Alexandre Gramfort

icon indicating a website link http://alexandre.gramfort.net icon indicating an email [email protected]

icon company Meta icon location Paris, France icon location Machine Learning researcher @ Meta Reality Labs, scikit-learn core dev, pythonista interested in neuroscience applications

Results 758 comments of Alexandre Gramfort

Too few channels. autoreject is unlikely to be effective

yes 3 channels is too little. I would have said > 30 channels but it's just an educated guess... nothing quantitative

use better defaults for n_interpolates

would this break any existing code to change this? did you evaluate / quantify the impact? >

FIX: bound thresholds better

did you ask @thht for the dataset that trigged the issue?

multi channel support for ransac

lol :)

deal with flat channels

it's unclear how. But as we said in the paper flat channels are trivial to detect. Altough we don't have a function for that...

deal with flat channels

true :( there is a PR though to have epoch wise interpolation. Maybe this could become a priority to fix...

get_rejection_threshold stops optimizing threshold as soon as one threshold leads to rejection of all channels in at least one fold

yes we could give bounds in optim grid to avoid this case before staring the opt. >

Autoreject cleans only some epochs, not all

hum.... not sure what you mean with no cross-val but if you mean fitting and testing on the same data you can use cv object with a scikit-learn API that...

Autoreject cleans only some epochs, not all

it would not be allowed with sklearn but you can do if cv=1 use a splitter like this: ``` import numpy as np from sklearn.model_selection import BaseCrossValidator class NoSplitter(BaseCrossValidator): def...

[ENH] Add option for dropping epochs where there are more bads than n_interpolates_

to me it becomes an empirical question on some data where a new option appears useful (ideally quantitatively) > Message ID: ***@***.***> >