I was thinking that it would be useful to have an automatic versioning in learners so that if you training too far and overfit, you can just roll back to the previous version. I think it should either keep a version between each epoch, or each time fit/fit_one_cycle is called. I think this would just be a good idea for convenience’s sake; if you trained something overnight, and then realized you overfitted you could just roll back to before instead of having to create a new learner and do all of the training you did previously. I know you can call .save but sometimes you forget, or you’re just messing around and then you realized you made something very good.
As an added note, I can’t imagine this would be too hard to implement.
Hi Max. I think you do this with SaveModelCallback, by setting every_epoch=True.
Or perhaps a small modification to that code would exactly suit your needs.
HTH, 
That’s great! Thanks!