BCIC IV 2a examples without local path about braindecode HOT 6 CLOSED

Here are results I get atm:
https://gist.github.com/robintibor/37779c941ad43f6329b7d8ab13a435fe

Minor variations due to different machine and GPU indeterminism are okay

from braindecode.

from braindecode.

on CPU, should be too big. results see my last comment

from braindecode.

from braindecode.

ok fair enough. training on full train set and testing on test set one can reach ~83% on this dataset with our old pipeline (be aware, the random-seed-based variance here can be huge on this dataset. The 83% is averaged across 100 seeds in an internal reproducibility/variance estimation test. But here we are not doing that yet, just train/valid test on train.

Maybe someone want to have a try to modify example towards running with 1 epoch per default, using moabb to fetch the data and, training on full train, and evaluating on eval/test? Without changing any of the remaining/preprocessing code for now to make it easier to pinpoint differences later @sliwy @gemeinl @hubertjb @sbbrandt @TonioBall ? Also find out if we get different epochs from MOABB, e.g., if they throw out epochs marked as artifact-contaminated for BCIC IV 2a. I forgot what we found out there already.

from braindecode.

Done

from braindecode.