BrainDecode Plot

[carraraig]

Add an example of usage of BrainDecode

[bruAristimunha]

Hi @carraraig,

After carefully considering our problem, I have concluded that the best approach is to avoid modifying the braindecode side.

Upon studying the InputerSetter callback, I have discerned that it presents the most viable alternative. Accordingly, I have devised a new callback that allows us to access the relevant shape information. We may also develop a series of model-specific functions to streamline the process further.

While I acknowledge that there may be room for improvement, I think that we can optimise the parser and setter functions.

[carraraig]

Very nice idea for that! I've only see that if I create the EEGNetV4 as

model = EEGNetv4(
in_chans=1,
n_classes=1,
input_window_samples=100,
)

and as

model = EEGNetv4(
in_chans=X.shape[1],
n_classes=len(events),
input_window_samples=X.shape[2],
)

we get slightly different result, you know why?

[bruAristimunha]

Hi @carraraig,

It is a good idea set the EEGNetV4 with these dummy values, and it makes sense. I am still thinking about the second problem of defining the pytorch model with a "lazy" approach.

I don't know why )= maybe some deterministic flag?

[bruAristimunha]

Hi @carraraig and @sylvchev,

Now with InputSetterEEG, we are able to run the ShallowNet. I don't know if this is a good solution, but it is working.

[sylvchev]

Ok nice, I'll check this

[bruAristimunha]

Very nice idea for that! I've only see that if I create the EEGNetV4 as

model = EEGNetv4( in_chans=1, n_classes=1, input_window_samples=100, )

and as

model = EEGNetv4( in_chans=X.shape[1], n_classes=len(events), input_window_samples=X.shape[2], )

we get slightly different result, you know why?

@carraraig, I change the inputSetterEEG to reinitialise the model instead of changing the input layers. Can you test if the behaviour you reported is gone?

[carraraig]

Very nice idea for that! I've only see that if I create the EEGNetV4 as
model = EEGNetv4( in_chans=1, n_classes=1, input_window_samples=100, )
and as
model = EEGNetv4( in_chans=X.shape[1], n_classes=len(events), input_window_samples=X.shape[2], )
we get slightly different result, you know why?

@carraraig, I change the inputSetterEEG to reinitialise the model instead of changing the input layers. Can you test if the behaviour you reported is gone?

I've tryed and with the first option the code is not running. Is the same for you?