Hendrik/xgboost
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
xgboost/src/tree/updater_colmaker.cc
Jiaming Yuan f2f7dd87b8
Use view for SparsePage exclusively. (#6590)
2021-01-11 18:04:55 +08:00

630 lines
24 KiB
C++
Raw Blame History

/*!
* Copyright 2014-2019 by Contributors
* \file updater_colmaker.cc
* \brief use columnwise update to construct a tree
* \author Tianqi Chen
*/
#include <rabit/rabit.h>
#include <memory>
#include <vector>
#include <cmath>
#include <algorithm>
#include "xgboost/parameter.h"
#include "xgboost/tree_updater.h"
#include "xgboost/logging.h"
#include "xgboost/json.h"
#include "param.h"
#include "constraints.h"
#include "../common/random.h"
#include "split_evaluator.h"
namespace xgboost {
namespace tree {
DMLC_REGISTRY_FILE_TAG(updater_colmaker);
struct ColMakerTrainParam : XGBoostParameter<ColMakerTrainParam> {
// speed optimization for dense column
float opt_dense_col;
DMLC_DECLARE_PARAMETER(ColMakerTrainParam) {
DMLC_DECLARE_FIELD(opt_dense_col)
.set_range(0.0f, 1.0f)
.set_default(1.0f)
.describe("EXP Param: speed optimization for dense column.");
}
/*! \brief whether need forward small to big search: default right */
inline bool NeedForwardSearch(int default_direction, float col_density,
bool indicator) const {
return default_direction == 2 ||
(default_direction == 0 && (col_density < opt_dense_col) &&
!indicator);
}
/*! \brief whether need backward big to small search: default left */
inline bool NeedBackwardSearch(int default_direction) const {
return default_direction != 2;
}
};
DMLC_REGISTER_PARAMETER(ColMakerTrainParam);
/*! \brief column-wise update to construct a tree */
class ColMaker: public TreeUpdater {
public:
void Configure(const Args& args) override {
param_.UpdateAllowUnknown(args);
colmaker_param_.UpdateAllowUnknown(args);
}
void LoadConfig(Json const& in) override {
auto const& config = get<Object const>(in);
FromJson(config.at("train_param"), &this->param_);
FromJson(config.at("colmaker_train_param"), &this->colmaker_param_);
}
void SaveConfig(Json* p_out) const override {
auto& out = *p_out;
out["train_param"] = ToJson(param_);
out["colmaker_train_param"] = ToJson(colmaker_param_);
}
char const* Name() const override {
return "grow_colmaker";
}
void LazyGetColumnDensity(DMatrix *dmat) {
// Finds densities if we don't already have them
if (column_densities_.empty()) {
std::vector<size_t> column_size(dmat->Info().num_col_);
for (const auto &batch : dmat->GetBatches<SortedCSCPage>()) {
auto page = batch.GetView();
for (auto i = 0u; i < batch.Size(); i++) {
column_size[i] += page[i].size();
}
}
column_densities_.resize(column_size.size());
for (auto i = 0u; i < column_densities_.size(); i++) {
size_t nmiss = dmat->Info().num_row_ - column_size[i];
column_densities_[i] =
1.0f - (static_cast<float>(nmiss)) / dmat->Info().num_row_;
}
}
}
void Update(HostDeviceVector<GradientPair> *gpair,
DMatrix* dmat,
const std::vector<RegTree*> &trees) override {
if (rabit::IsDistributed()) {
LOG(FATAL) << "Updater `grow_colmaker` or `exact` tree method doesn't "
"support distributed training.";
}
this->LazyGetColumnDensity(dmat);
// rescale learning rate according to size of trees
float lr = param_.learning_rate;
param_.learning_rate = lr / trees.size();
interaction_constraints_.Configure(param_, dmat->Info().num_row_);
// build tree
for (auto tree : trees) {
Builder builder(
param_,
colmaker_param_,
interaction_constraints_, column_densities_);
builder.Update(gpair->ConstHostVector(), dmat, tree);
}
param_.learning_rate = lr;
}
protected:
// training parameter
TrainParam param_;
ColMakerTrainParam colmaker_param_;
// SplitEvaluator that will be cloned for each Builder
std::vector<float> column_densities_;
FeatureInteractionConstraintHost interaction_constraints_;
// data structure
/*! \brief per thread x per node entry to store tmp data */
struct ThreadEntry {
/*! \brief statistics of data */
GradStats stats;
/*! \brief last feature value scanned */
bst_float last_fvalue { 0 };
/*! \brief current best solution */
SplitEntry best;
// constructor
ThreadEntry() = default;
};
struct NodeEntry {
/*! \brief statics for node entry */
GradStats stats;
/*! \brief loss of this node, without split */
bst_float root_gain { 0.0f };
/*! \brief weight calculated related to current data */
bst_float weight { 0.0f };
/*! \brief current best solution */
SplitEntry best;
// constructor
NodeEntry() = default;
};
// actual builder that runs the algorithm
class Builder {
public:
// constructor
explicit Builder(const TrainParam& param,
const ColMakerTrainParam& colmaker_train_param,
FeatureInteractionConstraintHost _interaction_constraints,
const std::vector<float> &column_densities)
: param_(param), colmaker_train_param_{colmaker_train_param},
nthread_(omp_get_max_threads()),
tree_evaluator_(param_, column_densities.size(), GenericParameter::kCpuId),
interaction_constraints_{std::move(_interaction_constraints)},
column_densities_(column_densities) {}
// update one tree, growing
virtual void Update(const std::vector<GradientPair>& gpair,
DMatrix* p_fmat,
RegTree* p_tree) {
std::vector<int> newnodes;
this->InitData(gpair, *p_fmat);
this->InitNewNode(qexpand_, gpair, *p_fmat, *p_tree);
for (int depth = 0; depth < param_.max_depth; ++depth) {
this->FindSplit(depth, qexpand_, gpair, p_fmat, p_tree);
this->ResetPosition(qexpand_, p_fmat, *p_tree);
this->UpdateQueueExpand(*p_tree, qexpand_, &newnodes);
this->InitNewNode(newnodes, gpair, *p_fmat, *p_tree);
for (auto nid : qexpand_) {
if ((*p_tree)[nid].IsLeaf()) {
continue;
}
int cleft = (*p_tree)[nid].LeftChild();
int cright = (*p_tree)[nid].RightChild();
tree_evaluator_.AddSplit(nid, cleft, cright, snode_[nid].best.SplitIndex(),
snode_[cleft].weight, snode_[cright].weight);
interaction_constraints_.Split(nid, snode_[nid].best.SplitIndex(), cleft, cright);
}
qexpand_ = newnodes;
// if nothing left to be expand, break
if (qexpand_.size() == 0) break;
}
// set all the rest expanding nodes to leaf
for (const int nid : qexpand_) {
(*p_tree)[nid].SetLeaf(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);
}
}
protected:
// initialize temp data structure
inline void InitData(const std::vector<GradientPair>& gpair,
const DMatrix& fmat) {
{
// setup position
position_.resize(gpair.size());
CHECK_EQ(fmat.Info().num_row_, position_.size());
std::fill(position_.begin(), position_.end(), 0);
// mark delete for the deleted datas
for (size_t ridx = 0; ridx < position_.size(); ++ridx) {
if (gpair[ridx].GetHess() < 0.0f) position_[ridx] = ~position_[ridx];
}
// mark subsample
if (param_.subsample < 1.0f) {
CHECK_EQ(param_.sampling_method, TrainParam::kUniform)
<< "Only uniform sampling is supported, "
<< "gradient-based sampling is only support by GPU Hist.";
std::bernoulli_distribution coin_flip(param_.subsample);
auto& rnd = common::GlobalRandom();
for (size_t ridx = 0; ridx < position_.size(); ++ridx) {
if (gpair[ridx].GetHess() < 0.0f) continue;
if (!coin_flip(rnd)) position_[ridx] = ~position_[ridx];
}
}
}
{
column_sampler_.Init(fmat.Info().num_col_,
fmat.Info().feature_weigths.ConstHostVector(),
param_.colsample_bynode, param_.colsample_bylevel,
param_.colsample_bytree);
}
{
// setup temp space for each thread
// reserve a small space
stemp_.clear();
stemp_.resize(this->nthread_, std::vector<ThreadEntry>());
for (auto& i : stemp_) {
i.clear(); i.reserve(256);
}
snode_.reserve(256);
}
{
// expand query
qexpand_.reserve(256); qexpand_.clear();
qexpand_.push_back(0);
}
}
/*!
* \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<GradientPair>& gpair,
const DMatrix& fmat,
const RegTree& tree) {
{
// setup statistics space for each tree node
for (auto& i : stemp_) {
i.resize(tree.param.num_nodes, ThreadEntry());
}
snode_.resize(tree.param.num_nodes, NodeEntry());
}
const MetaInfo& info = fmat.Info();
// setup position
const auto ndata = static_cast<bst_omp_uint>(info.num_row_);
#pragma omp parallel for schedule(static)
for (bst_omp_uint ridx = 0; ridx < ndata; ++ridx) {
const int tid = omp_get_thread_num();
if (position_[ridx] < 0) continue;
stemp_[tid][position_[ridx]].stats.Add(gpair[ridx]);
}
// sum the per thread statistics together
for (int nid : qexpand) {
GradStats stats;
for (auto& s : stemp_) {
stats.Add(s[nid].stats);
}
// update node statistics
snode_[nid].stats = stats;
}
auto evaluator = tree_evaluator_.GetEvaluator();
// calculating the weights
for (int nid : qexpand) {
bst_node_t parentid = tree[nid].Parent();
snode_[nid].weight = static_cast<float>(
evaluator.CalcWeight(parentid, param_, snode_[nid].stats));
snode_[nid].root_gain = static_cast<float>(
evaluator.CalcGain(parentid, param_, snode_[nid].stats));
}
}
/*! \brief update queue expand add in new leaves */
inline void UpdateQueueExpand(const RegTree& tree,
const std::vector<int> &qexpand,
std::vector<int>* p_newnodes) {
p_newnodes->clear();
for (int nid : qexpand) {
if (!tree[ nid ].IsLeaf()) {
p_newnodes->push_back(tree[nid].LeftChild());
p_newnodes->push_back(tree[nid].RightChild());
}
}
}
// update enumeration solution
inline void UpdateEnumeration(
int nid, GradientPair gstats, bst_float fvalue, int d_step,
bst_uint fid, GradStats &c, std::vector<ThreadEntry> &temp, // NOLINT(*)
TreeEvaluator::SplitEvaluator<TrainParam> const &evaluator) const {
// 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 (fvalue != e.last_fvalue &&
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 {0};
if (d_step == -1) {
loss_chg = static_cast<bst_float>(
evaluator.CalcSplitGain(param_, nid, fid, c, e.stats) -
snode_[nid].root_gain);
bst_float proposed_split = (fvalue + e.last_fvalue) * 0.5f;
if ( proposed_split == fvalue ) {
e.best.Update(loss_chg, fid, e.last_fvalue,
d_step == -1, c, e.stats);
} else {
e.best.Update(loss_chg, fid, proposed_split,
d_step == -1, c, e.stats);
}
} else {
loss_chg = static_cast<bst_float>(
evaluator.CalcSplitGain(param_, nid, fid, e.stats, c) -
snode_[nid].root_gain);
bst_float proposed_split = (fvalue + e.last_fvalue) * 0.5f;
if ( proposed_split == fvalue ) {
e.best.Update(loss_chg, fid, e.last_fvalue,
d_step == -1, e.stats, c);
} else {
e.best.Update(loss_chg, fid, proposed_split,
d_step == -1, e.stats, c);
}
}
}
}
// update the statistics
e.stats.Add(gstats);
e.last_fvalue = fvalue;
}
}
// same as EnumerateSplit, with cacheline prefetch optimization
void EnumerateSplit(
const Entry *begin, const Entry *end, int d_step, bst_uint fid,
const std::vector<GradientPair> &gpair,
std::vector<ThreadEntry> &temp, // NOLINT(*)
TreeEvaluator::SplitEvaluator<TrainParam> const &evaluator) const {
CHECK(param_.cache_opt) << "Support for `cache_opt' is removed in 1.0.0";
const std::vector<int> &qexpand = qexpand_;
// clear all the temp statistics
for (auto nid : qexpand) {
temp[nid].stats = GradStats();
}
// left statistics
GradStats c;
// local cache buffer for position and gradient pair
constexpr int kBuffer = 32;
int buf_position[kBuffer] = {};
GradientPair buf_gpair[kBuffer] = {};
// aligned ending position
const 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 Entry *it;
const int align_step = d_step * kBuffer;
// internal cached loop
for (it = begin; it != align_end; it += align_step) {
const 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 || !interaction_constraints_.Query(nid, fid)) { continue; }
this->UpdateEnumeration(nid, buf_gpair[i],
p->fvalue, d_step,
fid, c, temp, evaluator);
}
}
// 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 || !interaction_constraints_.Query(nid, fid)) { continue; }
this->UpdateEnumeration(nid, buf_gpair[i],
it->fvalue, d_step,
fid, c, temp, evaluator);
}
// finish updating all statistics, check if it is possible to include all sum statistics
for (int nid : qexpand) {
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;
const bst_float gap = std::abs(e.last_fvalue) + kRtEps;
const bst_float delta = d_step == +1 ? gap: -gap;
if (d_step == -1) {
loss_chg = static_cast<bst_float>(
evaluator.CalcSplitGain(param_, nid, fid, c, e.stats) -
snode_[nid].root_gain);
e.best.Update(loss_chg, fid, e.last_fvalue + delta, d_step == -1, c,
e.stats);
} else {
loss_chg = static_cast<bst_float>(
evaluator.CalcSplitGain(param_, nid, fid, e.stats, c) -
snode_[nid].root_gain);
e.best.Update(loss_chg, fid, e.last_fvalue + delta, d_step == -1,
e.stats, c);
}
}
}
}
// update the solution candidate
virtual void UpdateSolution(const SparsePage &batch,
const std::vector<bst_feature_t> &feat_set,
const std::vector<GradientPair> &gpair,
DMatrix*) {
// start enumeration
const auto num_features = static_cast<bst_omp_uint>(feat_set.size());
#if defined(_OPENMP)
const int batch_size = // NOLINT
std::max(static_cast<int>(num_features / this->nthread_ / 32), 1);
#endif // defined(_OPENMP)
{
dmlc::OMPException omp_handler;
auto page = batch.GetView();
#pragma omp parallel for schedule(dynamic, batch_size)
for (bst_omp_uint i = 0; i < num_features; ++i) {
omp_handler.Run([&]() {
auto evaluator = tree_evaluator_.GetEvaluator();
bst_feature_t const fid = feat_set[i];
int32_t const tid = omp_get_thread_num();
auto c = page[fid];
const bool ind =
c.size() != 0 && c[0].fvalue == c[c.size() - 1].fvalue;
if (colmaker_train_param_.NeedForwardSearch(
param_.default_direction, column_densities_[fid], ind)) {
this->EnumerateSplit(c.data(), c.data() + c.size(), +1, fid,
gpair, stemp_[tid], evaluator);
}
if (colmaker_train_param_.NeedBackwardSearch(
param_.default_direction)) {
this->EnumerateSplit(c.data() + c.size() - 1, c.data() - 1, -1,
fid, gpair, stemp_[tid], evaluator);
}
});
}
omp_handler.Rethrow();
}
}
// find splits at current level, do split per level
inline void FindSplit(int depth,
const std::vector<int> &qexpand,
const std::vector<GradientPair> &gpair,
DMatrix *p_fmat,
RegTree *p_tree) {
auto evaluator = tree_evaluator_.GetEvaluator();
auto feat_set = column_sampler_.GetFeatureSet(depth);
for (const auto &batch : p_fmat->GetBatches<SortedCSCPage>()) {
this->UpdateSolution(batch, feat_set->HostVector(), gpair, p_fmat);
}
// 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 (int nid : qexpand) {
NodeEntry const &e = snode_[nid];
// now we know the solution in snode[nid], set split
if (e.best.loss_chg > kRtEps) {
bst_float left_leaf_weight =
evaluator.CalcWeight(nid, param_, e.best.left_sum) *
param_.learning_rate;
bst_float right_leaf_weight =
evaluator.CalcWeight(nid, param_, e.best.right_sum) *
param_.learning_rate;
p_tree->ExpandNode(nid, e.best.SplitIndex(), e.best.split_value,
e.best.DefaultLeft(), e.weight, left_leaf_weight,
right_leaf_weight, e.best.loss_chg,
e.stats.sum_hess,
e.best.left_sum.GetHess(), e.best.right_sum.GetHess(),
0);
} else {
(*p_tree)[nid].SetLeaf(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,
DMatrix* p_fmat,
const RegTree& tree) {
// set the positions in the nondefault
this->SetNonDefaultPosition(qexpand, p_fmat, tree);
// set rest of instances to default position
// 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 auto ndata = static_cast<bst_omp_uint>(p_fmat->Info().num_row_);
#pragma omp parallel for schedule(static)
for (bst_omp_uint ridx = 0; ridx < ndata; ++ridx) {
CHECK_LT(ridx, position_.size())
<< "ridx exceed bound " << "ridx="<< ridx << " pos=" << position_.size();
const int nid = this->DecodePosition(ridx);
if (tree[nid].IsLeaf()) {
// mark finish when it is not a fresh leaf
if (tree[nid].RightChild() == -1) {
position_[ridx] = ~nid;
}
} else {
// push to default branch
if (tree[nid].DefaultLeft()) {
this->SetEncodePosition(ridx, tree[nid].LeftChild());
} else {
this->SetEncodePosition(ridx, tree[nid].RightChild());
}
}
}
}
// customization part
// synchronize the best solution of each node
virtual void SyncBestSolution(const std::vector<int> &qexpand) {
for (int nid : qexpand) {
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,
DMatrix *p_fmat,
const RegTree &tree) {
// step 1, classify the non-default data into right places
std::vector<unsigned> fsplits;
for (int nid : qexpand) {
if (!tree[nid].IsLeaf()) {
fsplits.push_back(tree[nid].SplitIndex());
}
}
std::sort(fsplits.begin(), fsplits.end());
fsplits.resize(std::unique(fsplits.begin(), fsplits.end()) - fsplits.begin());
for (const auto &batch : p_fmat->GetBatches<SortedCSCPage>()) {
auto page = batch.GetView();
for (auto fid : fsplits) {
auto col = page[fid];
const auto ndata = static_cast<bst_omp_uint>(col.size());
#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 bst_float fvalue = col[j].fvalue;
// go back to parent, correct those who are not default
if (!tree[nid].IsLeaf() && tree[nid].SplitIndex() == fid) {
if (fvalue < tree[nid].SplitCond()) {
this->SetEncodePosition(ridx, tree[nid].LeftChild());
} else {
this->SetEncodePosition(ridx, tree[nid].RightChild());
}
}
}
}
}
}
// 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_;
const ColMakerTrainParam& colmaker_train_param_;
// number of omp thread used during training
const int nthread_;
common::ColumnSampler column_sampler_;
// 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_;
TreeEvaluator tree_evaluator_;
FeatureInteractionConstraintHost interaction_constraints_;
const std::vector<float> &column_densities_;
};
};
XGBOOST_REGISTER_TREE_UPDATER(ColMaker, "grow_colmaker")
.describe("Grow tree with parallelization over columns.")
.set_body([]() {
return new ColMaker();
});
} // namespace tree
} // namespace xgboost
Reference in New Issue View Git Blame Copy Permalink