I am still seeing strange things that I don´t understand.
This first model was trained with fit_once_cycle(10, slice(1e-5,1e-4)):
This second model was trained with fit_once_cycle(20, slice(1e-5,1e-4)):
Why they aren´t the same losses??? Looks like number of epoch is afecting the results. I have looked into the code of fit_one_cycle and doesn´t look that it affects.
If you change the number of epochs, the schedule of the learning rate is not the same, hence the different losses. See the doc of one cycle for more information (GitHub pages are down so I can’t point the docs directly, sorry).
ohhhh, it is an interesting behaviour that didn´t know. Good to know! With doc you mean dev.fast.ai?
I have executed a notebook of a week ago after updating fastai2 and I am getting different plots and errors!!! Has been any important update to fastai 2?