On Sat, Oct 29, 2011 at 12:50 AM, Gael Varoquaux
<[email protected]> wrote:

> Could you give an example of what you have in mind? I am probably just
> revealing my lack of knowledge here.

I already gave 2 examples: early stopping (tuning the number of
iterations) and regularization parameter tuning (e.g. using Bottou's
trick for SGD). In general, any kind of hyperparameter tuning, which
can be broken do into a part which is common to each split and a part
which is specific to each split, is inefficient with a generic
GridSearchCV, as GridSearchCV recomputes the common part again and
again (of course joblib may help). I think that Olivier's idea of
having percentage_val in the constructor is reasonable and it doesn't
change the fit API.

Regarding GridSearchValidation. My feeling here is that just to avoid
adding a new helper class or function, you're bending the GridSearchCV
API in a way which is really user unfriendly. At first glance, I'm not
really convinced by your tricks: they seem hard to read and to require
fairly advance knowledge of what's going on in GridSearchCV. In
constrast, Olivier's IPython snippet speaks for itself.

Digression. In my experience, in machine learning, generic code can be
much slower than specific code: you just cannot beat an algorithm
which takes advantage of a problem's peculiarities. For example, a
generic one-vs-rest implementation is probably near optimal if the
base classifier is a linear one but it will be really sub-optimal if
the base classifier is a kernel one. In the kernel case, I could
design my outer and inner loops to maximize kernel cache reuse across
the classes.

I should really go back to work, otherwise there's no point staying
late in the lab on a friday night :)

Cheers,
Mathieu

------------------------------------------------------------------------------
The demand for IT networking professionals continues to grow, and the
demand for specialized networking skills is growing even more rapidly.
Take a complimentary Learning@Cisco Self-Assessment and learn 
about Cisco certifications, training, and career opportunities. 
http://p.sf.net/sfu/cisco-dev2dev
_______________________________________________
Scikit-learn-general mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general

Reply via email to