Dropout for Marian MT models?

I’m finetuning a Marian mt model for a translation task, and I want to use dropout to avoid overfitting, but I cant figure out how to do it. Does anyone know how to do it?

Hey @TheaMT

You should be able to control dropout through the configuration file – see here.

Please note that it has to be defined at model initialization time! If you’re using from_pretrained, you can pass the (modified) configuration as an input – see here

I hope this helps :hugs: