intermediate tokens

[wangyf8848]

Hi，

I have got the mdoel using : 'dinov2_vitl14 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vitl14')'.
And how can I get the output tokens of specific intermediate layers when inferencing.

[aeinsara]

hi, do you find a solution?

[barbolo]

x = torch.rand((1, 3, 224, 224))
dinov2 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vits14_reg')
features = dinov2.forward_features(x)

features is a dict with keys:

['x_norm_clstoken', 'x_norm_regtokens', 'x_norm_patchtokens', 'x_prenorm', 'masks']

[paulgay]

x = torch.rand(1, 3, 224, 224)
x = dinov2_vitb14_reg_lc.backbone.patch_embed(x)
blocks = []
with torch.no_grad():
for m in dinov2_vitb14_reg_lc.backbone.blocks:
x = m(x)
blocks.append(x)

See also get_intermediate_layers function in dinov2/models/vision_transformer.py

[qasfb]

get_intermediate_layers is indeed what you need !