Deprecated the deep learning model

[bruAristimunha]

trafficstars

We have put a lot of effort into building a deep learning module, but clearly, this is not the best environment to maintain a deep learning library. I am proposing to deprecate the entire deep learning module and keep only the support for braindecode.

Currently, everything is already integrated into braindecode (EEGNeX, shallow, deep, eegnet and eegtcn), and it does not make much sense to have duplicate code.

Before officially removing the code duplication, a comparative study should be done and approved.

cc. @sylvchev and @carraraig

[bruAristimunha]

cc. @tomMoral too

[carraraig]

Hi @bruAristimunha ! I think on my side is the direction to go, since brain decode is well maintained. However I think can be useful to keep the current implementation in Tensorflow in order to ensure the full replicability of the MOABB benchmark. Maybe we can keep them in a separated folder. What do you think?

[PierreGtch]

Even if we remove it, the benchmark will be replicable by using the correct release. Maybe we could simply add a tag in the repo on the commit people should use to replicate this benchmark?

[bruAristimunha]

hey @carraraig,

Scikit-keras is deprecated: https://github.com/adriangb/scikeras/pull/334

And it is not working with the new version of scikit-learn. There is a complicated dependency issue, so we can not downgrade scikit-learn. Looks like the way to go is to use the scikit-learn API wrapper API that is back to Keras:

Can you work on this? https://keras.io/api/utils/sklearn_wrappers/

If you can't, I think it's best to remove this moabb dependency and focus on braindecode.

[bruAristimunha]

I'm holding moabb with sklearn < 1.16 for one week to give you time to answer.

[Hivenet-Igor]

Hi @bruAristimunha ! I think that it is better to follow the direction of Braindecode, which is better maintained.

[bruAristimunha]

okay @Hivenet-Igor!