Should support training

[ralfbiedert]

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.

[rth]

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.

[ralfbiedert]

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 ...

[fumoboy007]

Support for training would be awesome!

Do you think the speedup would be similar to prediction? 2.5×–14× faster training would be crazy.

[ralfbiedert]

Do you think the speedup would be similar to prediction? 2.5×–14× faster training would be crazy.

I haven't looked into SVM training in a couple of years, so this is hard to tell. Generally, if libSVM training is similar (on a CPU profiling level) to classification, some speedup should be possible.