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

[Breaking] Accept multi-dim meta info. (#7405)

Browse Source 
This PR changes base_margin into a 3-dim array, with one of them being reserved for multi-target classification. Also, a breaking change is made for binary serialization due to extra dimension along with a fix for saving the feature weights. Lastly, it unifies the prediction initialization between CPU and GPU. After this PR, the meta info setter in Python will be based on array interface.
This commit is contained in:
Jiaming Yuan
2021-11-18 23:02:54 +08:00
committed by GitHub
parent 9fb4338964
commit d33854af1b
25 changed files with 545 additions and 256 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
tests/cpp/data/test_metainfo.cc
View File

@@ -1,4 +1,6 @@
// Copyright 2016-2020 by Contributors
// Copyright 2016-2021 by Contributors
#include "test_metainfo.h"
#include <dmlc/io.h>
#include <dmlc/filesystem.h>
#include <xgboost/data.h>
@@ -122,7 +124,10 @@ TEST(MetaInfo, SaveLoadBinary) {
EXPECT_EQ(inforead.labels_.HostVector(), info.labels_.HostVector());
EXPECT_EQ(inforead.group_ptr_, info.group_ptr_);
EXPECT_EQ(inforead.weights_.HostVector(), info.weights_.HostVector());
EXPECT_EQ(inforead.base_margin_.HostVector(), info.base_margin_.HostVector());
auto orig_margin = info.base_margin_.View(xgboost::GenericParameter::kCpuId);
auto read_margin = inforead.base_margin_.View(xgboost::GenericParameter::kCpuId);
EXPECT_TRUE(std::equal(orig_margin.cbegin(), orig_margin.cend(), read_margin.cbegin()));
EXPECT_EQ(inforead.feature_type_names.size(), kCols);
EXPECT_EQ(inforead.feature_types.Size(), kCols);
@@ -254,10 +259,10 @@ TEST(MetaInfo, Validate) {
xgboost::HostDeviceVector<xgboost::bst_group_t> d_groups{groups};
d_groups.SetDevice(0);
d_groups.DevicePointer(); // pull to device
auto arr_interface = xgboost::GetArrayInterface(&d_groups, 64, 1);
std::string arr_interface_str;
xgboost::Json::Dump(arr_interface, &arr_interface_str);
EXPECT_THROW(info.SetInfo("group", arr_interface_str), dmlc::Error);
std::string arr_interface_str{
xgboost::linalg::MakeVec(d_groups.ConstDevicePointer(), d_groups.Size(), 0)
.ArrayInterfaceStr()};
EXPECT_THROW(info.SetInfo("group", xgboost::StringView{arr_interface_str}), dmlc::Error);
#endif // defined(XGBOOST_USE_CUDA)
}
@@ -292,3 +297,7 @@ TEST(MetaInfo, HostExtend) {
ASSERT_EQ(lhs.group_ptr_.at(i), per_group * i);
}
}
namespace xgboost {
TEST(MetaInfo, CPUStridedData) { TestMetaInfoStridedData(GenericParameter::kCpuId); }
} // namespace xgboost

16
tests/cpp/data/test_metainfo.cu
View File

@@ -3,10 +3,13 @@
#include <gtest/gtest.h>
#include <xgboost/data.h>
#include <xgboost/json.h>
#include <xgboost/generic_parameters.h>
#include <thrust/device_vector.h>
#include "test_array_interface.h"
#include "../../../src/common/device_helpers.cuh"
#include "test_metainfo.h"
namespace xgboost {
template <typename T>
@@ -23,7 +26,7 @@ std::string PrepareData(std::string typestr, thrust::device_vector<T>* out, cons
std::vector<Json> j_shape {Json(Integer(static_cast<Integer::Int>(kRows)))};
column["shape"] = Array(j_shape);
column["strides"] = Array(std::vector<Json>{Json(Integer(static_cast<Integer::Int>(sizeof(T))))});
column["version"] = Integer(static_cast<Integer::Int>(1));
column["version"] = 3;
column["typestr"] = String(typestr);
auto p_d_data = d_data.data().get();
@@ -31,6 +34,7 @@ std::string PrepareData(std::string typestr, thrust::device_vector<T>* out, cons
Json(Integer(reinterpret_cast<Integer::Int>(p_d_data))),
Json(Boolean(false))};
column["data"] = j_data;
column["stream"] = nullptr;
Json array(std::vector<Json>{column});
std::string str;
@@ -49,6 +53,7 @@ TEST(MetaInfo, FromInterface) {
info.SetInfo("label", str.c_str());
auto const& h_label = info.labels_.HostVector();
ASSERT_EQ(h_label.size(), d_data.size());
for (size_t i = 0; i < d_data.size(); ++i) {
ASSERT_EQ(h_label[i], d_data[i]);
}
@@ -60,9 +65,10 @@ TEST(MetaInfo, FromInterface) {
}
info.SetInfo("base_margin", str.c_str());
auto const& h_base_margin = info.base_margin_.HostVector();
auto const h_base_margin = info.base_margin_.View(GenericParameter::kCpuId);
ASSERT_EQ(h_base_margin.Size(), d_data.size());
for (size_t i = 0; i < d_data.size(); ++i) {
ASSERT_EQ(h_base_margin[i], d_data[i]);
ASSERT_EQ(h_base_margin(i), d_data[i]);
}
thrust::device_vector<int> d_group_data;
@@ -76,6 +82,10 @@ TEST(MetaInfo, FromInterface) {
EXPECT_EQ(info.group_ptr_, expected_group_ptr);
}
TEST(MetaInfo, GPUStridedData) {
TestMetaInfoStridedData(0);
}
TEST(MetaInfo, Group) {
cudaSetDevice(0);
MetaInfo info;

82
tests/cpp/data/test_metainfo.h Normal file
View File

@@ -0,0 +1,82 @@
/*!
* Copyright 2021 by XGBoost Contributors
*/
#ifndef XGBOOST_TESTS_CPP_DATA_TEST_METAINFO_H_
#define XGBOOST_TESTS_CPP_DATA_TEST_METAINFO_H_
#include <gtest/gtest.h>
#include <xgboost/data.h>
#include <xgboost/host_device_vector.h>
#include <xgboost/linalg.h>
#include <numeric>
#include "../../../src/data/array_interface.h"
#include "../../../src/common/linalg_op.h"
namespace xgboost {
inline void TestMetaInfoStridedData(int32_t device) {
MetaInfo info;
{
// label
HostDeviceVector<float> labels;
labels.Resize(64);
auto& h_labels = labels.HostVector();
std::iota(h_labels.begin(), h_labels.end(), 0.0f);
bool is_gpu = device >= 0;
if (is_gpu) {
labels.SetDevice(0);
}
auto t = linalg::TensorView<float const, 2>{
is_gpu ? labels.ConstDeviceSpan() : labels.ConstHostSpan(), {32, 2}, device};
auto s = t.Slice(linalg::All(), 0);
auto str = s.ArrayInterfaceStr();
ASSERT_EQ(s.Size(), 32);
info.SetInfo("label", StringView{str});
auto const& h_result = info.labels_.HostVector();
ASSERT_EQ(h_result.size(), 32);
for (auto v : h_result) {
ASSERT_EQ(static_cast<int32_t>(v) % 2, 0);
}
}
{
// qid
linalg::Tensor<uint64_t, 2> qid;
qid.Reshape(32, 2);
auto& h_qid = qid.Data()->HostVector();
std::iota(h_qid.begin(), h_qid.end(), 0);
auto s = qid.View(device).Slice(linalg::All(), 0);
auto str = s.ArrayInterfaceStr();
info.SetInfo("qid", StringView{str});
auto const& h_result = info.group_ptr_;
ASSERT_EQ(h_result.size(), s.Size() + 1);
}
{
// base margin
linalg::Tensor<float, 4> base_margin;
base_margin.Reshape(4, 3, 2, 3);
auto& h_margin = base_margin.Data()->HostVector();
std::iota(h_margin.begin(), h_margin.end(), 0.0);
auto t_margin = base_margin.View(device).Slice(linalg::All(), linalg::All(), 0, linalg::All());
ASSERT_EQ(t_margin.Shape().size(), 3);
info.SetInfo("base_margin", StringView{t_margin.ArrayInterfaceStr()});
auto const& h_result = info.base_margin_.View(-1);
ASSERT_EQ(h_result.Shape().size(), 3);
auto in_margin = base_margin.View(-1);
linalg::ElementWiseKernelHost(h_result, omp_get_max_threads(), [&](size_t i, float v_0) {
auto tup = linalg::UnravelIndex(i, h_result.Shape());
auto i0 = std::get<0>(tup);
auto i1 = std::get<1>(tup);
auto i2 = std::get<2>(tup);
// Sliced at 3^th dimension.
auto v_1 = in_margin(i0, i1, 0, i2);
CHECK_EQ(v_0, v_1);
return v_0;
});
}
}
} // namespace xgboost
#endif // XGBOOST_TESTS_CPP_DATA_TEST_METAINFO_H_

10
tests/cpp/data/test_simple_dmatrix.cc
View File

@@ -253,8 +253,8 @@ TEST(SimpleDMatrix, Slice) {
std::iota(lower.begin(), lower.end(), 0.0f);
std::iota(upper.begin(), upper.end(), 1.0f);
auto& margin = p_m->Info().base_margin_.HostVector();
margin.resize(kRows * kClasses);
auto& margin = p_m->Info().base_margin_;
margin = linalg::Tensor<float, 3>{{kRows, kClasses}, GenericParameter::kCpuId};
std::array<int32_t, 3> ridxs {1, 3, 5};
std::unique_ptr<DMatrix> out { p_m->Slice(ridxs) };
@@ -284,10 +284,10 @@ TEST(SimpleDMatrix, Slice) {
ASSERT_EQ(p_m->Info().weights_.HostVector().at(ridx),
out->Info().weights_.HostVector().at(i));
auto& out_margin = out->Info().base_margin_.HostVector();
auto out_margin = out->Info().base_margin_.View(GenericParameter::kCpuId);
auto in_margin = margin.View(GenericParameter::kCpuId);
for (size_t j = 0; j < kClasses; ++j) {
auto in_beg = ridx * kClasses;
ASSERT_EQ(out_margin.at(i * kClasses + j), margin.at(in_beg + j));
ASSERT_EQ(out_margin(i, j), in_margin(ridx, j));
}
}
}

10
tests/cpp/data/test_simple_dmatrix.cu
View File

@@ -122,13 +122,13 @@ TEST(SimpleDMatrix, FromColumnarWithEmptyRows) {
col["data"] = j_data;
std::vector<Json> j_shape{Json(Integer(static_cast<Integer::Int>(kRows)))};
col["shape"] = Array(j_shape);
col["version"] = Integer(static_cast<Integer::Int>(1));
col["version"] = 3;
col["typestr"] = String("<f4");
// Construct the mask object.
col["mask"] = Object();
auto& j_mask = col["mask"];
j_mask["version"] = Integer(static_cast<Integer::Int>(1));
j_mask["version"] = 3;
auto& mask_storage = column_bitfields[i];
mask_storage.resize(16); // 16 bytes
@@ -220,7 +220,7 @@ TEST(SimpleCSRSource, FromColumnarSparse) {
for (size_t c = 0; c < kCols; ++c) {
auto& column = j_columns[c];
column = Object();
column["version"] = Integer(static_cast<Integer::Int>(1));
column["version"] = 3;
column["typestr"] = String("<f4");
auto p_d_data = raw_pointer_cast(columns_data[c].data());
std::vector<Json> j_data {
@@ -229,12 +229,12 @@ TEST(SimpleCSRSource, FromColumnarSparse) {
column["data"] = j_data;
std::vector<Json> j_shape {Json(Integer(static_cast<Integer::Int>(kRows)))};
column["shape"] = Array(j_shape);
column["version"] = Integer(static_cast<Integer::Int>(1));
column["version"] = 3;
column["typestr"] = String("<f4");
column["mask"] = Object();
auto& j_mask = column["mask"];
j_mask["version"] = Integer(static_cast<Integer::Int>(1));
j_mask["version"] = 3;
j_mask["data"] = std::vector<Json>{
Json(Integer(reinterpret_cast<Integer::Int>(column_bitfields[c].data().get()))),
Json(Boolean(false))};