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

Automatically remove nan from input data when it is sparse. (#2062)

Browse Source 
* [DATALoad] Automatically remove Nan when load from sparse matrix

* add log
This commit is contained in:
Tianqi Chen 2017-02-25 08:59:17 -08:00 committed by GitHub
parent 5d093a7f4c
commit fd19b7a188
4 changed files with 41 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
NEWS.md
View File

@ -6,6 +6,8 @@ This file records the changes in xgboost library in reverse chronological order.
## in progress version ## in progress version
* Refactored gbm to allow more friendly cache strategy * Refactored gbm to allow more friendly cache strategy
- Specialized some prediction routine - Specialized some prediction routine
* Automatically remove nan from input data when it is sparse.
- This can solve some of user reported problem of istart != hist.size
## v0.6 (2016.07.29) ## v0.6 (2016.07.29)
* Version 0.5 is skipped due to major improvements in the core * Version 0.5 is skipped due to major improvements in the core

43
src/c_api/c_api.cc
View File

@ -238,22 +238,31 @@ XGB_DLL int XGDMatrixCreateFromCSREx(const size_t* indptr,
API_BEGIN(); API_BEGIN();
data::SimpleCSRSource& mat = *source; data::SimpleCSRSource& mat = *source;
mat.row_ptr_.resize(nindptr); mat.row_ptr_.reserve(nindptr);
for (size_t i = 0; i < nindptr; ++i) { mat.row_data_.reserve(nelem);
mat.row_ptr_[i] = indptr[i]; mat.row_ptr_.resize(1);
} mat.row_ptr_[0] = 0;
mat.row_data_.resize(nelem); size_t num_column = 0;
for (size_t i = 0; i < nelem; ++i) { for (size_t i = 1; i < nindptr; ++i) {
mat.row_data_[i] = RowBatch::Entry(indices[i], data[i]); for (size_t j = indptr[i - 1]; j < indptr[i]; ++j) {
mat.info.num_col = std::max(mat.info.num_col, if (!common::CheckNAN(data[j])) {
static_cast<uint64_t>(indices[i] + 1)); // automatically skip nan.
mat.row_data_.emplace_back(RowBatch::Entry(indices[j], data[j]));
num_column = std::max(num_column, static_cast<size_t>(indices[j] + 1));
}
}
mat.row_ptr_.push_back(mat.row_data_.size());
} }
mat.info.num_col = num_column;
if (num_col > 0) { if (num_col > 0) {
CHECK_LE(mat.info.num_col, num_col); CHECK_LE(mat.info.num_col, num_col)
<< "num_col=" << num_col << " vs " << mat.info.num_col;
mat.info.num_col = num_col; mat.info.num_col = num_col;
} }
mat.info.num_row = nindptr - 1; mat.info.num_row = nindptr - 1;
mat.info.num_nonzero = nelem; LOG(INFO) << "num_row=" << mat.info.num_row;
mat.info.num_nonzero = mat.row_data_.size();
*out = new std::shared_ptr<DMatrix>(DMatrix::Create(std::move(source))); *out = new std::shared_ptr<DMatrix>(DMatrix::Create(std::move(source)));
API_END(); API_END();
} }
@ -291,7 +300,9 @@ XGB_DLL int XGDMatrixCreateFromCSCEx(const size_t* col_ptr,
for (omp_ulong i = 0; i < static_cast<omp_ulong>(ncol); ++i) { // NOLINT(*) for (omp_ulong i = 0; i < static_cast<omp_ulong>(ncol); ++i) { // NOLINT(*)
int tid = omp_get_thread_num(); int tid = omp_get_thread_num();
for (size_t j = col_ptr[i]; j < col_ptr[i+1]; ++j) { for (size_t j = col_ptr[i]; j < col_ptr[i+1]; ++j) {
builder.AddBudget(indices[j], tid); if (!common::CheckNAN(data[j])) {
builder.AddBudget(indices[j], tid);
}
} }
} }
builder.InitStorage(); builder.InitStorage();
@ -299,9 +310,11 @@ XGB_DLL int XGDMatrixCreateFromCSCEx(const size_t* col_ptr,
for (omp_ulong i = 0; i < static_cast<omp_ulong>(ncol); ++i) { // NOLINT(*) for (omp_ulong i = 0; i < static_cast<omp_ulong>(ncol); ++i) { // NOLINT(*)
int tid = omp_get_thread_num(); int tid = omp_get_thread_num();
for (size_t j = col_ptr[i]; j < col_ptr[i+1]; ++j) { for (size_t j = col_ptr[i]; j < col_ptr[i+1]; ++j) {
builder.Push(indices[j], if (!common::CheckNAN(data[j])) {
RowBatch::Entry(static_cast<bst_uint>(i), data[j]), builder.Push(indices[j],
tid); RowBatch::Entry(static_cast<bst_uint>(i), data[j]),
tid);
}
} }
} }
mat.info.num_row = mat.row_ptr_.size() - 1; mat.info.num_row = mat.row_ptr_.size() - 1;

5
src/tree/updater_colmaker.cc
View File

@ -670,9 +670,8 @@ class ColMaker: public TreeUpdater {
#pragma omp parallel for schedule(static) #pragma omp parallel for schedule(static)
for (bst_omp_uint i = 0; i < ndata; ++i) { for (bst_omp_uint i = 0; i < ndata; ++i) {
const bst_uint ridx = rowset[i]; const bst_uint ridx = rowset[i];
if (ridx >= position.size()) { CHECK_LT(ridx, position.size())
LOG(INFO) << "ridx exceed bound\n"; << "ridx exceed bound " << "ridx="<< ridx << " pos=" << position.size();
}
const int nid = this->DecodePosition(ridx); const int nid = this->DecodePosition(ridx);
if (tree[nid].is_leaf()) { if (tree[nid].is_leaf()) {
// mark finish when it is not a fresh leaf // mark finish when it is not a fresh leaf

11
src/tree/updater_histmaker.cc
View File

@ -298,8 +298,15 @@ class CQHistMaker: public HistMaker<TStats> {
hist.data[istart].Add(gstats); hist.data[istart].Add(gstats);
} else { } else {
while (istart < hist.size && !(fv < hist.cut[istart])) ++istart; while (istart < hist.size && !(fv < hist.cut[istart])) ++istart;
CHECK_NE(istart, hist.size); if (istart != hist.size) {
hist.data[istart].Add(gstats); hist.data[istart].Add(gstats);
} else {
LOG(INFO) << "fv=" << fv << ", hist.size=" << hist.size;
for (size_t i = 0; i < hist.size; ++i) {
LOG(INFO) << "hist[" << i << "]=" << hist.cut[i];
}
LOG(FATAL) << "fv=" << fv << ", hist.last=" << hist.cut[hist.size - 1];
}
} }
} }
}; };