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

Handle missing values in dataframe with category dtype. (#7331)

Browse Source 
* Replace -1 in pandas initializer.
* Unify `IsValid` functor.
* Mimic pandas data handling in cuDF glue code.
* Check invalid categories.
* Fix DDM sketching.
This commit is contained in:
Jiaming Yuan
2021-10-28 03:33:54 +08:00
committed by GitHub
parent 2eee87423c
commit ac9bfaa4f2
13 changed files with 301 additions and 103 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
src/common/categorical.h
View File

@@ -1,5 +1,5 @@
/*!
* Copyright 2020 by XGBoost Contributors
* Copyright 2020-2021 by XGBoost Contributors
* \file categorical.h
*/
#ifndef XGBOOST_COMMON_CATEGORICAL_H_
@@ -42,6 +42,11 @@ inline XGBOOST_DEVICE bool Decision(common::Span<uint32_t const> cats, bst_cat_t
return !s_cats.Check(cat);
}
inline void CheckCat(bst_cat_t cat) {
CHECK_GE(cat, 0) << "Invalid categorical value detected. Categorical value "
"should be non-negative.";
}
struct IsCatOp {
XGBOOST_DEVICE bool operator()(FeatureType ft) {
return ft == FeatureType::kCategorical;

1
src/common/hist_util.cu
View File

@@ -133,6 +133,7 @@ void RemoveDuplicatedCategories(
int32_t device, MetaInfo const &info, Span<bst_row_t> d_cuts_ptr,
dh::device_vector<Entry> *p_sorted_entries,
dh::caching_device_vector<size_t> *p_column_sizes_scan) {
info.feature_types.SetDevice(device);
auto d_feature_types = info.feature_types.ConstDeviceSpan();
CHECK(!d_feature_types.empty());
auto &column_sizes_scan = *p_column_sizes_scan;

31
src/common/hist_util.cuh
View File

@@ -124,6 +124,11 @@ void MakeEntriesFromAdapter(AdapterBatch const& batch, BatchIter batch_iter,
void SortByWeight(dh::device_vector<float>* weights,
dh::device_vector<Entry>* sorted_entries);
void RemoveDuplicatedCategories(
int32_t device, MetaInfo const &info, Span<bst_row_t> d_cuts_ptr,
dh::device_vector<Entry> *p_sorted_entries,
dh::caching_device_vector<size_t> *p_column_sizes_scan);
} // namespace detail
// Compute sketch on DMatrix.
@@ -132,9 +137,10 @@ HistogramCuts DeviceSketch(int device, DMatrix* dmat, int max_bins,
size_t sketch_batch_num_elements = 0);
template <typename AdapterBatch>
void ProcessSlidingWindow(AdapterBatch const& batch, int device, size_t columns,
size_t begin, size_t end, float missing,
SketchContainer* sketch_container, int num_cuts) {
void ProcessSlidingWindow(AdapterBatch const &batch, MetaInfo const &info,
int device, size_t columns, size_t begin, size_t end,
float missing, SketchContainer *sketch_container,
int num_cuts) {
// Copy current subset of valid elements into temporary storage and sort
dh::device_vector<Entry> sorted_entries;
dh::caching_device_vector<size_t> column_sizes_scan;
@@ -142,6 +148,7 @@ void ProcessSlidingWindow(AdapterBatch const& batch, int device, size_t columns,
thrust::make_counting_iterator(0llu),
[=] __device__(size_t idx) { return batch.GetElement(idx); });
HostDeviceVector<SketchContainer::OffsetT> cuts_ptr;
cuts_ptr.SetDevice(device);
detail::MakeEntriesFromAdapter(batch, batch_iter, {begin, end}, missing,
columns, num_cuts, device,
&cuts_ptr,
@@ -151,8 +158,14 @@ void ProcessSlidingWindow(AdapterBatch const& batch, int device, size_t columns,
thrust::sort(thrust::cuda::par(alloc), sorted_entries.begin(),
sorted_entries.end(), detail::EntryCompareOp());
auto const& h_cuts_ptr = cuts_ptr.ConstHostVector();
if (sketch_container->HasCategorical()) {
auto d_cuts_ptr = cuts_ptr.DeviceSpan();
detail::RemoveDuplicatedCategories(device, info, d_cuts_ptr,
&sorted_entries, &column_sizes_scan);
}
auto d_cuts_ptr = cuts_ptr.DeviceSpan();
auto const &h_cuts_ptr = cuts_ptr.HostVector();
// Extract the cuts from all columns concurrently
sketch_container->Push(dh::ToSpan(sorted_entries),
dh::ToSpan(column_sizes_scan), d_cuts_ptr,
@@ -222,6 +235,12 @@ void ProcessWeightedSlidingWindow(Batch batch, MetaInfo const& info,
detail::SortByWeight(&temp_weights, &sorted_entries);
if (sketch_container->HasCategorical()) {
auto d_cuts_ptr = cuts_ptr.DeviceSpan();
detail::RemoveDuplicatedCategories(device, info, d_cuts_ptr,
&sorted_entries, &column_sizes_scan);
}
auto const& h_cuts_ptr = cuts_ptr.ConstHostVector();
auto d_cuts_ptr = cuts_ptr.DeviceSpan();
@@ -274,8 +293,8 @@ void AdapterDeviceSketch(Batch batch, int num_bins,
device, num_cuts_per_feature, false);
for (auto begin = 0ull; begin < batch.Size(); begin += sketch_batch_num_elements) {
size_t end = std::min(batch.Size(), size_t(begin + sketch_batch_num_elements));
ProcessSlidingWindow(batch, device, num_cols,
begin, end, missing, sketch_container, num_cuts_per_feature);
ProcessSlidingWindow(batch, info, device, num_cols, begin, end, missing,
sketch_container, num_cuts_per_feature);
}
}
}