Would features_only model work as backbone for torchvision model

[AnhMinhTran]

Anyone here has tried putting using the feature extractor arg to make a backbone for torch.vision model and did it work for you guys?

[rwightman]

@AnhMinhTran no, it would not work via a factory method... if you created a .feature_info (list of dicts) like the models here that marks the module names for each feature level, you could wrap it in one of my feature extractors. Not sure it's worth the effort though, most models that are there are also here. And often have newer / better trained weights here.

[AnhMinhTran]

I might just do as you suggested for the backbone, thanks for your tip for faster RCNN on torchvision. By the way, I saw that you also have a faster RCNN repo but havent been updated for 2 years. Would timm model work on it as well?