-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathhnsw.cpp
199 lines (195 loc) · 7.97 KB
/
hnsw.cpp
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
#include "lib/DataSet.cpp"
#include "lib/Config.cpp"
#include "./res/hnswlib/hnswlib/hnswlib.h"
#include "./res/kmeans/kmeans.cpp"
std::string solveName(int dataset, int M, int ef_construction, int zero) {
std::string index = "./dataset/index/"
+ std::to_string(dataset)
+ "_" + std::to_string(M)
+ "_" + std::to_string(ef_construction)
+ "_" + std::to_string(zero) + ".bin";
return index;
}
std::vector<float> norm_vector(std::vector<float> data) {
std::vector<float> norm_array;
float norm = 0.0f;
for (auto & item : data)
norm += item * item;
norm = 1.0f / (sqrtf(norm) + 1e-30f);
for (auto & item : data)
norm_array.push_back(item * norm);
return norm_array;
}
int main(int argc, char **argv ){
if (argc != 3 && argc != 5) {
std::cout << "Usage: ./hnsw <M> <ef_construction> \n or ./hnsw <M> <ef_construction> <CLUSTER_NUM> <CLUSTER_CORE>" << std::endl;
return 0;
}
M = atoi(argv[1]);
ef_construction = atoi(argv[2]);
if (argc == 5) {
CLUSTER_NUM = atoi(argv[3]);
CLUSTER_CORE = atoi(argv[4]);
}
std::ios::sync_with_stdio(false);
std::cin.tie(0);
#if DatabaseSelect > 6
HDF5DataSet<FILETYPE> *ds = new HDF5DataSet<FILETYPE>(baseFileName);
#else
SIFTDataSet<FILETYPE> *ds = new SIFTDataSet<FILETYPE>(baseFileName, queryFileName, ansFileName);
#endif
DataSet<FILETYPE> *dataSet = ds;
// Initing index
hnswlib::L2Space space(D); // D Dimension of the elements
hnswlib::HierarchicalNSW<float>* alg_hnsw;
// cosin space
hnswlib::InnerProductSpace space_cos(D);
std::string index;
#ifdef ZERO
index = solveName(DatabaseSelect, M, ef_construction, 1);
#else
index = solveName(DatabaseSelect, M, ef_construction, 0);
#endif
std::ifstream file(index);
// std::cout << "index: " << index << std::endl;
if (file.is_open()) {
if (DatabaseSelect >= 10) {
alg_hnsw = new hnswlib::HierarchicalNSW<float>(&space_cos, index.c_str());
} else alg_hnsw = new hnswlib::HierarchicalNSW<float>(&space, index.c_str());
} else {
// maxbaseNum-> Maximum number of elements, should be known beforehand
if (DatabaseSelect >= 10) {
alg_hnsw = new hnswlib::HierarchicalNSW<float>(&space_cos, maxbaseNum, M, ef_construction);
} else alg_hnsw = new hnswlib::HierarchicalNSW<float>(&space, maxbaseNum, M, ef_construction);
for (int i = 0; i < dataSet->baseData.size(); i++) {
if (DatabaseSelect == 5) {
std::vector<float> temp;
for (auto & item : dataSet->baseData[i].vectors) {
temp.push_back(item);
}
alg_hnsw->addPoint(temp.data(), i);
} else if (DatabaseSelect >= 10) {
std::vector<float> temp = norm_vector(dataSet->baseData[i].vectors);
alg_hnsw->addPoint(temp.data(), i);
}else
alg_hnsw->addPoint(dataSet->baseData[i].vectors.data(), i);
}
alg_hnsw->saveIndex(index.c_str());
}
// Query the elements for themselves and measure recall
float correct = 0;
double allTime = 0;
#ifdef CLUSTER
int num_clusters = dataSet->queryData.size() / CLUSTER_NUM;
std::vector<ClusterItem> clusters;
std::vector<Point> points;
for (int i = 0; i < dataSet->queryData.size(); i++) {
std::vector<double> temp;
for (auto & item : dataSet->queryData[i].vectors) {
temp.push_back(item);
}
Point p(temp);
p.id = i;
points.push_back(p);
}
KMeans kmeans(num_clusters);
kmeans.init(points);
kmeans.run();
for (int i = 0; i < num_clusters; i++) {
clusters.emplace_back();
clusters[i].clusterId = i;
auto & temp = kmeans.getMeans()[i].data_;
for (auto & item : temp) {
clusters[i].clusterCenter.vectors.push_back(item);
}
}
for (auto & item : kmeans.getPoints()) {
clusters[item.cluster_].clusterItem.push_back(item.id);
clusters[item.cluster_].clusterNum++;
}
for (auto & cluster : clusters) {
// sort with -
std::sort(cluster.clusterItem.begin(), cluster.clusterItem.end(), [&](int a, int b) {
return dataSet->queryData[a] - cluster.clusterCenter < dataSet->queryData[b] - cluster.clusterCenter;
});
}
std::vector<QUERYANS> queryAns;
std::mutex add;
// #pragma omp parallel for num_threads(THREAD_CONFIG)
for (auto & cluster : clusters) {
std::vector<int> beginVector;
for (auto & i : cluster.clusterItem) {
std::priority_queue<std::pair<float, hnswlib::labeltype>> result;
auto start = std::chrono::steady_clock::now();
if (DatabaseSelect == 5) {
std::vector<float> temp;
for (auto & item : dataSet->queryData[i].vectors) {
temp.push_back(item);
}
result = alg_hnsw->searchKnn(temp.data(), K, nullptr, beginVector);
} else if (DatabaseSelect >= 10) {
std::vector<float> temp = norm_vector(dataSet->queryData[i].vectors);
result = alg_hnsw->searchKnn(temp.data(), K, nullptr, beginVector);
} else {
result = alg_hnsw->searchKnn(dataSet->queryData[i].vectors.data(), K, nullptr, beginVector);
}
auto end = std::chrono::steady_clock::now();
add.lock();
allTime += (end - start)/ 1us;
if (beginVector.size() < CLUSTER_CORE) beginVector.push_back(result.top().second);
while (!result.empty()) {
int ans = result.top().second;
result.pop();
for (auto & ansItem : dataSet->ansData[i].vectors) {
if (ans == ansItem) {
correct++;
break;
}
}
}
add.unlock();
}
}
#else
for (int i = 0; i < dataSet->queryData.size(); i++) {
std::priority_queue<std::pair<float, hnswlib::labeltype>> result;
auto start = std::chrono::steady_clock::now();
if (DatabaseSelect == 5) {
std::vector<float> temp;
for (auto & item : dataSet->queryData[i].vectors) {
temp.push_back(item);
}
result = alg_hnsw->searchKnn(temp.data(), K);
} else if (DatabaseSelect >= 10) {
std::vector<float> temp = norm_vector(dataSet->queryData[i].vectors);
result = alg_hnsw->searchKnn(temp.data(), K);
} else
result = alg_hnsw->searchKnn(dataSet->queryData[i].vectors.data(), K);
auto end = std::chrono::steady_clock::now();
allTime += (end - start)/ 1us;
while (!result.empty()) {
int ans = result.top().second;
result.pop();
for (auto & ansItem : dataSet->ansData[i].vectors) {
if (ans == ansItem) {
correct++;
break;
}
}
}
}
#endif
float recall = correct / (dataSet->queryData.size() * K);
std::cout
<< "Dataset=" << DatabaseSelect << "\t"
<< "M=" << M << "\t"
<< "ef_construction=" << ef_construction << "\t"
<< "CLUSTER_NUM=" << CLUSTER_NUM << "\t"
<< "CLUSTER_CORE=" << CLUSTER_CORE << "\t"
<< "Recall=" << recall << "\t"
<< "avgTime=" << allTime / dataSet->queryData.size() << "us"
<< std::endl;
delete alg_hnsw;
delete ds;
return 0;
}