I’ve followed the new nlp tutorial here: https://docs.fast.ai/text.html to finetune a language model, and add a custom head for sentiment classification.
I then save the models to disk and reload them later in a new script.
I’ve seen a fews posts around the forum where people mention that you have to turn off training mode on the model, otherwise it will still apply regularisations such as dropout.
Should I turn on eval mode for the loaded model with:
learner.model.eval() # or is it learner.model.eval = True
Or is this applied when using predict by default?