[pymvpa] Sensitivity analysis fails

[Sergi Costafreda]

Thanks for fixing this so quickly!

I installed pymvpa-0.4.1 from its tarball on a Mac Os X 10.5.6.

To refresh from your version in git would it be:
$ git clone git://git.debian.org/git/pkg-exppsy/pymvpa.git
?

Also, thanks for your comments on the interpretation of remapping sens into
original feature space:

> so, sens itself is mapped back into original feature space, but I am not
> sure what that would actually mean here (interpretations could vary)
> since it effectively replaces (IIRC) singular values with
> sensitivites, ie changes contribution of each of SVD components.

Could we not interpret it as which elements of the original feature space
are more influential on the classification?





On 05/02/2009 18:08, "Yaroslav Halchenko" <debian at onerussian.com> wrote:

> o o -- you've hit a bug ;) sorry about that
> 
> I pushed a fix in my (yoh/master) branch in our git repository,
> but what do you run it on (OS?) and from (package, git repository,
> source tarball)? would you be able to use pymvpa directly from the
> sources in git repository?
> 
> More about fix:
> 
> since this moment we have 2 separate kinds of analyzers (for
> regular mapped or featureselection classifiers), but that is irrelevant
> at this user level. But also I've added a state variable
> clf_sensitivities for sensitivity analyzers of proxy-classifiers, so
> now, in your case I have:
> 
> clf=MappedClassifier(LinearCSVMC(), SVDMapper())
> sensana=clf.getSensitivityAnalyzer()
> sensana.states.enable(['clf_sensitivities'])
> sens=sensana(dataset)
> print sens.shape
> print sensana.clf_sensitivities.shape
> 
> and if it is ran as a part of our svdclf example on sample dataset I
> see:
> 
> Dataset / float32 216 x 530 uniq: 12 chunks 2 labels
> (530,)
> (216, 1)
> 
> 
> so, sens itself is mapped back into original feature space, but I am not
> sure what that would actually mean here (interpretations could vary)
> since it effectively replaces (IIRC) singular values with
> sensitivites, ie changes contribution of each of SVD components.
> 
> but sensana.clf_sensitivities is smth easy to make sense of -- that
> would be the sensitivities of the 'slave' classifier (ie
> LinearCSVMC) which was ran on remapped data... from it you could assess
> what dimensions of eigenspace are more influential on classification
> than the others
> 
> On Thu, 05 Feb 2009, Sergi Costafreda wrote:
> 
>>    Hi all
>>    Just starting with PyMVPA - thanks for a great tool!
>>    I am trying to obtain a sensitivity analysis for a SVD+SVM mapped
>>    classifier, by adapting the code of sensanas.py. When I try to run
>>    (1) >>> clf=MappedClassifier(LinearCSVMC(), SVDMapper())
>>    (2) >>> sensana=clf.getSensitivityAnalyzer()
>>    (3) >>> sens=sensana(data)
>>    Line (2) returns the following error:
>>>>> "AttributeError: 'MappedClassifier' object has no attribute
>>    '_MappedClassifier__clf'"
>>    I'd be grateful for any help with this!
>>    Best,
>>    Sergi
> 
>> _______________________________________________
>> Pkg-ExpPsy-PyMVPA mailing list
>> Pkg-ExpPsy-PyMVPA at lists.alioth.debian.org
>> http://lists.alioth.debian.org/mailman/listinfo/pkg-exppsy-pymvpa
>