Feature Extraction Losses

[tristan-deep]

Current losses supported by Keras are documented here. They include probabilistic, regression and hinge losses. Something that would be nice to add are feature extraction losses, such as LPIPS or FID (afaik these are not included in Keras). Alternatively these could be added to the metrics.

As a starter I have a working Keras implementation of the LPIPS loss. It relies on:

vgg16 = keras.applications.vgg16.VGG16(include_top=False, weights="imagenet")

with some additional layers and it produces the exact same values as the original lpips package (of the authors) as well as the torch metrics implementation.

I can make a PR for adding LPIPS if useful, however I need some pointers on how to host the weights (currently I have them on Hugging Face).