[pymvpa] sphere_gnbsearchlight & Monte Carlo Testing

Yaroslav Halchenko debian at onerussian.com
Tue Oct 23 20:08:12 UTC 2012

Thanks for the patience... unfortunately for now (after staring at the
code for a while in the evening and then having Aha moment outside) I
just disabled (it would throw an exception) providing any partitioners
which would change the targets in gnb or m1nn adhoc-searchlights.  The
reason is simple - they were coded with an assumption of targets
being "constant" across partitioning schemes and that only partitions
would change...  That is why you were getting the same values -- since
all values in the dist_samples were the same as the original
classification performance.

I will try to get into fixing it "properly" later today/tomorrow. 

Meanwhile, on your data, could you confirm that there is some variance
in the obtained performances in a "good/lengthy" code, i.e. that values

np.std(distr_est.ca.dist_samples.samples[0], axis=1)

are not all 0s (or nearby)... either my testing dataset is too small or there
might be yet another bug hiding ;-)

On Mon, 22 Oct 2012, wolf zinke wrote:

> Hi Yaroslav,

> I hope you did survive the SfN well. If you have any ideas regarding
> the constant value I get with my Monte Carlo testing of the
> gnb_searchlight, I would be very happy. SO far I dod not try any
> further to solve this issue, but tried a common Searchlight in
> combination with Monte Carlo testing. And this takes an awful time
> to get some results...

> cheers,
> wolf

> On 10/11/2012 06:23 PM, Yaroslav Halchenko wrote:
> >On Thu, 11 Oct 2012, wolf zinke wrote:

> >>    Hi,
> >>    With some great input from Michael I figured out now, that contrary to the
> >>    common sphere_searchlight, the sphere_gnbsearchlight already implements a
> >sorry for being silent since I guess blame should be on me here:
> >sphere_gnbsearchlight and sphere_m1nnsearchlight indeed somewhat of a
> >crafty beasts to achieve super-fast performance with those "naive"
> >classification schemes -- they do splitting internally BUT should be
> >parametrized with a partitioner to be instructed on how to "split".

> >I will try looking in details of your case later on whenever rush of
> >preparation for sfn goes down... sorry about the delay

> >cheers

> _______________________________________________
> Pkg-ExpPsy-PyMVPA mailing list
> Pkg-ExpPsy-PyMVPA at lists.alioth.debian.org
> http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/pkg-exppsy-pymvpa

Yaroslav O. Halchenko
Postdoctoral Fellow,   Department of Psychological and Brain Sciences
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        

More information about the Pkg-ExpPsy-PyMVPA mailing list