Rust-ML WG Meeting 00035

Meeting Info

Date: 2022-01-14

Start time: 1100ET

Some smaller updates on Linfa
- Removed reqwest from dependencies
- Integrating NN into AppxDbscan
Looking into using Linfa without BLAS systems libraries
- ndarray-linalg is, in particular, affected
  - Largely using the Cholesky decomposition
- Could use nalgebra to provide access to the methods that are currently gatekeeping that switch
  - https://docs.rs/nalgebra/0.3.2/src/nalgebra/.cargo/registry/src/github.com-1ecc6299db9ec823/nalgebra-0.3.2/src/linalg/decompositions.rs.html#264-306
- Could also create ndarray-based lib, using nalgebra's implementations, to provide those methods
  - Create secondary libary doing this first, then trying to upstream afterwards?
- Want to use upstream ndarray-linalg traits, which shouldn't create linker error without LAPACK
Should we think about splitting linfa-clustering into individual crates?
- Doesn't largely effect the dependency tree, but might affect build times
- Something similar about Nearest Neighbors
Yuhan did initial review of the PR for Multinomial Naive Bayes
- Suggested using traits rather than enums
- Lorenz is going to take a look as well, especially on the actual algorithm implementation
Doing benchmarks with linfa-python
- https://github.com/LukeMathWalker/linfa-python
- Using PyO3
- Switching from poetry to maturin per PyO3
- Get started with KMeans
- chrism

chrism will open issue around adding Python bindings and benchmarks to the linfa repo
Lorenz to take a look at the Multinomial Naive Bayes PR as well, with a focus on the actual algorithm implementation