mih-rs
Rust implementation of multi-index hashing (MIH) for neighbor searches on binary codes in the Hamming space, described in the paper
Norouzi, Punjani, and Fleet, Fast exact search in Hamming space with multi-index hashing, IEEE TPAMI, 36(6):1107– 1119, 2014.
As the benchmark result shows, on 10 million 64-bit codes, mih-rs
can perform top-k searches 19−94 times faster than linear search when k = 1..100.
Features
-
Two types of neighbor searches:
mih-rs
provides the two search operations:- Range search finds neighbor codes whose Hamming distances to a given code are within a radius.
- Top-K search finds the top-K codes that are closest to a given code.
-
Fast and memory-efficient implementation: The data structure is built on sparse hash tables, following the original implementation.
-
Parameter free:
mih-rs
automatically sets an optimal parameter of MIH depending on a given database (although you can also set this manually). -
Serialization:
mih-rs
supports to serialize/deserialize the index.
Example
use mih_rs::Index;
// Database of codes
let codes: Vec<u64> = vec![
0b1111111111111111111111011111111111111111111111111011101111111111, // #zeros = 3
0b1111111111111111111111111111111101111111111011111111111111111111, // #zeros = 2
0b1111111011011101111111111111111101111111111111111111111111111111, // #zeros = 4
0b1111111111111101111111111111111111111000111111111110001111111110, // #zeros = 8
0b1101111111111111111111111111111111111111111111111111111111111111, // #zeros = 1
0b1111111111111111101111111011111111111111111101001110111111111111, // #zeros = 6
0b1111111111111111111111111111111111101111111111111111011111111111, // #zeros = 2
0b1110110101011011011111111111111101111111111111111000011111111111, // #zeros = 11
];
// Query code
let qcode: u64 = 0b1111111111111111111111111111111111111111111111111111111111111111; // #zeros = 0
// Construct the index
let index = Index::new(codes).unwrap();
// Find the ids of neighbor codes whose Hamming distances are within 2
let mut searcher = index.range_searcher();
let answers = searcher.run(qcode, 2);
assert_eq!(answers, vec![1, 4, 6]);
// Find the ids of the top-4 nearest neighbor codes
let mut searcher = index.topk_searcher();
let answers = searcher.run(qcode, 4);
assert_eq!(answers, vec![4, 1, 6, 0]);
// Serialization/Deserialization
let mut data = vec![];
index.serialize_into(&mut data).unwrap();
let other = Index::<u64>::deserialize_from(&data[..]).unwrap();
assert_eq!(index, other);
Binary code types
mih_rs::Index
can be built from a vector of type mih_rs::CodeInt
that is a primitive integer trait supporting a popcount operation. Currently, this library defines mih_rs::CodeInt
for u8
, u16
, u32
, and u64
.
Benchmark
timeperf_topk.rs
offers the benchmark of top-K search for MIH and LinearSearch algorithms on binary code types u32
and u64
.
The following table shows the result of average search times in milliseconds per query, in the settings:
- Database: N random codes from a uniform distribution.
- Query set: 100 random codes from a uniform distribution.
- Machine: MacBook Pro (2019) of Quad-Core Intel Core i5 @2.4 GHz with 16 GB of RAM.
- Library version: v0.2.0
u32
Result for Algorithm | N=10,000 | N=100,000 | N=1,000,000 | N=10,000,000 |
---|---|---|---|---|
MIH (K=1) | 0.01 | 0.02 | 0.07 | 0.38 |
MIH (K=10) | 0.04 | 0.08 | 0.30 | 1.06 |
MIH (K=100) | 0.13 | 0.22 | 1.22 | 4.35 |
LinearSearch | 0.36 | 4.40 | 50.96 | 626.87 |
u64
Result for Algorithm | N=10,000 | N=100,000 | N=1,000,000 | N=10,000,000 |
---|---|---|---|---|
MIH (K=1) | 0.10 | 0.36 | 1.46 | 6.7 |
MIH (K=10) | 0.20 | 0.76 | 3.72 | 14.8 |
MIH (K=100) | 0.41 | 1.53 | 7.02 | 33.2 |
LinearSearch | 0.36 | 4.36 | 52.28 | 629.1 |
Licensing
This library is free software provided under MIT.