So I’ve recently been playing with Callbacks and wondering why I haven’t been using these in the past, less time spending running models for unnecessary epochs.
My question is about when to use callbacks in the learner versuses the fitting e.g. fine_tune?
I understand some aren’t meant for learner/fitting but those that are when is it best to use them learner or fitter?
You want to use callbacks on the Learner when you always want it present. You should have it in your fit when you want it just during training.
A common confusion is passing in MixUp() into Learner. What this means is MixUp is always present, even when we get predictions on normal images. We obviously do not want this as if we’re getting just predictions we don’t have labels, etc. so instead it should be passed to fit. (Most callbacks should be passed to fit)