I'm not sure if you can do this automatically, but I've done this manually by creating 'pseudo-runs' and using that attribute as the basis for partition. This gives you arbitrary amounts of control -- for instance, I wanted to make it so that each pseudo-run incorporated elements from every real run; but if you really don't care about that, you could just generate an array with the number of events, divide that by the number of samples you want per fold, create the runs, then permute that array, assign it as a sample attribute, and tell NFoldPartitioner to use that attribute for the folds, ie:
In [*1*]: samples = 20 In [*2*]: folds = 4 In [*3*]: samples_per_fold = samples / folds In [*7*]: pseudo_runs = [_ for _ in range(folds) for x in range(samples_per_fold)] In [*8*]: pseudo_runs Out[*8*]: [0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3] In [*11*]: random.shuffle(pseudo_runs) In [*12*]: pseudo_runs Out[*12*]: [2, 2, 3, 1, 3, 2, 0, 0, 0, 1, 1, 0, 0, 1, 3, 2, 3, 1, 2, 3] then do my_dataset.sa['pseudo_runs'] = psuedo_runs, etc. Note that some people will object vehemently about this method of cross-validation however, due to temporal correlation and general non-independence of the samples. S
_______________________________________________ Pkg-ExpPsy-PyMVPA mailing list [email protected] http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/pkg-exppsy-pymvpa
