Does calling lr_find or lr_find2 actually update the model's weights?

Looking at the code, calling lr_find or lr_find2 does indeed seem to update model weights … but can anyone confirm that this is, or is not, true?

If so, I’m wondering what the impact of calling these methods over and over again, with different hyperparameter combinations, might have on providing us with accurate information regarding what values would be best. Maybe the best practice is to always re-create the model before running the LR finder methods?

1 Like

Both methods save the model (in a file named tmp.h5) and load it again when they’re finished so there is no change to the model.


I see it now.

The methods just re-load the model as it was before calling lr_find/2() so that your model is in the same state it was before the finder method was called.

Thanks for the reply.

btw, you guys sticking with use_clr_beta for 1-cycle or are you updating it to something like 1-cycle? I remember Jeremy mentioning this would probably be changed but don’t know if that is the case.

I think the name will change in fastai v1.