>
> On Thu, Apr 11, 2013 at 12:33 PM, Gael Varoquaux <
> [email protected]> wrote:
>
>> On Thu, Apr 11, 2013 at 10:45:23AM -0400, James Bergstra wrote:
>> > * Develop a SearchCV-style wrapper around as much of the scikit as
>> > possible, in order to provide very high-level base estimators like
>> > "VectorClassification" that search through the configurations of single
>> > or multiple algorithms
>>
>> Before that, could we not have a GridSearchCV replacement that is a bit
>> more clever? Ie something a bit generic with regard to the estimator? Of
>> course it won't achieve the same efficiency as objects dedicated to
>> specific models, but it would be very useful, and too many people are
>> stuck on the GridSearchCV.
>
>
I'm still not sure what you mean here by "more clever" (or "a bit
generic"). There seem to be a few issues:

* more sophisticated search algorithms, rather than the predetermined
sequences currently used. This is easy to implement given a search function
that generates sequences in response to previous outputs (as in
https://github.com/jnothman/scikit-learn/tree/param_search_callback). It's
perhaps not so easy to make efficient, or interactive (as in checkpointing
and resuming).
* more expressive parameter setting. This may be largely solved by
https://github.com/scikit-learn/scikit-learn/pull/1769 which allows one to
switch the underlying estimator using a pipeline.
* something that can exploit properties of the estimator to perform the
search efficiently, in particular to reuse fit models when changing
parameter settings that only affect prediction.
* underlying assumptions like the use of fit() then score() that is most
appropriate in supervised settings.

But no, I will not be eligible as a GSoC candidate.
------------------------------------------------------------------------------
Precog is a next-generation analytics platform capable of advanced
analytics on semi-structured data. The platform includes APIs for building
apps and a phenomenal toolset for data science. Developers can use
our toolset for easy data analysis & visualization. Get a free account!
http://www2.precog.com/precogplatform/slashdotnewsletter
_______________________________________________
Scikit-learn-general mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general

Reply via email to