[pymvpa] searchlight gets stuck, issues error

[Yaroslav Halchenko]

make sure that your mask doesn't include any voxel with 0s (i.e. you've
masked them out prior loading with different mask).  Or just use 0.4.4
-- that one shouldn't scale C in such cases which would avoid 'getting
stuck'

On Tue, 02 Feb 2010, Sergey.V.Fogelson at Dartmouth.edu wrote:

> Hello,

> Im running the searchlight algorithm on some data and it gets stuck
> right around 68% complete.

> Ive zscored and detrended the dataset before preprocessing. the
> error that is issued says that the SVM classifier cannot converge
> because of zero norm.

> here's the code ive used to preprocess the data:

> if __debug__:
>     debug.active += ["SLC"]

> attr = SampleAttributes('FFA_class_25jan10sf0')

> dataset = NiftiDataset(samples='25jan10sf.nii', labels=attr.labels,
> chunks=attr.chunks, mask = 'mask.nii')

> detrend(dataset, perchunk=True, model="linear", polyord = 1)

> zscore(dataset, perchunk=True,targetdtype= 'float16')

> dataset = dataset.select( labels = [1,2] )

> clf = LinearCSVMC()

> cv = CrossValidatedTransferError(TransferError(clf), NFoldSplitter())
> s1 = Searchlight(cv, radius=9)


> s1_map = s1(dataset)
> orig_s1_map = dataset.mapReverse(N.array(s1_map))
> data_hdr = dataset.niftihdr
> NiftiImage(orig_s1_map,header=data_hdr).save('9mmSL_SVM_tr1_facevshouse.nii.gz')


-- 
                                  .-.
=------------------------------   /v\  ----------------------------=
Keep in touch                    // \\     (yoh@|www.)onerussian.com
Yaroslav Halchenko              /(   )\               ICQ#: 60653192
                   Linux User    ^^-^^    [175555]