problem training EQTransformer with original_compatible=True

[barcheck]

I have encountered a problem implementing transfer learning with the original EQTransformer model (original_compatible=True and .from_pretrained("original")). When I train starting with the "original" model, everything seems to work fine, the weights change, etc. But when I run, for example, eval.py to use the resulting model to predict, all of the "score_detection" values for the predictions are identical for each trace. Other values do change for each trace.

This does not happen if I train starting with .from_pretrained("stead" or "instance")