Set `warm_start=True` by default about lightning HOT 7 CLOSED

This looks like a bug. warm_start=True is not supposed to be faster than warm_start=False when the estimator is fitted for the first time.

from lightning.

Oops. seems like my mind was clogged with the cv models in sklearn. Do you want me to have a look, if you confirm that there is something fishy?

from lightning.

I just remembered that the multiclass option is actually only used when penalty='l1/l2'. The multiclass option is currently silently ignored for other penalties. This is really bad, an exception needs to be raised instead. So your results above are actually for OVR logistic regression. But regardless, there is definitely something fishy regarding training time when warm_start=False. You can have a look but it's probably a tough bug. I was planning to redesign / simplify how warm_start works in lightning. Perhaps this is a good opportunity.

from lightning.

Okay. I've figured out why. This is because when %timeit is used in IPython across 1000 loops, it effectively warm starts!

I'm not sure if its a bug in IPython or not. Should the previous attributes be destroyed when %timeit is used, when doing multiple loops to make them independent of each other, or is the present behavior correct?

from lightning.

I opened a bug report in IPython ipython/ipython#6471. So closing this.

from lightning.

Indeed, good catch.

from lightning.

The consensus has been to print some sort of warning if such a thing happens. I've submitted a PR here, ipython/ipython#6479

from lightning.