Add SRM experiment #82
Conversation
|
Maybe I can do the merge now ? As it is a single file, independent of the rest and all tests are passing ? |
|
I'd like to give Elizabeth a chance. In general it's good to have two pairs of eyes on the PR. |
|
Sorry for the slow follow-up ; I was sick last week and am still catching up. I'll try to take a look this afternoon |
There was a problem hiding this comment.
Thank you, @denisfouchard ! This will be great to have !!
I have a few small comments, but my larger question is about how we're handling training and testing data. In our real-data experiments, we had used add_subjects to add in the target_train data after fitting on all of the sources_train data. Here, it seems like we're fitting on all training data (including target_train) and then extracting the target subject's basis in the fitted W. This strikes me as a little harder to follow for the naive reader (since we end up having to ignore the last image in multiple other lines like L95, L144, etc).
Could you confirm if my understanding is correct, and if there's a strong reason to do it this way ?
| from nilearn.image import concat_imgs | ||
|
|
||
| template_train = [] | ||
| for i in range(6): |
There was a problem hiding this comment.
Is there a reason why we want to have sub-07 in template_train at all, rather than only in target_train and target_test ? I know we exclude the last entry later so that it's not actually included in e.g., average_img, but that's one extra bit of overhead for the reader to keep in mind.
There was a problem hiding this comment.
I don't understand ? We need the data from sub-07 in the training to get its basis, so we need to concat the imgs at some point. Maybe you mean changing the same of this list to simply imgs ? But it would be inconsistant with the INT experiment ?
There was a problem hiding this comment.
What would be your suggestion @emdupre ?
There was a problem hiding this comment.
Oh, I see it would be to use add_subjects. @denisfouchard can you try and do that ?
Co-authored-by: Elizabeth DuPre <[email protected]>
Co-authored-by: Elizabeth DuPre <[email protected]>
As suggested by @bthirion I open this separate PR for an addition of an SRM experiment.
This is co-smoothing on IBC-Contrast (same as INT), inspired by Hugo Richard's code.