[pymvpa] Searchlight again

[Valentin Haenel]

Hi,

this concerns my previous patch for the searchlight. The one posten to the ML on
Friday 24 April has been superseeded. For ease of review i put the changes into
the branch val/new-sl on alioth.

Note however it is still unclear to me if there is not a better way to run a
searchlight on a list of datasets.

I can kinda of see how to "prep a new metric" but i'm a bit unsure where to
start looking. It would be great if you could point me in the right direction,
either for code or documentation.

Thanks.

V-

Just to summarize here are the two comments from Yaroslav:

--------------------------------------------------------------------------------

> also why not just prep a new metric, which would take care about taking
> a single time-point... ie originally you have 5D dataset:
> samples x FIR-timepoints x 3D volume
> (instead of current default samples x 3D volume), which is similar in
> nature to what we get now from BoxcarMapper when we take multiple
> volumes after some onsets.

> Later on those datasets get mapped into 2D but if original dataset's
> mapper has a proper metric to select neighbors it can spit out only
> neighbors within the same time-offset.   

--------------------------------------------------------------------------------

> I remember Michael had recently expanded metrics support -- so it might
> be we don't need anything more there ;)

> the question is in what shape do you have your FIRs? just a separate
> dataset per each FIR delay?

Yes, i have beta volumes, one per timebin per condition per run.

> then I guess you could simply create a new
> dataset by combining all those 4D to get 5D volume, create MaskedDataset
> based on that one, then see how its mapper.metric could be adjusted so
> whenever you use mapper.getNeighbors you obtain only the ones in the
> given FIR time-offset

--------------------------------------------------------------------------------