how we should decide the ‘lr’ and train once again, and one more query this how to freeze and how to unfreeze the things in fastai, please explain someone
for example:
lr = 1.5e-03
learn.fit_one_cycle(10, slice(lr))
how we should decide the ‘lr’ and train once again, and one more query this how to freeze and how to unfreeze the things in fastai, please explain someone
for example:
lr = 1.5e-03
learn.fit_one_cycle(10, slice(lr))
My hunch would be using the learning rate after the spike. But I think you should test both learning rates - before and after the spike - with one epoch, and then decide which direction to head in.
What is that min numerical gradient?? Is that the top of the steepest slope?
why don’t my plots have those??
I think you have to use the suggestion parameter to see the min numerical gradient.
learner.recorder.plot(suggestion=True)
I think the lr = 1e-04 bro…