[ENH]: start refactoring pipelines module

[dengemann]

Do not merge.
This is still broken but you can see already a bit where it's going.
Addresses multiple issues.

[dengemann]

I am still having some difficulty debugging the test_pipelines.py @DavidSabbagh @agramfort
I absolutely want things to be indexible by column names, not integer indices.
The ExpandFeatures class must be probably updated to return a DatafFrame.
The code inside ExpandFeatures is the problem, all passes when not expanding interaction effects.
As it is in this PR, scikit-learn complains becaus –– I guess –– it cannot drop the indicator colum based on a name, as ExpanFeagures returns an array, not a DataFrame.
Will push/investigate more tomorrow.

[DavidSabbagh]

This is exactly the problem I encountered and why I gave up and finally indexed with integer indices. I did not wanted to loose time and make things work. But i agree it would be much cleaner Le ven. 23 avr. 2021 à 23:13, Denis A. Engemann ***@***.***> a écrit :

…

I am still having some difficulty debugging the test_pipelines.py @DavidSabbagh <https://github.com/DavidSabbagh> @agramfort <https://github.com/agramfort> I absolutely want things to be indexible by column names, not integer indices. The ExpandFeatures class must be probably updated to return a DatafFrame. The code inside ExpandFeatures is the problem, all passes when not expanding interaction effects. As it is in this PR, scikit-learn complains becaus –– I guess –– it cannot drop the indicator colum based on a name, as ExpanFeagures returns an array, not a DataFrame. Will push/investigate more tomorrow. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#14 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AIC2QGTJLQNGOBS5MWDQG2LTKHPGRANCNFSM43LF33RA> .

[dengemann]

I think we have to make clear design choices here and require that the code here only supports DataFrames all the way down. It will simplify things; working on a fix in the meantime.

[dengemann]

@agramfort @DavidSabbagh there is more to be done (especially tests) but I think this is already something. If we can do the necessary to merge this, I think, we make a step ahead.

[agramfort]

I am wondering if we should not make 3 functions

make_filter_bank_transformer
make_filter_bank_regressor
make_filter_bank_classifier

where make_filter_bank_transformer returns the first steps of the pipeline without the scaler and the RIdgeCV

[dengemann]

I was thinking in that direction too. make_filter_bank_regressor/classifier would then just be high-level convenience functions that work API-wise like the current make_filter_bank_model?

[agramfort]

yes !

…

[dengemann]

@agramfort next milestone unlocked :)

[agramfort]

besides LGTM

[agramfort]

@dengemann merge if happy