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
234
|
/**
* Copyright 2021-2024, XGBoost contributors.
*/
#include <gtest/gtest.h>
#include <xgboost/tree_updater.h> // for TreeUpdater
#include <algorithm> // for transform
#include <memory> // for unique_ptr
#include <vector> // for vector
#include "../../../src/tree/common_row_partitioner.h"
#include "../../../src/tree/param.h" // for TrainParam
#include "../collective/test_worker.h" // for TestDistributedGlobal
#include "../helpers.h"
#include "test_column_split.h" // for TestColumnSplit
#include "test_partitioner.h"
#include "xgboost/tree_model.h" // for RegTree
namespace xgboost::tree {
namespace {
std::vector<float> GenerateHess(size_t n_samples) {
auto grad = GenerateRandomGradients(n_samples);
std::vector<float> hess(grad.Size());
std::transform(grad.HostVector().cbegin(), grad.HostVector().cend(), hess.begin(),
[](auto gpair) { return gpair.GetHess(); });
return hess;
}
} // anonymous namespace
TEST(Approx, Partitioner) {
size_t n_samples = 1024, n_features = 1, base_rowid = 0;
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 const Xy = RandomDataGenerator{n_samples, n_features, 0}.GenerateDMatrix(true);
auto hess = GenerateHess(n_samples);
std::vector<CPUExpandEntry> candidates{{0, 0}};
candidates.front().split.loss_chg = 0.4;
for (auto const& page : Xy->GetBatches<GHistIndexMatrix>(&ctx, {64, hess, true})) {
bst_feature_t const split_ind = 0;
{
auto min_value = page.cut.MinValues()[split_ind];
RegTree tree;
CommonRowPartitioner partitioner{&ctx, n_samples, base_rowid, false};
GetSplit(&tree, min_value, &candidates);
partitioner.UpdatePosition(&ctx, page, 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 = page.cut.Ptrs()[split_ind + 1];
float split_value = page.cut.Values().at(ptr / 2);
RegTree tree;
GetSplit(&tree, split_value, &candidates);
partitioner.UpdatePosition(&ctx, page, candidates, &tree);
{
auto left_nidx = tree[RegTree::kRoot].LeftChild();
auto const& elem = partitioner[left_nidx];
ASSERT_LT(elem.Size(), n_samples);
ASSERT_GT(elem.Size(), 1);
for (auto& it : elem) {
auto value = page.cut.Values().at(page.index[it]);
ASSERT_LE(value, split_value);
}
}
{
auto right_nidx = tree[RegTree::kRoot].RightChild();
auto const& elem = partitioner[right_nidx];
for (auto& it : elem) {
auto value = page.cut.Values().at(page.index[it]);
ASSERT_GT(value, split_value) << it;
}
}
}
}
}
TEST(Approx, InteractionConstraint) {
auto constexpr kRows = 32;
auto constexpr kCols = 16;
auto p_dmat = GenerateCatDMatrix(kRows, kCols, 0.6f, false);
Context ctx;
linalg::Matrix<GradientPair> gpair({kRows}, ctx.Device());
gpair.Data()->Copy(GenerateRandomGradients(kRows));
ObjInfo task{ObjInfo::kRegression};
{
// With constraints
RegTree tree{1, kCols};
std::unique_ptr<TreeUpdater> updater{TreeUpdater::Create("grow_histmaker", &ctx, &task)};
TrainParam param;
param.UpdateAllowUnknown(
Args{{"interaction_constraints", "[[0, 1]]"}, {"num_feature", std::to_string(kCols)}});
std::vector<HostDeviceVector<bst_node_t>> position(1);
updater->Configure(Args{});
updater->Update(¶m, &gpair, p_dmat.get(), position, {&tree});
ASSERT_EQ(tree.NumExtraNodes(), 4);
ASSERT_EQ(tree[0].SplitIndex(), 1);
ASSERT_EQ(tree[tree[0].LeftChild()].SplitIndex(), 0);
ASSERT_EQ(tree[tree[0].RightChild()].SplitIndex(), 0);
}
{
// Without constraints
RegTree tree{1u, kCols};
std::unique_ptr<TreeUpdater> updater{TreeUpdater::Create("grow_histmaker", &ctx, &task)};
std::vector<HostDeviceVector<bst_node_t>> position(1);
TrainParam param;
param.Init(Args{});
updater->Configure(Args{});
updater->Update(¶m, &gpair, p_dmat.get(), position, {&tree});
ASSERT_EQ(tree.NumExtraNodes(), 10);
ASSERT_EQ(tree[0].SplitIndex(), 1);
ASSERT_NE(tree[tree[0].LeftChild()].SplitIndex(), 0);
ASSERT_NE(tree[tree[0].RightChild()].SplitIndex(), 0);
}
}
namespace {
void TestColumnSplitPartitioner(size_t n_samples, size_t base_rowid, std::shared_ptr<DMatrix> Xy,
std::vector<float>* hess, float min_value, float mid_value,
CommonRowPartitioner const& expected_mid_partitioner) {
auto dmat =
std::unique_ptr<DMatrix>{Xy->SliceCol(collective::GetWorldSize(), collective::GetRank())};
std::vector<CPUExpandEntry> candidates{{0, 0}};
candidates.front().split.loss_chg = 0.4;
Context ctx;
ctx.InitAllowUnknown(Args{});
for (auto const& page : dmat->GetBatches<GHistIndexMatrix>(&ctx, {64, *hess, true})) {
{
RegTree tree;
CommonRowPartitioner partitioner{&ctx, n_samples, base_rowid, true};
GetSplit(&tree, min_value, &candidates);
partitioner.UpdatePosition(&ctx, page, 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, true};
RegTree tree;
GetSplit(&tree, mid_value, &candidates);
partitioner.UpdatePosition(&ctx, page, candidates, &tree);
{
auto left_nidx = tree[RegTree::kRoot].LeftChild();
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[RegTree::kRoot].RightChild();
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);
}
}
}
}
}
} // anonymous namespace
TEST(Approx, PartitionerColumnSplit) {
size_t n_samples = 1024, n_features = 16, base_rowid = 0;
auto const Xy = RandomDataGenerator{n_samples, n_features, 0}.GenerateDMatrix(true);
auto hess = GenerateHess(n_samples);
std::vector<CPUExpandEntry> candidates{{0, 0}};
candidates.front().split.loss_chg = 0.4;
float min_value, mid_value;
Context ctx;
ctx.InitAllowUnknown(Args{});
CommonRowPartitioner mid_partitioner{&ctx, n_samples, base_rowid, false};
for (auto const& page : Xy->GetBatches<GHistIndexMatrix>(&ctx, {64, hess, true})) {
bst_feature_t const split_ind = 0;
min_value = page.cut.MinValues()[split_ind];
auto ptr = page.cut.Ptrs()[split_ind + 1];
mid_value = page.cut.Values().at(ptr / 2);
RegTree tree;
GetSplit(&tree, mid_value, &candidates);
mid_partitioner.UpdatePosition(&ctx, page, candidates, &tree);
}
auto constexpr kWorkers = 4;
collective::TestDistributedGlobal(kWorkers, [&] {
TestColumnSplitPartitioner(n_samples, base_rowid, Xy, &hess, min_value, mid_value,
mid_partitioner);
});
}
namespace {
class TestApproxColumnSplit : public ::testing::TestWithParam<std::tuple<bool, float>> {
public:
void Run() {
auto [categorical, sparsity] = GetParam();
TestColumnSplit(1u, categorical, "grow_histmaker", sparsity);
}
};
} // namespace
TEST_P(TestApproxColumnSplit, Basic) { this->Run(); }
INSTANTIATE_TEST_SUITE_P(ColumnSplit, TestApproxColumnSplit, ::testing::ValuesIn([]() {
std::vector<std::tuple<bool, float>> params;
for (auto categorical : {true, false}) {
for (auto sparsity : {0.0f, 0.6f}) {
params.emplace_back(categorical, sparsity);
}
}
return params;
}()));
} // namespace xgboost::tree
|
