Loading Pretrained Models(dinov2_vits14_pretrain.pth, inov2_vitb14/dinov2_vitb14_pretrain.pth)

[kongminseok]

Hi, How can I initiate the pre-trained Dinov2 model by using load_state_dict? I've seen issue #91, but there is no information about vit_small and vit_large arch to load the state of pre-trained models.

from dinov2.models.vision_transformer import vit_small, vit_base, vit_large 

model_small = vit_small(
   ???
)

model_base = vit_base(
   ???
)

model_large = vit_large(
    patch_size=14,
    img_size=526,
    init_values=1.0,
    block_chunks=0
 )

model_small.load_state_dict(torch.load('dinov2_vits14_pretrain.pth'), strict=True)
model_base.load_state_dict(torch.load('dinov2_vitb14_pretrain.pth'), strict=True)
model_large.load_state_dict(torch.load('dinov2_vitl14_pretrain.pth'), strict=True)

[wangh09]

didn't verify but the following should work

model = vit_x(
    patch_size=14,
    img_size=518,
    init_values=1.0,
    block_chunks=0
 )
x = small/base/large

518 should be the correct img_size but 526 also works as 526//14 == 518//14

[qasfb]

yes
essentially the best way is to do what is done with torch.hub.load :
https://github.com/facebookresearch/dinov2/blob/main/hubconf.py