Running ridge_cv with a pre-defined parcellation yields ValueError

[emdupre]

Running ridge_cv with a predefined parcellation such as the BASC parcellation yields an error:

ValueError: Cannot have number of splits n_splits=4 greater than the number of samples: n_samples=1.

This seems to be because the cv value for the RidgeAlignment estimator is hard-coded at 4. Changing to another value such as None for LOO allows ridge_cv to run to completion.

Would it make sense to expose cv as an attribute in the user-facing PairwiseAlignment class ?

[thomasbazeille]

PairwiseAlignment also accepts instances of estimators in method so here, you could *from fmralign.alignment_methods import RidgeAlignment* *custom_ridge = RidgeAlignment(cv=1)* *alignment = PairwiseAlignment(method=custom_ridge, ...)* But still it seems weird to me that clustering on a parcellation changes anything for Ridge cross-validation, since Ridge should be applied on as many samples as you have. Can you provide replication code ? I'd like to have a look. Le lun. 2 déc. 2019 à 16:38, Elizabeth DuPre <notifications@github.com> a écrit :

…

Running ridge_cv with a predefined parcellation such as the BASC parcellation <https://nilearn.github.io/modules/generated/nilearn.datasets.fetch_atlas_basc_multiscale_2015.html> yields an error: ValueError: Cannot have number of splits n_splits=4 greater than the number of samples: n_samples=1. This seems to be because the cv value for the RidgeAlignment estimator is hard-coded at 4 <https://github.com/Parietal-INRIA/fmralign/blob/5f3380d94f52652ebb8e643d4bd2b6bad4bc4d32/fmralign/alignment_methods.py#L255>. Changing to another value such as None for LOO allows ridge_cv to run to completion. Would it make sense to expose cv as an attribute in the user-facing PairwiseAlignment class ? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#36?email_source=notifications&email_token=ABR7RNDAYRLNJOJDK2AIKADQWUTXLA5CNFSM4JTXXAS2YY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4H5JTLYA>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABR7RNH6ZNE33CA5KUVXNMDQWUTXLANCNFSM4JTXXASQ> .

[emdupre]

Here's the code I was trying.

But that's a great idea ! What do you think if I create another example in the documentation that includes that ?

[thomasbazeille]

Hello I answered this one in #37 after running your code. The problem here is that you try to fit ridge_cv on only one sample and it's built-in cross validation scheme requires at least one image per fold.

Anyway, it's always a bad idea to learn an alignment on only one sample because it'll be very noisy.

+1 for underlining the ability to use a non-default estimator in the docs though, if you have an idea on where it should take place it'd be great.