[pymvpa] questions

[Yaroslav Halchenko]

On Wed, 23 Mar 2011, Jonas Kaplan wrote:
>  have successfully installed 0.6rc2 on OSX, which I consider a great achievement. :)

congratulations ;)

> I miss the old nperlabel option in the NFoldSplitter... is there
> an easy way to recreate this behavior with the new NFoldPartitioner? 

in the quest for the modularization and thus flexibility, all such
options were pulled out of *Splitters and have their own classes which
you would need to chain to achieve virtually any effect you like:

e.g. for nperlabel: Balancer
http://www.pymvpa.org/generated/mvpa.generators.resampling.Balancer.html

so, just chain the NFoldPartitioner an Balancer with something like:

            partitioner = ChainNode([NFoldPartitioner(), Balancer()],
                                     space='partitions')

Such chaining, although might be looking like an unnecessary inconvenience,
allows to greater flexibility, e.g. see
http://www.pymvpa.org/examples/permutation_test.html
for permutations generators

> Also, I've been wanting to give RFE a try... I have been reading
> through all the messages on this list about it and I *think* I am starting to
> understand the many layers.  However, all of the code examples seem to be
> written with the older syntax.  It would be helpful to see a working version of
oy -- which ones? 

I guess these guys:
$> ls -l doc/source/*txt
-rw-rw-r-- 1 yoh yoh  1990 Mar  5 15:36 doc/source/authors.txt
-rw-rw-r-- 1 yoh yoh 26629 Mar  4 17:38 doc/source/classifiers.txt
-rw-rw-r-- 1 yoh yoh 15677 Mar  4 17:38 doc/source/datasets.txt
-rw-rw-r-- 1 yoh yoh 10696 Mar  4 17:38 doc/source/featsel.txt
-rw-rw-r-- 1 yoh yoh   797 Mar  5 15:36 doc/source/mappers.txt
-rw-rw-r-- 1 yoh yoh  9099 Mar  5 15:36 doc/source/measures.txt

which managed to escape our daily testing because of old-fashioned .txt
extension (instead of .rst) so they were not part of the daily tests

we thought to convert (nearly) or deprecate all of them.

> RFE written under .6, does anything like that exist? 

well, if documentation might be lagging behind in covering the features, the
unittests are staying quite in sync;  so you could find plenty of constructs in
there with a simple query:

$> git grep RFE mvpa/tests
mvpa/tests/test_datameasure.py:from mvpa.featsel.rfe import RFE
mvpa/tests/test_datameasure.py:        # Proper RFE: aggregate sensitivities across multiple splits,
mvpa/tests/test_datameasure.py:        rfe = RFE(sensitivity_analyzer=
mvpa/tests/test_rfe.py:from mvpa.featsel.rfe import RFE
mvpa/tests/test_rfe.py:class RFETests(unittest.TestCase):
mvpa/tests/test_rfe.py:        rfe = RFE(sens_ana,
mvpa/tests/test_rfe.py:            RFE(rfesvm_split.get_sensitivity_analyzer(),
mvpa/tests/test_rfe.py:            # on features selected via RFE
mvpa/tests/test_rfe.py:            self.fail('CrossValidation cannot handle classifier with RFE '
mvpa/tests/test_rfe.py:            feature_selection = RFE(
mvpa/tests/test_rfe.py:    return unittest.makeSuite(RFETests)

could you have a look at mvpa/tests/test_rfe.py ?
if you do not have a local clone, then see
https://github.com/PyMVPA/PyMVPA/blob/master/mvpa/tests/test_rfe.py

-- 
=------------------------------------------------------------------=
Keep in touch                                     www.onerussian.com
Yaroslav Halchenko                 www.ohloh.net/accounts/yarikoptic