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

Support categorical data for dask functional interface and DQM. (#7043)

Browse Source 
* Support categorical data for dask functional interface and DQM.

* Implement categorical data support for GPU GK-merge.
* Add support for dask functional interface.
* Add support for DQM.

* Get newer cupy.
This commit is contained in:
Jiaming Yuan
2021-06-18 13:06:52 +08:00
committed by GitHub
parent 7dd29ffd47
commit 86715e4cd4
16 changed files with 364 additions and 167 deletions
Download Patch File Download Diff File Expand all files Collapse all files

50
src/data/ellpack_page.cu
View File

@@ -134,17 +134,20 @@ struct WriteCompressedEllpackFunctor {
const common::CompressedBufferWriter& writer,
AdapterBatchT batch,
EllpackDeviceAccessor accessor,
common::Span<FeatureType const> feature_types,
const data::IsValidFunctor& is_valid)
: d_buffer(buffer),
writer(writer),
batch(std::move(batch)),
accessor(std::move(accessor)),
feature_types(std::move(feature_types)),
is_valid(is_valid) {}
common::CompressedByteT* d_buffer;
common::CompressedBufferWriter writer;
AdapterBatchT batch;
EllpackDeviceAccessor accessor;
common::Span<FeatureType const> feature_types;
data::IsValidFunctor is_valid;
using Tuple = thrust::tuple<size_t, size_t, size_t>;
@@ -154,7 +157,12 @@ struct WriteCompressedEllpackFunctor {
// -1 because the scan is inclusive
size_t output_position =
accessor.row_stride * e.row_idx + out.get<1>() - 1;
auto bin_idx = accessor.SearchBin(e.value, e.column_idx);
uint32_t bin_idx = 0;
if (common::IsCat(feature_types, e.column_idx)) {
bin_idx = accessor.SearchBin<true>(e.value, e.column_idx);
} else {
bin_idx = accessor.SearchBin<false>(e.value, e.column_idx);
}
writer.AtomicWriteSymbol(d_buffer, bin_idx, output_position);
}
return 0;
@@ -184,8 +192,9 @@ class TypedDiscard : public thrust::discard_iterator<T> {
// Here the data is already correctly ordered and simply needs to be compacted
// to remove missing data
template <typename AdapterBatchT>
void CopyDataToEllpack(const AdapterBatchT& batch, EllpackPageImpl* dst,
int device_idx, float missing) {
void CopyDataToEllpack(const AdapterBatchT &batch,
common::Span<FeatureType const> feature_types,
EllpackPageImpl *dst, int device_idx, float missing) {
// Some witchcraft happens here
// The goal is to copy valid elements out of the input to an ELLPACK matrix
// with a given row stride, using no extra working memory Standard stream
@@ -220,7 +229,8 @@ void CopyDataToEllpack(const AdapterBatchT& batch, EllpackPageImpl* dst,
// We redirect the scan output into this functor to do the actual writing
WriteCompressedEllpackFunctor<AdapterBatchT> functor(
d_compressed_buffer, writer, batch, device_accessor, is_valid);
d_compressed_buffer, writer, batch, device_accessor, feature_types,
is_valid);
TypedDiscard<Tuple> discard;
thrust::transform_output_iterator<
WriteCompressedEllpackFunctor<AdapterBatchT>, decltype(discard)>
@@ -263,22 +273,22 @@ template <typename AdapterBatch>
EllpackPageImpl::EllpackPageImpl(AdapterBatch batch, float missing, int device,
bool is_dense, int nthread,
common::Span<size_t> row_counts_span,
common::Span<FeatureType const> feature_types,
size_t row_stride, size_t n_rows, size_t n_cols,
common::HistogramCuts const& cuts) {
dh::safe_cuda(cudaSetDevice(device));
*this = EllpackPageImpl(device, cuts, is_dense, row_stride, n_rows);
CopyDataToEllpack(batch, this, device, missing);
CopyDataToEllpack(batch, feature_types, this, device, missing);
WriteNullValues(this, device, row_counts_span);
}
#define ELLPACK_BATCH_SPECIALIZE(__BATCH_T) \
template EllpackPageImpl::EllpackPageImpl( \
__BATCH_T batch, float missing, int device, \
bool is_dense, int nthread, \
common::Span<size_t> row_counts_span, \
size_t row_stride, size_t n_rows, size_t n_cols, \
common::HistogramCuts const& cuts);
#define ELLPACK_BATCH_SPECIALIZE(__BATCH_T) \
template EllpackPageImpl::EllpackPageImpl( \
__BATCH_T batch, float missing, int device, bool is_dense, int nthread, \
common::Span<size_t> row_counts_span, \
common::Span<FeatureType const> feature_types, size_t row_stride, \
size_t n_rows, size_t n_cols, common::HistogramCuts const &cuts);
ELLPACK_BATCH_SPECIALIZE(data::CudfAdapterBatch)
ELLPACK_BATCH_SPECIALIZE(data::CupyAdapterBatch)
@@ -467,11 +477,17 @@ size_t EllpackPageImpl::MemCostBytes(size_t num_rows, size_t row_stride,
return compressed_size_bytes;
}
EllpackDeviceAccessor EllpackPageImpl::GetDeviceAccessor(int device) const {
EllpackDeviceAccessor EllpackPageImpl::GetDeviceAccessor(
int device, common::Span<FeatureType const> feature_types) const {
gidx_buffer.SetDevice(device);
return EllpackDeviceAccessor(
device, cuts_, is_dense, row_stride, base_rowid, n_rows,
common::CompressedIterator<uint32_t>(gidx_buffer.ConstDevicePointer(),
NumSymbols()));
return {device,
cuts_,
is_dense,
row_stride,
base_rowid,
n_rows,
common::CompressedIterator<uint32_t>(gidx_buffer.ConstDevicePointer(),
NumSymbols()),
feature_types};
}
} // namespace xgboost

36
src/data/ellpack_page.cuh
View File

@@ -10,6 +10,7 @@
#include "../common/compressed_iterator.h"
#include "../common/device_helpers.cuh"
#include "../common/hist_util.h"
#include "../common/categorical.h"
#include <thrust/binary_search.h>
namespace xgboost {
@@ -31,13 +32,17 @@ struct EllpackDeviceAccessor {
/*! \brief Histogram cut values. Size equals to (bins per feature * number of features). */
common::Span<const bst_float> gidx_fvalue_map;
common::Span<const FeatureType> feature_types;
EllpackDeviceAccessor(int device, const common::HistogramCuts& cuts,
bool is_dense, size_t row_stride, size_t base_rowid,
size_t n_rows,common::CompressedIterator<uint32_t> gidx_iter)
size_t n_rows,common::CompressedIterator<uint32_t> gidx_iter,
common::Span<FeatureType const> feature_types)
: is_dense(is_dense),
row_stride(row_stride),
base_rowid(base_rowid),
n_rows(n_rows) ,gidx_iter(gidx_iter){
n_rows(n_rows) ,gidx_iter(gidx_iter),
feature_types{feature_types} {
cuts.cut_values_.SetDevice(device);
cuts.cut_ptrs_.SetDevice(device);
cuts.min_vals_.SetDevice(device);
@@ -64,12 +69,23 @@ struct EllpackDeviceAccessor {
return gidx;
}
template <bool is_cat>
__device__ uint32_t SearchBin(float value, size_t column_id) const {
auto beg = feature_segments[column_id];
auto end = feature_segments[column_id + 1];
auto it =
thrust::upper_bound(thrust::seq, gidx_fvalue_map.cbegin()+ beg, gidx_fvalue_map.cbegin() + end, value);
uint32_t idx = it - gidx_fvalue_map.cbegin();
uint32_t idx = 0;
if (is_cat) {
auto it = dh::MakeTransformIterator<bst_cat_t>(
gidx_fvalue_map.cbegin(), [](float v) { return common::AsCat(v); });
idx = thrust::lower_bound(thrust::seq, it + beg, it + end,
common::AsCat(value)) -
it;
} else {
auto it = thrust::upper_bound(thrust::seq, gidx_fvalue_map.cbegin() + beg,
gidx_fvalue_map.cbegin() + end, value);
idx = it - gidx_fvalue_map.cbegin();
}
if (idx == end) {
idx -= 1;
}
@@ -134,10 +150,12 @@ class EllpackPageImpl {
explicit EllpackPageImpl(DMatrix* dmat, const BatchParam& parm);
template <typename AdapterBatch>
explicit EllpackPageImpl(AdapterBatch batch, float missing, int device, bool is_dense, int nthread,
explicit EllpackPageImpl(AdapterBatch batch, float missing, int device,
bool is_dense, int nthread,
common::Span<size_t> row_counts_span,
common::Span<FeatureType const> feature_types,
size_t row_stride, size_t n_rows, size_t n_cols,
common::HistogramCuts const& cuts);
common::HistogramCuts const &cuts);
/*! \brief Copy the elements of the given ELLPACK page into this page.
*
@@ -176,7 +194,9 @@ class EllpackPageImpl {
* not found). */
size_t NumSymbols() const { return cuts_.TotalBins() + 1; }
EllpackDeviceAccessor GetDeviceAccessor(int device) const;
EllpackDeviceAccessor
GetDeviceAccessor(int device,
common::Span<FeatureType const> feature_types = {}) const;
private:
/*!

8
src/data/iterative_device_dmatrix.cu
View File

@@ -148,9 +148,13 @@ void IterativeDeviceDMatrix::Initialize(DataIterHandle iter_handle, float missin
return GetRowCounts(value, row_counts_span, get_device(), missing);
});
auto is_dense = this->IsDense();
proxy->Info().feature_types.SetDevice(get_device());
auto d_feature_types = proxy->Info().feature_types.ConstDeviceSpan();
auto new_impl = Dispatch(proxy, [&](auto const &value) {
return EllpackPageImpl(value, missing, get_device(), is_dense, nthread,
row_counts_span, row_stride, rows, cols, cuts);
return EllpackPageImpl(value, missing, get_device(), is_dense, nthread,
row_counts_span, d_feature_types, row_stride, rows,
cols, cuts);
});
size_t num_elements = page_->Impl()->Copy(get_device(), &new_impl, offset);
offset += num_elements;