Should support training

Open ralfbiedert opened this issue 7 years ago • 4 comments

Blocked by

[ ] Need high-performance matrix library that abstracts away sparse and dense (SIMD) matrices.
[ ] In turn blocked by num-traits abstraction to handle scalar and packed types / math transparently.

Sep 02 '18 14:09 ralfbiedert

Need high-performance matrix library that abstracts away sparse and dense (SIMD) matrices.

For sparse maybe https://github.com/vbarrielle/sprs ? They would be happy to accept SIMD improvements I think.

Apr 30 '20 16:04 rth

Sorry, I just now noticed your comment!

Random thoughts:

sprs has too many dependencies for my taste. If safe transmutes and packed SIMD are ever going to be a thing I'd probably want to have no dependencies for the core library
That said I could imagine training with extra dependencies under a feature flag
I probably won't implement training myself though; as I don't have a need for it
I'll happily accept PRs from anyone seriously wanting to add that

Edit - omg ... for 2 years that issue label was wrong ...

Jan 15 '21 22:01 ralfbiedert