The filter bank paradigm does not correspond to the new abstraction that depends only on the MNE

[bruAristimunha]

If you have time after the ICLR review period, it would be great if you could take a look.

[PierreGtch]

@bruAristimunha do you have a bit more details? like a bit of code

First thought that comes to mind: could this be related to edge effects while filtering on epoched data?

[bruAristimunha]

Thanks for replying @PierreGtch, I was thinking about it while looking at the code, and you're right. We have two things here: if we have a paradigm that creates filters correctly, e.g, if we pass the frequency list, we create a bank on the data as a whole.

moabb/moabb/paradigms/motor_imagery.py

Lines 158 to 167 in d2edb29

	class FilterBank(BaseMotorImagery):
	"""Filter Bank MI."""
	def __init__(
	self,
	filters=([8, 12], [12, 16], [16, 20], [20, 24], [24, 28], [28, 32]),
	**kwargs,
	):
	"""init."""
	super().__init__(filters=filters, **kwargs)

And we have a mixer/function that filters the data that has already been epoched, applying the filter at training time, which doesn't necessarily use mne to filter. In my experience, this can lead to errors and difficult checking.

moabb/moabb/pipelines/utils.py

Line 216 in d2edb29

class FilterBank(BaseEstimator, TransformerMixin):

moabb/moabb/pipelines/utils.py

Line 267 in d2edb29

def filterbank(X, sfreq, idx_fb, peaks):

Besides improving documentation, the question now is, what should we do here?

[PierreGtch]

I think it’s fine to keep both methods, as different users might have different habits. Because of the edge effects, they will not give the same results, so we don’t have much to test here.
The first method (filtering raw signals) gives the most reliable results, so we should present it in the doc and tutorials as the default method for filter banks.
Maybe the second one can simply be listed in the API references, with a disclaimer and a recommendation to use the other one. Also, we could remove this tutorial which uses it:
https://github.com/NeuroTechX/moabb/blob/develop/moabb/pipelines/classification.py

[tomMoral]

the core question is: does this affect performances of the different pipelines?
In any case, it would be nice to make an example comparing both and explaining the differences.

[PierreGtch]

This could be a good onboarding project for a Bachelor/Master student :)