Nearest neighbors and approximate clustering about linfa

[LukeMathWalker]

I would definitely be interested in adding it, first of all, as a variant of the nearest neighbors algorithm (similarly to LSHForest in scikit-learn).

A couple of questions though concerning SIMD: how much of it leaks to the public interface? In other words, do I - as a user of the algorithm - have to be aware of SIMD-specific datatype when using the algorithm? Can that implementation complexity be hidden away from me?

[vadixidav]

@LukeMathWalker The reason that SIMD types are used in the public API as vectors for nearest neighbor is mostly due to alignment issues. By using the SIMD types I can ensure that Rust will always put them in vectors, structs, and the stack at the correct alignment. This could be hidden from the user, but I would need to specially take care of alignment issues manually in the code. I could create a wrapper type that just took in an array that creates a SIMD wrapper internally as necessary.

[LukeMathWalker]

That would be ideal @vadixidav - I'd like to avoid each model forcing the user to learn a different input type. It would significantly complicate the usage of the overall crate.

[vadixidav]

Alright, I am a bit focused on some other open source stuff currently, but I will come back to this and simplify that API. Technically it could be used now easily given a speed penalty for not using SIMD, but I can make wrappers so we can get speed without exposing the SIMD types.

[vadixidav]

Alright, I have done my best to make it as easy as possible to interact with arrays from HNSW. It now has extensive impls of From and Into, including From to convert from slices of u8 and f32. While you do have to worry about how many bytes or floats you have, you now no longer need to interact with SIMD types beyond adding the packed_simd crate so you can write HNSW<Euclidean<[u8x64; 2]>> and typing out array.into() to convert your array/slice types of f32 or u8 into Euclidean or Hamming types respectively.

https://docs.rs/hnsw/0.4.0/hnsw/struct.Hamming.html

If desired, I can even make type aliases to avoid needing the packed_simd crate for specifying which HNSW you want. That would not be a breaking change.

Let me know if you need anything else to be done.

[YuhanLiin]

Pretty sure this issue can be closed, since we already have #119 for nearest-neighbours.