[feature proposal] U-Nets with pretrained torchvision backbones

I was thinking about extending torchvision models with an U-Net builder for segmentation, that takes pre-trained torchvision classification models as backbone architectures in the encoder path of the U-Net, and builds a decoder on top of it, using features from specified layers of the backbone model.

I already implemented this for ResNet, DenseNet and VGG models in a separate module:
https://github.com/mkisantal/backboned-unet

Now I'm thinking about integrating it directly with torchvision. Do you think it would be a useful new feature?

It's not an addition to the available torchvision models in the traditional sense, as it just transforms the available models, does not work out of the box but requires training. But it can make torchvision easier to use for segmentation problems.

The proposal links to the separate backboned-unet module and names its ResNet, DenseNet, and VGG implementations. Start by reviewing that module and the relevant torchvision model interfaces; the integration scope and acceptance criteria would need to be decided before implementation.