Starlento
Starlento
> That's surprising! Do you happen to still have the full notebook transcript? The change is minor, basically the code from experiments.ipynb. Change model_name = "mistralai/Mistral-7B-Instruct-v0.1" to "Qwen/Qwen1.5-7B-Chat". user_tag, asst_tag...
For VIT, it supports dynamic resolution: https://discuss.huggingface.co/t/fine-tuning-vit-with-more-patches-higher-resolution/18731. For Dinov2, it originally support this, so that I copy the code to CLIP model file. Maybe it should be a config which...
And I am new to open-source community, so I apologize for not ticking anything above and any unproper behaviors... Feel free to comment and I am glad to learn.
> Thanks for adding @Starlento! > > Could you protect this behaviour behind an `interpolate_pos_encoding` argument flag, as done in other models e.g. [like here for ViT](https://github.com/huggingface/transformers/blob/b86c54d9ff4504b7287d95a215f8d6fa9388761f/src/transformers/models/vit/modeling_vit.py#L136)? That's a good...
> To resolve the code quality tests, run `make fix-copies` and then `make fixup` and push the changes Thank you for the suggestion. I just run the commands, and I...
> > Should I also solve the check_repository_consistency thing? > > @Starlento Yep, we want the changes CLIP to applied across to any models copying its architecture. The thing I...
> @Starlento Don't worry! The reason is that not all of CLIPs architecture will have been copied across for all models. Looking at the diff, you'll need to copy directly...
> Hi @Starlento, any update on this PR? It would be great to add this to the library! Hi, sorry that I am quite busy recently. I would like to...