Add ConditionalDETR to HuggingFace Transformers

[NielsRogge]

Hi!

As ConditionalDETR seems like some (relatively) minor modifications to the original DETR, it might make sense to add ConditionalDETR to HuggingFace Transformers to increase visibility and adoption. We do have the original DETR in the library, found here: https://huggingface.co/docs/transformers/model_doc/detr. This also comes with nice inference widgets on the hub, check out this one for instance (on the right -> you can directly try out DETR in the browser!): https://huggingface.co/facebook/detr-resnet-50.

The Python implementation is made in a single python script, found here: https://github.com/huggingface/transformers/blob/main/src/transformers/models/detr/modeling_detr.py.

So, if we want to add ConditionalDETR, it would have to be implemented in modeling_conditional_detr.py, which includes the modifications compared to modeling_detr.py.

Are you interested in adding this model to the library?

Kind regards,

Niels, ML Engineer @ HuggingFace

[DeppMeng]

Hi Niels,

Thank you for your message! That will be great, we are very interested in. I am kind of busy today. I will check out the implementation of DETR in HuggingFace and implement Conditional DETR based on it in the next couple days.

If I understand correctly, we will need to provide both the single-script implementation and the pretrained weights, right?

Thanks, Depu

[NielsRogge]

Yes that's correct! Maybe it might be easiest to set up a slack channel for easier communication. Is that ok for you?

[DeppMeng]

Sure. I am not very familiar with Slack. Can you invite me to a slack channel? My email address is mdpustc@gmail.com