What distinguishes Unda from other Rust packages?

[yoderj]

In reviewing other "neural network" packages higher in the list of rust ML libraries, sorted by"best", it appears this may be the "best" library that is focused on "general-purpose" ML which is written entirely in rust rather than wrapping an external library. The only other library that I saw that seems to seek the same goal is darjeeling.

Is this correct? Is this what differentiates this library, and are you aware of other Rust libraries with similar goals?

[BradenEverson]

Thank you for asking! Unda actually does wrap the XLA library using a Rust crate xla-rs https://github.com/LaurentMazare/xla-rs, but I believe that you're more referring to wrapping libraries such as TensorFlow. Unda is currently in a pretty beta stage of development, as we are rearchitecting the library to utilize XLA for GPU/CPU acceleration and automatic differentiation. The version you can find on crates.io right now is a fairly unoptomized CPU solution, and in the future with the next stable release the new XLA changes will become accessible instead.

That being said, I do know of a few other libraries that achieve what we are doing as well. Most notably is Burn. From what I've seen Burn is a general Rust solution and I would recommend giving them a check out. As Unda matures a little more, I would certainly recommend checking back.

I also noticed that you are an MSOE professor, I am a sophmore Computer Science student myself and would be happy to meet with you in person to possibly discuss this furhter!