search

NeuraLegion / shainet

SHAInet - a pure Crystal machine learning library
MIT License
183 stars 19 forks source link

Ameba fixes and Apatite::Matrix use #90

Closed rmarronnier closed 5 years ago

rmarronnier commented 5 years ago

I'm trying to learn about neural networks and this library is incredibly educational. Thanks for sharing this ! I couldn't help myself but made some fixes (mostly Ameba suggested) and replaced the PtrMatrix with the one from Apatite. I saw no performance issues and could make Shainet benefit from already written methods.

Hope you find this weird PR useful :-)

rmarronnier commented 5 years ago

I have no idea why the specs fail on Travis :/

bararchy commented 5 years ago

I guess we haven't really touched the CI part in a while, let me try to pass specs locally

bararchy commented 5 years ago

Seems it fails on the rprop run, not the most stable one for sure.... we might want to remote Rprop from specs as it's a bad optimzer anyway, @ArtLinkov what say you?

rmarronnier commented 5 years ago

Thanks for looking into this. All specs (except the ones already commented out in master) pass for me locally. It's strange that this one fails on Travis.

ArtLinkov commented 5 years ago

@rmarronnier first, let me say I'm happy you find SHAInet useful and thank you for contributing, Regarding the specs, if I'm not mistaken the rprop spec is the one that fails. In that case, you can just comment it out because it is quite volatile (converges fast but sometime to local minima) and doesn't work very well with small datasets

rmarronnier commented 5 years ago

Thanks @ArtLinkov for the explanation. I commented out the failing spec. I hope it's not too essential and that it was not a safeguard against any possible regression.

bararchy commented 5 years ago

@rmarronnier I believe it should be OK now. Merging.