Forced weight download for dinov2 from torch hub

[barwojcik]

Issue

When model is initialized from hugging face like this:

model = MapAnything.from_pretrained("facebook/map-anything").to(device)

it forces weights download for dinov2_vitg14 from torch hub. It's redundant, as the weights are already included in hugging face checkpoint.

Cause

It looks like DINOv2Encoder from uniception always loads weights from torch hub, and forces the download when they are missing from hub's cache. It seems that there is no way to pass pretrained=False to torch.hub.load() inside this class. I'm going to raise the issue at https://github.com/castacks/UniCeption