Converting My pretrained EVA or EVA2 model to timms weights

[ayushnangia]

I'm working with the EVA and EVA2 models and would like to request a feature to convert the pretrained weights from these models to a format compatible with the timm (pytorch-image-models) library.

Background:
EVA (Exploring the Limits of Masked Visual Representation Learning at Scale) and its successor EVA2 are powerful vision-centric foundation models. Currently, I have pretrained weights for these models in the mmpretrain format, but I'd like to use them with the timm library for broader compatibility and integration with other workflows.

Feature Request:

1. Add a conversion script or function to transform EVA/EVA2 weights from mmpretrain format to timm format.
2. Ensure that the converted weights are compatible with timm's model loading and inference pipelines.
3. If possible, include documentation or examples demonstrating how to use the converted weights with timm.

Potential Benefits:

• Increased interoperability between mmpretrain and timm ecosystems
• Easier integration of EVA/EVA2 models into existing timm-based workflows
• Broader adoption and testing of EVA/EVA2 models in various applications

Questions:

1. Is this conversion feasible given the current architectures of EVA/EVA2 and timm's model structures?
2. Are there any known challenges or limitations in converting between these formats that should be considered?

Thank you for considering this feature request. I believe it would be valuable for the community and enhance the usability of these powerful models.

Let me know if you need any additional information or clarification.