Loading pretrained EVA-02 weights using RoPE at different resolution

[CardenB]

Hi,

Does anyone have any guidance on how to load pretrained weights using RoPE for different resolution input image for fine tuning?

It's unclear to me how EVA-02 went for 1024 MIM checkpoints fine tuned at 1536 for o365. I'd personally like to do the inverse.

What are the primary issues to avoid? Is it the changing size for the global attention indexes?

Appreciate any insight people have to offer. Let me know if I can share more information to make the discussion easier.

Thanks