[pymvpa] custom cross-validation procedure: train on individual blocks, test on averaged blocks?

[Yaroslav Halchenko]

now that it is morning I see it clearer ;)

so... it has nothing to do you with disbalanced set in your case -- it is
due to assigned by me unique 'custom' sample attribute per each sample in
training so they do not get meanned...

what I have ovelooked though that by the time measure obtains the
dataset for training or testing it is already split, so we can fulfill
your need in more straightforward way which would not confuse the
partitioner:

class MeanTesting(Mapper):
    def _forward_dataset(self, ds):
        # Here we get already split training or testing dataset
        # so it should have a single partition
        assert(len(ds.sa['partitions'].unique) == 1)
        if np.all(ds.sa.partitions == 1):
            # so we got training split
            # thus nothing really todo
            return ds
        else:
            # and for testing mean the beast per each target
            return mean_group_sample(['targets'])(ds)



cv2 = CrossValidation(ChainMapper(
                          [MeanTesting(),
                           CLASSIFIER], space='targets'),
                     NFoldPartitioner(1),
                     descr='custom-CV2')


is this one clearer?  results must be identical

>    print dataset.summary() yields:
>    Counts of targets in each chunk:
>      chunks\targets  HI  HO
>                     --- ---
>           0.0        3   2
-- 
=------------------------------------------------------------------=
Keep in touch                                     www.onerussian.com
Yaroslav Halchenko                 www.ohloh.net/accounts/yarikoptic