fine_tune has default values: (see doc here)
Learner.fine_tune(epochs, base_lr=0.002, freeze_epochs=1, lr_mult=100, pct_start=0.3, div=5.0, lr_max=None, div_final=100000.0, wd=None, moms=None, cbs=None, reset_opt=False)
if you do not specify any lr, it will use base_lr=0.002
. But this may not be optimum. You are beter off trying lr_find()
and then define an optimum lr.
1 Like
I try with another data and my error_rate when using discriminative is bigger than error_rate not using it. What happens with my model and how can I fix it?
This may depend on many things:
- the lr you use (if not optimum for the specific network and dataset, will fare poorly)
- other hyperparameters
- the number of epochs
Also, you have a training error and a validation error, how do they evolve?
Refer to fastbook so see whether you are overfitting or not and how to pick the right lr, …
Hard to give an absolute answer like that. Sorry
1 Like
thanks a lot. I think my data is bad. I create data from Bing Search API.