Open zparis99 opened 2 weeks ago
I have a temporary fix for this where we just skip over any examples that result in nan loss. The issue seems to come from this file: dataset_full/derivatives/preprocessed/sub-04/car/sub-04_task-part004chunk14_desc-preproc_ieeg.edf where the signal is largely a flat line so there's no frequency info. We can come up with a better long term fix later but this works for now.
For some reason there are nan values appearing in the model parameters. Figure out what's going wrong in training.