On Wed, Dec 3, 2014 at 10:26 AM, Alexandre Gramfort <
alexandre.gramf...@m4x.org> wrote:

> > A compromise would be to just implement the Cython routine in a separate
> > file, while sharing the same file for the pure Python side.
>
> sounds reasonable.
>
> > That said, using a separate class for Adagrad would allow to get rid of
> > irrelevant hyper-parameters.
>
> +1


> > Some code from the SGD module can probably be
> > factorized and OVR should ideally re-use functions from the multiclass
> > module.
>
> can you be more specific?
>

I think the SGD code could be refactored to call plain_sgd/average_sgd just
once throughout the entire file (stochastic_gradient.py) instead of twice
(once for regressor, once for classifier). Similar to what's being done in
my SAG branch. I think this would make the code a lot clearer. I haven't
looked at OVR but I'm sure it could benefit. I'm not sure if this was what
Mathieu was talking about though...


> >> It could later be improved and extended with schemes that also use
> feature
> >> specific learning rates.
> >
> > Do you have specific examples in mind? If not I would just call the class
> > AdaGrad. Early (name) optimization is the root of evil :)
>
> AdaDelta ? :)
>
> > PS: SGD is not strictly speaking a descent method so optimization people
> now
> > usually refer to it as just the stochastic gradient method
>
> sure.
>
> Alex
>
>
> ------------------------------------------------------------------------------
> Download BIRT iHub F-Type - The Free Enterprise-Grade BIRT Server
> from Actuate! Instantly Supercharge Your Business Reports and Dashboards
> with Interactivity, Sharing, Native Excel Exports, App Integration & more
> Get technology previously reserved for billion-dollar corporations, FREE
>
> http://pubads.g.doubleclick.net/gampad/clk?id=164703151&iu=/4140/ostg.clktrk
> _______________________________________________
> Scikit-learn-general mailing list
> Scikit-learn-general@lists.sourceforge.net
> https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
>
------------------------------------------------------------------------------
Download BIRT iHub F-Type - The Free Enterprise-Grade BIRT Server
from Actuate! Instantly Supercharge Your Business Reports and Dashboards
with Interactivity, Sharing, Native Excel Exports, App Integration & more
Get technology previously reserved for billion-dollar corporations, FREE
http://pubads.g.doubleclick.net/gampad/clk?id=164703151&iu=/4140/ostg.clktrk
_______________________________________________
Scikit-learn-general mailing list
Scikit-learn-general@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general

Reply via email to