[pymvpa] NFoldSplitter question

[Emanuele Olivetti]

I've added another implementation that works directly on
combinations of a given size and uses generators. It is
in eo/master. Should work well also for n>20 but obviously
there is a limit... I think no one would run more than 10000
times a learning algorithm... ;)

Please have a look and merge into main if it is OK.

E.

Yaroslav Halchenko wrote:
> indeed ;) me is also swamped with work, so no diving into combinatorics
> or recursive code comprehensions.
>
> Hence, for now, we provided a fix which should work for cvtype>2 BUT for
> a sane number of unique chunks (ie < 20). Otherwise it would fail at
> number of unique chunk ids > 20 I guess... didn't check carefully yet ;)
> But otherwise it works fine now. Fix is present in yoh/master and
> maint/0.4 branches.
>
> enjoy -- and thanks again for spotting and reporting evil things ;)
>
> On Wed, 04 Mar 2009, Michael Hanke wrote:
>
>> What can I say -- it looks like a bug.
>
>> The function that does it is getUniqueLengthNCombinations() in
>> mvpa.misc.support. It looks like it is broken for k>2. Never done that
>> before and hence never noticed. Thanks for spotting.
>
>
>> However, that function operates recursively and my blood-sugar level is
>> too low for combinatorics and recursive function together -- need some food
>> first...
>
>
>> Michael