[MNT] Eegnetv4 fix v2

[bruAristimunha]

No description provided.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 87.25%. Comparing base (1c1b845) to head (f5caf9b).
Report is 8 commits behind head on master.

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #680   +/-   ##
=======================================
  Coverage   87.25%   87.25%           
=======================================
  Files          76       76           
  Lines        6833     6833           
=======================================
  Hits         5962     5962           
  Misses        871      871           

[bruAristimunha]

I talked again with Victor from the speechbrain team. We checked the model diff again again based on the torchinfo summary, and the only remaining difference is the convolution order that needs to be re-checked.

[bruAristimunha]

The benchmark table @tomMoral

[bruAristimunha]

Hey @tomMoral,

I did a more in-depth study for EEGNet because the benchmark I initially did was slightly different from the original result. (different evaluation)

In the original article, we have interfered from the barplot that the accuracy is 68% +-[70%-67%] (Within-Subject, mixing two sessions and doing cross-validation with four folds (conceptually wrong for me)) and 38% +- [42%-35%] (inter-subjects).

Using benchopt, I investigated the impact of different band frequencies (one of the key differences following our table):

Using only the closest frequency range [4Hz, 38Hz], we have better results.

objective_name	use_linear_layer=False	use_linear_layer=True	original
inter_sessions	44.5+-19.96	59.38+-15.63
inter_subjects	51.15+-6.31	49.31+-8.51	38% +- [42%-35%]
intra_session	47.09+-23.21	62.87+-20.77
within subject	-		68% +-[70%-67%]

There are several interesting things about this contrast of different evaluations, more or less filtering and domain generalization, and possibly, because of this, the implementation has not been corrected before. Since linear is better when we filter less, I will keep it in my ideas box to explore in the future, somewhat along the lines of our article in the CNS.

We should integrate this code that adds this parameter, and during the benchmark, we validate whether we leave it as linear as default or not.

[PierreGtch]

Nice job @bruAristimunha !! I'm just not sure about the parameter name and the default. WDYT?

[bruAristimunha]

Hi @PierreGtch and @tomMoral,

I appreciate your review @PierreGtch. I have incorporated everything. I really like the new name, thank you for that. As for changing the name, I would rather not. Let's wait for a release to do that.

[sylvchev]

Thanks @bruAristimunha !