[pymvpa] Searchlights and additional measures
john.clithero at gmail.com
Sun Aug 23 16:36:44 UTC 2009
Thanks for the two ideas. I will certainly put #2 to task and get back to you.
Another naive (but somewhat related) question:
If I ran a searchlight analysis like the one I proposed, what would be
the best way to save the peak performing (or other voxels of interest)
voxel's coordinate/index and the index for the searchlight's
neighboring voxels? I could then, I suppose, rerun the classifier on
that searchlight ROI (without using the searchlight function) and
harvest other properties (e.g., weights and confusion matrix) that I
Again, many thanks for this, I am trying to learn quickly!
On Tue, Aug 18, 2009 at 11:08 PM, Yaroslav
Halchenko<debian at onerussian.com> wrote:
>> I am new to PyMVPA, but so far I am really enjoying it.
> Glad to hear it! Especially since you are coming with concrete questions
> on somewhat novel usecases, not with bug reports that something crashed
> upon you ;)
>> (2) Construct confusion matrix for each searchlight and save them all.
>> (3) Save averaged (across the folds) SVM weights for all of the searchlights.
> I could only wonder what you would do with those ;) but yeah -- we don't
> have such functionality available "out of the box".
> But there are 2 possible ways "to accomplish the mission"
> #1 with our intervention: adjust searchlight's __call__ to provide
> callbacks to your custom function or similar harvest_attribs,
> which would simply collect anything you like from results of running cv
> on each sphere
> #2 without our intervention: just look at the code of searchlight -- it
> is very very short and self-explanatory. So, you could craft that loop
> yourself by simply copy/pasting it and extending the way you like (can't
> resist giving a pride to benefits of scripting and open-source)
> I guess, in the long run, #1 is preferable, but #2 could be the faster
> track for you just to get the job done.
> P.S. to look at the code of searchlight, if you use ipython, the
> simplest way is to call
> Let us know if we should proceed with #1 ;)
> =------------------------------ /v\ ----------------------------=
> Keep in touch // \\ (yoh@|www.)onerussian.com
> Yaroslav Halchenko /( )\ ICQ#: 60653192
> Linux User ^^-^^ 
> Pkg-ExpPsy-PyMVPA mailing list
> Pkg-ExpPsy-PyMVPA at lists.alioth.debian.org
More information about the Pkg-ExpPsy-PyMVPA