[pymvpa] searchlight analysis

[Nick Oosterhof]

Some minor comments inserted:

> On 8 Sep 2017, at 17:52, Pegah Kassraian Fard <pegahkf at gmail.com> wrote:
> 
> 
> from glob import glob
> import os
> import numpy as np
> 
> from mvpa2.suite import *
> 
> %matplotlib inline
> 
> 
> # enable debug output for searchlight call
> if __debug__:
>     debug.active += ["SLC"]
> 
> 
> # change working directory to 'WB'
> os.chdir('mypath/WB')
> 
> # use glob to get the filenames of .nii data into a list
> nii_fns = glob('beta*.nii')
> 
> # read data
> 
> labels = [
>     1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 
>     2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 
>     7, 7, 7, 7, 7, 7, 7, 
>     1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 
>     2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 
>     7, 7, 7, 7, 7, 7, 7
> ]
> grps = np.repeat([0, 1], 37, axis=0) # used for `chuncks`
> 
> db = mvpa2.datasets.mri.fmri_dataset(
>     nii_fns, targets=labels, chunks=grps, mask=None, sprefix='vxl', tprefix='tpref', add_fa=None
> )

Is there a reason not to use a mask? At least a brain mask to avoid stuff stuff like skull and air? 

> 
> # use only the samples of which labels are 1 or 2
> db12 = db[np.array([label in [1, 2] for label in labels], dtype='bool')]
> 
> # in-place z-score normalization
> zscore(db12)
> 
> # choose classifier
> clf = LinearNuSVMC()

Have you tried a different classifier, for example Naive Bayes? That one is simpler (though usually a bit less sensitive than SVM / LDA in my experience)?

> 
> # setup measure to be computed by Searchlight
> # cross-validated mean transfer using an N-fold dataset splitter
> cv = CrossValidation(clf, NFoldPartitioner())
> 
> # define searchlight methods
> radius_ = 1

That's a tiny radius - why not use something like 3?