Incorrect Interval for BNCI2014001

[dcwil]

Someone with better knowledge of the codebase can maybe weigh in here but I think since #447 at least BNCI2014001 has been cutting trials without the shift for task onset/offset. Currently for BNCI2014_001 (the dataset I noticed this with) I think you need to set tmin=2 and tmax=6 in the paradigm to actually cut the correct trial. I assume by default it should tmin=0 should be relative to the dataset interval [2, 6] (so by default, it cuts the correct trial).

[dcwil]

@bruAristimunha @PierreGtch Could you look at this/confirm? I think if left unnoticed by a user (aka me haha) this can have a pretty big impact on classification performance without throwing any errors

[PierreGtch]

Hi @dcwil, could you provide an example of code that changed behaviour since #447?

[dcwil]

I'll try and find something, although I'm not sure what the intent is/was and I might have the PR wrong.

What I can say now is:
The events found by

moabb/moabb/datasets/preprocessing.py

Line 75 in 4b297b9

events = mne.find_events(raw, shortest_event=0, verbose=False)

for BNCI2014001 are the fixation markers, so 2s before the actual task onset, hence the [2, 6] interval, but as far as I can tell the interval isn't applied anywhere in the prepro (if it is can you point me in the right direction?). If this is the intended functionality then the only thing that needs changing is stuff in examples that should probably use the interval to cut the trial:

moabb/examples/plot_braindecode.py

Lines 73 to 74 in 4b297b9

	tmin = 0
	tmax = None

[PierreGtch]

I'not sure I understand where the problem is, [2,6] is the default interval in the BNCI2014_001 class:

moabb/moabb/datasets/bnci.py

Line 813 in 4b297b9

interval=[2, 6],

In case the preprocessing is a bit unclear, do you know the paradigms have a make_process_pipelines() method?
For example:

from moabb.datasets import BNCI2014_001
from moabb.paradigms import MotorImagery
print(MotorImagery().make_process_pipelines(BNCI2014_001()))

[Pipeline(steps=[(<StepType.RAW: 'raw'>,
                 SetRawAnnotations(durations=4,
                                   event_id={'feet': 3, 'left_hand': 1,
                                             'right_hand': 2, 'tongue': 4})),
                (<StepType.RAW: 'raw'>,
                 FunctionTransformer(func=operator.methodcaller('filter', l_freq=8, h_freq=32, method='iir', picks='eeg', verbose=False))),
                (<StepType.EPOCHS: 'epochs'>,
                 Pipeline(steps=[('epoching',
                                  Pipeline(steps=[('forkpip...
                                                   RawToEpochs(baseline=None,
                                                               event_id={'feet': 3,
                                                                         'left_hand': 1,
                                                                         'right_hand': 2,
                                                                         'tongue': 4},
                                                               tmax=6,
                                                               tmin=2.0))]))])),
                (<StepType.ARRAY: 'array'>,
                 ForkPipelines(transformers=[('X',
                                              Pipeline(steps=[('get_data',
                                                               FunctionTransformer(func=operator.methodcaller('get_data'))),
                                                              ('scaling',
                                                               FunctionTransformer(func=operator.methodcaller('__mul__', 1000000.0)))])),
                                             ('events', EpochsToEvents())]))])]

Printing this scikit learn Pipeline is not very pretty but you can inspect it in detail.
Here you see the interval for epoching is [2-6]s.

[dcwil]

Ah i see - sorry for the confusion, I was a little too hasty!

[PierreGtch]

no problem @dcwil 😌