1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
|
/**
* Copyright 2018-2024, XGBoost Contributors
*/
#include <gtest/gtest.h>
#include <xgboost/host_device_vector.h>
#include <xgboost/tree_updater.h>
#include <cstddef> // for size_t
#include <string>
#include <vector>
#include "../../../src/tree/common_row_partitioner.h"
#include "../../../src/tree/hist/expand_entry.h" // for MultiExpandEntry, CPUExpandEntry
#include "../collective/test_worker.h" // for TestDistributedGlobal
#include "../helpers.h"
#include "test_column_split.h" // for TestColumnSplit
#include "test_partitioner.h"
#include "xgboost/data.h"
namespace xgboost::tree {
namespace {
template <typename ExpandEntry>
void TestPartitioner(bst_target_t n_targets) {
std::size_t n_samples = 1024, base_rowid = 0;
bst_feature_t n_features = 1;
Context ctx;
ctx.InitAllowUnknown(Args{});
CommonRowPartitioner partitioner{&ctx, n_samples, base_rowid, false};
ASSERT_EQ(partitioner.base_rowid, base_rowid);
ASSERT_EQ(partitioner.Size(), 1);
ASSERT_EQ(partitioner.Partitions()[0].Size(), n_samples);
auto Xy = RandomDataGenerator{n_samples, n_features, 0}.GenerateDMatrix(true);
std::vector<ExpandEntry> candidates{{0, 0}};
candidates.front().split.loss_chg = 0.4;
auto cuts = common::SketchOnDMatrix(&ctx, Xy.get(), 64);
for (auto const& page : Xy->GetBatches<SparsePage>()) {
GHistIndexMatrix gmat(page, {}, cuts, 64, true, 0.5, ctx.Threads());
bst_feature_t const split_ind = 0;
common::ColumnMatrix column_indices;
column_indices.InitFromSparse(page, gmat, 0.5, ctx.Threads());
{
auto min_value = gmat.cut.MinValues()[split_ind];
RegTree tree{n_targets, n_features};
CommonRowPartitioner partitioner{&ctx, n_samples, base_rowid, false};
if constexpr (std::is_same_v<ExpandEntry, CPUExpandEntry>) {
GetSplit(&tree, min_value, &candidates);
} else {
GetMultiSplitForTest(&tree, min_value, &candidates);
}
partitioner.UpdatePosition<false, true>(&ctx, gmat, column_indices, candidates, &tree);
ASSERT_EQ(partitioner.Size(), 3);
ASSERT_EQ(partitioner[1].Size(), 0);
ASSERT_EQ(partitioner[2].Size(), n_samples);
}
{
CommonRowPartitioner partitioner{&ctx, n_samples, base_rowid, false};
auto ptr = gmat.cut.Ptrs()[split_ind + 1];
float split_value = gmat.cut.Values().at(ptr / 2);
RegTree tree{n_targets, n_features};
if constexpr (std::is_same_v<ExpandEntry, CPUExpandEntry>) {
GetSplit(&tree, split_value, &candidates);
} else {
GetMultiSplitForTest(&tree, split_value, &candidates);
}
partitioner.UpdatePosition<false, true>(&ctx, gmat, column_indices, candidates, &tree);
{
auto left_nidx = tree.LeftChild(RegTree::kRoot);
auto const& elem = partitioner[left_nidx];
ASSERT_LT(elem.Size(), n_samples);
ASSERT_GT(elem.Size(), 1);
for (auto& it : elem) {
auto value = gmat.cut.Values().at(gmat.index[it]);
ASSERT_LE(value, split_value);
}
}
{
auto right_nidx = tree.RightChild(RegTree::kRoot);
auto const& elem = partitioner[right_nidx];
for (auto& it : elem) {
auto value = gmat.cut.Values().at(gmat.index[it]);
ASSERT_GT(value, split_value);
}
}
}
}
}
} // anonymous namespace
TEST(QuantileHist, Partitioner) { TestPartitioner<CPUExpandEntry>(1); }
TEST(QuantileHist, MultiPartitioner) { TestPartitioner<MultiExpandEntry>(3); }
namespace {
template <typename ExpandEntry>
void VerifyColumnSplitPartitioner(bst_target_t n_targets, size_t n_samples,
bst_feature_t n_features, size_t base_rowid,
std::shared_ptr<DMatrix> Xy, float min_value, float mid_value,
CommonRowPartitioner const& expected_mid_partitioner) {
auto dmat =
std::unique_ptr<DMatrix>{Xy->SliceCol(collective::GetWorldSize(), collective::GetRank())};
Context ctx;
ctx.InitAllowUnknown(Args{});
std::vector<ExpandEntry> candidates{{0, 0}};
candidates.front().split.loss_chg = 0.4;
auto cuts = common::SketchOnDMatrix(&ctx, dmat.get(), 64);
for (auto const& page : Xy->GetBatches<SparsePage>()) {
GHistIndexMatrix gmat(page, {}, cuts, 64, true, 0.5, ctx.Threads());
common::ColumnMatrix column_indices;
column_indices.InitFromSparse(page, gmat, 0.5, ctx.Threads());
{
RegTree tree{n_targets, n_features};
CommonRowPartitioner partitioner{&ctx, n_samples, base_rowid, true};
if constexpr (std::is_same_v<ExpandEntry, CPUExpandEntry>) {
GetSplit(&tree, min_value, &candidates);
} else {
GetMultiSplitForTest(&tree, min_value, &candidates);
}
partitioner.UpdatePosition<false, true>(&ctx, gmat, column_indices, candidates, &tree);
ASSERT_EQ(partitioner.Size(), 3);
ASSERT_EQ(partitioner[1].Size(), 0);
ASSERT_EQ(partitioner[2].Size(), n_samples);
}
{
RegTree tree{n_targets, n_features};
CommonRowPartitioner partitioner{&ctx, n_samples, base_rowid, true};
if constexpr (std::is_same_v<ExpandEntry, CPUExpandEntry>) {
GetSplit(&tree, mid_value, &candidates);
} else {
GetMultiSplitForTest(&tree, mid_value, &candidates);
}
auto left_nidx = tree.LeftChild(RegTree::kRoot);
partitioner.UpdatePosition<false, true>(&ctx, gmat, column_indices, candidates, &tree);
{
auto const& elem = partitioner[left_nidx];
ASSERT_LT(elem.Size(), n_samples);
ASSERT_GT(elem.Size(), 1);
auto const& expected_elem = expected_mid_partitioner[left_nidx];
ASSERT_EQ(elem.Size(), expected_elem.Size());
for (auto it = elem.begin(), eit = expected_elem.begin(); it != elem.end(); ++it, ++eit) {
ASSERT_EQ(*it, *eit);
}
}
{
auto right_nidx = tree.RightChild(RegTree::kRoot);
auto const& elem = partitioner[right_nidx];
auto const& expected_elem = expected_mid_partitioner[right_nidx];
ASSERT_EQ(elem.Size(), expected_elem.Size());
for (auto it = elem.begin(), eit = expected_elem.begin(); it != elem.end(); ++it, ++eit) {
ASSERT_EQ(*it, *eit);
}
}
}
}
}
template <typename ExpandEntry>
void TestColumnSplitPartitioner(bst_target_t n_targets) {
std::size_t n_samples = 1024, base_rowid = 0;
bst_feature_t n_features = 16;
auto Xy = RandomDataGenerator{n_samples, n_features, 0}.GenerateDMatrix(true);
std::vector<ExpandEntry> candidates{{0, 0}};
candidates.front().split.loss_chg = 0.4;
Context ctx;
ctx.InitAllowUnknown(Args{});
auto cuts = common::SketchOnDMatrix(&ctx, Xy.get(), 64);
float min_value, mid_value;
CommonRowPartitioner mid_partitioner{&ctx, n_samples, base_rowid, false};
for (auto const& page : Xy->GetBatches<SparsePage>()) {
GHistIndexMatrix gmat(page, {}, cuts, 64, true, 0.5, ctx.Threads());
bst_feature_t const split_ind = 0;
common::ColumnMatrix column_indices;
column_indices.InitFromSparse(page, gmat, 0.5, ctx.Threads());
min_value = gmat.cut.MinValues()[split_ind];
auto ptr = gmat.cut.Ptrs()[split_ind + 1];
mid_value = gmat.cut.Values().at(ptr / 2);
RegTree tree{n_targets, n_features};
if constexpr (std::is_same_v<ExpandEntry, CPUExpandEntry>) {
GetSplit(&tree, mid_value, &candidates);
} else {
GetMultiSplitForTest(&tree, mid_value, &candidates);
}
mid_partitioner.UpdatePosition<false, true>(&ctx, gmat, column_indices, candidates, &tree);
}
auto constexpr kWorkers = 4;
collective::TestDistributedGlobal(kWorkers, [&] {
VerifyColumnSplitPartitioner<ExpandEntry>(n_targets, n_samples, n_features, base_rowid, Xy,
min_value, mid_value, mid_partitioner);
});
}
} // anonymous namespace
TEST(QuantileHist, PartitionerColumnSplit) { TestColumnSplitPartitioner<CPUExpandEntry>(1); }
TEST(QuantileHist, MultiPartitionerColumnSplit) { TestColumnSplitPartitioner<MultiExpandEntry>(3); }
namespace {
class TestHistColumnSplit : public ::testing::TestWithParam<std::tuple<bst_target_t, bool, float>> {
public:
void Run() {
auto [n_targets, categorical, sparsity] = GetParam();
TestColumnSplit(n_targets, categorical, "grow_quantile_histmaker", sparsity);
}
};
} // anonymous namespace
TEST_P(TestHistColumnSplit, Basic) { this->Run(); }
INSTANTIATE_TEST_SUITE_P(ColumnSplit, TestHistColumnSplit, ::testing::ValuesIn([]() {
std::vector<std::tuple<bst_target_t, bool, float>> params;
for (auto categorical : {true, false}) {
for (auto sparsity : {0.0f, 0.6f}) {
for (bst_target_t n_targets : {1u, 3u}) {
params.emplace_back(n_targets, categorical, sparsity);
}
}
}
return params;
}()));
} // namespace xgboost::tree
|
