search

scikit-learn-contrib / lightning

Large-scale linear classification, regression and ranking in Python
https://contrib.scikit-learn.org/lightning/
1.73k stars 214 forks source link

ENH: allow to pass a custom Penalty object to FISTA #81

Closed fabianp closed 8 years ago

mblondel commented 8 years ago

Great, merged!

Did you need this for your research?

fabianp commented 8 years ago

Also, but this time its for some examples I'm preparing for pydata Paris next week.

Also, I've ported to numba Condat's super-fast fussed lasso proximal operator: https://gist.github.com/fabianp/8b62d905ae254fadf763cf1909904e3a . According to my benchmarks it's marginally faster than Condat's C code distributed with the paper.

What is your opinion on adding this proximal operator to lightning and a soft dependency on numba (i.e. without numba the code would still run but would be much slower) ? I think this would be a very nice addition to numba that is not found in other ML packages.

mblondel commented 8 years ago

It would be a nice addition but I would rather convert the code to Cython for now.

We can discuss making numba a dependency later.

fabianp commented 8 years ago

OK. Note that numba would be a soft dependency, ie the code would still run (slower) without it. On Jun 10, 2016 10:59 AM, "Mathieu Blondel" notifications@github.com wrote:

It would be a nice addition but I would rather convert the code to Cython for now.

We can discuss making numba a dependency later.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/scikit-learn-contrib/lightning/pull/81#issuecomment-225128860, or mute the thread https://github.com/notifications/unsubscribe/AAQ8h28CwXA3jL825IpPu33Zdpx-2RvRks5qKSdTgaJpZM4IyM5a .

mblondel commented 8 years ago

I know but the performance without Numba would be horrible.

In the future, we can try to port lightning to Numba.