VGG models for perceptual loss / texture loss

[daniel-j-h]

Hey hey 👋 amazing project 🚀 🌔

For some use cases, vgg still works best. For example perceptual loss / texture loss or similar feature based losses where we only use the first few layers and then compute a loss based on the features for inputs and targets.

For some reason vgg's stem and initial layers still give better results in some cases - therefore it would be awesome to also have vgg (and its variants e.g. with BN) in here for consistency, for strong pre-trainined weights, and to benefit from this awesome project.

This feature is not urgent, the workaround is to simply use e.g. torchvision's pre-trained vgg16_bn.

[rwightman]

@daniel-j-h I had a local branch with this on it a year ago... was intending to include it since it's a default backbone for unet style networks. However, that got lost somewhere... I'll consider adding it again in todos. Probably the BN variant only. The non BN is a big PITA to train.

[zakajd]

Hey, I'm also interested in tuned VGG models.
Here is the recent paper + code that shows their potential
https://github.com/DingXiaoH/RepVGG

[rwightman]

VGG and RepVGG are in #419 ... not yet merged. VGG is essentially torchvision version with a few tweaks, RepVGG sits on a diff backbone.

[rwightman]

Merged