[FEATURE] DINOv2 pretrained vit weights

[seefun]

Is your feature request related to a problem? Please describe.
dinov2 has been open sourced with the pretrained vit weight. Could it be added to timm?

[rwightman]

yeah, it's fairly easy to add, especially with the recent eva.py variant as it includes the swiglu, just needs a slight remap as most of the model they use is derived from timm at some point, will do but not urgent level priority since it's a non-commercial license and all :)

[leng-yue]

As I am also relying on DINOv2 for my work, I will o implement it and submit a pull request within the next few days.

[leng-yue]

I have implemented the small, base, and large models, and they appear to be functioning correctly. However, I need to fix the SwiGLU before proceeding with the implementation of the giant model.

Specifically, the code below attempts to initialize a non-existent linear fc1a, which results in an error:

pytorch-image-models/timm/layers/mlp.py

Lines 133 to 134 in 7326470

	nn.init.ones_(self.fc1a.bias)
	nn.init.normal_(self.fc1a.weight, std=1e-6)

Additionally, the official repository supports different image sizes than 518 by dynamically resizing the positional embedding during inference. I am uncertain about the best approach to incorporate this feature into timm.

[leng-yue]

Update: I fixed the SwiGLU bug and updated the giant model.

[rwightman]

added thanks to @leng-yue ... will update README shortly before next release