If the loss function is set as a default argument it will be loaded even
if not used, which is not desired.
FeatureLoss is used only for training, but was being loaded also for
inference, using a little more memory than required due VGG16 features.
This will not cause any harm because training and inference usage always
set gen_loss argument.
It will be better to do:
from deoldify import visualize
Than:
from fasterai import visualize
The PyPI package will be called DeOldify, in this case
it makes more sense to have an import name that matches
the package name.
Also, fasterai resembles fastai library, and DeOldify
is not a library to use with fastai, it's built on
top of it.
Gareth Davidson