Hi Andy,

according to [1] "The multiclass support is handled according to a
one-vs-one scheme."
That's why I was using the wrapper.
For now I am doing multi-class single-label, although that might change
in the future.
Maybe it is not the MetaEstimatorMixin that needs fixing but just the
docs? :P

Best
/Daan

[1]: http://scikit-learn.org/stable/modules/generated/sklearn.svm.SVC.html

On 11/09/15 16:43, Andreas Mueller wrote:
> Hi Daan.
> This looks like a bug, please open a pull request.
> Are you doing multi-label classification? For multi-class, SVC already 
> has built-in OvR support.
>
> Best,
> Andy
>
>
> On 09/11/2015 04:17 AM, Daan Wynen wrote:
>> Hi all,
>>
>> I recently was in the position of wanting to use a OneVsRest SVC for ~50
>> classes.
>> It turned out that this was not easily possible with sklearn because the
>> GridSearchCV class queries the classifier's _pairwise property to see if
>> the data is a square matrix and should be sliced in both dimensions or
>> if it is a matrix that only needs to be sliced in the first dimension.
>>
>> What works for me is adding these lines to the MetaEstimatorMixin in
>> ./base.py:
>>
>> class MetaEstimatorMixin(object):
>>      """Mixin class for all meta estimators in scikit-learn."""
>>     
>>      @property
>>      def _pairwise(self):
>>          # Used by cross_val_score.
>>          return getattr(self.estimator, "_pairwise", False)
>>
>>
>> OneVsRestClassifier uses this mixin which so far was just a flag.
>> Now it takes the property from the wrapped class.
>> Is there a good reason why this was not the case before?Any pitfalls I
>> don't see coming at the moment?
>>
>> Thanks for your input
>> /Daan
>>
>>
>> ------------------------------------------------------------------------------
>> _______________________________________________
>> Scikit-learn-general mailing list
>> Scikit-learn-general@lists.sourceforge.net
>> https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
>
> ------------------------------------------------------------------------------
> _______________________________________________
> Scikit-learn-general mailing list
> Scikit-learn-general@lists.sourceforge.net
> https://lists.sourceforge.net/lists/listinfo/scikit-learn-general


------------------------------------------------------------------------------
_______________________________________________
Scikit-learn-general mailing list
Scikit-learn-general@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general

Reply via email to