SearchlightHyperalignment is incapable of handling datasets with varying # of features

[beausievers]

I have a dataset where one of the subjects is completely missing variance in data in one of the searchlights. That leads SearchlightHyperalignment to fail in ProcrusteanMapper

> /afs/dbic.dartmouth.edu/usr/pkg/python/python_v2.7/amd64_linux26/lib/python2.7/site-packages/mvpa2/mappers/procrustean.py(123)_train()
    122                                        * sn * means[i] )**2)):
--> 123                 raise ValueError, "For now do not handle invariant in time datasets"

but if I remove invariant features in the datasets, that leads them to have different # of features, and then it fails during forward mapping.

/afs/dbic.dartmouth.edu/usr/pkg/python/python_v2.7/amd64_linux26/lib/python2.7/site-packages/scipy/sparse/base.pyc in __mul__(self, other)
    403 
    404             if other.shape[0] != self.shape[1]:
--> 405                 raise ValueError('dimension mismatch')
    406 
    407             result = self._mul_multivector(np.asarray(other))

ValueError: dimension mismatch

[mohmdrezk]

Hi, by any chance, did you find a solution to this problem? I am having the same issue. Thank you in advance.

[beausievers]

I worked around this issue by adding very low amplitude Gaussian noise to the invariant features before calculating the hyperalignment maps. Seems to work just fine.

[mohmdrezk]

Thank you very much. I will give that a try. Best regards.

[yarikoptic]

Fwiw, please use mvpa2.seed in your script to make results reproducible

[mohmdrezk]

Thank you all. It worked fine. All the best.

[soligschlager]

Hi, I'm encountering the same issue and was wondering if you added the low amplitude Gaussian noise to the invariant features before or after zscoring? Many thanks in advance!