[pymvpa] Train and test on different classes from a dataset

[Jan Derrfuss]

Dear Michael,

Thank you very much for your reply! I'm afraid that I might still need 
some help after having looked at the Sifter examples...

So, as a first step, you're suggesting this, right? I create a new 
sample attribute for a and c (say, "e"), and another one for b and d 
("f"). Thus, I would now have something like this (l1 = old labels, l2 = 
new labels, ch = chunk):

l1 l2 ch
a  e  1
a  e  1
b  f  1
b  f  1
c  e  1
c  e  1
d  f  1
d  f  1
-------
a  e  2
a  e  2
b  f  2
b  f  2
c  e  2
c  e  2
d  f  2
d  f  2
-------
a  e  3
a  e  3
b  f  3
b  f  3
c  e  3
c  e  3
d  f  3
d  f  3

Now let's assume the current fold involves training on chunk 1 and 2, 
and testing on 3. What I'm not sure about is how to tell the Sifter to 
remove labels c and d from chunk 1 and 2, but a and b from chunk 3 (and 
how to change that with every new fold).

It would be great if you could give me another hint!

Thanks again,
Jan