Alexandre Gramfort

you have more test in mne-python see test_eeglab_infomax.py it's comparing with commonly used matlab version.

> Basically need some clarification on this: > There is a file test_eeglab_infomax.py which needs to be ported. This file is specifically to test with scikit-learn results to compare with...

a good first step for me is: if sample_weight is passed to fit in pipeline then **all** steps need to support it. then we can allow estimator__sample_weights in a next...

> > In practice this means that we would need to add sample weight to > preprocessors and ColumnTransformer to make this useful. > sure but it can be done...

I like the approach from https://github.com/scikit-learn/scikit-learn/pull/13565/files +1 to fall back to has_fit_parameter('sample_weight') by default in BaseEstimator if supports_sample_weight is not set. We could also warn that supports_sample_weight is required if...

indeed ! > Message ID: ***@***.***> >

autoreject is slow on your data?

can you share a snippet and a file so we can profile?

yes it can happen. I see 2 options - fix the seed - use more random splits and draws in bayes opt if you have such a variability I suspect...

can this actually work on circleci?