Adding Braindecode pipeline

[bruAristimunha]

Hi @sylvchev!

This is a work in progress to add the braindecode object to moabb. I need your assistance with this task.

In the file moabb/pipelines/braindecode.py, I create two custom classes for working with the braindecode library and scikit-learn classifiers.

The CreateBraindecodeDataset class serves as a transformer that converts a BaseEpochs object from the mne library into a WindowsDataset object from the braindecode library. We can evolve this first class/transformation to handle the raw data, and window sliding becomes a parameter.

The BraindecodeClassifierModel class is a classifier that wraps any classifier from EEGClassifier, providing a consistent interface for fitting the classifier to a WindowsDataset, making predictions, and computing predicted probabilities. In theory, this wrapper is not necessary, as long as we make this code sequence with the object returned by the EEGClassifier. It's a design decision.

from functools import partial
clf.fit = partial(clf.fit, epochs=10)

My points of concern for this part are mainly contained in the second file, pipelines/braindecode.py. To generate the EEGClassifier object, we need to have the n_classes, n_chans and the input_window_samples. And I understand this is handled more dynamically with moabb.

[bruAristimunha]

Hi @agramfort!

You are more than welcome to review the code if you have some space in your schedule.

I understand that the benefit of this wrapper is tremendous for braindecode—an extremely simplified way to benchmark multiple moabb datasets with any braindecode model.

[agramfort]

I don't have enough background on moabb but I am super excited to see this happening !

thx @bruAristimunha 🙏

[bruAristimunha]

Hi @agramfort!

I really appreciate your message, and I understand that it may be different.

Insisting just a bit, I wanted to share a positive aspect of your revision - the library seems similar to scikit-learn and pyriemann, with classes like the transformations and pipelines in sklearn.

Would you mind looking at a tutorial on how braindecode fits in? I can put together something similar to what I showed earlier: braindecode/braindecode#46 (comment)

I think it will be something very close to your revisions in braindecode and pyriemann.

I greatly respect your time and expertise, and I'd really appreciate your help with this, but I totally understand if that's not possible.

[bruAristimunha]

I think, we need to change the poetry command to install everything using the new optional libraries.

[sylvchev]

Ok, I wanted to propose a version that relies on braindecode.datasets.create_from_mne_raw or braindecode.preprocessing.create_windows_from_events but the crop approach seems broken for now (see braindecode/braindecode#422 (comment)) so the braindecode.datasets.create_from_X_y seems the best option right now.

[sylvchev]

We should identify the origin of the predict_proba error

[sylvchev]

Thanks @bruAristimunha and @carraraig ! 🙏