Search nearest neighbour vectors in n-dimensional space with hashes. There are no dependencies in this package.
The algorithm is based on the assumption that two real numbers can be considered equal within certain equality distance. Then quantization is used for comparison. To make sure points near or at quantization borders are also comparable, a vector can be discretized into more than one hash, as described here (also as PDF). The method indirectly clusters given vectors by hypercubes.
- Provided a float vector []float64, use
CubeSetandCentralCubefunctions to generate hypercube coordinates []int. The difference between the two functions is that one corresponds to hash-table record and the other to a query or vice versa, depending on performance/memory preference. HashSetandDecimalHash/FNV1aHashare used to get corresponding hash set and central hash from the hypercube coordinates above. There are 2 alternative hash functions: DecimalHash and FNV1aHash. DecimalHash does not have collisions, but is not suitable for cases with large number of buckets or dimensions. FNV1aHash is applicable for all cases.
Example for similar image search and clustering.
Go doc for full code documentation.