bench
HNSW
faiss/benchs/bench_hnsw.py
本程序用于对比不同索引算法在 SIFT1M 数据集上的性能表现，帮助用户评估搜索速度与准确率。测试方法包括 HNSW、IVF 等多种常见索引类型。
使用步骤
-
基本命令格式
python 程序名.py [k值] [测试方法1] [测试方法2]... -
参数说明
- k值 (必需): 指定每个查询要返回的最近邻数量
- 测试方法 (可选): 指定要运行的算法测试，支持以下选项：
- hnsw (HNSW 平面索引)
- hnsw_sq (量化HNSW)
- ivf (IVF 平面索引)
- ivf_hnsw_quantizer (HNSW量化的IVF)
- kmeans (K均值基准测试)
- kmeans_hnsw (HNSW辅助的K均值)
- nsg (NSG 图索引)
若不指定测试方法，默认运行所有测试
-
运行示例
# 返回10个最近邻，测试HNSW和IVF python demo.py 10 hnsw ivf # 返回5个最近邻，测试所有方法 python demo.py 5 -
输出说明 程序运行后会显示每个测试方法的以下指标：
- ms per query: 每个查询的平均耗时（毫秒）
- R@1: 前1个结果的召回率（准确率）
- missing rate: 无效结果比例（-1表示未找到）
示例输出：
Testing HNSW Flat add... search... efSearch 16 bounded_queue True 0.153 ms per query, R@1 0.9210, missing rate 0.0000 efSearch 16 bounded_queue False 0.162 ms per query, R@1 0.9210, missing rate 0.0000 ...
Note
Note
在代码中，missing rate 是通过以下方式计算的：
missing_rate = (I == -1).sum() / float(k * nq)
I: 是搜索结果的索引矩阵，形状为(nq, k)，其中nq是查询的数量，k是每个查询返回的最近邻数量。- 如果某个查询的某个最近邻未被找到，
I中对应的值会被设置为-1。
- 如果某个查询的某个最近邻未被找到，
(I == -1).sum(): 统计所有查询结果中未找到的最近邻数量（即I中值为-1的元素总数）。k * nq: 表示理论上应该返回的总结果数量（每个查询返回k个结果，共有nq个查询）。missing_rate: 未找到的结果占总结果的比例。
直观上看，missing rate表现了要搜索k个最近邻但是并没有找到那么多的最近邻的情况。
运行结果
以下是对实验结果的详细分析，按测试方法分类总结性能指标（搜索速度、召回率、缺失率）和参数影响：
HNSW Flat (分层可导航小世界图)
| efSearch | Bounded Queue | 耗时/查询 | R@1 | 缺失率 |
|---|---|---|---|---|
| 16 | True | 0.006ms | 0.9125 | 0% |
| 16 | False | 0.006ms | 0.9124 | 0% |
| 32 | True | 0.010ms | 0.9620 | 0% |
| 32 | False | 0.010ms | 0.9620 | 0% |
| 64 | True | 0.017ms | 0.9868 | 0% |
| 128 | True | 0.029ms | 0.9966 | 0% |
| 256 | True | 0.055ms | 0.9992 | 0% |
- 速度与召回率的权衡：
efSearch增大时，召回率显著提升（从 0.9125 到 0.9992），但搜索耗时线性增加（从 0.006ms 到 0.055ms）。- 推荐参数：
- 高精度场景：
efSearch=256（R@1=99.92%），耗时 0.055ms/查询。 - 平衡场景：
efSearch=64（R@1=98.68%），耗时 0.017ms/查询。
- 高精度场景：
- Bounded Queue 影响：
- Bounded Queue 开启与否对召回率和速度几乎无影响，可默认开启以节省内存。
HNSW with Scalar Quantizer (标量量化)
| efSearch | 耗时/查询 | R@1 | 缺失率 |
|---|---|---|---|
| 16 | 0.002ms | 0.7993 | 0% |
| 32 | 0.003ms | 0.8877 | 0% |
| 64 | 0.006ms | 0.9453 | 0% |
| 128 | 0.010ms | 0.9780 | 0% |
| 256 | 0.017ms | 0.9932 | 0% |
- 量化对性能的影响：
- 相比 HNSW Flat，量化后搜索速度更快（
efSearch=256时 0.017ms vs 0.055ms），但召回率下降（99.32% vs 99.92%）。 - 适用场景：
- 内存敏感场景：量化可减少内存占用，但需接受轻微精度损失。
- 相比 HNSW Flat，量化后搜索速度更快（
- 参数选择：
- 若需接近原始精度，选择
efSearch=256（R@1=99.32%）。
- 若需接近原始精度，选择
IVF Flat (倒排文件索引)
| nprobe | 耗时/查询 | R@1 | 缺失率 |
|---|---|---|---|
| 1 | 0.004ms | 0.4139 | 3.0% |
| 4 | 0.005ms | 0.6368 | 0% |
| 16 | 0.014ms | 0.8322 | 0% |
| 64 | 0.046ms | 0.9562 | 0% |
| 256 | 0.132ms | 0.9961 | 0% |
- nprobe 的重要性：
nprobe=1时缺失率 3%，召回率仅 41.39%，表明搜索覆盖的聚类中心不足。nprobe=16时召回率 83.22%，耗时 0.014ms，性价比最高。
- 与 HNSW 对比：
- 相同召回率下（如 99%），IVF 耗时更高（0.132ms vs HNSW 的 0.055ms）。
IVF with HNSW Quantizer
| nprobe | 耗时/查询 | R@1 | 缺失率 |
|---|---|---|---|
| 1 | 0.003ms | 0.4095 | 3.16% |
| 4 | 0.005ms | 0.6324 | 0% |
| 16 | 0.013ms | 0.8291 | 0% |
| 64 | 0.041ms | 0.9540 | 0% |
| 256 | 0.136ms | 0.9922 | 0% |
- HNSW Quantizer 的影响：
- 性能与传统 IVF 几乎一致，但训练速度可能更快（代码中未展示具体时间）。
- 仍需设置
nprobe≥16以避免高缺失率。
KMeans 聚类测试
-
Baseline KMeans：
- 目标函数值
3.8508e+10，聚类不均衡度1.227。
- 目标函数值
-
HNSW-Assisted KMeans：
- 目标函数值略高（
3.85126e+10），但训练时间显著减少（24.75s vs 53.52s）。
- 目标函数值略高（
-
HNSW 加速效果：
- HNSW 辅助的 KMeans 将训练时间减少 50%+，适合大规模数据聚类。
NSG (导航小世界图)
| search_L | 耗时/查询 | R@1 | 缺失率 |
|---|---|---|---|
| -1 | 0.004ms | 0.8150 | 0% |
| 16 | 0.006ms | 0.8852 | 0% |
| 32 | 0.009ms | 0.9533 | 0% |
| 64 | 0.015ms | 0.9832 | 0% |
| 128 | 0.027ms | 0.9959 | 0% |
| 256 | 0.050ms | 0.9994 | 0% |
- search_L 的影响：
search_L=256时召回率 99.94%，与 HNSW Flat（99.92%）相当，但耗时稍高（0.050ms vs 0.055ms）。- 优势：NSG 在中等参数下（如
search_L=64）即可达到 98.32% 召回率，适合对精度和速度均衡要求较高的场景。
综合对比与推荐
| 索引类型 | 最佳参数 | 耗时/查询 | R@1 | 适用场景 |
|---|---|---|---|---|
| HNSW Flat | efSearch=64 | 0.017ms | 98.68% | 高精度实时搜索 |
| HNSW SQ | efSearch=256 | 0.017ms | 99.32% | 内存敏感场景 |
| IVF Flat | nprobe=64 | 0.046ms | 95.62% | 高吞吐量批量查询 |
| NSG | search_L=64 | 0.015ms | 98.32% | 均衡精度与速度 |
| IVF+HNSQ Quant | nprobe=64 | 0.041ms | 95.40% | 需要快速训练的 IVF 场景 |
- 精度优先：选择 HNSW Flat（efSearch=256） 或 NSG（search_L=256）。
- 速度优先：选择 HNSW SQ（efSearch=16） 或 IVF Flat（nprobe=16）。
- 内存敏感：使用 HNSW SQ 或 IVF 索引。
- 训练时间敏感：尝试 HNSW-Assisted KMeans 加速聚类。
实验输出
实验输出
(faiss_env) gpu@gpu-node:~/faiss/faiss/benchs$ python bench_hnsw.py 10
load data
Testing HNSW Flat
add
hnsw_add_vertices: adding 1000000 elements on top of 0 (preset_levels=0)
max_level = 4
Adding 1 elements at level 4
Adding 26 elements at level 3
Adding 951 elements at level 2
Adding 30276 elements at level 1
Adding 968746 elements at level 0
Done in 10194.740 ms
search
efSearch 16 bounded queue True 0.006 ms per query, R@1 0.9125, missing rate 0.0000
efSearch 16 bounded queue False 0.006 ms per query, R@1 0.9124, missing rate 0.0000
efSearch 32 bounded queue True 0.010 ms per query, R@1 0.9620, missing rate 0.0000
efSearch 32 bounded queue False 0.010 ms per query, R@1 0.9620, missing rate 0.0000
efSearch 64 bounded queue True 0.017 ms per query, R@1 0.9868, missing rate 0.0000
efSearch 64 bounded queue False 0.017 ms per query, R@1 0.9868, missing rate 0.0000
efSearch 128 bounded queue True 0.029 ms per query, R@1 0.9966, missing rate 0.0000
efSearch 128 bounded queue False 0.030 ms per query, R@1 0.9966, missing rate 0.0000
efSearch 256 bounded queue True 0.055 ms per query, R@1 0.9992, missing rate 0.0000
efSearch 256 bounded queue False 0.055 ms per query, R@1 0.9992, missing rate 0.0000
Testing HNSW with a scalar quantizer
training
add
hnsw_add_vertices: adding 1000000 elements on top of 0 (preset_levels=0)
max_level = 5
Adding 1 elements at level 5
Adding 15 elements at level 4
Adding 194 elements at level 3
Adding 3693 elements at level 2
Adding 58500 elements at level 1
Adding 937597 elements at level 0
Done in 3322.349 ms
search
efSearch 16 0.002 ms per query, R@1 0.7993, missing rate 0.0000
efSearch 32 0.003 ms per query, R@1 0.8877, missing rate 0.0000
efSearch 64 0.006 ms per query, R@1 0.9453, missing rate 0.0000
efSearch 128 0.010 ms per query, R@1 0.9780, missing rate 0.0000
efSearch 256 0.017 ms per query, R@1 0.9932, missing rate 0.0000
Testing IVF Flat (baseline)
training
Training level-1 quantizer
Training level-1 quantizer on 100000 vectors in 128D
Training IVF residual
IndexIVF: no residual training
add
IndexIVFFlat::add_core: added 1000000 / 1000000 vectors
search
nprobe 1 0.004 ms per query, R@1 0.4139, missing rate 0.0300
nprobe 4 0.005 ms per query, R@1 0.6368, missing rate 0.0000
nprobe 16 0.014 ms per query, R@1 0.8322, missing rate 0.0000
nprobe 64 0.046 ms per query, R@1 0.9562, missing rate 0.0000
nprobe 256 0.132 ms per query, R@1 0.9961, missing rate 0.0000
Testing IVF Flat with HNSW quantizer
training
Training level-1 quantizer
Training L2 quantizer on 100000 vectors in 128D
Adding centroids to quantizer
Training IVF residual
IndexIVF: no residual training
add
IndexIVFFlat::add_core: added 1000000 / 1000000 vectors
search
nprobe 1 0.003 ms per query, R@1 0.4095, missing rate 0.0316
nprobe 4 0.005 ms per query, R@1 0.6324, missing rate 0.0000
nprobe 16 0.013 ms per query, R@1 0.8291, missing rate 0.0000
nprobe 64 0.041 ms per query, R@1 0.9540, missing rate 0.0000
nprobe 256 0.136 ms per query, R@1 0.9922, missing rate 0.0000
Performing kmeans on sift1M database vectors (baseline)
Clustering 1000000 points in 128D to 16384 clusters, redo 1 times, 10 iterations
Preprocessing in 0.15 s
Iteration 9 (53.52 s, search 53.23 s): objective=3.8508e+10 imbalance=1.227 nsplit=0
Performing kmeans on sift1M using HNSW assignment
Clustering 1000000 points in 128D to 16384 clusters, redo 1 times, 10 iterations
Preprocessing in 0.15 s
Iteration 9 (24.75 s, search 24.04 s): objective=3.85126e+10 imbalance=1.227 nsplit=0
Testing NSG Flat
add
IndexNSG::add 1000000 vectors
Build knn graph with NNdescent S=10 R=100 L=114 niter=10
Parameters: K=64, S=10, R=100, L=114, iter=10
Iter: 0, recall@64: 0.000000
Iter: 1, recall@64: 0.002813
Iter: 2, recall@64: 0.024688
Iter: 3, recall@64: 0.117656
Iter: 4, recall@64: 0.334219
Iter: 5, recall@64: 0.579063
Iter: 6, recall@64: 0.759375
Iter: 7, recall@64: 0.867188
Iter: 8, recall@64: 0.928594
Iter: 9, recall@64: 0.959063
Added 1000000 points into the index
Check the knn graph
nsg building
NSG::build R=32, L=64, C=132
Degree Statistics: Max = 32, Min = 1, Avg = 20.055180
Attached nodes: 0
search
search_L -1 0.004 ms per query, R@1 0.8150, missing rate 0.0000
search_L 16 0.006 ms per query, R@1 0.8852, missing rate 0.0000
search_L 32 0.009 ms per query, R@1 0.9533, missing rate 0.0000
search_L 64 0.015 ms per query, R@1 0.9832, missing rate 0.0000
search_L 128 0.027 ms per query, R@1 0.9959, missing rate 0.0000
search_L 256 0.050 ms per query, R@1 0.9994, missing rate 0.0000
运行结果（WSL）
HNSW Flat 性能
| 参数 | 构建时间 (ms) | 搜索时间 (ms/query) | R@1 | 缺失率 |
|---|---|---|---|---|
| efSearch=16 (bounded) | 28,095.8 | 0.015 | 0.9069 | 0.0000 |
| efSearch=16 (unbounded) | 28,095.8 | 0.015 | 0.9069 | 0.0000 |
| efSearch=32 (bounded) | 28,095.8 | 0.023 | 0.9614 | 0.0000 |
| efSearch=32 (unbounded) | 28,095.8 | 0.025 | 0.9614 | 0.0000 |
| efSearch=64 (bounded) | 28,095.8 | 0.043 | 0.9868 | 0.0000 |
| efSearch=64 (unbounded) | 28,095.8 | 0.041 | 0.9868 | 0.0000 |
| efSearch=128 (bounded) | 28,095.8 | 0.069 | 0.9964 | 0.0000 |
| efSearch=256 (bounded) | 28,095.8 | 0.126 | 0.9995 | 0.0000 |
- 搜索效率与召回率正相关：
efSearch值越大，搜索时间增加，但召回率显著提升（如efSearch=256时 R@1 达 0.9995）。 - 队列类型影响小：
bounded与unbounded队列对搜索性能无明显差异。
HNSW 标量量化（SQ）性能
| 参数 | 构建时间 (ms) | 搜索时间 (ms/query) | R@1 | 缺失率 |
|---|---|---|---|---|
| efSearch=16 | 12,376.8 | 0.007 | 0.7884 | 0.0000 |
| efSearch=32 | 12,376.8 | 0.011 | 0.8827 | 0.0000 |
| efSearch=64 | 12,376.8 | 0.020 | 0.9419 | 0.0000 |
| efSearch=128 | 12,376.8 | 0.033 | 0.9751 | 0.0000 |
| efSearch=256 | 12,376.8 | 0.055 | 0.9912 | 0.0000 |
- 量化显著加速构建：构建时间缩短至 12,376.8 ms（相比 HNSW Flat 的 28,095.8 ms）。
- 召回率略降：相同
efSearch下，量化后 R@1 低于原始 HNSW（如efSearch=256时 R@1=0.9912 vs. 0.9995）。
IVF Flat 基准性能
| 参数 | 搜索时间 (ms/query) | R@1 | 缺失率 |
|---|---|---|---|
| nprobe=1 | 0.022 | 0.4139 | 0.0300 |
| nprobe=4 | 0.018 | 0.6368 | 0.0000 |
| nprobe=16 | 0.035 | 0.8322 | 0.0000 |
| nprobe=64 | 0.103 | 0.9562 | 0.0000 |
| nprobe=256 | 0.332 | 0.9961 | 0.0000 |
- 需高
nprobe保证召回率：nprobe=256时 R@1 接近 HNSW，但搜索时间显著增加（0.332 ms vs. HNSW 的 0.126 ms）。
NSG Flat 性能
| 参数 | 搜索时间 (ms/query) | R@1 | 缺失率 |
|---|---|---|---|
| search_L=-1 | 0.008 | 0.8152 | 0.0000 |
| search_L=16 | 0.011 | 0.8849 | 0.0000 |
| search_L=32 | 0.017 | 0.9528 | 0.0000 |
| search_L=64 | 0.032 | 0.9843 | 0.0000 |
| search_L=128 | 0.051 | 0.9959 | 0.0000 |
| search_L=256 | 0.086 | 0.9993 | 0.0000 |
- 高效搜索与高召回率：在
search_L=256时，R@1 达 0.9993，且搜索时间（0.086 ms）优于 HNSW Flat 的 0.126 ms。
综合对比
| 索引类型 | 构建时间 (ms) | 最佳 R@1 | 对应搜索时间 (ms) |
|---|---|---|---|
| HNSW Flat | 28,095.8 | 0.9995 | 0.126 |
| HNSW SQ | 12,376.8 | 0.9912 | 0.055 |
| IVF Flat | - | 0.9961 | 0.332 |
| NSG Flat | - | 0.9993 | 0.086 |
- HNSW 系列：
- 适合高召回率场景，但构建时间较长；量化版本（HNSW SQ）可显著加速构建，但牺牲少量召回率。
- IVF Flat：
- 需要高
nprobe达到高召回率，但搜索效率较低。
- 需要高
- NSG Flat：
- 在搜索速度与召回率间取得最佳平衡（
search_L=256时 R@1=0.9993，搜索时间仅 0.086 ms）。
- 在搜索速度与召回率间取得最佳平衡（
- 实时搜索：优先选择 HNSW SQ 或 NSG Flat；
- 高精度需求：选择 HNSW Flat 或 NSG Flat；
- 快速构建：选择 HNSW SQ。
实验输出
实验输出
(faiss_env) root@Quaternijkon:~/faiss/benchs# python bench_hnsw.py 10
load data
Testing HNSW Flat
add
hnsw_add_vertices: adding 1000000 elements on top of 0 (preset_levels=0)
max_level = 4
Adding 1 elements at level 4
Adding 26 elements at level 3
Adding 951 elements at level 2
Adding 30276 elements at level 1
Adding 968746 elements at level 0
Done in 28095.800 ms
search
efSearch 16 bounded queue True 0.015 ms per query, R@1 0.9069, missing rate 0.0000
efSearch 16 bounded queue False 0.015 ms per query, R@1 0.9069, missing rate 0.0000
efSearch 32 bounded queue True 0.023 ms per query, R@1 0.9614, missing rate 0.0000
efSearch 32 bounded queue False 0.025 ms per query, R@1 0.9614, missing rate 0.0000
efSearch 64 bounded queue True 0.043 ms per query, R@1 0.9868, missing rate 0.0000
efSearch 64 bounded queue False 0.041 ms per query, R@1 0.9868, missing rate 0.0000
efSearch 128 bounded queue True 0.069 ms per query, R@1 0.9964, missing rate 0.0000
efSearch 128 bounded queue False 0.069 ms per query, R@1 0.9964, missing rate 0.0000
efSearch 256 bounded queue True 0.126 ms per query, R@1 0.9995, missing rate 0.0000
efSearch 256 bounded queue False 0.122 ms per query, R@1 0.9995, missing rate 0.0000
Testing HNSW with a scalar quantizer
training
add
hnsw_add_vertices: adding 1000000 elements on top of 0 (preset_levels=0)
max_level = 5
Adding 1 elements at level 5
Adding 15 elements at level 4
Adding 194 elements at level 3
Adding 3693 elements at level 2
Adding 58500 elements at level 1
Adding 937597 elements at level 0
Done in 12376.822 ms
search
efSearch 16 0.007 ms per query, R@1 0.7884, missing rate 0.0000
efSearch 32 0.011 ms per query, R@1 0.8827, missing rate 0.0000
efSearch 64 0.020 ms per query, R@1 0.9419, missing rate 0.0000
efSearch 128 0.033 ms per query, R@1 0.9751, missing rate 0.0000
efSearch 256 0.055 ms per query, R@1 0.9912, missing rate 0.0000
Testing IVF Flat (baseline)
training
Training level-1 quantizer
Training level-1 quantizer on 100000 vectors in 128D
Training IVF residual
IndexIVF: no residual training
add
IndexIVFFlat::add_core: added 1000000 / 1000000 vectors
search
nprobe 1 0.022 ms per query, R@1 0.4139, missing rate 0.0300
nprobe 4 0.018 ms per query, R@1 0.6368, missing rate 0.0000
nprobe 16 0.035 ms per query, R@1 0.8322, missing rate 0.0000
nprobe 64 0.103 ms per query, R@1 0.9562, missing rate 0.0000
nprobe 256 0.332 ms per query, R@1 0.9961, missing rate 0.0000
Testing IVF Flat with HNSW quantizer
training
Training level-1 quantizer
Training L2 quantizer on 100000 vectors in 128D
Adding centroids to quantizer
Training IVF residual
IndexIVF: no residual training
add
IndexIVFFlat::add_core: added 1000000 / 1000000 vectors
search
nprobe 1 0.009 ms per query, R@1 0.4102, missing rate 0.0323
nprobe 4 0.014 ms per query, R@1 0.6327, missing rate 0.0000
nprobe 16 0.030 ms per query, R@1 0.8288, missing rate 0.0000
nprobe 64 0.094 ms per query, R@1 0.9545, missing rate 0.0000
nprobe 256 0.332 ms per query, R@1 0.9925, missing rate 0.0000
Performing kmeans on sift1M database vectors (baseline)
Clustering 1000000 points in 128D to 16384 clusters, redo 1 times, 10 iterations
Preprocessing in 0.07 s
Iteration 0 (36.15 s, search 36.12 s): objective=5.58803e+10 imbalance=1.737 nsplit=5 Iteration 1 (71.32 s, search 71.25 s): objective=4.11771e+10 imbalance=1.378 nsplit=0 Iteration 2 (107.25 s, search 107.15 s): objective=3.9914e+10 imbalance=1.301 nsplit= Iteration 3 (142.60 s, search 142.47 s): objective=3.93716e+10 imbalance=1.269 nsplit Iteration 4 (178.17 s, search 178.01 s): objective=3.9066e+10 imbalance=1.253 nsplit= Iteration 5 (214.30 s, search 214.11 s): objective=3.8872e+10 imbalance=1.243 nsplit= Iteration 6 (250.13 s, search 249.92 s): objective=3.87377e+10 imbalance=1.237 nsplit Iteration 7 (285.57 s, search 285.33 s): objective=3.86397e+10 imbalance=1.232 nsplit Iteration 8 (321.46 s, search 321.19 s): objective=3.85653e+10 imbalance=1.229 nsplit Iteration 9 (357.11 s, search 356.81 s): objective=3.8508e+10 imbalance=1.227 nsplit=0
Performing kmeans on sift1M using HNSW assignment
Clustering 1000000 points in 128D to 16384 clusters, redo 1 times, 10 iterations
Preprocessing in 0.07 s
Iteration 0 (11.13 s, search 10.96 s): objective=5.58887e+10 imbalance=1.737 nsplit=0 Iteration 1 (21.37 s, search 21.07 s): objective=4.11817e+10 imbalance=1.379 nsplit=2 Iteration 2 (31.33 s, search 30.89 s): objective=3.99504e+10 imbalance=1.301 nsplit=0 Iteration 3 (41.08 s, search 40.47 s): objective=3.93767e+10 imbalance=1.269 nsplit=2 Iteration 4 (50.91 s, search 50.17 s): objective=3.90675e+10 imbalance=1.253 nsplit=1 Iteration 5 (60.60 s, search 59.72 s): objective=3.88936e+10 imbalance=1.243 nsplit=0 Iteration 6 (70.48 s, search 69.46 s): objective=3.87889e+10 imbalance=1.236 nsplit=0 Iteration 7 (80.54 s, search 79.37 s): objective=3.8648e+10 imbalance=1.232 nsplit=1 Iteration 8 (90.33 s, search 89.03 s): objective=3.85683e+10 imbalance=1.229 nsplit=3 Iteration 9 (100.14 s, search 98.69 s): objective=3.85264e+10 imbalance=1.227 nsplit=1
Testing NSG Flat
add
IndexNSG::add 1000000 vectors
Build knn graph with NNdescent S=10 R=100 L=114 niter=10
Parameters: K=64, S=10, R=100, L=114, iter=10
Iter: 0, recall@64: 0.000000
Iter: 1, recall@64: 0.002969
Iter: 2, recall@64: 0.024531
Iter: 3, recall@64: 0.110781
Iter: 4, recall@64: 0.332656
Iter: 5, recall@64: 0.559375
Iter: 6, recall@64: 0.751094
Iter: 7, recall@64: 0.862188
Iter: 8, recall@64: 0.922656
Iter: 9, recall@64: 0.953906
Added 1000000 points into the index
Check the knn graph
nsg building
NSG::build R=32, L=64, C=132
Degree Statistics: Max = 32, Min = 1, Avg = 20.056442
Attached nodes: 0
search
search_L -1 0.008 ms per query, R@1 0.8152, missing rate 0.0000
search_L 16 0.011 ms per query, R@1 0.8849, missing rate 0.0000
search_L 32 0.017 ms per query, R@1 0.9528, missing rate 0.0000
search_L 64 0.032 ms per query, R@1 0.9843, missing rate 0.0000
search_L 128 0.051 ms per query, R@1 0.9959, missing rate 0.0000
search_L 256 0.086 ms per query, R@1 0.9993, missing rate 0.0000