Getting NaN values when trying to finetune

#2
by imerad7 - opened

Hello,
I'm trying to finetune this pretrained model on a classification task by adding a classification head. However, the training quickly starts giving NaN values after a few iterations. Do you know how this could be fixed ?
Thanks

I found the issue. The classification head parameters were not properly initialized and had absurd values from the start which threw the training off. Manually initializing them solved it.

Sign up or log in to comment