[pymvpa] Help with map2nifti()

[Yaroslav Halchenko]

On Mon, 04 Nov 2013, Taku Ito wrote:

>    You're right, thank you, I was looking through the incorrect data set - My
>    original data set (the one I wrote to right after fmri_dataset), has the
>    following mappers:
>    In [104]: data[1].a.mapper
>    Out[104]: ChainMapper(nodes=[FlattenMapper(shape=(59, 69, 59),
>    auto_train=True, space='voxel_indices'),
>    StaticFeatureSelection(dshape=(240189,), slicearg=array([False, False,
>    False, ..., False, False, False], dtype=bool))])
>    However, after this data set, I constructed a sub-data set by just
>    selecting certain samples with the target values I wanted, splitting the
>    datasets accordingly, and then reconcatenating them into a new data set.

that is where probably the mapper got "lost" ;-)

>     (I then ran the ZScoreMapper() on this new data set, since I only wanted
>    to Z score the samples I was going to use in the searchlight analysis.)
>    In [105]: data_stack[1].a.mapper
>    Out[105]: ZScoreMapper()
>    Finally, I ran the searchlight analysis which resulted in:
>    In [106]: res[1].a.mapper
>    Out[106]: ZScoreMapper()
>    Would it then be accurate for me to simply use the ChainMapper from the
>    original data set (exclude the ZScoreMapper from the map2nifti function),
>    like so:
>    map2nifti(data[1], res[1].samples)

yes -- should work

>    Or would it be more accurate of me to construct a new ChainMapper which
>    includes the ZScoreMapper, append it to the ChainMapper of the original
>    data set, and then re-attribute it to the searchlight output data set? 

nah -- probably not needed/not worth the hassle.  If you did no action
which would change 'features' origin, taking original dataset would be
the easiest and most straightforward 

>    Thanks again!  This has been a huge hassle, and you've been a great help
>    (though I realize that much of the troubles were my own carelessness
>    now... Sorry!).

well -- we should bear our portion of blame here too ;)  may be we
should issue a warning whenever hstacking/etc datasets with the mappers
so they do not get lost without notice (actually current master might
keep them... I would need to check).

Cheers,
-- 
Yaroslav O. Halchenko, Ph.D.
http://neuro.debian.net http://www.pymvpa.org http://www.fail2ban.org
Senior Research Associate,     Psychological and Brain Sciences Dept.
Dartmouth College, 419 Moore Hall, Hinman Box 6207, Hanover, NH 03755
Phone: +1 (603) 646-9834                       Fax: +1 (603) 646-1419
WWW:   http://www.linkedin.com/in/yarik