mode = 0;
updater_mode = 0;
pred_buffer_size = 0;
seed_per_iteration = 0;
seed = 0;
save_base64 = 0;
}
virtual ~BoostLearner(void) {
if (obj_ != NULL) delete obj_;
if (gbm_ != NULL) delete gbm_;
}
/*!
* \brief add internal cache space for mat, this can speedup prediction for matrix,
* please cache prediction for training and eval data
* warning: if the model is loaded from file from some previous training history
* set cache data must be called with exactly SAME
* data matrices to continue training otherwise it will cause error
* \param mats array of pointers to matrix whose prediction result need to be cached
*/
inline void SetCacheData(const std::vector<DMatrix*>& mats) {
utils::Assert(cache_.size() == 0, "can only call cache data once");
// assign buffer index
size_t buffer_size = 0;
for (size_t i = 0; i < mats.size(); ++i) {
bool dupilicate = false;
for (size_t j = 0; j < i; ++j) {
if (mats[i] == mats[j]) dupilicate = true;
}
if (dupilicate) continue;
// set mats[i]'s cache learner pointer to this
mats[i]->cache_learner_ptr_ = this;
cache_.push_back(CacheEntry(mats[i], buffer_size, mats[i]->info.num_row()));
buffer_size += mats[i]->info.num_row();
}
char str_temp[25];
utils::SPrintf(str_temp, sizeof(str_temp), "%lu",
static_cast<unsigned long>(buffer_size)); // NOLINT(*)
this->SetParam("num_pbuffer", str_temp);
this->pred_buffer_size = buffer_size;
}
/*!
* \brief set parameters from outside
* \param name name of the parameter
* \param val value of the parameter
*/
inline void SetParam(const char *name, const char *val) {
using namespace std;
// in this version, bst: prefix is no longer required
if (strncmp(name, "bst:", 4) != 0) {
std::string n = "bst:"; n += name;
this->SetParam(n.c_str(), val);
}
if (!strcmp(name, "silent")) silent = atoi(val);
if (!strcmp(name, "dsplit")) {
if (!strcmp(val, "col")) {
this->SetParam("updater", "distcol");
distributed_mode = 1;
} else if (!strcmp(val, "row")) {
this->SetParam("updater", "grow_histmaker,prune");
distributed_mode = 2;
} else {
utils::Error("%s is invalid value for dsplit, should be row or col", val);
}
}
if (!strcmp(name, "updater_mode")) updater_mode = atoi(val);
if (!strcmp(name, "prob_buffer_row")) {
prob_buffer_row = static_cast<float>(atof(val));
utils::Check(distributed_mode == 0,
"prob_buffer_row can only be used in single node mode so far");
this->SetParam("updater", "grow_colmaker,refresh,prune");
}
if (!strcmp(name, "eval_metric")) evaluator_.AddEval(val);
if (!strcmp("seed", name)) {
seed = atoi(val); random::Seed(seed);
}
if (!strcmp("seed_per_iter", name)) seed_per_iteration = atoi(val);
if (!strcmp("save_base64", name)) save_base64 = atoi(val);
if (!strcmp(name, "num_class")) {
this->SetParam("num_output_group", val);
}
if (!strcmp(name, "nthread")) {
omp_set_num_threads(atoi(val));
}
if (gbm_ == NULL) {
if (!strcmp(name, "objective")) name_obj_ = val;
if (!strcmp(name, "booster")) name_gbm_ = val;
mparam.SetParam(name, val);
}
if (gbm_ != NULL) gbm_->SetParam(name, val);
if (obj_ != NULL) obj_->SetParam(name, val);
if (gbm_ == NULL || obj_ == NULL) {
cfg_.push_back(std::make_pair(std::string(name), std::string(val)));
}
}
// this is an internal function
// initialize the trainer, called at InitModel and LoadModel
inline void InitTrainer(bool calc_num_feature = true) {
if (calc_num_feature) {
// estimate feature bound
unsigned num_feature = 0;
for (size_t i = 0; i < cache_.size(); ++i) {
num_feature = std::max(num_feature,
static_cast<unsigned>(cache_[i].mat_->info.num_col()));
}
// run allreduce on num_feature to find the maximum value
rabit::Allreduce<rabit::op::Max>(&num_feature, 1);
if (num_feature > mparam.num_feature) mparam.num_feature = num_feature;
}
char str_temp[25];
utils::SPrintf(str_temp, sizeof(str_temp), "%d", mparam.num_feature);
this->SetParam("bst:num_feature", str_temp);
}
/*!
* \brief initialize the model
*/
inline void InitModel(void) {
this->InitTrainer();
// initialize model
this->InitObjGBM();
// reset the base score
mparam.base_score = obj_->ProbToMargin(mparam.base_score);
// initialize GBM model
gbm_->InitModel();
}
/*!
* \brief load model from stream
* \param fi input stream
* \param calc_num_feature whether call InitTrainer with calc_num_feature
*/
inline void LoadModel(utils::IStream &fi, // NOLINT(*)
bool calc_num_feature = true) {
utils::Check(fi.Read(&mparam, sizeof(ModelParam)) != 0,
"BoostLearner: wrong model format");
{
// backward compatibility code for compatible with old model type
// for new model, Read(&name_obj_) is suffice
uint64_t len;
utils::Check(fi.Read(&len, sizeof(len)) != 0, "BoostLearner: wrong model format");
if (len >= std::numeric_limits<unsigned>::max()) {
int gap;
utils::Check(fi.Read(&gap, sizeof(gap)) != 0, "BoostLearner: wrong model format");
len = len >> static_cast<uint64_t>(32UL);
}
if (len != 0) {
name_obj_.resize(len);
utils::Check(fi.Read(&name_obj_[0], len) != 0, "BoostLearner: wrong model format");
}
}
utils::Check(fi.Read(&name_gbm_), "BoostLearner: wrong model format");
// delete existing gbm if any
if (obj_ != NULL) delete obj_;
if (gbm_ != NULL) delete gbm_;
this->InitTrainer(calc_num_feature);
this->InitObjGBM();
char tmp[32];
utils::SPrintf(tmp, sizeof(tmp), "%u", mparam.num_class);
obj_->SetParam("num_class", tmp);
gbm_->LoadModel(fi, mparam.saved_with_pbuffer != 0);
if (mparam.saved_with_pbuffer == 0) {
gbm_->ResetPredBuffer(pred_buffer_size);
}
}
// rabit load model from rabit checkpoint
virtual void Load(rabit::Stream *fi) {
// for row split, we should not keep pbuffer
this->LoadModel(*fi, false);
}
// rabit save model to rabit checkpoint
virtual void Save(rabit::Stream *fo) const {
// for row split, we should not keep pbuffer
this->SaveModel(*fo, distributed_mode != 2);
}
/*!
* \brief load model from file
* \param fname file name
*/
inline void LoadModel(const char *fname) {
utils::IStream *fi = utils::IStream::Create(fname, "r");
std::string header; header.resize(4);
// check header for different binary encode
// can be base64 or binary
utils::Check(fi->Read(&header[0], 4) != 0, "invalid model");
// base64 format
if (header == "bs64") {
utils::Base64InStream bsin(fi);
bsin.InitPosition();
this->LoadModel(bsin, true);
} else if (header == "binf") {
this->LoadModel(*fi, true);
} else {
delete fi;
fi = utils::IStream::Create(fname, "r");
this->LoadModel(*fi, true);
}
delete fi;
}
inline void SaveModel(utils::IStream &fo, bool with_pbuffer) const { // NOLINT(*)
ModelParam p = mparam;
p.saved_with_pbuffer = static_cast<int>(with_pbuffer);
fo.Write(&p, sizeof(ModelParam));
fo.Write(name_obj_);
fo.Write(name_gbm_);
gbm_->SaveModel(fo, with_pbuffer);
}
/*!
* \brief save model into file
* \param fname file name
* \param with_pbuffer whether save pbuffer together
*/
inline void SaveModel(const char *fname, bool with_pbuffer) const {
utils::IStream *fo = utils::IStream::Create(fname, "w");
if (save_base64 != 0 || !strcmp(fname, "stdout")) {
fo->Write("bs64\t", 5);
utils::Base64OutStream bout(fo);
this->SaveModel(bout, with_pbuffer);
bout.Finish('\n');
} else {
fo->Write("binf", 4);
this->SaveModel(*fo, with_pbuffer);
}
delete fo;
}
/*!
* \brief check if data matrix is ready to be used by training,
* if not initialize it
* \param p_train pointer to the matrix used by training
*/
inline void CheckInit(DMatrix *p_train) {
int ncol = static_cast<int>(p_train->info.info.num_col);
std::vector<bool> enabled(ncol, true);
// set max row per batch to limited value
// in distributed mode, use safe choice otherwise
size_t max_row_perbatch = std::numeric_limits<size_t>::max();
if (updater_mode != 0 || distributed_mode == 2) {
max_row_perbatch = 32UL << 10UL;
}
// initialize column access
p_train->fmat()->InitColAccess(enabled,
prob_buffer_row,
max_row_perbatch);
const int kMagicPage = 0xffffab02;
// check, if it is DMatrixPage, then use hist maker
if (p_train->magic == kMagicPage) {
this->SetParam("updater", "grow_histmaker,prune");
}
}
/*!
* \brief update the model for one iteration
* \param iter current iteration number
* \param train reference to the data matrix
*/
inline void UpdateOneIter(int iter, const DMatrix &train) {
if (seed_per_iteration != 0 || rabit::IsDistributed()) {
random::Seed(this->seed * kRandSeedMagic + iter);
}
this->PredictRaw(train, &preds_);
obj_->GetGradient(preds_, train.info, iter, &gpair_);
gbm_->DoBoost(train.fmat(), this->FindBufferOffset(train), train.info.info, &gpair_);
}
/*!
* \brief whether model allow lazy checkpoint
*/
inline bool AllowLazyCheckPoint(void) const {
return gbm_->AllowLazyCheckPoint();
}
/*!
* \brief evaluate the model for specific iteration
* \param iter iteration number
* \param evals datas i want to evaluate
* \param evname name of each dataset
* \return a string corresponding to the evaluation result
*/
inline std::string EvalOneIter(int iter,
const std::vector<const DMatrix*> &evals,
const std::vector<std::string> &evname) {
std::string res;
char tmp[256];
utils::SPrintf(tmp, sizeof(tmp), "[%d]", iter);
res = tmp;
for (size_t i = 0; i < evals.size(); ++i) {
this->PredictRaw(*evals[i], &preds_);
obj_->EvalTransform(&preds_);
res += evaluator_.Eval(evname[i].c_str(), preds_, evals[i]->info, distributed_mode == 2);
}
return res;
}
/*!
* \brief simple evaluation function, with a specified metric
* \param data input data
* \param metric name of metric
* \return a pair of <evaluation name, result>
*/
std::pair<std::string, float> Evaluate(const DMatrix &data, std::string metric) {
if (metric == "auto") metric = obj_->DefaultEvalMetric();
IEvaluator *ev = CreateEvaluator(metric.c_str());
this->PredictRaw(data, &preds_);
obj_->EvalTransform(&preds_);
float res = ev->Eval(preds_, data.info);
delete ev;
return std::make_pair(metric, res);
}
/*!
* \brief get prediction
* \param data input data
* \param output_margin whether to only predict margin value instead of transformed prediction
* \param out_preds output vector that stores the prediction
* \param ntree_limit limit number of trees used for boosted tree
* predictor, when it equals 0, this means we are using all the trees
* \param pred_leaf whether to only predict the leaf index of each tree in a boosted tree predictor
*/
inline void Predict(const DMatrix &data,
bool output_margin,
std::vector<float> *out_preds,
unsigned ntree_limit = 0,
bool pred_leaf = false) const {
if (pred_leaf) {
gbm_->PredictLeaf(data.fmat(), data.info.info, out_preds, ntree_limit);
} else {
this->PredictRaw(data, out_preds, ntree_limit);
if (!output_margin) {
obj_->PredTransform(out_preds);
}
}
}
/*!
* \brief online prediction function, predict score for one instance at a time
* NOTE: use the batch prediction interface if possible, batch prediction is usually
* more efficient than online prediction
* This function is NOT threadsafe, make sure you only call from one thread
*
* \param inst the instance you want to predict
* \param output_margin whether to only predict margin value instead of transformed prediction
* \param out_preds output vector to hold the predictions
* \param ntree_limit limit the number of trees used in prediction
* \sa Predict
*/
inline void Predict(const SparseBatch::Inst &inst,
bool output_margin,
std::vector<float> *out_preds,
unsigned ntree_limit = 0) const {
gbm_->Predict(inst, out_preds, ntree_limit);
if (out_preds->size() == 1) {
(*out_preds)[0] += mparam.base_score;
}
if (!output_margin) {
obj_->PredTransform(out_preds);
}
}
/*! \brief dump model out */
inline std::vector<std::string> DumpModel(const utils::FeatMap& fmap, int option) {
return gbm_->DumpModel(fmap, option);
}
protected:
/*!
* \brief initialize the objective function and GBM,
* if not yet done
*/
inline void InitObjGBM(void) {
if (obj_ != NULL) return;
utils::Assert(gbm_ == NULL, "GBM and obj should be NULL");
obj_ = CreateObjFunction(name_obj_.c_str());
gbm_ = gbm::CreateGradBooster(name_gbm_.c_str());
this->InitAdditionDefaultParam();
// set parameters
for (size_t i = 0; i < cfg_.size(); ++i) {
obj_->SetParam(cfg_[i].first.c_str(), cfg_[i].second.c_str());
gbm_->SetParam(cfg_[i].first.c_str(), cfg_[i].second.c_str());
}
if (evaluator_.Size() == 0) {
evaluator_.AddEval(obj_->DefaultEvalMetric());
}
}
/*!
* \brief additional default value for specific objs
*/
inline void InitAdditionDefaultParam(void) {
if (name_obj_ == "count:poisson") {
obj_->SetParam("max_delta_step", "0.7");
gbm_->SetParam("max_delta_step", "0.7");
}
}
/*!
* \brief get un-transformed prediction
* \param data training data matrix
* \param out_preds output vector that stores the prediction
* \param ntree_limit limit number of trees used for boosted tree
* predictor, when it equals 0, this means we are using all the trees
*/
inline void PredictRaw(const DMatrix &data,
std::vector<float> *out_preds,
unsigned ntree_limit = 0) const {
gbm_->Predict(data.fmat(), this->FindBufferOffset(data),
data.info.info, out_preds, ntree_limit);
// add base margin
std::vector<float> &preds = *out_preds;
const bst_omp_uint ndata = static_cast<bst_omp_uint>(preds.size());
if (data.info.base_margin.size() != 0) {
utils::Check(preds.size() == data.info.base_margin.size(),
"base_margin.size does not match with prediction size");
#pragma omp parallel for schedule(static)
for (bst_omp_uint j = 0; j < ndata; ++j) {
preds[j] += data.info.base_margin[j];
}
} else {
#pragma omp parallel for schedule(static)
for (bst_omp_uint j = 0; j < ndata; ++j) {
preds[j] += mparam.base_score;
}
}
}
/*! \brief training parameter for regression */
struct ModelParam{
/* \brief global bias */
float base_score;
/* \brief number of features */
unsigned num_feature;
/* \brief number of classes, if it is multi-class classification */
int num_class;
/*! \brief whether the model itself is saved with pbuffer */
int saved_with_pbuffer;
/*! \brief reserved field */
int reserved[30];
/*! \brief constructor */
ModelParam(void) {
std::memset(this, 0, sizeof(ModelParam));
base_score = 0.5f;
num_feature = 0;
num_class = 0;
saved_with_pbuffer = 0;
}
/*!
* \brief set parameters from outside
* \param name name of the parameter
* \param val value of the parameter
*/
inline void SetParam(const char *name, const char *val) {
using namespace std;
if (!strcmp("base_score", name)) base_score = static_cast<float>(atof(val));
if (!strcmp("num_class", name)) num_class = atoi(val);
if (!strcmp("bst:num_feature", name)) num_feature = atoi(val);
}
};
// data fields
// stored random seed
int seed;
// whether seed the PRNG each iteration
// this is important for restart from existing iterations
// default set to no, but will auto switch on in distributed mode
int seed_per_iteration;
// save model in base64 encoding
int save_base64;
// silent during training
int silent;
// distributed learning mode, if any, 0:none, 1:col, 2:row
int distributed_mode;
// updater mode, 0:normal, reserved for internal test
int updater_mode;
// cached size of predict buffer
size_t pred_buffer_size;
// maximum buffered row value
float prob_buffer_row;
// evaluation set
EvalSet evaluator_;
// model parameter
ModelParam mparam;
// gbm model that back everything
gbm::IGradBooster *gbm_;
// name of gbm model used for training
std::string name_gbm_;
// objective function
IObjFunction *obj_;
// name of objective function
std::string name_obj_;
// configurations
std::vector< std::pair<std::string, std::string> > cfg_;
// temporal storages for prediction
std::vector<float> preds_;
// gradient pairs
std::vector<bst_gpair> gpair_;
protected:
// magic number to transform random seed
static const int kRandSeedMagic = 127;
// cache entry object that helps handle feature caching
struct CacheEntry {
const DMatrix *mat_;
size_t buffer_offset_;
size_t num_row_;
CacheEntry(const DMatrix *mat, size_t buffer_offset, size_t num_row)
:mat_(mat), buffer_offset_(buffer_offset), num_row_(num_row) {}
};
// find internal buffer offset for certain matrix, if not exist, return -1
inline int64_t FindBufferOffset(const DMatrix &mat) const {
for (size_t i = 0; i < cache_.size(); ++i) {
if (cache_[i].mat_ == &mat && mat.cache_learner_ptr_ == this) {
if (cache_[i].num_row_ == mat.info.num_row()) {
return static_cast<int64_t>(cache_[i].buffer_offset_);
}
}
}
return -1;
}
// data structure field
/*! \brief the entries indicates that we have internal prediction cache */
std::vector<CacheEntry> cache_;
};
} // namespace learner
} // namespace xgboost
#endif // XGBOOST_LEARNER_LEARNER_INL_HPP_

642
src/learner/objective-inl.hpp
View File

@@ -1,642 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file objective-inl.hpp
* \brief objective function implementations
* \author Tianqi Chen, Kailong Chen
*/
#ifndef XGBOOST_LEARNER_OBJECTIVE_INL_HPP_
#define XGBOOST_LEARNER_OBJECTIVE_INL_HPP_
#include <vector>
#include <algorithm>
#include <utility>
#include <cmath>
#include <functional>
#include "../data.h"
#include "./objective.h"
#include "./helper_utils.h"
#include "../utils/random.h"
#include "../utils/omp.h"
namespace xgboost {
namespace learner {
/*! \brief defines functions to calculate some commonly used functions */
struct LossType {
/*! \brief indicate which type we are using */
int loss_type;
// list of constants
static const int kLinearSquare = 0;
static const int kLogisticNeglik = 1;
static const int kLogisticClassify = 2;
static const int kLogisticRaw = 3;
/*!
* \brief transform the linear sum to prediction
* \param x linear sum of boosting ensemble
* \return transformed prediction
*/
inline float PredTransform(float x) const {
switch (loss_type) {
case kLogisticRaw:
case kLinearSquare: return x;
case kLogisticClassify:
case kLogisticNeglik: return 1.0f / (1.0f + std::exp(-x));
default: utils::Error("unknown loss_type"); return 0.0f;
}
}
/*!
* \brief check if label range is valid
*/
inline bool CheckLabel(float x) const {
if (loss_type != kLinearSquare) {
return x >= 0.0f && x <= 1.0f;
}
return true;
}
/*!
* \brief error message displayed when check label fail
*/
inline const char * CheckLabelErrorMsg(void) const {
if (loss_type != kLinearSquare) {
return "label must be in [0,1] for logistic regression";
} else {
return "";
}
}
/*!
* \brief calculate first order gradient of loss, given transformed prediction
* \param predt transformed prediction
* \param label true label
* \return first order gradient
*/
inline float FirstOrderGradient(float predt, float label) const {
switch (loss_type) {
case kLinearSquare: return predt - label;
case kLogisticRaw: predt = 1.0f / (1.0f + std::exp(-predt));
case kLogisticClassify:
case kLogisticNeglik: return predt - label;
default: utils::Error("unknown loss_type"); return 0.0f;
}
}
/*!
* \brief calculate second order gradient of loss, given transformed prediction
* \param predt transformed prediction
* \param label true label
* \return second order gradient
*/
inline float SecondOrderGradient(float predt, float label) const {
// cap second order gradient to positive value
const float eps = 1e-16f;
switch (loss_type) {
case kLinearSquare: return 1.0f;
case kLogisticRaw: predt = 1.0f / (1.0f + std::exp(-predt));
case kLogisticClassify:
case kLogisticNeglik: return std::max(predt * (1.0f - predt), eps);
default: utils::Error("unknown loss_type"); return 0.0f;
}
}
/*!
* \brief transform probability value back to margin
*/
inline float ProbToMargin(float base_score) const {
if (loss_type == kLogisticRaw ||
loss_type == kLogisticClassify ||
loss_type == kLogisticNeglik ) {
utils::Check(base_score > 0.0f && base_score < 1.0f,
"base_score must be in (0,1) for logistic loss");
base_score = -std::log(1.0f / base_score - 1.0f);
}
return base_score;
}
/*! \brief get default evaluation metric for the objective */
inline const char *DefaultEvalMetric(void) const {
if (loss_type == kLogisticClassify) return "error";
if (loss_type == kLogisticRaw) return "auc";
return "rmse";
}
};
/*! \brief objective function that only need to */
class RegLossObj : public IObjFunction {
public:
explicit RegLossObj(int loss_type) {
loss.loss_type = loss_type;
scale_pos_weight = 1.0f;
}
virtual ~RegLossObj(void) {}
virtual void SetParam(const char *name, const char *val) {
using namespace std;
if (!strcmp("scale_pos_weight", name)) {
scale_pos_weight = static_cast<float>(atof(val));
}
}
virtual void GetGradient(const std::vector<float> &preds,
const MetaInfo &info,
int iter,
std::vector<bst_gpair> *out_gpair) {
utils::Check(info.labels.size() != 0, "label set cannot be empty");
utils::Check(preds.size() % info.labels.size() == 0,
"labels are not correctly provided");
std::vector<bst_gpair> &gpair = *out_gpair;
gpair.resize(preds.size());
// check if label in range
bool label_correct = true;
// start calculating gradient
const unsigned nstep = static_cast<unsigned>(info.labels.size());
const bst_omp_uint ndata = static_cast<bst_omp_uint>(preds.size());
#pragma omp parallel for schedule(static)
for (bst_omp_uint i = 0; i < ndata; ++i) {
const unsigned j = i % nstep;
float p = loss.PredTransform(preds[i]);
float w = info.GetWeight(j);
if (info.labels[j] == 1.0f) w *= scale_pos_weight;
if (!loss.CheckLabel(info.labels[j])) label_correct = false;
gpair[i] = bst_gpair(loss.FirstOrderGradient(p, info.labels[j]) * w,
loss.SecondOrderGradient(p, info.labels[j]) * w);
}
utils::Check(label_correct, loss.CheckLabelErrorMsg());
}
virtual const char* DefaultEvalMetric(void) const {
return loss.DefaultEvalMetric();
}
virtual void PredTransform(std::vector<float> *io_preds) {
std::vector<float> &preds = *io_preds;
const bst_omp_uint ndata = static_cast<bst_omp_uint>(preds.size());
#pragma omp parallel for schedule(static)
for (bst_omp_uint j = 0; j < ndata; ++j) {
preds[j] = loss.PredTransform(preds[j]);
}
}
virtual float ProbToMargin(float base_score) const {
return loss.ProbToMargin(base_score);
}
protected:
float scale_pos_weight;
LossType loss;
};
// poisson regression for count
class PoissonRegression : public IObjFunction {
public:
PoissonRegression(void) {
max_delta_step = 0.0f;
}
virtual ~PoissonRegression(void) {}
virtual void SetParam(const char *name, const char *val) {
using namespace std;
if (!strcmp("max_delta_step", name)) {
max_delta_step = static_cast<float>(atof(val));
}
}
virtual void GetGradient(const std::vector<float> &preds,
const MetaInfo &info,
int iter,
std::vector<bst_gpair> *out_gpair) {
utils::Check(max_delta_step != 0.0f,
"PoissonRegression: need to set max_delta_step");
utils::Check(info.labels.size() != 0, "label set cannot be empty");
utils::Check(preds.size() == info.labels.size(),
"labels are not correctly provided");
std::vector<bst_gpair> &gpair = *out_gpair;
gpair.resize(preds.size());
// check if label in range
bool label_correct = true;
// start calculating gradient
const long ndata = static_cast<bst_omp_uint>(preds.size()); // NOLINT(*)
#pragma omp parallel for schedule(static)
for (long i = 0; i < ndata; ++i) { // NOLINT(*)
float p = preds[i];
float w = info.GetWeight(i);
float y = info.labels[i];
if (y >= 0.0f) {
gpair[i] = bst_gpair((std::exp(p) - y) * w,
std::exp(p + max_delta_step) * w);
} else {
label_correct = false;
}
}
utils::Check(label_correct,
"PoissonRegression: label must be nonnegative");
}
virtual void PredTransform(std::vector<float> *io_preds) {
std::vector<float> &preds = *io_preds;
const long ndata = static_cast<long>(preds.size()); // NOLINT(*)
#pragma omp parallel for schedule(static)
for (long j = 0; j < ndata; ++j) { // NOLINT(*)
preds[j] = std::exp(preds[j]);
}
}
virtual void EvalTransform(std::vector<float> *io_preds) {
PredTransform(io_preds);
}
virtual float ProbToMargin(float base_score) const {
return std::log(base_score);
}
virtual const char* DefaultEvalMetric(void) const {
return "poisson-nloglik";
}
private:
float max_delta_step;
};
// softmax multi-class classification
class SoftmaxMultiClassObj : public IObjFunction {
public:
explicit SoftmaxMultiClassObj(int output_prob)
: output_prob(output_prob) {
nclass = 0;
}
virtual ~SoftmaxMultiClassObj(void) {}
virtual void SetParam(const char *name, const char *val) {
using namespace std;
if (!strcmp( "num_class", name )) nclass = atoi(val);
}
virtual void GetGradient(const std::vector<float> &preds,
const MetaInfo &info,
int iter,
std::vector<bst_gpair> *out_gpair) {
utils::Check(nclass != 0, "must set num_class to use softmax");
utils::Check(info.labels.size() != 0, "label set cannot be empty");
utils::Check(preds.size() % (static_cast<size_t>(nclass) * info.labels.size()) == 0,
"SoftmaxMultiClassObj: label size and pred size does not match");
std::vector<bst_gpair> &gpair = *out_gpair;
gpair.resize(preds.size());
const unsigned nstep = static_cast<unsigned>(info.labels.size() * nclass);
const bst_omp_uint ndata = static_cast<bst_omp_uint>(preds.size() / nclass);
int label_error = 0;
#pragma omp parallel
{
std::vector<float> rec(nclass);
#pragma omp for schedule(static)
for (bst_omp_uint i = 0; i < ndata; ++i) {
for (int k = 0; k < nclass; ++k) {
rec[k] = preds[i * nclass + k];
}
Softmax(&rec);
const unsigned j = i % nstep;
int label = static_cast<int>(info.labels[j]);
if (label < 0 || label >= nclass) {
label_error = label; label = 0;
}
const float wt = info.GetWeight(j);
for (int k = 0; k < nclass; ++k) {
float p = rec[k];
const float h = 2.0f * p * (1.0f - p) * wt;
if (label == k) {
gpair[i * nclass + k] = bst_gpair((p - 1.0f) * wt, h);
} else {
gpair[i * nclass + k] = bst_gpair(p* wt, h);
}
}
}
}
utils::Check(label_error >= 0 && label_error < nclass,
"SoftmaxMultiClassObj: label must be in [0, num_class),"\
" num_class=%d but found %d in label", nclass, label_error);
}
virtual void PredTransform(std::vector<float> *io_preds) {
this->Transform(io_preds, output_prob);
}
virtual void EvalTransform(std::vector<float> *io_preds) {
this->Transform(io_preds, 1);
}
virtual const char* DefaultEvalMetric(void) const {
return "merror";
}
private:
inline void Transform(std::vector<float> *io_preds, int prob) {
utils::Check(nclass != 0, "must set num_class to use softmax");
std::vector<float> &preds = *io_preds;
std::vector<float> tmp;
const bst_omp_uint ndata = static_cast<bst_omp_uint>(preds.size()/nclass);
if (prob == 0) tmp.resize(ndata);
#pragma omp parallel
{
std::vector<float> rec(nclass);
#pragma omp for schedule(static)
for (bst_omp_uint j = 0; j < ndata; ++j) {
for (int k = 0; k < nclass; ++k) {
rec[k] = preds[j * nclass + k];
}
if (prob == 0) {
tmp[j] = static_cast<float>(FindMaxIndex(rec));
} else {
Softmax(&rec);
for (int k = 0; k < nclass; ++k) {
preds[j * nclass + k] = rec[k];
}
}
}
}
if (prob == 0) preds = tmp;
}
// data field
int nclass;
int output_prob;
};
/*! \brief objective for lambda rank */
class LambdaRankObj : public IObjFunction {
public:
LambdaRankObj(void) {
loss.loss_type = LossType::kLogisticRaw;
fix_list_weight = 0.0f;
num_pairsample = 1;
}
virtual ~LambdaRankObj(void) {}
virtual void SetParam(const char *name, const char *val) {
using namespace std;
if (!strcmp( "loss_type", name )) loss.loss_type = atoi(val);
if (!strcmp( "fix_list_weight", name)) fix_list_weight = static_cast<float>(atof(val));
if (!strcmp( "num_pairsample", name)) num_pairsample = atoi(val);
}
virtual void GetGradient(const std::vector<float> &preds,
const MetaInfo &info,
int iter,
std::vector<bst_gpair> *out_gpair) {
utils::Check(preds.size() == info.labels.size(), "label size predict size not match");
std::vector<bst_gpair> &gpair = *out_gpair;
gpair.resize(preds.size());
// quick consistency when group is not available
std::vector<unsigned> tgptr(2, 0); tgptr[1] = static_cast<unsigned>(info.labels.size());
const std::vector<unsigned> &gptr = info.group_ptr.size() == 0 ? tgptr : info.group_ptr;
utils::Check(gptr.size() != 0 && gptr.back() == info.labels.size(),
"group structure not consistent with #rows");
const bst_omp_uint ngroup = static_cast<bst_omp_uint>(gptr.size() - 1);
#pragma omp parallel
{
// parall construct, declare random number generator here, so that each
// thread use its own random number generator, seed by thread id and current iteration
random::Random rnd; rnd.Seed(iter* 1111 + omp_get_thread_num());
std::vector<LambdaPair> pairs;
std::vector<ListEntry> lst;
std::vector< std::pair<float, unsigned> > rec;
#pragma omp for schedule(static)
for (bst_omp_uint k = 0; k < ngroup; ++k) {
lst.clear(); pairs.clear();
for (unsigned j = gptr[k]; j < gptr[k+1]; ++j) {
lst.push_back(ListEntry(preds[j], info.labels[j], j));
gpair[j] = bst_gpair(0.0f, 0.0f);
}
std::sort(lst.begin(), lst.end(), ListEntry::CmpPred);
rec.resize(lst.size());
for (unsigned i = 0; i < lst.size(); ++i) {
rec[i] = std::make_pair(lst[i].label, i);
}
std::sort(rec.begin(), rec.end(), CmpFirst);
// enumerate buckets with same label, for each item in the lst, grab another sample randomly
for (unsigned i = 0; i < rec.size(); ) {
unsigned j = i + 1;
while (j < rec.size() && rec[j].first == rec[i].first) ++j;
// bucket in [i,j), get a sample outside bucket
unsigned nleft = i, nright = static_cast<unsigned>(rec.size() - j);
if (nleft + nright != 0) {
int nsample = num_pairsample;
while (nsample --) {
for (unsigned pid = i; pid < j; ++pid) {
unsigned ridx = static_cast<unsigned>(rnd.RandDouble() * (nleft+nright));
if (ridx < nleft) {
pairs.push_back(LambdaPair(rec[ridx].second, rec[pid].second));
} else {
pairs.push_back(LambdaPair(rec[pid].second, rec[ridx+j-i].second));
}
}
}
}
i = j;
}
// get lambda weight for the pairs
this->GetLambdaWeight(lst, &pairs);
// rescale each gradient and hessian so that the lst have constant weighted
float scale = 1.0f / num_pairsample;
if (fix_list_weight != 0.0f) {
scale *= fix_list_weight / (gptr[k+1] - gptr[k]);
}
for (size_t i = 0; i < pairs.size(); ++i) {
const ListEntry &pos = lst[pairs[i].pos_index];
const ListEntry &neg = lst[pairs[i].neg_index];
const float w = pairs[i].weight * scale;
float p = loss.PredTransform(pos.pred - neg.pred);
float g = loss.FirstOrderGradient(p, 1.0f);
float h = loss.SecondOrderGradient(p, 1.0f);
// accumulate gradient and hessian in both pid, and nid
gpair[pos.rindex].grad += g * w;
gpair[pos.rindex].hess += 2.0f * w * h;
gpair[neg.rindex].grad -= g * w;
gpair[neg.rindex].hess += 2.0f * w * h;
}
}
}
}
virtual const char* DefaultEvalMetric(void) const {
return "map";
}
protected:
/*! \brief helper information in a list */
struct ListEntry {
/*! \brief the predict score we in the data */
float pred;
/*! \brief the actual label of the entry */
float label;
/*! \brief row index in the data matrix */
unsigned rindex;
// constructor
ListEntry(float pred, float label, unsigned rindex)
: pred(pred), label(label), rindex(rindex) {}
// comparator by prediction
inline static bool CmpPred(const ListEntry &a, const ListEntry &b) {
return a.pred > b.pred;
}
// comparator by label
inline static bool CmpLabel(const ListEntry &a, const ListEntry &b) {
return a.label > b.label;
}
};
/*! \brief a pair in the lambda rank */
struct LambdaPair {
/*! \brief positive index: this is a position in the list */
unsigned pos_index;
/*! \brief negative index: this is a position in the list */
unsigned neg_index;
/*! \brief weight to be filled in */
float weight;
// constructor
LambdaPair(unsigned pos_index, unsigned neg_index)
: pos_index(pos_index), neg_index(neg_index), weight(1.0f) {}
};
/*!
* \brief get lambda weight for existing pairs
* \param list a list that is sorted by pred score
* \param io_pairs record of pairs, containing the pairs to fill in weights
*/
virtual void GetLambdaWeight(const std::vector<ListEntry> &sorted_list,
std::vector<LambdaPair> *io_pairs) = 0;
private:
// loss function
LossType loss;
// number of samples peformed for each instance
int num_pairsample;
// fix weight of each elements in list
float fix_list_weight;
};
class PairwiseRankObj: public LambdaRankObj{
public:
virtual ~PairwiseRankObj(void) {}
protected:
virtual void GetLambdaWeight(const std::vector<ListEntry> &sorted_list,
std::vector<LambdaPair> *io_pairs) {}
};
// beta version: NDCG lambda rank
class LambdaRankObjNDCG : public LambdaRankObj {
public:
virtual ~LambdaRankObjNDCG(void) {}
protected:
virtual void GetLambdaWeight(const std::vector<ListEntry> &sorted_list,
std::vector<LambdaPair> *io_pairs) {
std::vector<LambdaPair> &pairs = *io_pairs;
float IDCG;
{
std::vector<float> labels(sorted_list.size());
for (size_t i = 0; i < sorted_list.size(); ++i) {
labels[i] = sorted_list[i].label;
}
std::sort(labels.begin(), labels.end(), std::greater<float>());
IDCG = CalcDCG(labels);
}
if (IDCG == 0.0) {
for (size_t i = 0; i < pairs.size(); ++i) {
pairs[i].weight = 0.0f;
}
} else {
IDCG = 1.0f / IDCG;
for (size_t i = 0; i < pairs.size(); ++i) {
unsigned pos_idx = pairs[i].pos_index;
unsigned neg_idx = pairs[i].neg_index;
float pos_loginv = 1.0f / std::log(pos_idx + 2.0f);
float neg_loginv = 1.0f / std::log(neg_idx + 2.0f);
int pos_label = static_cast<int>(sorted_list[pos_idx].label);
int neg_label = static_cast<int>(sorted_list[neg_idx].label);
float original =
((1 << pos_label) - 1) * pos_loginv + ((1 << neg_label) - 1) * neg_loginv;
float changed =
((1 << neg_label) - 1) * pos_loginv + ((1 << pos_label) - 1) * neg_loginv;
float delta = (original - changed) * IDCG;
if (delta < 0.0f) delta = - delta;
pairs[i].weight = delta;
}
}
}
inline static float CalcDCG(const std::vector<float> &labels) {
double sumdcg = 0.0;
for (size_t i = 0; i < labels.size(); ++i) {
const unsigned rel = static_cast<unsigned>(labels[i]);
if (rel != 0) {
sumdcg += ((1 << rel) - 1) / std::log(static_cast<float>(i + 2));
}
}
return static_cast<float>(sumdcg);
}
};
class LambdaRankObjMAP : public LambdaRankObj {
public:
virtual ~LambdaRankObjMAP(void) {}
protected:
struct MAPStats {
/*! \brief the accumulated precision */
float ap_acc;
/*!
* \brief the accumulated precision,
* assuming a positive instance is missing
*/
float ap_acc_miss;
/*!
* \brief the accumulated precision,
* assuming that one more positive instance is inserted ahead
*/
float ap_acc_add;
/* \brief the accumulated positive instance count */
float hits;
MAPStats(void) {}
MAPStats(float ap_acc, float ap_acc_miss, float ap_acc_add, float hits)
: ap_acc(ap_acc), ap_acc_miss(ap_acc_miss), ap_acc_add(ap_acc_add), hits(hits) {}
};
/*!
* \brief Obtain the delta MAP if trying to switch the positions of instances in index1 or index2
* in sorted triples
* \param sorted_list the list containing entry information
* \param index1,index2 the instances switched
* \param map_stats a vector containing the accumulated precisions for each position in a list
*/
inline float GetLambdaMAP(const std::vector<ListEntry> &sorted_list,
int index1, int index2,
std::vector<MAPStats> *p_map_stats) {
std::vector<MAPStats> &map_stats = *p_map_stats;
if (index1 == index2 || map_stats[map_stats.size() - 1].hits == 0) {
return 0.0f;
}
if (index1 > index2) std::swap(index1, index2);
float original = map_stats[index2].ap_acc;
if (index1 != 0) original -= map_stats[index1 - 1].ap_acc;
float changed = 0;
float label1 = sorted_list[index1].label > 0.0f ? 1.0f : 0.0f;
float label2 = sorted_list[index2].label > 0.0f ? 1.0f : 0.0f;
if (label1 == label2) {
return 0.0;
} else if (label1 < label2) {
changed += map_stats[index2 - 1].ap_acc_add - map_stats[index1].ap_acc_add;
changed += (map_stats[index1].hits + 1.0f) / (index1 + 1);
} else {
changed += map_stats[index2 - 1].ap_acc_miss - map_stats[index1].ap_acc_miss;
changed += map_stats[index2].hits / (index2 + 1);
}
float ans = (changed - original) / (map_stats[map_stats.size() - 1].hits);
if (ans < 0) ans = -ans;
return ans;
}
/*
* \brief obtain preprocessing results for calculating delta MAP
* \param sorted_list the list containing entry information
* \param map_stats a vector containing the accumulated precisions for each position in a list
*/
inline void GetMAPStats(const std::vector<ListEntry> &sorted_list,
std::vector<MAPStats> *p_map_acc) {
std::vector<MAPStats> &map_acc = *p_map_acc;
map_acc.resize(sorted_list.size());
float hit = 0, acc1 = 0, acc2 = 0, acc3 = 0;
for (size_t i = 1; i <= sorted_list.size(); ++i) {
if (sorted_list[i - 1].label > 0.0f) {
hit++;
acc1 += hit / i;
acc2 += (hit - 1) / i;
acc3 += (hit + 1) / i;
}
map_acc[i - 1] = MAPStats(acc1, acc2, acc3, hit);
}
}
virtual void GetLambdaWeight(const std::vector<ListEntry> &sorted_list,
std::vector<LambdaPair> *io_pairs) {
std::vector<LambdaPair> &pairs = *io_pairs;
std::vector<MAPStats> map_stats;
GetMAPStats(sorted_list, &map_stats);
for (size_t i = 0; i < pairs.size(); ++i) {
pairs[i].weight =
GetLambdaMAP(sorted_list, pairs[i].pos_index,
pairs[i].neg_index, &map_stats);
}
}
};
} // namespace learner
} // namespace xgboost
#endif // XGBOOST_LEARNER_OBJECTIVE_INL_HPP_

89
src/learner/objective.h
View File

@@ -1,89 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file objective.h
* \brief interface of objective function used for gradient boosting
* \author Tianqi Chen, Kailong Chen
*/
#ifndef XGBOOST_LEARNER_OBJECTIVE_H_
#define XGBOOST_LEARNER_OBJECTIVE_H_
#include <vector>
#include "./dmatrix.h"
namespace xgboost {
namespace learner {
/*! \brief interface of objective function */
class IObjFunction{
public:
/*! \brief virtual destructor */
virtual ~IObjFunction(void) {}
/*!
* \brief set parameters from outside
* \param name name of the parameter
* \param val value of the parameter
*/
virtual void SetParam(const char *name, const char *val) = 0;
/*!
* \brief get gradient over each of predictions, given existing information
* \param preds prediction of current round
* \param info information about labels, weights, groups in rank
* \param iter current iteration number
* \param out_gpair output of get gradient, saves gradient and second order gradient in
*/
virtual void GetGradient(const std::vector<float> &preds,
const MetaInfo &info,
int iter,
std::vector<bst_gpair> *out_gpair) = 0;
/*! \return the default evaluation metric for the objective */
virtual const char* DefaultEvalMetric(void) const = 0;
// the following functions are optional, most of time default implementation is good enough
/*!
* \brief transform prediction values, this is only called when Prediction is called
* \param io_preds prediction values, saves to this vector as well
*/
virtual void PredTransform(std::vector<float> *io_preds) {}
/*!
* \brief transform prediction values, this is only called when Eval is called,
* usually it redirect to PredTransform
* \param io_preds prediction values, saves to this vector as well
*/
virtual void EvalTransform(std::vector<float> *io_preds) {
this->PredTransform(io_preds);
}
/*!
* \brief transform probability value back to margin
* this is used to transform user-set base_score back to margin
* used by gradient boosting
* \return transformed value
*/
virtual float ProbToMargin(float base_score) const {
return base_score;
}
};
} // namespace learner
} // namespace xgboost
// this are implementations of objective functions
#include "objective-inl.hpp"
// factory function
namespace xgboost {
namespace learner {
/*! \brief factory function to create objective function by name */
inline IObjFunction* CreateObjFunction(const char *name) {
using namespace std;
if (!strcmp("reg:linear", name)) return new RegLossObj(LossType::kLinearSquare);
if (!strcmp("reg:logistic", name)) return new RegLossObj(LossType::kLogisticNeglik);
if (!strcmp("binary:logistic", name)) return new RegLossObj(LossType::kLogisticClassify);
if (!strcmp("binary:logitraw", name)) return new RegLossObj(LossType::kLogisticRaw);
if (!strcmp("count:poisson", name)) return new PoissonRegression();
if (!strcmp("multi:softmax", name)) return new SoftmaxMultiClassObj(0);
if (!strcmp("multi:softprob", name)) return new SoftmaxMultiClassObj(1);
if (!strcmp("rank:pairwise", name )) return new PairwiseRankObj();
if (!strcmp("rank:ndcg", name)) return new LambdaRankObjNDCG();
if (!strcmp("rank:map", name)) return new LambdaRankObjMAP();
utils::Error("unknown objective function type: %s", name);
return NULL;
}
} // namespace learner
} // namespace xgboost
#endif // XGBOOST_LEARNER_OBJECTIVE_H_

13
src/sync/sync.h
View File

@@ -1,13 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file sync.h
* \brief the synchronization module of rabit
* redirects to subtree rabit header
* \author Tianqi Chen
*/
#ifndef XGBOOST_SYNC_SYNC_H_
#define XGBOOST_SYNC_SYNC_H_
#include "../../subtree/rabit/include/rabit.h"
#include "../../subtree/rabit/include/rabit/timer.h"
#endif // XGBOOST_SYNC_SYNC_H_

573
src/tree/model.h
View File

@@ -1,573 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file model.h
* \brief model structure for tree
* \author Tianqi Chen
*/
#ifndef XGBOOST_TREE_MODEL_H_
#define XGBOOST_TREE_MODEL_H_
#include <string>
#include <cstring>
#include <sstream>
#include <limits>
#include <algorithm>
#include <vector>
#include <cmath>
#include "../utils/io.h"
#include "../utils/fmap.h"
#include "../utils/utils.h"
namespace xgboost {
namespace tree {
/*!
* \brief template class of TreeModel
* \tparam TSplitCond data type to indicate split condition
* \tparam TNodeStat auxiliary statistics of node to help tree building
*/
template<typename TSplitCond, typename TNodeStat>
class TreeModel {
public:
/*! \brief data type to indicate split condition */
typedef TNodeStat NodeStat;
/*! \brief auxiliary statistics of node to help tree building */
typedef TSplitCond SplitCond;
/*! \brief parameters of the tree */
struct Param{
/*! \brief number of start root */
int num_roots;
/*! \brief total number of nodes */
int num_nodes;
/*!\brief number of deleted nodes */
int num_deleted;
/*! \brief maximum depth, this is a statistics of the tree */
int max_depth;
/*! \brief number of features used for tree construction */
int num_feature;
/*!
* \brief leaf vector size, used for vector tree
* used to store more than one dimensional information in tree
*/
int size_leaf_vector;
/*! \brief reserved part */
int reserved[31];
/*! \brief constructor */
Param(void) {
max_depth = 0;
size_leaf_vector = 0;
std::memset(reserved, 0, sizeof(reserved));
}
/*!
* \brief set parameters from outside
* \param name name of the parameter
* \param val value of the parameter
*/
inline void SetParam(const char *name, const char *val) {
using namespace std;
if (!strcmp("num_roots", name)) num_roots = atoi(val);
if (!strcmp("num_feature", name)) num_feature = atoi(val);
if (!strcmp("size_leaf_vector", name)) size_leaf_vector = atoi(val);
}
};
/*! \brief tree node */
class Node {
public:
Node(void) : sindex_(0) {}
/*! \brief index of left child */
inline int cleft(void) const {
return this->cleft_;
}
/*! \brief index of right child */
inline int cright(void) const {
return this->cright_;
}
/*! \brief index of default child when feature is missing */
inline int cdefault(void) const {
return this->default_left() ? this->cleft() : this->cright();
}
/*! \brief feature index of split condition */
inline unsigned split_index(void) const {
return sindex_ & ((1U << 31) - 1U);
}
/*! \brief when feature is unknown, whether goes to left child */
inline bool default_left(void) const {
return (sindex_ >> 31) != 0;
}
/*! \brief whether current node is leaf node */
inline bool is_leaf(void) const {
return cleft_ == -1;
}
/*! \brief get leaf value of leaf node */
inline float leaf_value(void) const {
return (this->info_).leaf_value;
}
/*! \brief get split condition of the node */
inline TSplitCond split_cond(void) const {
return (this->info_).split_cond;
}
/*! \brief get parent of the node */
inline int parent(void) const {
return parent_ & ((1U << 31) - 1);
}
/*! \brief whether current node is left child */
inline bool is_left_child(void) const {
return (parent_ & (1U << 31)) != 0;
}
/*! \brief whether this node is deleted */
inline bool is_deleted(void) const {
return sindex_ == std::numeric_limits<unsigned>::max();
}
/*! \brief whether current node is root */
inline bool is_root(void) const {
return parent_ == -1;
}
/*!
* \brief set the right child
* \param nide node id to right child
*/
inline void set_right_child(int nid) {
this->cright_ = nid;
}
/*!
* \brief set split condition of current node
* \param split_index feature index to split
* \param split_cond split condition
* \param default_left the default direction when feature is unknown
*/
inline void set_split(unsigned split_index, TSplitCond split_cond,
bool default_left = false) {
if (default_left) split_index |= (1U << 31);
this->sindex_ = split_index;
(this->info_).split_cond = split_cond;
}
/*!
* \brief set the leaf value of the node
* \param value leaf value
* \param right right index, could be used to store
* additional information
*/
inline void set_leaf(float value, int right = -1) {
(this->info_).leaf_value = value;
this->cleft_ = -1;
this->cright_ = right;
}
/*! \brief mark that this node is deleted */
inline void mark_delete(void) {
this->sindex_ = std::numeric_limits<unsigned>::max();
}
private:
friend class TreeModel<TSplitCond, TNodeStat>;
/*!
* \brief in leaf node, we have weights, in non-leaf nodes,
* we have split condition
*/
union Info{
float leaf_value;
TSplitCond split_cond;
};
// pointer to parent, highest bit is used to
// indicate whether it's a left child or not
int parent_;
// pointer to left, right
int cleft_, cright_;
// split feature index, left split or right split depends on the highest bit
unsigned sindex_;
// extra info
Info info_;
// set parent
inline void set_parent(int pidx, bool is_left_child = true) {
if (is_left_child) pidx |= (1U << 31);
this->parent_ = pidx;
}
};
protected:
// vector of nodes
std::vector<Node> nodes;
// free node space, used during training process
std::vector<int> deleted_nodes;
// stats of nodes
std::vector<TNodeStat> stats;
// leaf vector, that is used to store additional information
std::vector<bst_float> leaf_vector;
// allocate a new node,
// !!!!!! NOTE: may cause BUG here, nodes.resize
inline int AllocNode(void) {
if (param.num_deleted != 0) {
int nd = deleted_nodes.back();
deleted_nodes.pop_back();
--param.num_deleted;
return nd;
}
int nd = param.num_nodes++;
utils::Check(param.num_nodes < std::numeric_limits<int>::max(),
"number of nodes in the tree exceed 2^31");
nodes.resize(param.num_nodes);
stats.resize(param.num_nodes);
leaf_vector.resize(param.num_nodes * param.size_leaf_vector);
return nd;
}
// delete a tree node, keep the parent field to allow trace back
inline void DeleteNode(int nid) {
utils::Assert(nid >= param.num_roots, "can not delete root");
deleted_nodes.push_back(nid);
nodes[nid].mark_delete();
++param.num_deleted;
}
public:
/*!
* \brief change a non leaf node to a leaf node, delete its children
* \param rid node id of the node
* \param new leaf value
*/
inline void ChangeToLeaf(int rid, float value) {
utils::Assert(nodes[nodes[rid].cleft() ].is_leaf(),
"can not delete a non termial child");
utils::Assert(nodes[nodes[rid].cright()].is_leaf(),
"can not delete a non termial child");
this->DeleteNode(nodes[rid].cleft());
this->DeleteNode(nodes[rid].cright());
nodes[rid].set_leaf(value);
}
/*!
* \brief collapse a non leaf node to a leaf node, delete its children
* \param rid node id of the node
* \param new leaf value
*/
inline void CollapseToLeaf(int rid, float value) {
if (nodes[rid].is_leaf()) return;
if (!nodes[nodes[rid].cleft() ].is_leaf()) {
CollapseToLeaf(nodes[rid].cleft(), 0.0f);
}
if (!nodes[nodes[rid].cright() ].is_leaf()) {
CollapseToLeaf(nodes[rid].cright(), 0.0f);
}
this->ChangeToLeaf(rid, value);
}
public:
/*! \brief model parameter */
Param param;
/*! \brief constructor */
TreeModel(void) {
param.num_nodes = 1;
param.num_roots = 1;
param.num_deleted = 0;
nodes.resize(1);
}
/*! \brief get node given nid */
inline Node &operator[](int nid) {
return nodes[nid];
}
/*! \brief get node given nid */
inline const Node &operator[](int nid) const {
return nodes[nid];
}
/*! \brief get node statistics given nid */
inline NodeStat &stat(int nid) {
return stats[nid];
}
/*! \brief get leaf vector given nid */
inline bst_float* leafvec(int nid) {
if (leaf_vector.size() == 0) return NULL;
return &leaf_vector[nid * param.size_leaf_vector];
}
/*! \brief get leaf vector given nid */
inline const bst_float* leafvec(int nid) const {
if (leaf_vector.size() == 0) return NULL;
return &leaf_vector[nid * param.size_leaf_vector];
}
/*! \brief initialize the model */
inline void InitModel(void) {
param.num_nodes = param.num_roots;
nodes.resize(param.num_nodes);
stats.resize(param.num_nodes);
leaf_vector.resize(param.num_nodes * param.size_leaf_vector, 0.0f);
for (int i = 0; i < param.num_nodes; i ++) {
nodes[i].set_leaf(0.0f);
nodes[i].set_parent(-1);
}
}
/*!
* \brief load model from stream
* \param fi input stream
*/
inline void LoadModel(utils::IStream &fi) { // NOLINT(*)
utils::Check(fi.Read(&param, sizeof(Param)) > 0,
"TreeModel: wrong format");
nodes.resize(param.num_nodes); stats.resize(param.num_nodes);
utils::Assert(param.num_nodes != 0, "invalid model");
utils::Check(fi.Read(BeginPtr(nodes), sizeof(Node) * nodes.size()) > 0,
"TreeModel: wrong format");
utils::Check(fi.Read(BeginPtr(stats), sizeof(NodeStat) * stats.size()) > 0,
"TreeModel: wrong format");
if (param.size_leaf_vector != 0) {
utils::Check(fi.Read(&leaf_vector), "TreeModel: wrong format");
}
// chg deleted nodes
deleted_nodes.resize(0);
for (int i = param.num_roots; i < param.num_nodes; ++i) {
if (nodes[i].is_deleted()) deleted_nodes.push_back(i);
}
utils::Assert(static_cast<int>(deleted_nodes.size()) == param.num_deleted,
"number of deleted nodes do not match, num_deleted=%d, dnsize=%lu, num_nodes=%d",
param.num_deleted, deleted_nodes.size(), param.num_nodes);
}
/*!
* \brief save model to stream
* \param fo output stream
*/
inline void SaveModel(utils::IStream &fo) const { // NOLINT(*)
utils::Assert(param.num_nodes == static_cast<int>(nodes.size()),
"TreeModel::SaveModel");
utils::Assert(param.num_nodes == static_cast<int>(stats.size()),
"TreeModel::SaveModel");
fo.Write(&param, sizeof(Param));
utils::Assert(param.num_nodes != 0, "invalid model");
fo.Write(BeginPtr(nodes), sizeof(Node) * nodes.size());
fo.Write(BeginPtr(stats), sizeof(NodeStat) * nodes.size());
if (param.size_leaf_vector != 0) fo.Write(leaf_vector);
}
/*!
* \brief add child nodes to node
* \param nid node id to add childs
*/
inline void AddChilds(int nid) {
int pleft = this->AllocNode();
int pright = this->AllocNode();
nodes[nid].cleft_ = pleft;
nodes[nid].cright_ = pright;
nodes[nodes[nid].cleft() ].set_parent(nid, true);
nodes[nodes[nid].cright()].set_parent(nid, false);
}
/*!
* \brief only add a right child to a leaf node
* \param node id to add right child
*/
inline void AddRightChild(int nid) {
int pright = this->AllocNode();
nodes[nid].right = pright;
nodes[nodes[nid].right].set_parent(nid, false);
}
/*!
* \brief get current depth
* \param nid node id
* \param pass_rchild whether right child is not counted in depth
*/
inline int GetDepth(int nid, bool pass_rchild = false) const {
int depth = 0;
while (!nodes[nid].is_root()) {
if (!pass_rchild || nodes[nid].is_left_child()) ++depth;
nid = nodes[nid].parent();
}
return depth;
}
/*!
* \brief get maximum depth
* \param nid node id
*/
inline int MaxDepth(int nid) const {
if (nodes[nid].is_leaf()) return 0;
return std::max(MaxDepth(nodes[nid].cleft())+1,
MaxDepth(nodes[nid].cright())+1);
}
/*!
* \brief get maximum depth
*/
inline int MaxDepth(void) {
int maxd = 0;
for (int i = 0; i < param.num_roots; ++i) {
maxd = std::max(maxd, MaxDepth(i));
}
return maxd;
}
/*! \brief number of extra nodes besides the root */
inline int num_extra_nodes(void) const {
return param.num_nodes - param.num_roots - param.num_deleted;
}
/*!
* \brief dump model to text string
* \param fmap feature map of feature types
* \param with_stats whether dump out statistics as well
* \return the string of dumped model
*/
inline std::string DumpModel(const utils::FeatMap& fmap, bool with_stats) {
std::stringstream fo("");
for (int i = 0; i < param.num_roots; ++i) {
this->Dump(i, fo, fmap, 0, with_stats);
}
return fo.str();
}
private:
void Dump(int nid, std::stringstream &fo, // NOLINT(*)
const utils::FeatMap& fmap, int depth, bool with_stats) {
for (int i = 0; i < depth; ++i) {
fo << '\t';
}
if (nodes[nid].is_leaf()) {
fo << nid << ":leaf=" << nodes[nid].leaf_value();
if (with_stats) {
stat(nid).Print(fo, true);
}
fo << '\n';
} else {
// right then left,
TSplitCond cond = nodes[nid].split_cond();
const unsigned split_index = nodes[nid].split_index();
if (split_index < fmap.size()) {
switch (fmap.type(split_index)) {
case utils::FeatMap::kIndicator: {
int nyes = nodes[nid].default_left() ?
nodes[nid].cright() : nodes[nid].cleft();
fo << nid << ":[" << fmap.name(split_index) << "] yes=" << nyes
<< ",no=" << nodes[nid].cdefault();
break;
}
case utils::FeatMap::kInteger: {
fo << nid << ":[" << fmap.name(split_index) << "<"
<< int(float(cond)+1.0f)
<< "] yes=" << nodes[nid].cleft()
<< ",no=" << nodes[nid].cright()
<< ",missing=" << nodes[nid].cdefault();
break;
}
case utils::FeatMap::kFloat:
case utils::FeatMap::kQuantitive: {
fo << nid << ":[" << fmap.name(split_index) << "<"<< float(cond)
<< "] yes=" << nodes[nid].cleft()
<< ",no=" << nodes[nid].cright()
<< ",missing=" << nodes[nid].cdefault();
break;
}
default: utils::Error("unknown fmap type");
}
} else {
fo << nid << ":[f" << split_index << "<"<< float(cond)
<< "] yes=" << nodes[nid].cleft()
<< ",no=" << nodes[nid].cright()
<< ",missing=" << nodes[nid].cdefault();
}
if (with_stats) {
stat(nid).Print(fo, false);
}
fo << '\n';
this->Dump(nodes[nid].cleft(), fo, fmap, depth+1, with_stats);
this->Dump(nodes[nid].cright(), fo, fmap, depth+1, with_stats);
}
}
};
/*! \brief node statistics used in regression tree */
struct RTreeNodeStat {
/*! \brief loss change caused by current split */
float loss_chg;
/*! \brief sum of hessian values, used to measure coverage of data */
float sum_hess;
/*! \brief weight of current node */
float base_weight;
/*! \brief number of child that is leaf node known up to now */
int leaf_child_cnt;
/*! \brief print information of current stats to fo */
inline void Print(std::stringstream &fo, bool is_leaf) const { // NOLINT(*)
if (!is_leaf) {
fo << ",gain=" << loss_chg << ",cover=" << sum_hess;
} else {
fo << ",cover=" << sum_hess;
}
}
};
/*! \brief define regression tree to be the most common tree model */
class RegTree: public TreeModel<bst_float, RTreeNodeStat>{
public:
/*!
* \brief dense feature vector that can be taken by RegTree
* to do traverse efficiently
* and can be construct from sparse feature vector
*/
struct FVec {
/*!
* \brief a union value of value and flag
* when flag == -1, this indicate the value is missing
*/
union Entry{
float fvalue;
int flag;
};
std::vector<Entry> data;
/*! \brief initialize the vector with size vector */
inline void Init(size_t size) {
Entry e; e.flag = -1;
data.resize(size);
std::fill(data.begin(), data.end(), e);
}
/*! \brief fill the vector with sparse vector */
inline void Fill(const RowBatch::Inst &inst) {
for (bst_uint i = 0; i < inst.length; ++i) {
if (inst[i].index >= data.size()) continue;
data[inst[i].index].fvalue = inst[i].fvalue;
}
}
/*! \brief drop the trace after fill, must be called after fill */
inline void Drop(const RowBatch::Inst &inst) {
for (bst_uint i = 0; i < inst.length; ++i) {
if (inst[i].index >= data.size()) continue;
data[inst[i].index].flag = -1;
}
}
/*! \brief get ith value */
inline float fvalue(size_t i) const {
return data[i].fvalue;
}
/*! \brief check whether i-th entry is missing */
inline bool is_missing(size_t i) const {
return data[i].flag == -1;
}
};
/*!
* \brief get the leaf index
* \param feat dense feature vector, if the feature is missing the field is set to NaN
* \param root_id starting root index of the instance
* \return the leaf index of the given feature
*/
inline int GetLeafIndex(const FVec &feat, unsigned root_id = 0) const {
// start from groups that belongs to current data
int pid = static_cast<int>(root_id);
// traverse tree
while (!(*this)[ pid ].is_leaf()) {
unsigned split_index = (*this)[pid].split_index();
pid = this->GetNext(pid, feat.fvalue(split_index), feat.is_missing(split_index));
}
return pid;
}
/*!
* \brief get the prediction of regression tree, only accepts dense feature vector
* \param feats dense feature vector, if the feature is missing the field is set to NaN
* \param root_id starting root index of the instance
* \return the leaf index of the given feature
*/
inline float Predict(const FVec &feat, unsigned root_id = 0) const {
int pid = this->GetLeafIndex(feat, root_id);
return (*this)[pid].leaf_value();
}
/*! \brief get next position of the tree given current pid */
inline int GetNext(int pid, float fvalue, bool is_unknown) const {
float split_value = (*this)[pid].split_cond();
if (is_unknown) {
return (*this)[pid].cdefault();
} else {
if (fvalue < split_value) {
return (*this)[pid].cleft();
} else {
return (*this)[pid].cright();
}
}
}
};
} // namespace tree
} // namespace xgboost
#endif // XGBOOST_TREE_MODEL_H_

429
src/tree/param.h
View File

@@ -1,429 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file param.h
* \brief training parameters, statistics used to support tree construction
* \author Tianqi Chen
*/
#ifndef XGBOOST_TREE_PARAM_H_
#define XGBOOST_TREE_PARAM_H_
#include <vector>
#include <cstring>
#include "../data.h"
namespace xgboost {
namespace tree {
/*! \brief training parameters for regression tree */
struct TrainParam{
// learning step size for a time
float learning_rate;
// minimum loss change required for a split
float min_split_loss;
// maximum depth of a tree
int max_depth;
//----- the rest parameters are less important ----
// minimum amount of hessian(weight) allowed in a child
float min_child_weight;
// L2 regularization factor
float reg_lambda;
// L1 regularization factor
float reg_alpha;
// default direction choice
int default_direction;
// maximum delta update we can add in weight estimation
// this parameter can be used to stabilize update
// default=0 means no constraint on weight delta
float max_delta_step;
// whether we want to do subsample
float subsample;
// whether to subsample columns each split, in each level
float colsample_bylevel;
// whether to subsample columns during tree construction
float colsample_bytree;
// speed optimization for dense column
float opt_dense_col;
// accuracy of sketch
float sketch_eps;
// accuracy of sketch
float sketch_ratio;
// leaf vector size
int size_leaf_vector;
// option for parallelization
int parallel_option;
// option to open cacheline optimization
int cache_opt;
// number of threads to be used for tree construction,
// if OpenMP is enabled, if equals 0, use system default
int nthread;
/*! \brief constructor */
TrainParam(void) {
learning_rate = 0.3f;
min_split_loss = 0.0f;
min_child_weight = 1.0f;
max_delta_step = 0.0f;
max_depth = 6;
reg_lambda = 1.0f;
reg_alpha = 0.0f;
default_direction = 0;
subsample = 1.0f;
colsample_bytree = 1.0f;
colsample_bylevel = 1.0f;
opt_dense_col = 1.0f;
nthread = 0;
size_leaf_vector = 0;
// enforce parallel option to 0 for now, investigate the other strategy
parallel_option = 0;
sketch_eps = 0.1f;
sketch_ratio = 2.0f;
cache_opt = 1;
}
/*!
* \brief set parameters from outside
* \param name name of the parameter
* \param val value of the parameter
*/
inline void SetParam(const char *name, const char *val) {
using namespace std;
// sync-names
if (!strcmp(name, "gamma")) min_split_loss = static_cast<float>(atof(val));
if (!strcmp(name, "eta")) learning_rate = static_cast<float>(atof(val));
if (!strcmp(name, "lambda")) reg_lambda = static_cast<float>(atof(val));
if (!strcmp(name, "alpha")) reg_alpha = static_cast<float>(atof(val));
if (!strcmp(name, "learning_rate")) learning_rate = static_cast<float>(atof(val));
if (!strcmp(name, "min_child_weight")) min_child_weight = static_cast<float>(atof(val));
if (!strcmp(name, "min_split_loss")) min_split_loss = static_cast<float>(atof(val));
if (!strcmp(name, "max_delta_step")) max_delta_step = static_cast<float>(atof(val));
if (!strcmp(name, "reg_lambda")) reg_lambda = static_cast<float>(atof(val));
if (!strcmp(name, "reg_alpha")) reg_alpha = static_cast<float>(atof(val));
if (!strcmp(name, "subsample")) subsample = static_cast<float>(atof(val));
if (!strcmp(name, "colsample_bylevel")) colsample_bylevel = static_cast<float>(atof(val));
if (!strcmp(name, "colsample_bytree")) colsample_bytree = static_cast<float>(atof(val));
if (!strcmp(name, "sketch_eps")) sketch_eps = static_cast<float>(atof(val));
if (!strcmp(name, "sketch_ratio")) sketch_ratio = static_cast<float>(atof(val));
if (!strcmp(name, "opt_dense_col")) opt_dense_col = static_cast<float>(atof(val));
if (!strcmp(name, "size_leaf_vector")) size_leaf_vector = atoi(val);
if (!strcmp(name, "cache_opt")) cache_opt = atoi(val);
if (!strcmp(name, "max_depth")) max_depth = atoi(val);
if (!strcmp(name, "nthread")) nthread = atoi(val);
if (!strcmp(name, "parallel_option")) parallel_option = atoi(val);
if (!strcmp(name, "default_direction")) {
if (!strcmp(val, "learn")) default_direction = 0;
if (!strcmp(val, "left")) default_direction = 1;
if (!strcmp(val, "right")) default_direction = 2;
}
}
// calculate the cost of loss function
inline double CalcGain(double sum_grad, double sum_hess) const {
if (sum_hess < min_child_weight) return 0.0;
if (max_delta_step == 0.0f) {
if (reg_alpha == 0.0f) {
return Sqr(sum_grad) / (sum_hess + reg_lambda);
} else {
return Sqr(ThresholdL1(sum_grad, reg_alpha)) / (sum_hess + reg_lambda);
}
} else {
double w = CalcWeight(sum_grad, sum_hess);
double ret = sum_grad * w + 0.5 * (sum_hess + reg_lambda) * Sqr(w);
if (reg_alpha == 0.0f) {
return - 2.0 * ret;
} else {
return - 2.0 * (ret + reg_alpha * std::abs(w));
}
}
}
// calculate cost of loss function with four statistics
inline double CalcGain(double sum_grad, double sum_hess,
double test_grad, double test_hess) const {
double w = CalcWeight(sum_grad, sum_hess);
double ret = test_grad * w + 0.5 * (test_hess + reg_lambda) * Sqr(w);
if (reg_alpha == 0.0f) {
return - 2.0 * ret;
} else {
return - 2.0 * (ret + reg_alpha * std::abs(w));
}
}
// calculate weight given the statistics
inline double CalcWeight(double sum_grad, double sum_hess) const {
if (sum_hess < min_child_weight) return 0.0;
double dw;
if (reg_alpha == 0.0f) {
dw = -sum_grad / (sum_hess + reg_lambda);
} else {
dw = -ThresholdL1(sum_grad, reg_alpha) / (sum_hess + reg_lambda);
}
if (max_delta_step != 0.0f) {
if (dw > max_delta_step) dw = max_delta_step;
if (dw < -max_delta_step) dw = -max_delta_step;
}
return dw;
}
/*! \brief whether need forward small to big search: default right */
inline bool need_forward_search(float col_density, bool indicator) const {
return this->default_direction == 2 ||
(default_direction == 0 && (col_density < opt_dense_col) && !indicator);
}
/*! \brief whether need backward big to small search: default left */
inline bool need_backward_search(float col_density, bool indicator) const {
return this->default_direction != 2;
}
/*! \brief given the loss change, whether we need to invoke pruning */
inline bool need_prune(double loss_chg, int depth) const {
return loss_chg < this->min_split_loss;
}
/*! \brief whether we can split with current hessian */
inline bool cannot_split(double sum_hess, int depth) const {
return sum_hess < this->min_child_weight * 2.0;
}
/*! \brief maximum sketch size */
inline unsigned max_sketch_size(void) const {
unsigned ret = static_cast<unsigned>(sketch_ratio / sketch_eps);
utils::Check(ret > 0, "sketch_ratio/sketch_eps must be bigger than 1");
return ret;
}
protected:
// functions for L1 cost
inline static double ThresholdL1(double w, double lambda) {
if (w > +lambda) return w - lambda;
if (w < -lambda) return w + lambda;
return 0.0;
}
inline static double Sqr(double a) {
return a * a;
}
};
/*! \brief core statistics used for tree construction */
struct GradStats {
/*! \brief sum gradient statistics */
double sum_grad;
/*! \brief sum hessian statistics */
double sum_hess;
/*!
* \brief whether this is simply statistics and we only need to call
* Add(gpair), instead of Add(gpair, info, ridx)
*/
static const int kSimpleStats = 1;
/*! \brief constructor, the object must be cleared during construction */
explicit GradStats(const TrainParam &param) {
this->Clear();
}
/*! \brief clear the statistics */
inline void Clear(void) {
sum_grad = sum_hess = 0.0f;
}
/*! \brief check if necessary information is ready */
inline static void CheckInfo(const BoosterInfo &info) {
}
/*!
* \brief accumulate statistics
* \param p the gradient pair
*/
inline void Add(bst_gpair p) {
this->Add(p.grad, p.hess);
}
/*!
* \brief accumulate statistics, more complicated version
* \param gpair the vector storing the gradient statistics
* \param info the additional information
* \param ridx instance index of this instance
*/
inline void Add(const std::vector<bst_gpair> &gpair,
const BoosterInfo &info,
bst_uint ridx) {
const bst_gpair &b = gpair[ridx];
this->Add(b.grad, b.hess);
}
/*! \brief calculate leaf weight */
inline double CalcWeight(const TrainParam &param) const {
return param.CalcWeight(sum_grad, sum_hess);
}
/*! \brief calculate gain of the solution */
inline double CalcGain(const TrainParam &param) const {
return param.CalcGain(sum_grad, sum_hess);
}
/*! \brief add statistics to the data */
inline void Add(const GradStats &b) {
this->Add(b.sum_grad, b.sum_hess);
}
/*! \brief same as add, reduce is used in All Reduce */
inline static void Reduce(GradStats &a, const GradStats &b) { // NOLINT(*)
a.Add(b);
}
/*! \brief set current value to a - b */
inline void SetSubstract(const GradStats &a, const GradStats &b) {
sum_grad = a.sum_grad - b.sum_grad;
sum_hess = a.sum_hess - b.sum_hess;
}
/*! \return whether the statistics is not used yet */
inline bool Empty(void) const {
return sum_hess == 0.0;
}
/*! \brief set leaf vector value based on statistics */
inline void SetLeafVec(const TrainParam &param, bst_float *vec) const {
}
// constructor to allow inheritance
GradStats(void) {}
/*! \brief add statistics to the data */
inline void Add(double grad, double hess) {
sum_grad += grad; sum_hess += hess;
}
};
/*! \brief vectorized cv statistics */
template<unsigned vsize>
struct CVGradStats : public GradStats {
// additional statistics
GradStats train[vsize], valid[vsize];
// constructor
explicit CVGradStats(const TrainParam &param) {
utils::Check(param.size_leaf_vector == vsize,
"CVGradStats: vsize must match size_leaf_vector");
this->Clear();
}
/*! \brief check if necessary information is ready */
inline static void CheckInfo(const BoosterInfo &info) {
utils::Check(info.fold_index.size() != 0,
"CVGradStats: require fold_index");
}
/*! \brief clear the statistics */
inline void Clear(void) {
GradStats::Clear();
for (unsigned i = 0; i < vsize; ++i) {
train[i].Clear(); valid[i].Clear();
}
}
inline void Add(const std::vector<bst_gpair> &gpair,
const BoosterInfo &info,
bst_uint ridx) {
GradStats::Add(gpair[ridx].grad, gpair[ridx].hess);
const size_t step = info.fold_index.size();
for (unsigned i = 0; i < vsize; ++i) {
const bst_gpair &b = gpair[(i + 1) * step + ridx];
if (info.fold_index[ridx] == i) {
valid[i].Add(b.grad, b.hess);
} else {
train[i].Add(b.grad, b.hess);
}
}
}
/*! \brief calculate gain of the solution */
inline double CalcGain(const TrainParam &param) const {
double ret = 0.0;
for (unsigned i = 0; i < vsize; ++i) {
ret += param.CalcGain(train[i].sum_grad,
train[i].sum_hess,
vsize * valid[i].sum_grad,
vsize * valid[i].sum_hess);
}
return ret / vsize;
}
/*! \brief add statistics to the data */
inline void Add(const CVGradStats &b) {
GradStats::Add(b);
for (unsigned i = 0; i < vsize; ++i) {
train[i].Add(b.train[i]);
valid[i].Add(b.valid[i]);
}
}
/*! \brief same as add, reduce is used in All Reduce */
inline static void Reduce(CVGradStats &a, const CVGradStats &b) { // NOLINT(*)
a.Add(b);
}
/*! \brief set current value to a - b */
inline void SetSubstract(const CVGradStats &a, const CVGradStats &b) {
GradStats::SetSubstract(a, b);
for (int i = 0; i < vsize; ++i) {
train[i].SetSubstract(a.train[i], b.train[i]);
valid[i].SetSubstract(a.valid[i], b.valid[i]);
}
}
/*! \brief set leaf vector value based on statistics */
inline void SetLeafVec(const TrainParam &param, bst_float *vec) const{
for (int i = 0; i < vsize; ++i) {
vec[i] = param.learning_rate *
param.CalcWeight(train[i].sum_grad, train[i].sum_hess);
}
}
};
/*!
* \brief statistics that is helpful to store
* and represent a split solution for the tree
*/
struct SplitEntry{
/*! \brief loss change after split this node */
bst_float loss_chg;
/*! \brief split index */
unsigned sindex;
/*! \brief split value */
float split_value;
/*! \brief constructor */
SplitEntry(void) : loss_chg(0.0f), sindex(0), split_value(0.0f) {}
/*!
* \brief decides whether we can replace current entry with the given statistics
* This function gives better priority to lower index when loss_chg == new_loss_chg.
* Not the best way, but helps to give consistent result during multi-thread execution.
* \param new_loss_chg the loss reduction get through the split
* \param split_index the feature index where the split is on
*/
inline bool NeedReplace(bst_float new_loss_chg, unsigned split_index) const {
if (this->split_index() <= split_index) {
return new_loss_chg > this->loss_chg;
} else {
return !(this->loss_chg > new_loss_chg);
}
}
/*!
* \brief update the split entry, replace it if e is better
* \param e candidate split solution
* \return whether the proposed split is better and can replace current split
*/
inline bool Update(const SplitEntry &e) {
if (this->NeedReplace(e.loss_chg, e.split_index())) {
this->loss_chg = e.loss_chg;
this->sindex = e.sindex;
this->split_value = e.split_value;
return true;
} else {
return false;
}
}
/*!
* \brief update the split entry, replace it if e is better
* \param new_loss_chg loss reduction of new candidate
* \param split_index feature index to split on
* \param new_split_value the split point
* \param default_left whether the missing value goes to left
* \return whether the proposed split is better and can replace current split
*/
inline bool Update(bst_float new_loss_chg, unsigned split_index,
float new_split_value, bool default_left) {
if (this->NeedReplace(new_loss_chg, split_index)) {
this->loss_chg = new_loss_chg;
if (default_left) split_index |= (1U << 31);
this->sindex = split_index;
this->split_value = new_split_value;
return true;
} else {
return false;
}
}
/*! \brief same as update, used by AllReduce*/
inline static void Reduce(SplitEntry &dst, const SplitEntry &src) { // NOLINT(*)
dst.Update(src);
}
/*!\return feature index to split on */
inline unsigned split_index(void) const {
return sindex & ((1U << 31) - 1U);
}
/*!\return whether missing value goes to left branch */
inline bool default_left(void) const {
return (sindex >> 31) != 0;
}
};
} // namespace tree
} // namespace xgboost
#endif // XGBOOST_TREE_PARAM_H_

35
src/tree/updater.cpp
View File

@@ -1,35 +0,0 @@
// Copyright 2014 by Contributors
#define _CRT_SECURE_NO_WARNINGS
#define _CRT_SECURE_NO_DEPRECATE
#define NOMINMAX
#include <cstring>
#include "./updater.h"
#include "./updater_prune-inl.hpp"
#include "./updater_refresh-inl.hpp"
#include "./updater_colmaker-inl.hpp"
#ifndef XGBOOST_STRICT_CXX98_
#include "./updater_sync-inl.hpp"
#include "./updater_distcol-inl.hpp"
#include "./updater_histmaker-inl.hpp"
#include "./updater_skmaker-inl.hpp"
#endif
namespace xgboost {
namespace tree {
IUpdater* CreateUpdater(const char *name) {
using namespace std;
if (!strcmp(name, "prune")) return new TreePruner();
if (!strcmp(name, "refresh")) return new TreeRefresher<GradStats>();
if (!strcmp(name, "grow_colmaker")) return new ColMaker<GradStats>();
#ifndef XGBOOST_STRICT_CXX98_
if (!strcmp(name, "sync")) return new TreeSyncher();
if (!strcmp(name, "grow_histmaker")) return new CQHistMaker<GradStats>();
if (!strcmp(name, "grow_skmaker")) return new SketchMaker();
if (!strcmp(name, "distcol")) return new DistColMaker<GradStats>();
#endif
utils::Error("unknown updater:%s", name);
return NULL;
}
} // namespace tree
} // namespace xgboost

63
src/tree/updater.h
View File

@@ -1,63 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file updater.h
* \brief interface to update the tree
* \author Tianqi Chen
*/
#ifndef XGBOOST_TREE_UPDATER_H_
#define XGBOOST_TREE_UPDATER_H_
#include <vector>
#include "../data.h"
#include "./model.h"
namespace xgboost {
namespace tree {
/*!
* \brief interface of tree update module, that performs update of a tree
*/
class IUpdater {
public:
/*!
* \brief set parameters from outside
* \param name name of the parameter
* \param val value of the parameter
*/
virtual void SetParam(const char *name, const char *val) = 0;
/*!
* \brief perform update to the tree models
* \param gpair the gradient pair statistics of the data
* \param p_fmat feature matrix that provide access to features
* \param info extra side information that may be need, such as root index
* \param trees references the trees to be updated, updater will change the content of trees
* note: all the trees in the vector are updated, with the same statistics,
* but maybe different random seeds, usually one tree is passed in at a time,
* there can be multiple trees when we train random forest style model
*/
virtual void Update(const std::vector<bst_gpair> &gpair,
IFMatrix *p_fmat,
const BoosterInfo &info,
const std::vector<RegTree*> &trees) = 0;
/*!
* \brief this is simply a function for optimizing performance
* this function asks the updater to return the leaf position of each instance in the p_fmat,
* if it is cached in the updater, if it is not available, return NULL
* \return array of leaf position of each instance in the last updated tree
*/
virtual const int* GetLeafPosition(void) const {
return NULL;
}
// destructor
virtual ~IUpdater(void) {}
};
/*!
* \brief create an updater based on name
* \param name name of updater
* \return return the updater instance
*/
IUpdater* CreateUpdater(const char *name);
} // namespace tree
} // namespace xgboost
#endif // XGBOOST_TREE_UPDATER_H_

427
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_

732
src/tree/updater_colmaker-inl.hpp
View File

@@ -1,732 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file updater_colmaker-inl.hpp
* \brief use columnwise update to construct a tree
* \author Tianqi Chen
*/
#ifndef XGBOOST_TREE_UPDATER_COLMAKER_INL_HPP_
#define XGBOOST_TREE_UPDATER_COLMAKER_INL_HPP_
#include <vector>
#include <cmath>
#include <algorithm>
#include "./param.h"
#include "./updater.h"
#include "../utils/omp.h"
#include "../utils/random.h"
namespace xgboost {
namespace tree {
/*! \brief column-wise update to construct a tree */
template<typename TStats>
class ColMaker: public IUpdater {
public:
virtual ~ColMaker(void) {}
// set training parameter
virtual void SetParam(const char *name, const char *val) {
param.SetParam(name, val);
}
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) {
Builder builder(param);
builder.Update(gpair, p_fmat, info, trees[i]);
}
param.learning_rate = lr;
}
protected:
// training parameter
TrainParam param;
// data structure
/*! \brief per thread x per node entry to store tmp data */
struct ThreadEntry {
/*! \brief statistics of data */
TStats stats;
/*! \brief extra statistics of data */
TStats stats_extra;
/*! \brief last feature value scanned */
float last_fvalue;
/*! \brief first feature value scanned */
float first_fvalue;
/*! \brief current best solution */
SplitEntry best;
// constructor
explicit ThreadEntry(const TrainParam &param)
: stats(param), stats_extra(param) {
}
};
struct NodeEntry {
/*! \brief statics for node entry */
TStats stats;
/*! \brief loss of this node, without split */
bst_float root_gain;
/*! \brief weight calculated related to current data */
float weight;
/*! \brief current best solution */
SplitEntry best;
// constructor
explicit NodeEntry(const TrainParam &param)
: stats(param), root_gain(0.0f), weight(0.0f){
}
};
// actual builder that runs the algorithm
struct Builder{
public:
// constructor
explicit Builder(const TrainParam &param) : param(param) {}
// update one tree, growing
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->InitNewNode(qexpand_, gpair, *p_fmat, info, *p_tree);
for (int depth = 0; depth < param.max_depth; ++depth) {
this->FindSplit(depth, qexpand_, gpair, p_fmat, info, p_tree);
this->ResetPosition(qexpand_, p_fmat, *p_tree);
this->UpdateQueueExpand(*p_tree, &qexpand_);
this->InitNewNode(qexpand_, gpair, *p_fmat, info, *p_tree);
// if nothing left to be expand, break
if (qexpand_.size() == 0) break;
}
// set all the rest expanding nodes to leaf
for (size_t i = 0; i < qexpand_.size(); ++i) {
const int nid = qexpand_[i];
(*p_tree)[nid].set_leaf(snode[nid].weight * param.learning_rate);
}
// remember auxiliary statistics in the tree node
for (int nid = 0; nid < p_tree->param.num_nodes; ++nid) {
p_tree->stat(nid).loss_chg = snode[nid].best.loss_chg;
p_tree->stat(nid).base_weight = snode[nid].weight;
p_tree->stat(nid).sum_hess = static_cast<float>(snode[nid].stats.sum_hess);
snode[nid].stats.SetLeafVec(param, p_tree->leafvec(nid));
}
}
protected:
// 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,
"ColMaker: can only grow new tree");
const std::vector<bst_uint> &rowset = fmat.buffered_rowset();
{
// setup position
position.resize(gpair.size());
if (root_index.size() == 0) {
for (size_t i = 0; i < rowset.size(); ++i) {
position[rowset[i]] = 0;
}
} else {
for (size_t i = 0; i < rowset.size(); ++i) {
const bst_uint ridx = rowset[i];
position[ridx] = root_index[ridx];
utils::Assert(root_index[ridx] < (unsigned)tree.param.num_roots,
"root index exceed setting");
}
}
// mark delete for the deleted datas
for (size_t i = 0; i < rowset.size(); ++i) {
const bst_uint ridx = rowset[i];
if (gpair[ridx].hess < 0.0f) position[ridx] = ~position[ridx];
}
// mark subsample
if (param.subsample < 1.0f) {
for (size_t i = 0; i < rowset.size(); ++i) {
const bst_uint ridx = rowset[i];
if (gpair[ridx].hess < 0.0f) continue;
if (random::SampleBinary(param.subsample) == 0) position[ridx] = ~position[ridx];
}
}
}
{
// initialize feature index
unsigned ncol = static_cast<unsigned>(fmat.NumCol());
for (unsigned i = 0; i < ncol; ++i) {
if (fmat.GetColSize(i) != 0) {
feat_index.push_back(i);
}
}
unsigned n = static_cast<unsigned>(param.colsample_bytree * feat_index.size());
random::Shuffle(feat_index);
utils::Check(n > 0, "colsample_bytree=%g is too small that no feature can be included",
param.colsample_bytree);
feat_index.resize(n);
}
{
// setup temp space for each thread
#pragma omp parallel
{
this->nthread = omp_get_num_threads();
}
// reserve a small space
stemp.clear();
stemp.resize(this->nthread, std::vector<ThreadEntry>());
for (size_t i = 0; i < stemp.size(); ++i) {
stemp[i].clear(); stemp[i].reserve(256);
}
snode.reserve(256);
}
{
// expand query
qexpand_.reserve(256); qexpand_.clear();
for (int i = 0; i < tree.param.num_roots; ++i) {
qexpand_.push_back(i);
}
}
}
/*!
* \brief initialize the base_weight, root_gain,
* and NodeEntry for all the new nodes in qexpand
*/
inline void InitNewNode(const std::vector<int> &qexpand,
const std::vector<bst_gpair> &gpair,
const IFMatrix &fmat,
const BoosterInfo &info,
const RegTree &tree) {
{
// setup statistics space for each tree node
for (size_t i = 0; i < stemp.size(); ++i) {
stemp[i].resize(tree.param.num_nodes, ThreadEntry(param));
}
snode.resize(tree.param.num_nodes, NodeEntry(param));
}
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 tid = omp_get_thread_num();
if (position[ridx] < 0) continue;
stemp[tid][position[ridx]].stats.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 stats(param);
for (size_t tid = 0; tid < stemp.size(); ++tid) {
stats.Add(stemp[tid][nid].stats);
}
// update node statistics
snode[nid].stats = stats;
snode[nid].root_gain = static_cast<float>(stats.CalcGain(param));
snode[nid].weight = static_cast<float>(stats.CalcWeight(param));
}
}
/*! \brief update queue expand add in new leaves */
inline void UpdateQueueExpand(const RegTree &tree, std::vector<int> *p_qexpand) {
std::vector<int> &qexpand = *p_qexpand;
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;
}
// parallel find the best split of current fid
// this function does not support nested functions
inline void ParallelFindSplit(const ColBatch::Inst &col,
bst_uint fid,
const IFMatrix &fmat,
const std::vector<bst_gpair> &gpair,
const BoosterInfo &info) {
const bool ind = col.length != 0 && col.data[0].fvalue == col.data[col.length - 1].fvalue;
bool need_forward = param.need_forward_search(fmat.GetColDensity(fid), ind);
bool need_backward = param.need_backward_search(fmat.GetColDensity(fid), ind);
const std::vector<int> &qexpand = qexpand_;
#pragma omp parallel
{
const int tid = omp_get_thread_num();
std::vector<ThreadEntry> &temp = stemp[tid];
// cleanup temp statistics
for (size_t j = 0; j < qexpand.size(); ++j) {
temp[qexpand[j]].stats.Clear();
}
nthread = omp_get_num_threads();
bst_uint step = (col.length + nthread - 1) / nthread;
bst_uint end = std::min(col.length, step * (tid + 1));
for (bst_uint i = tid * step; i < end; ++i) {
const bst_uint ridx = col[i].index;
const int nid = position[ridx];
if (nid < 0) continue;
const float fvalue = col[i].fvalue;
if (temp[nid].stats.Empty()) {
temp[nid].first_fvalue = fvalue;
}
temp[nid].stats.Add(gpair, info, ridx);
temp[nid].last_fvalue = fvalue;
}
}
// start collecting the partial sum statistics
bst_omp_uint nnode = static_cast<bst_omp_uint>(qexpand.size());
#pragma omp parallel for schedule(static)
for (bst_omp_uint j = 0; j < nnode; ++j) {
const int nid = qexpand[j];
TStats sum(param), tmp(param), c(param);
for (int tid = 0; tid < nthread; ++tid) {
tmp = stemp[tid][nid].stats;
stemp[tid][nid].stats = sum;
sum.Add(tmp);
if (tid != 0) {
std::swap(stemp[tid - 1][nid].last_fvalue, stemp[tid][nid].first_fvalue);
}
}
for (int tid = 0; tid < nthread; ++tid) {
stemp[tid][nid].stats_extra = sum;
ThreadEntry &e = stemp[tid][nid];
float fsplit;
if (tid != 0) {
if (std::abs(stemp[tid - 1][nid].last_fvalue - e.first_fvalue) > rt_2eps) {
fsplit = (stemp[tid - 1][nid].last_fvalue - e.first_fvalue) * 0.5f;
} else {
continue;
}
} else {
fsplit = e.first_fvalue - rt_eps;
}
if (need_forward && tid != 0) {
c.SetSubstract(snode[nid].stats, e.stats);
if (c.sum_hess >= param.min_child_weight &&
e.stats.sum_hess >= param.min_child_weight) {
bst_float loss_chg = static_cast<bst_float>(e.stats.CalcGain(param) +
c.CalcGain(param) - snode[nid].root_gain);
e.best.Update(loss_chg, fid, fsplit, false);
}
}
if (need_backward) {
tmp.SetSubstract(sum, e.stats);
c.SetSubstract(snode[nid].stats, tmp);
if (c.sum_hess >= param.min_child_weight &&
tmp.sum_hess >= param.min_child_weight) {
bst_float loss_chg = static_cast<bst_float>(tmp.CalcGain(param) +
c.CalcGain(param) - snode[nid].root_gain);
e.best.Update(loss_chg, fid, fsplit, true);
}
}
}
if (need_backward) {
tmp = sum;
ThreadEntry &e = stemp[nthread-1][nid];
c.SetSubstract(snode[nid].stats, tmp);
if (c.sum_hess >= param.min_child_weight &&
tmp.sum_hess >= param.min_child_weight) {
bst_float loss_chg = static_cast<bst_float>(tmp.CalcGain(param) +
c.CalcGain(param) - snode[nid].root_gain);
e.best.Update(loss_chg, fid, e.last_fvalue + rt_eps, true);
}
}
}
// rescan, generate candidate split
#pragma omp parallel
{
TStats c(param), cright(param);
const int tid = omp_get_thread_num();
std::vector<ThreadEntry> &temp = stemp[tid];
nthread = static_cast<bst_uint>(omp_get_num_threads());
bst_uint step = (col.length + nthread - 1) / nthread;
bst_uint end = std::min(col.length, step * (tid + 1));
for (bst_uint i = tid * step; i < end; ++i) {
const bst_uint ridx = col[i].index;
const int nid = position[ridx];
if (nid < 0) continue;
const float fvalue = col[i].fvalue;
// get the statistics of nid
ThreadEntry &e = temp[nid];
if (e.stats.Empty()) {
e.stats.Add(gpair, info, ridx);
e.first_fvalue = fvalue;
} else {
// forward default right
if (std::abs(fvalue - e.first_fvalue) > rt_2eps) {
if (need_forward) {
c.SetSubstract(snode[nid].stats, e.stats);
if (c.sum_hess >= param.min_child_weight &&
e.stats.sum_hess >= param.min_child_weight) {
bst_float loss_chg = static_cast<bst_float>(e.stats.CalcGain(param) +
c.CalcGain(param) -
snode[nid].root_gain);
e.best.Update(loss_chg, fid, (fvalue + e.first_fvalue) * 0.5f, false);
}
}
if (need_backward) {
cright.SetSubstract(e.stats_extra, e.stats);
c.SetSubstract(snode[nid].stats, cright);
if (c.sum_hess >= param.min_child_weight &&
cright.sum_hess >= param.min_child_weight) {
bst_float loss_chg = static_cast<bst_float>(cright.CalcGain(param) +
c.CalcGain(param) -
snode[nid].root_gain);
e.best.Update(loss_chg, fid, (fvalue + e.first_fvalue) * 0.5f, true);
}
}
}
e.stats.Add(gpair, info, ridx);
e.first_fvalue = fvalue;
}
}
}
}
// update enumeration solution
inline void UpdateEnumeration(int nid, bst_gpair gstats,
float fvalue, int d_step, bst_uint fid,
TStats &c, std::vector<ThreadEntry> &temp) { // NOLINT(*)
// get the statistics of nid
ThreadEntry &e = temp[nid];
// test if first hit, this is fine, because we set 0 during init
if (e.stats.Empty()) {
e.stats.Add(gstats);
e.last_fvalue = fvalue;
} else {
// try to find a split
if (std::abs(fvalue - e.last_fvalue) > rt_2eps &&
e.stats.sum_hess >= param.min_child_weight) {
c.SetSubstract(snode[nid].stats, e.stats);
if (c.sum_hess >= param.min_child_weight) {
bst_float loss_chg = static_cast<bst_float>(e.stats.CalcGain(param) +
c.CalcGain(param) - snode[nid].root_gain);
e.best.Update(loss_chg, fid, (fvalue + e.last_fvalue) * 0.5f, d_step == -1);
}
}
// update the statistics
e.stats.Add(gstats);
e.last_fvalue = fvalue;
}
}
// same as EnumerateSplit, with cacheline prefetch optimization
inline void EnumerateSplitCacheOpt(const ColBatch::Entry *begin,
const ColBatch::Entry *end,
int d_step,
bst_uint fid,
const std::vector<bst_gpair> &gpair,
std::vector<ThreadEntry> &temp) { // NOLINT(*)
const std::vector<int> &qexpand = qexpand_;
// clear all the temp statistics
for (size_t j = 0; j < qexpand.size(); ++j) {
temp[qexpand[j]].stats.Clear();
}
// left statistics
TStats c(param);
// local cache buffer for position and gradient pair
const int kBuffer = 32;
int buf_position[kBuffer];
bst_gpair buf_gpair[kBuffer];
// aligned ending position
const ColBatch::Entry *align_end;
if (d_step > 0) {
align_end = begin + (end - begin) / kBuffer * kBuffer;
} else {
align_end = begin - (begin - end) / kBuffer * kBuffer;
}
int i;
const ColBatch::Entry *it;
const int align_step = d_step * kBuffer;
// internal cached loop
for (it = begin; it != align_end; it += align_step) {
const ColBatch::Entry *p;
for (i = 0, p = it; i < kBuffer; ++i, p += d_step) {
buf_position[i] = position[p->index];
buf_gpair[i] = gpair[p->index];
}
for (i = 0, p = it; i < kBuffer; ++i, p += d_step) {
const int nid = buf_position[i];
if (nid < 0) continue;
this->UpdateEnumeration(nid, buf_gpair[i],
p->fvalue, d_step,
fid, c, temp);
}
}
// finish up the ending piece
for (it = align_end, i = 0; it != end; ++i, it += d_step) {
buf_position[i] = position[it->index];
buf_gpair[i] = gpair[it->index];
}
for (it = align_end, i = 0; it != end; ++i, it += d_step) {
const int nid = buf_position[i];
if (nid < 0) continue;
this->UpdateEnumeration(nid, buf_gpair[i],
it->fvalue, d_step,
fid, c, temp);
}
// finish updating all statistics, check if it is possible to include all sum statistics
for (size_t i = 0; i < qexpand.size(); ++i) {
const int nid = qexpand[i];
ThreadEntry &e = temp[nid];
c.SetSubstract(snode[nid].stats, e.stats);
if (e.stats.sum_hess >= param.min_child_weight && c.sum_hess >= param.min_child_weight) {
bst_float loss_chg = static_cast<bst_float>(e.stats.CalcGain(param) +
c.CalcGain(param) - snode[nid].root_gain);
const float gap = std::abs(e.last_fvalue) + rt_eps;
const float delta = d_step == +1 ? gap: -gap;
e.best.Update(loss_chg, fid, e.last_fvalue + delta, d_step == -1);
}
}
}
// enumerate the split values of specific feature
inline void EnumerateSplit(const ColBatch::Entry *begin,
const ColBatch::Entry *end,
int d_step,
bst_uint fid,
const std::vector<bst_gpair> &gpair,
const BoosterInfo &info,
std::vector<ThreadEntry> &temp) { // NOLINT(*)
// use cacheline aware optimization
if (TStats::kSimpleStats != 0 && param.cache_opt != 0) {
EnumerateSplitCacheOpt(begin, end, d_step, fid, gpair, temp);
return;
}
const std::vector<int> &qexpand = qexpand_;
// clear all the temp statistics
for (size_t j = 0; j < qexpand.size(); ++j) {
temp[qexpand[j]].stats.Clear();
}
// left statistics
TStats c(param);
for (const ColBatch::Entry *it = begin; it != end; it += d_step) {
const bst_uint ridx = it->index;
const int nid = position[ridx];
if (nid < 0) continue;
// start working
const float fvalue = it->fvalue;
// get the statistics of nid
ThreadEntry &e = temp[nid];
// test if first hit, this is fine, because we set 0 during init
if (e.stats.Empty()) {
e.stats.Add(gpair, info, ridx);
e.last_fvalue = fvalue;
} else {
// try to find a split
if (std::abs(fvalue - e.last_fvalue) > rt_2eps &&
e.stats.sum_hess >= param.min_child_weight) {
c.SetSubstract(snode[nid].stats, e.stats);
if (c.sum_hess >= param.min_child_weight) {
bst_float loss_chg = static_cast<bst_float>(e.stats.CalcGain(param) +
c.CalcGain(param) - snode[nid].root_gain);
e.best.Update(loss_chg, fid, (fvalue + e.last_fvalue) * 0.5f, d_step == -1);
}
}
// update the statistics
e.stats.Add(gpair, info, ridx);
e.last_fvalue = fvalue;
}
}
// finish updating all statistics, check if it is possible to include all sum statistics
for (size_t i = 0; i < qexpand.size(); ++i) {
const int nid = qexpand[i];
ThreadEntry &e = temp[nid];
c.SetSubstract(snode[nid].stats, e.stats);
if (e.stats.sum_hess >= param.min_child_weight && c.sum_hess >= param.min_child_weight) {
bst_float loss_chg = static_cast<bst_float>(e.stats.CalcGain(param) +
c.CalcGain(param) - snode[nid].root_gain);
const float gap = std::abs(e.last_fvalue) + rt_eps;
const float delta = d_step == +1 ? gap: -gap;
e.best.Update(loss_chg, fid, e.last_fvalue + delta, d_step == -1);
}
}
}
// update the solution candidate
virtual void UpdateSolution(const ColBatch &batch,
const std::vector<bst_gpair> &gpair,
const IFMatrix &fmat,
const BoosterInfo &info) {
// start enumeration
const bst_omp_uint nsize = static_cast<bst_omp_uint>(batch.size);
#if defined(_OPENMP)
const int batch_size = std::max(static_cast<int>(nsize / this->nthread / 32), 1);
#endif
int poption = param.parallel_option;
if (poption == 2) {
poption = static_cast<int>(nsize) * 2 < nthread ? 1 : 0;
}
if (poption == 0) {
#pragma omp parallel for schedule(dynamic, batch_size)
for (bst_omp_uint i = 0; i < nsize; ++i) {
const bst_uint fid = batch.col_index[i];
const int tid = omp_get_thread_num();
const ColBatch::Inst c = batch[i];
const bool ind = c.length != 0 && c.data[0].fvalue == c.data[c.length - 1].fvalue;
if (param.need_forward_search(fmat.GetColDensity(fid), ind)) {
this->EnumerateSplit(c.data, c.data + c.length, +1,
fid, gpair, info, stemp[tid]);
}
if (param.need_backward_search(fmat.GetColDensity(fid), ind)) {
this->EnumerateSplit(c.data + c.length - 1, c.data - 1, -1,
fid, gpair, info, stemp[tid]);
}
}
} else {
for (bst_omp_uint i = 0; i < nsize; ++i) {
this->ParallelFindSplit(batch[i], batch.col_index[i],
fmat, gpair, info);
}
}
}
// find splits at current level, do split per level
inline void FindSplit(int depth,
const std::vector<int> &qexpand,
const std::vector<bst_gpair> &gpair,
IFMatrix *p_fmat,
const BoosterInfo &info,
RegTree *p_tree) {
std::vector<bst_uint> feat_set = feat_index;
if (param.colsample_bylevel != 1.0f) {
random::Shuffle(feat_set);
unsigned n = static_cast<unsigned>(param.colsample_bylevel * feat_index.size());
utils::Check(n > 0, "colsample_bylevel is too small that no feature can be included");
feat_set.resize(n);
}
utils::IIterator<ColBatch> *iter = p_fmat->ColIterator(feat_set);
while (iter->Next()) {
this->UpdateSolution(iter->Value(), gpair, *p_fmat, info);
}
// after this each thread's stemp will get the best candidates, aggregate results
this->SyncBestSolution(qexpand);
// get the best result, we can synchronize the solution
for (size_t i = 0; i < qexpand.size(); ++i) {
const int nid = qexpand[i];
NodeEntry &e = snode[nid];
// now we know the solution in snode[nid], set split
if (e.best.loss_chg > rt_eps) {
p_tree->AddChilds(nid);
(*p_tree)[nid].set_split(e.best.split_index(), e.best.split_value, e.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(e.weight * param.learning_rate);
}
}
}
// reset position of each data points after split is created in the tree
inline void ResetPosition(const std::vector<int> &qexpand,
IFMatrix *p_fmat, const RegTree &tree) {
// set the positions in the nondefault
this->SetNonDefaultPosition(qexpand, 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];
if (ridx >= position.size()) {
utils::Printf("ridx exceed bound\n");
}
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());
}
}
}
}
// customization part
// synchronize the best solution of each node
virtual void SyncBestSolution(const std::vector<int> &qexpand) {
for (size_t i = 0; i < qexpand.size(); ++i) {
const int nid = qexpand[i];
NodeEntry &e = snode[nid];
for (int tid = 0; tid < this->nthread; ++tid) {
e.best.Update(stemp[tid][nid].best);
}
}
}
virtual void SetNonDefaultPosition(const std::vector<int> &qexpand,
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 < qexpand.size(); ++i) {
const int nid = qexpand[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 int nid = this->DecodePosition(ridx);
const float fvalue = col[j].fvalue;
// 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());
}
}
}
}
}
}
// utils to get/set position, with encoded format
// 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;
}
}
// --data fields--
const TrainParam &param;
// number of omp thread used during training
int nthread;
// Per feature: shuffle index of each feature index
std::vector<bst_uint> feat_index;
// Instance Data: current node position in the tree of each instance
std::vector<int> position;
// PerThread x PerTreeNode: statistics for per thread construction
std::vector< std::vector<ThreadEntry> > stemp;
/*! \brief TreeNode Data: statistics for each constructed node */
std::vector<NodeEntry> snode;
/*! \brief queue of nodes to be expanded */
std::vector<int> qexpand_;
};
};
} // namespace tree
} // namespace xgboost
#endif // XGBOOST_TREE_UPDATER_COLMAKER_INL_HPP_

175
src/tree/updater_distcol-inl.hpp
View File

@@ -1,175 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file updater_distcol-inl.hpp
* \brief beta distributed version that takes a sub-column
* and construct a tree
* \author Tianqi Chen
*/
#ifndef XGBOOST_TREE_UPDATER_DISTCOL_INL_HPP_
#define XGBOOST_TREE_UPDATER_DISTCOL_INL_HPP_
#include <vector>
#include <algorithm>
#include "../sync/sync.h"
#include "../utils/bitmap.h"
#include "../utils/io.h"
#include "./updater_colmaker-inl.hpp"
#include "./updater_prune-inl.hpp"
namespace xgboost {
namespace tree {
template<typename TStats>
class DistColMaker : public ColMaker<TStats> {
public:
DistColMaker(void) : builder(param) {}
virtual ~DistColMaker(void) {}
// set training parameter
virtual void SetParam(const char *name, const char *val) {
param.SetParam(name, val);
pruner.SetParam(name, val);
}
virtual void Update(const std::vector<bst_gpair> &gpair,
IFMatrix *p_fmat,
const BoosterInfo &info,
const std::vector<RegTree*> &trees) {
TStats::CheckInfo(info);
utils::Check(trees.size() == 1, "DistColMaker: only support one tree at a time");
// build the tree
builder.Update(gpair, p_fmat, info, trees[0]);
//// prune the tree, note that pruner will sync the tree
pruner.Update(gpair, p_fmat, info, trees);
// update position after the tree is pruned
builder.UpdatePosition(p_fmat, *trees[0]);
}
virtual const int* GetLeafPosition(void) const {
return builder.GetLeafPosition();
}
private:
struct Builder : public ColMaker<TStats>::Builder {
public:
explicit Builder(const TrainParam &param)
: ColMaker<TStats>::Builder(param) {
}
inline void UpdatePosition(IFMatrix *p_fmat, const RegTree &tree) {
const std::vector<bst_uint> &rowset = p_fmat->buffered_rowset();
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];
int nid = this->DecodePosition(ridx);
while (tree[nid].is_deleted()) {
nid = tree[nid].parent();
utils::Assert(nid >=0, "distributed learning error");
}
this->position[ridx] = nid;
}
}
virtual const int* GetLeafPosition(void) const {
return BeginPtr(this->position);
}
protected:
virtual void SetNonDefaultPosition(const std::vector<int> &qexpand,
IFMatrix *p_fmat, const RegTree &tree) {
// step 2, classify the non-default data into right places
std::vector<unsigned> fsplits;
for (size_t i = 0; i < qexpand.size(); ++i) {
const int nid = qexpand[i];
if (!tree[nid].is_leaf()) {
fsplits.push_back(tree[nid].split_index());
}
}
// get the candidate split index
std::sort(fsplits.begin(), fsplits.end());
fsplits.resize(std::unique(fsplits.begin(), fsplits.end()) - fsplits.begin());
while (fsplits.size() != 0 && fsplits.back() >= p_fmat->NumCol()) {
fsplits.pop_back();
}
// bitmap is only word concurrent, set to bool first
{
bst_omp_uint ndata = static_cast<bst_omp_uint>(this->position.size());
boolmap.resize(ndata);
#pragma omp parallel for schedule(static)
for (bst_omp_uint j = 0; j < ndata; ++j) {
boolmap[j] = 0;
}
}
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);
if (!tree[nid].is_leaf() && tree[nid].split_index() == fid) {
if (fvalue < tree[nid].split_cond()) {
if (!tree[nid].default_left()) boolmap[ridx] = 1;
} else {
if (tree[nid].default_left()) boolmap[ridx] = 1;
}
}
}
}
}
bitmap.InitFromBool(boolmap);
// communicate bitmap
rabit::Allreduce<rabit::op::BitOR>(BeginPtr(bitmap.data), bitmap.data.size());
const std::vector<bst_uint> &rowset = p_fmat->buffered_rowset();
// get the new 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 = this->DecodePosition(ridx);
if (bitmap.Get(ridx)) {
utils::Assert(!tree[nid].is_leaf(), "inconsistent reduce information");
if (tree[nid].default_left()) {
this->SetEncodePosition(ridx, tree[nid].cright());
} else {
this->SetEncodePosition(ridx, tree[nid].cleft());
}
}
}
}
// synchronize the best solution of each node
virtual void SyncBestSolution(const std::vector<int> &qexpand) {
std::vector<SplitEntry> vec;
for (size_t i = 0; i < qexpand.size(); ++i) {
const int nid = qexpand[i];
for (int tid = 0; tid < this->nthread; ++tid) {
this->snode[nid].best.Update(this->stemp[tid][nid].best);
}
vec.push_back(this->snode[nid].best);
}
// TODO(tqchen) lazy version
// communicate best solution
reducer.Allreduce(BeginPtr(vec), vec.size());
// assign solution back
for (size_t i = 0; i < qexpand.size(); ++i) {
const int nid = qexpand[i];
this->snode[nid].best = vec[i];
}
}
private:
utils::BitMap bitmap;
std::vector<int> boolmap;
rabit::Reducer<SplitEntry, SplitEntry::Reduce> reducer;
};
// we directly introduce pruner here
TreePruner pruner;
// training parameter
TrainParam param;
// pointer to the builder
Builder builder;
};
} // namespace tree
} // namespace xgboost
#endif // XGBOOST_TREE_UPDATER_DISTCOL_INL_HPP_

769
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
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
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
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_

56
src/tree/updater_sync-inl.hpp
View File

@@ -1,56 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file updater_sync-inl.hpp
* \brief synchronize the tree in all distributed nodes
* \author Tianqi Chen
*/
#ifndef XGBOOST_TREE_UPDATER_SYNC_INL_HPP_
#define XGBOOST_TREE_UPDATER_SYNC_INL_HPP_
#include <vector>
#include <string>
#include <limits>
#include "../sync/sync.h"
#include "./updater.h"
namespace xgboost {
namespace tree {
/*!
* \brief syncher that synchronize the tree in all distributed nodes
* can implement various strategies, so far it is always set to node 0's tree
*/
class TreeSyncher: public IUpdater {
public:
virtual ~TreeSyncher(void) {}
virtual void SetParam(const char *name, const char *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) {
this->SyncTrees(trees);
}
private:
// synchronize the trees in different nodes, take tree from rank 0
inline void SyncTrees(const std::vector<RegTree *> &trees) {
if (rabit::GetWorldSize() == 1) return;
std::string s_model;
utils::MemoryBufferStream fs(&s_model);
int rank = rabit::GetRank();
if (rank == 0) {
for (size_t i = 0; i < trees.size(); ++i) {
trees[i]->SaveModel(fs);
}
}
fs.Seek(0);
rabit::Broadcast(&s_model, 0);
for (size_t i = 0; i < trees.size(); ++i) {
trees[i]->LoadModel(fs);
}
}
};
} // namespace tree
} // namespace xgboost
#endif // XGBOOST_TREE_UPDATER_SYNC_INL_HPP_

267
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_

68
src/utils/bitmap.h
View File

@@ -1,68 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file bitmap.h
* \brief a simple implement of bitmap
* NOTE: bitmap is only threadsafe per word access, remember this when using bitmap
* \author Tianqi Chen
*/
#ifndef XGBOOST_UTILS_BITMAP_H_
#define XGBOOST_UTILS_BITMAP_H_
#include <vector>
#include "./utils.h"
#include "./omp.h"
namespace xgboost {
namespace utils {
/*! \brief bit map that contains set of bit indicators */
struct BitMap {
/*! \brief internal data structure */
std::vector<uint32_t> data;
/*!
* \brief resize the bitmap to be certain size
* \param size the size of bitmap
*/
inline void Resize(size_t size) {
data.resize((size + 31U) >> 5, 0);
}
/*!
* \brief query the i-th position of bitmap
* \param i the position in
*/
inline bool Get(size_t i) const {
return (data[i >> 5] >> (i & 31U)) & 1U;
}
/*!
* \brief set i-th position to true
* \param i position index
*/
inline void SetTrue(size_t i) {
data[i >> 5] |= (1 << (i & 31U));
}
/*! \brief initialize the value of bit map from vector of bool*/
inline void InitFromBool(const std::vector<int> &vec) {
this->Resize(vec.size());
// parallel over the full cases
bst_omp_uint nsize = static_cast<bst_omp_uint>(vec.size() / 32);
#pragma omp parallel for schedule(static)
for (bst_omp_uint i = 0; i < nsize; ++i) {
uint32_t res = 0;
for (int k = 0; k < 32; ++k) {
int bit = vec[(i << 5) | k];
res |= (bit << k);
}
data[i] = res;
}
if (nsize != vec.size()) data.back() = 0;
for (size_t i = nsize; i < vec.size(); ++i) {
if (vec[i]) this->SetTrue(i);
}
}
/*! \brief clear the bitmap, set all places to false */
inline void Clear(void) {
std::fill(data.begin(), data.end(), 0U);
}
};
} // namespace utils
} // namespace xgboost
#endif // XGBOOST_UTILS_BITMAP_H_

194
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_

83
src/utils/fmap.h
View File

@@ -1,83 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file fmap.h
* \brief helper class that holds the feature names and interpretations
* \author Tianqi Chen
*/
#ifndef XGBOOST_UTILS_FMAP_H_
#define XGBOOST_UTILS_FMAP_H_
#include <vector>
#include <string>
#include <cstring>
#include "./utils.h"
namespace xgboost {
namespace utils {
/*! \brief helper class that holds the feature names and interpretations */
class FeatMap {
public:
enum Type {
kIndicator = 0,
kQuantitive = 1,
kInteger = 2,
kFloat = 3
};
// function definitions
/*! \brief load feature map from text format */
inline void LoadText(const char *fname) {
std::FILE *fi = utils::FopenCheck(fname, "r");
this->LoadText(fi);
std::fclose(fi);
}
/*! \brief load feature map from text format */
inline void LoadText(std::FILE *fi) {
int fid;
char fname[1256], ftype[1256];
while (std::fscanf(fi, "%d\t%[^\t]\t%s\n", &fid, fname, ftype) == 3) {
this->PushBack(fid, fname, ftype);
}
}
/*!\brief push back feature map */
inline void PushBack(int fid, const char *fname, const char *ftype) {
utils::Check(fid == static_cast<int>(names_.size()), "invalid fmap format");
names_.push_back(std::string(fname));
types_.push_back(GetType(ftype));
}
inline void Clear(void) {
names_.clear(); types_.clear();
}
/*! \brief number of known features */
size_t size(void) const {
return names_.size();
}
/*! \brief return name of specific feature */
const char* name(size_t idx) const {
utils::Assert(idx < names_.size(), "utils::FMap::name feature index exceed bound");
return names_[idx].c_str();
}
/*! \brief return type of specific feature */
const Type& type(size_t idx) const {
utils::Assert(idx < names_.size(), "utils::FMap::type feature index exceed bound");
return types_[idx];
}
private:
inline static Type GetType(const char *tname) {
using namespace std;
if (!strcmp("i", tname)) return kIndicator;
if (!strcmp("q", tname)) return kQuantitive;
if (!strcmp("int", tname)) return kInteger;
if (!strcmp("float", tname)) return kFloat;
utils::Error("unknown feature type, use i for indicator and q for quantity");
return kIndicator;
}
/*! \brief name of the feature */
std::vector<std::string> names_;
/*! \brief type of the feature */
std::vector<Type> types_;
};
} // namespace utils
} // namespace xgboost
#endif // XGBOOST_UTILS_FMAP_H_

114
src/utils/group_data.h
View File

@@ -1,114 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file group_data.h
* \brief this file defines utils to group data by integer keys
* Input: given input sequence (key,value), (k1,v1), (k2,v2)
* Ouptupt: an array of values data = [v1,v2,v3 .. vn]
* and a group pointer ptr,
* data[ptr[k]:ptr[k+1]] contains values that corresponds to key k
*
* This can be used to construct CSR/CSC matrix from un-ordered input
* The major algorithm is a two pass linear scan algorithm that requires two pass scan over the data
* \author Tianqi Chen
*/
#ifndef XGBOOST_UTILS_GROUP_DATA_H_
#define XGBOOST_UTILS_GROUP_DATA_H_
#include <vector>
namespace xgboost {
namespace utils {
/*!
* \brief multi-thread version of group builder
* \tparam ValueType type of entries in the sparse matrix
* \tparam SizeType type of the index range holder
*/
template<typename ValueType, typename SizeType = size_t>
struct ParallelGroupBuilder {
public:
// parallel group builder of data
ParallelGroupBuilder(std::vector<SizeType> *p_rptr,
std::vector<ValueType> *p_data)
: rptr(*p_rptr), data(*p_data), thread_rptr(tmp_thread_rptr) {
}
ParallelGroupBuilder(std::vector<SizeType> *p_rptr,
std::vector<ValueType> *p_data,
std::vector< std::vector<SizeType> > *p_thread_rptr)
: rptr(*p_rptr), data(*p_data), thread_rptr(*p_thread_rptr) {
}
public:
/*!
* \brief step 1: initialize the helper, with hint of number keys
* and thread used in the construction
* \param nkeys number of keys in the matrix, can be smaller than expected
* \param nthread number of thread that will be used in construction
*/
inline void InitBudget(size_t nkeys, int nthread) {
thread_rptr.resize(nthread);
for (size_t i = 0; i < thread_rptr.size(); ++i) {
thread_rptr[i].resize(nkeys);
std::fill(thread_rptr[i].begin(), thread_rptr[i].end(), 0);
}
}
/*!
* \brief step 2: add budget to each key
* \param key the key
* \param threadid the id of thread that calls this function
* \param nelem number of element budget add to this row
*/
inline void AddBudget(size_t key, int threadid, SizeType nelem = 1) {
std::vector<SizeType> &trptr = thread_rptr[threadid];
if (trptr.size() < key + 1) {
trptr.resize(key + 1, 0);
}
trptr[key] += nelem;
}
/*! \brief step 3: initialize the necessary storage */
inline void InitStorage(void) {
// set rptr to correct size
for (size_t tid = 0; tid < thread_rptr.size(); ++tid) {
if (rptr.size() <= thread_rptr[tid].size()) {
rptr.resize(thread_rptr[tid].size() + 1);
}
}
// initialize rptr to be beginning of each segment
size_t start = 0;
for (size_t i = 0; i + 1 < rptr.size(); ++i) {
for (size_t tid = 0; tid < thread_rptr.size(); ++tid) {
std::vector<SizeType> &trptr = thread_rptr[tid];
if (i < trptr.size()) {
size_t ncnt = trptr[i];
trptr[i] = start;
start += ncnt;
}
}
rptr[i + 1] = start;
}
data.resize(start);
}
/*!
* \brief step 4: add data to the allocated space,
* the calls to this function should be exactly match previous call to AddBudget
*
* \param key the key of
* \param threadid the id of thread that calls this function
*/
inline void Push(size_t key, ValueType value, int threadid) {
SizeType &rp = thread_rptr[threadid][key];
data[rp++] = value;
}
private:
/*! \brief pointer to the beginning and end of each continuous key */
std::vector<SizeType> &rptr;
/*! \brief index of nonzero entries in each row */
std::vector<ValueType> &data;
/*! \brief thread local data structure */
std::vector< std::vector<SizeType> > &thread_rptr;
/*! \brief local temp thread ptr, use this if not specified by the constructor */
std::vector< std::vector<SizeType> > tmp_thread_rptr;
};
} // namespace utils
} // namespace xgboost
#endif // XGBOOST_UTILS_GROUP_DATA_H_

59
src/utils/io.h
View File

@@ -1,59 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file io.h
* \brief general stream interface for serialization, I/O
* \author Tianqi Chen
*/
#ifndef XGBOOST_UTILS_IO_H_
#define XGBOOST_UTILS_IO_H_
#include <cstdio>
#include <vector>
#include <string>
#include <cstring>
#include "./utils.h"
#include "../sync/sync.h"
namespace xgboost {
namespace utils {
// reuse the definitions of streams
typedef rabit::Stream IStream;
typedef rabit::utils::SeekStream ISeekStream;
typedef rabit::utils::MemoryFixSizeBuffer MemoryFixSizeBuffer;
typedef rabit::utils::MemoryBufferStream MemoryBufferStream;
/*! \brief implementation of file i/o stream */
class FileStream : public ISeekStream {
public:
explicit FileStream(std::FILE *fp) : fp(fp) {}
FileStream(void) {
this->fp = NULL;
}
virtual size_t Read(void *ptr, size_t size) {
return std::fread(ptr, size, 1, fp);
}
virtual void Write(const void *ptr, size_t size) {
Check(std::fwrite(ptr, size, 1, fp) == 1, "FileStream::Write: fwrite error!");
}
virtual void Seek(size_t pos) {
std::fseek(fp, static_cast<long>(pos), SEEK_SET); // NOLINT(*)
}
virtual size_t Tell(void) {
return std::ftell(fp);
}
virtual bool AtEnd(void) const {
return std::feof(fp) != 0;
}
inline void Close(void) {
if (fp != NULL) {
std::fclose(fp); fp = NULL;
}
}
private:
std::FILE *fp;
};
} // namespace utils
} // namespace xgboost
#include "./base64-inl.h"
#endif // XGBOOST_UTILS_IO_H_

42
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_

45
src/utils/math.h
View File

@@ -1,45 +0,0 @@
/*!
* Copyright 2014 by Contributors
* \file math.h
* \brief support additional math
* \author Tianqi Chen
*/
#ifndef XGBOOST_UTILS_MATH_H_
#define XGBOOST_UTILS_MATH_H_
#include <cmath>
namespace xgboost {
namespace utils {
#ifdef XGBOOST_STRICT_CXX98_
// check nan
bool CheckNAN(double v);
double LogGamma(double v);
#else
template<typename T>
inline bool CheckNAN(T v) {
#ifdef _MSC_VER
return (_isnan(v) != 0);
#else
return isnan(v);
#endif
}
template<typename T>
inline T LogGamma(T v) {
#ifdef _MSC_VER
#if _MSC_VER >= 1800
return lgamma(v);
#else
#pragma message("Warning: lgamma function was not available until VS2013"\
", poisson regression will be disabled")
utils::Error("lgamma function was not available until VS2013");
return static_cast<T>(1.0);
#endif
#else
return lgamma(v);
#endif
}
#endif
} // namespace utils
} // namespace xgboost
#endif // XGBOOST_UTILS_MATH_H_

34
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
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
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
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
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
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_

335
src/xgboost_main.cpp
View File

@@ -1,335 +0,0 @@
// Copyright 2014 by Contributors
#define _CRT_SECURE_NO_WARNINGS
#define _CRT_SECURE_NO_DEPRECATE
#define NOMINMAX
#include <ctime>
#include <string>
#include <cstring>
#include <vector>
#include "./sync/sync.h"
#include "./io/io.h"
#include "./utils/utils.h"
#include "./utils/config.h"
#include "./learner/learner-inl.hpp"
namespace xgboost {
/*!
* \brief wrapping the training process
*/
class BoostLearnTask {
public:
inline int Run(int argc, char *argv[]) {
if (argc < 2) {
printf("Usage: <config>\n");
return 0;
}
utils::ConfigIterator itr(argv[1]);
while (itr.Next()) {
this->SetParam(itr.name(), itr.val());
}
for (int i = 2; i < argc; ++i) {
char name[256], val[256];
if (sscanf(argv[i], "%[^=]=%s", name, val) == 2) {
this->SetParam(name, val);
}
}
// do not save anything when save to stdout
if (model_out == "stdout" || name_pred == "stdout") {
this->SetParam("silent", "1");
save_period = 0;
}
// initialized the result
rabit::Init(argc, argv);
if (rabit::IsDistributed()) {
std::string pname = rabit::GetProcessorName();
fprintf(stderr, "start %s:%d\n", pname.c_str(), rabit::GetRank());
}
if (rabit::IsDistributed() && data_split == "NONE") {
this->SetParam("dsplit", "row");
}
if (rabit::GetRank() != 0) {
this->SetParam("silent", "2");
}
this->InitData();
if (task == "train") {
// if task is training, will try recover from checkpoint
this->TaskTrain();
return 0;
} else {
this->InitLearner();
}
if (task == "dump") {
this->TaskDump(); return 0;
}
if (task == "eval") {
this->TaskEval(); return 0;
}
if (task == "pred") {
this->TaskPred();
}
return 0;
}
inline void SetParam(const char *name, const char *val) {
if (!strcmp("silent", name)) silent = atoi(val);
if (!strcmp("use_buffer", name)) use_buffer = atoi(val);
if (!strcmp("num_round", name)) num_round = atoi(val);
if (!strcmp("pred_margin", name)) pred_margin = atoi(val);
if (!strcmp("ntree_limit", name)) ntree_limit = atoi(val);
if (!strcmp("save_period", name)) save_period = atoi(val);
if (!strcmp("eval_train", name)) eval_train = atoi(val);
if (!strcmp("task", name)) task = val;
if (!strcmp("data", name)) train_path = val;
if (!strcmp("test:data", name)) test_path = val;
if (!strcmp("model_in", name)) model_in = val;
if (!strcmp("model_out", name)) model_out = val;
if (!strcmp("model_dir", name)) model_dir_path = val;
if (!strcmp("fmap", name)) name_fmap = val;
if (!strcmp("name_dump", name)) name_dump = val;
if (!strcmp("name_pred", name)) name_pred = val;
if (!strcmp("dsplit", name)) data_split = val;
if (!strcmp("dump_stats", name)) dump_model_stats = atoi(val);
if (!strcmp("save_pbuffer", name)) save_with_pbuffer = atoi(val);
if (!strncmp("eval[", name, 5)) {
char evname[256];
utils::Assert(sscanf(name, "eval[%[^]]", evname) == 1,
"must specify evaluation name for display");
eval_data_names.push_back(std::string(evname));
eval_data_paths.push_back(std::string(val));
}
learner.SetParam(name, val);
}
public:
BoostLearnTask(void) {
// default parameters
silent = 0;
use_buffer = 1;
num_round = 10;
save_period = 0;
eval_train = 0;
pred_margin = 0;
ntree_limit = 0;
dump_model_stats = 0;
task = "train";
model_in = "NULL";
model_out = "NULL";
name_fmap = "NULL";
name_pred = "pred.txt";
name_dump = "dump.txt";
model_dir_path = "./";
data_split = "NONE";
load_part = 0;
save_with_pbuffer = 0;
data = NULL;
}
~BoostLearnTask(void) {
for (size_t i = 0; i < deval.size(); i++) {
delete deval[i];
}
if (data != NULL) delete data;
}
private:
inline void InitData(void) {
if (strchr(train_path.c_str(), '%') != NULL) {
char s_tmp[256];
utils::SPrintf(s_tmp, sizeof(s_tmp), train_path.c_str(), rabit::GetRank());
train_path = s_tmp;
load_part = 1;
}
bool loadsplit = data_split == "row";
if (name_fmap != "NULL") fmap.LoadText(name_fmap.c_str());
if (task == "dump") return;
if (task == "pred") {
data = io::LoadDataMatrix(test_path.c_str(), silent != 0, use_buffer != 0, loadsplit);
} else {
// training
data = io::LoadDataMatrix(train_path.c_str(),
silent != 0 && load_part == 0,
use_buffer != 0, loadsplit);
utils::Assert(eval_data_names.size() == eval_data_paths.size(), "BUG");
for (size_t i = 0; i < eval_data_names.size(); ++i) {
deval.push_back(io::LoadDataMatrix(eval_data_paths[i].c_str(),
silent != 0,
use_buffer != 0,
loadsplit));
devalall.push_back(deval.back());
}
std::vector<io::DataMatrix *> dcache(1, data);
for (size_t i = 0; i < deval.size(); ++i) {
dcache.push_back(deval[i]);
}
// set cache data to be all training and evaluation data
learner.SetCacheData(dcache);
// add training set to evaluation set if needed
if (eval_train != 0) {
devalall.push_back(data);
eval_data_names.push_back(std::string("train"));
}
}
}
inline void InitLearner(void) {
if (model_in != "NULL") {
learner.LoadModel(model_in.c_str());
} else {
utils::Assert(task == "train", "model_in not specified");
learner.InitModel();
}
}
inline void TaskTrain(void) {
int version = rabit::LoadCheckPoint(&learner);
if (version == 0) this->InitLearner();
const time_t start = time(NULL);
unsigned long elapsed = 0; // NOLINT(*)
learner.CheckInit(data);
bool allow_lazy = learner.AllowLazyCheckPoint();
for (int i = version / 2; i < num_round; ++i) {
elapsed = (unsigned long)(time(NULL) - start); // NOLINT(*)
if (version % 2 == 0) {
if (!silent) printf("boosting round %d, %lu sec elapsed\n", i, elapsed);
learner.UpdateOneIter(i, *data);
if (allow_lazy) {
rabit::LazyCheckPoint(&learner);
} else {
rabit::CheckPoint(&learner);
}
version += 1;
}
utils::Assert(version == rabit::VersionNumber(), "consistent check");
std::string res = learner.EvalOneIter(i, devalall, eval_data_names);
if (rabit::IsDistributed()) {
if (rabit::GetRank() == 0) {
rabit::TrackerPrintf("%s\n", res.c_str());
}
} else {
if (silent < 2) {
fprintf(stderr, "%s\n", res.c_str());
}
}
if (save_period != 0 && (i + 1) % save_period == 0) {
this->SaveModel(i);
}
if (allow_lazy) {
rabit::LazyCheckPoint(&learner);
} else {
rabit::CheckPoint(&learner);
}
version += 1;
utils::Assert(version == rabit::VersionNumber(), "consistent check");
elapsed = (unsigned long)(time(NULL) - start); // NOLINT(*)
}
// always save final round
if ((save_period == 0 || num_round % save_period != 0) && model_out != "NONE") {
if (model_out == "NULL") {
this->SaveModel(num_round - 1);
} else {
this->SaveModel(model_out.c_str());
}
}
if (!silent) {
printf("\nupdating end, %lu sec in all\n", elapsed);
}
}
inline void TaskEval(void) {
learner.EvalOneIter(0, devalall, eval_data_names);
}
inline void TaskDump(void) {
FILE *fo = utils::FopenCheck(name_dump.c_str(), "w");
std::vector<std::string> dump = learner.DumpModel(fmap, dump_model_stats != 0);
for (size_t i = 0; i < dump.size(); ++i) {
fprintf(fo, "booster[%lu]:\n", i);
fprintf(fo, "%s", dump[i].c_str());
}
fclose(fo);
}
inline void SaveModel(const char *fname) const {
if (rabit::GetRank() != 0) return;
learner.SaveModel(fname, save_with_pbuffer != 0);
}
inline void SaveModel(int i) const {
char fname[256];
utils::SPrintf(fname, sizeof(fname),
"%s/%04d.model", model_dir_path.c_str(), i + 1);
this->SaveModel(fname);
}
inline void TaskPred(void) {
std::vector<float> preds;
if (!silent) printf("start prediction...\n");
learner.Predict(*data, pred_margin != 0, &preds, ntree_limit);
if (!silent) printf("writing prediction to %s\n", name_pred.c_str());
FILE *fo;
if (name_pred != "stdout") {
fo = utils::FopenCheck(name_pred.c_str(), "w");
} else {
fo = stdout;
}
for (size_t i = 0; i < preds.size(); ++i) {
fprintf(fo, "%g\n", preds[i]);
}
if (fo != stdout) fclose(fo);
}
private:
/*! \brief whether silent */
int silent;
/*! \brief special load */
int load_part;
/*! \brief whether use auto binary buffer */
int use_buffer;
/*! \brief whether evaluate training statistics */
int eval_train;
/*! \brief number of boosting iterations */
int num_round;
/*! \brief the period to save the model, 0 means only save the final round model */
int save_period;
/*! \brief the path of training/test data set */
std::string train_path, test_path;
/*! \brief the path of test model file, or file to restart training */
std::string model_in;
/*! \brief the path of final model file, to be saved */
std::string model_out;
/*! \brief the path of directory containing the saved models */
std::string model_dir_path;
/*! \brief task to perform */
std::string task;
/*! \brief name of predict file */
std::string name_pred;
/*! \brief data split mode */
std::string data_split;
/*!\brief limit number of trees in prediction */
int ntree_limit;
/*!\brief whether to directly output margin value */
int pred_margin;
/*! \brief whether dump statistics along with model */
int dump_model_stats;
/*! \brief whether save prediction buffer */
int save_with_pbuffer;
/*! \brief name of feature map */
std::string name_fmap;
/*! \brief name of dump file */
std::string name_dump;
/*! \brief the paths of validation data sets */
std::vector<std::string> eval_data_paths;
/*! \brief the names of the evaluation data used in output log */
std::vector<std::string> eval_data_names;
private:
io::DataMatrix* data;
std::vector<io::DataMatrix*> deval;
std::vector<const io::DataMatrix*> devalall;
utils::FeatMap fmap;
learner::BoostLearner learner;
};
} // namespace xgboost
int main(int argc, char *argv[]) {
xgboost::BoostLearnTask tsk;
tsk.SetParam("seed", "0");
int ret = tsk.Run(argc, argv);
rabit::Finalize();
return ret;
}