Thanks to visit codestin.com
Credit goes to github.com

Skip to content

Pass pairwise attribute through MetaEstimatorMixin if present. #5393

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Closed
black-puppydog wants to merge 1 commit into from
Closed

Pass pairwise attribute through MetaEstimatorMixin if present. #5393

black-puppydog wants to merge 1 commit into from

Conversation

@black-puppydog
Copy link

@black-puppydog black-puppydog commented Oct 13, 2015

I raised this issue a while back on the mailing list.
When using GridSearchCV on a MetaEstimator wrapping an SVC (not LinearSVC) with a precomputed=True the Gram matrix does not get sliced correctly.
For me this was occurring because the default One-vs-One strategy for the SVC is not what I want.
This patch tells GridSearchCV that the input needs to be sliced along both dimensions.

As stated before, I am not deep enough into the code base to see if this will have any drawbacks but for me this has been working well for some time and it does not change the test results i get when calling make.

@amueller
Copy link
Member

amueller commented Oct 13, 2015

Which MetaEstimator? OneVsRest? I guess there is no harm in doing it here... But then it should be removed from the meta-estimators like GridSearchCV. And you should add a test fro OneVsRest if that was the problem.

@jnothman
Copy link
Member

jnothman commented Oct 13, 2015

We haven't explicitly stated that "estimator" is the right param...

On 14 October 2015 at 06:54, Andreas Mueller [email protected]
wrote:

Which MetaEstimator? OneVsRest? I guess there is no harm in doing it
here... But then it should be removed from the meta-estimators like
GridSearchCV. And you should add a test fro OneVsRest if that was the
problem.

—
Reply to this email directly or view it on GitHub
#5393 (comment)
.

@jnothman
Copy link
Member

jnothman commented Oct 13, 2015

I mean that all MetaEstimatorMixin users have a param named estimator.

On 14 October 2015 at 07:18, Joel Nothman [email protected] wrote:

We haven't explicitly stated that "estimator" is the right param...

On 14 October 2015 at 06:54, Andreas Mueller [email protected]
wrote:

Which MetaEstimator? OneVsRest? I guess there is no harm in doing it
here... But then it should be removed from the meta-estimators like
GridSearchCV. And you should add a test fro OneVsRest if that was the
problem.

—
Reply to this email directly or view it on GitHub
#5393 (comment)
.

@amueller
Copy link
Member

amueller commented Oct 13, 2015

whoops good catch.

@black-puppydog
Copy link
Author

black-puppydog commented Oct 14, 2015

Yes, I am using OneVsRest, to get around the default OneVsOne strategy in SVC.
I don't quite understand what you mean with removing it from GridSearchCV.
GridSearchCV itself also uses the MetaEstimatorMixin so with this patch it would also announce that the wrapped (Meta-)Estimator uses a pairwise matrix.
Is that a bad idea?

I'll have a look on how the tests are written.

@amueller
Copy link
Member

amueller commented Oct 14, 2015

Disregard my previous comment, @jnothman pointed out hat your fix doesn't really work. You should add it in OneVsRestClassifier and OneVsOneClassifier not MetaEstimatorMixin.

@amueller amueller added this to the 0.17 milestone Oct 14, 2015
@lesteve lesteve modified the milestones: 0.17, 0.18 Jul 27, 2016
@jnothman jnothman mentioned this pull request Aug 31, 2016
Closed
@amueller
Copy link
Member

amueller commented Sep 25, 2016

fixed in #7350.

@amueller amueller closed this Sep 25, 2016
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

0.18

Development

Successfully merging this pull request may close these issues.

4 participants

@black-puppydog @amueller @jnothman @lesteve