Integration with Detectron2

[jmspereira]

First of all, excellent work!

I know that probably the integration with detectron2 is not automatic since there are differences in the training architecture compared with the default detectron2 procedure. But there are any plans to integrate DETR in detectron2?

Thank you!

[fmassa]

Hi,

Thanks for the message!

There is some work needed in order to integrate DETR with detectron2, but this is something that should definitely be considered. My initial thinking was just to add a compatibility layer in a folder d2_compat/, which uses detectron2 dataset / transforms abstractions, while re-using most of DETR codebase for model definition.

cc @ppwwyyxx @wanyenlo for thoughts if we should integrate DETR directly inside detectron2, or use D2 as a library via a compatibility layer.

[jmspereira]

Thank you very much for the quick response!
I look forward to any news! 😄

[vantienpham]

Hi there, detectron2 brings me here :)

[poodarchu]

I've implemented DETR based on Detectron2 at:
https://github.com/poodarchu/DETR.detectron2

[poodarchu]

I've implemented DETR based on Detectron2 at:
https://github.com/poodarchu/DETR.detectron2

I'm glad to share the latest progress that I've make it usable and can reproduce the results in official implementation.

[asharakeh]

@poodarchu Thank you for this!

[alcinos]

Hi all, we are releasing the official Detectron2 wrapper (see #103), along with a conversion script.
Let us know if it fits your needs.

[jmspereira]

I do not have time to test the usage in the following days, but it seems to perfectly fit my needs.
Thank you again for the amazing work!

[poodarchu]

I think the MaskBackbone and Joiner are unnecessary and will introduce burdens of understanding.

[fmassa]

@poodarchu yes, the Joiner is unfortunate and reminiscent of previous (more complicated) implementation.
Removing it is totally possible, but care would need to be taken to avoid breaking backwards-compatibility with previously saved models, and for the sake of simplicity, we didn't do it before the release.

[alcinos]

Thanks for your feedback @poodarchu.
We took the option to design the wrapper to be as thin as possible so as to reduce maintenance cost for us, and ensure exact reproduction of the paper's results. In particular, our wrapper doesn't require rewriting the DETR's forward function, which is the approach you took in your codebase.
As you noted, this requires a tiny packaging (aka MaskBackbone) of Detectron2's backbones so that they return features along with the mask in the way expected by DETR. The Joiner is imported directly from DETR's codebase.

Both options are viable and have their own pros and cons, I hope this clarifies the motivation behind our choices.

Since the wrapper landed on master, I'm closing this. Feel free to reach out if you have further concerns.