[Feature] TorchEEG wants to add support for MoABB

[tczhangzhi]

Greeting,

Thank you for your awesome work. The benchmark dataset of MoABB is very helpful for research in the field of EEG data analysis. Recently, we are expanding our TorchEEG library to help more people easily analyze EEG signals using deep learning techniques.

We would like to add support for MoABB, which requires NO requirement for code changes on the MoABB side. We have completed the PR [1], please do not hesitate to contact us if you have any concerns.

We also noticed that the implementation of some datasets seems to be different from the comparative work, so we may add some different implementations of the same dataset in the future.

Thanks!

[1] https://github.com/torcheeg/torcheeg/pull/37/commits/259a5a8908a399882cf3cf11d310480df97b4625

[bruAristimunha]

Hello @tczhangzhi! 😊

I'm really enthusiastic and happy about this initiative! If there's anything I can do to assist with the integration process, please don't hesitate to reach out. I'll gladly conduct tests to ensure compatibility with all the moabb datasets; I will test the new class MOABBDataset from TorchEEG. However, please note that it might take me approximately one week to complete everything due to my other ongoing deadlines.

I've been contemplating reaching out to you for some time now because I believe it would be beneficial for us to have a discussion and collaborate to converge both libraries. Maybe a MOABB tutorial inside the TorchEEG or a TorchEEG inside MOABB. Btw, I'm a big fan of your library! If you're interested in converging our efforts, you can reach me via email at b.aristimunha@gmail.com or on Discord as b.aristimunha to talk a little just to arrange what we can do.

[tczhangzhi]

Thank you for your enthusiastic response, and for the kind words about our work (๑•̀ㅂ•́)و✧. I look forward to communicating with you. MoABB is indeed one of my favorite tools for EEG analysis, and your other library, Braindecode, is frequently used in my research on motor imagery.

I would appreciate your insights on two aspects:

1. Regarding dataset testing, we face the issue of slow CI. Implementing smoke tests on dataset classes could require time-consuming network data downloads or preprocessing. On the other hand, implementing unit tests on dataset classes would necessitate mocking real data, which involves simulating files with code.

2. I am also considering expanding the pipeline. Braindecode provides a Skorch-based deep-learning pipeline for MoABB. TorchEEG could potentially supplement this with a Pytorch-lightning-based implementation, making it suitable for other use cases. I'm thinking about how to provide pipeline's conversions in a non-intrusive way.

Thanks!

[tczhangzhi]

I sent an email to discuss other details (灬ºωº灬). My email is tczhangzhi@gmail.com.

[bruAristimunha]

Hi @tczhangzhi,

Amazing, I will return you with my ideas about the topics =)