Bugfix: Prevent model weights file duplicate to cache

[aselimc]

Hello,

First of all I use Windows, so this test was only done in Windows, not in Linux

REPO_DIR = r"C:\Users\aselimc\Repos\dinov3"
weights = r"C:\Users\aselimc\Repos\dinov3\checkpoints\dinov3_vit7b16_pretrain_sat493m-a6675841.pth"
dinov3_vit7b16 = torch.hub.load(REPO_DIR, 'dinov3_vit7b16', source='local', weights=weights)

The current implementation creates duplicate model weights in .cache/torch/ even if the weights have already been downloaded. This is caused by torch.hub.load_state_dict_from_url. As an example

Downloading: "file:///C:/Users/aselimc/Repos/dinov3/checkpoints/dinov3_vit7b16_pretrain_sat493m-a6675841.pth" to [C:\Users\aselimc/.cache\torch\hub\checkpoints\dinov3_vit7b16_pretrain_sat493m-a6675841.pth](file:///C:/Users/aselimc/.cache/torch/hub/checkpoints/dinov3_vit7b16_pretrain_sat493m-a6675841.pth)
100%|██████████| 25.0G/25.0G [00:10<00:00, 2.53GB/s]

my local file is copied into cache.

Therefore, I suggest using torch.load in case a local path is given to load state_dict.

[forensicmike]

Nevermind, I found them. Yeah, this should be fixed for sure!

[aselimc]

As I realized more into it, this behavior is not only for backbones but also for downstream heads.