When will gfocal loss+use_sigmoid=False be implemented?

open-mmlab / mmdetection

OpenMMLab Detection Toolbox and Benchmark

https://mmdetection.readthedocs.io

Apache License 2.0

28.5k stars 9.28k forks source link

When will gfocal loss+use_sigmoid=False be implemented? #6235

Open jackhu-bme opened 2 years ago

jackhu-bme commented 2 years ago

Describe the feature In gfocal loss, the use_sigmoind parameter must be set to True in latest versions. Or it will raise not implemented error.

Motivation In CVPR2021, GFocal loss v2 is published and it uses a new GFocalHead but an old GFocal loss with use_sigmoid=False.

Related resources You can look at this repo of GFocalV2 based on mmdetection 2.6.0, which is not supported in 2.16.0&2.17.0. https://github.com/implus/GFocalV2

Additional context It really worths to be implemented. After that you can simply add the config file in repo I mentioned above to support GFocalV2.

AronLin commented 2 years ago

We will pay attention to this feature, but this is not our plan for the time being. If you want to support this feature in MMDetection, PRs are definitely welcomed! We will also organize the PR review as soon as the PRs are created.

jackhu-bme commented 2 years ago

Thanks a lot! I 'm working on this recently. If I could accomplish this correctly, I'll surely submit my code!

shinya7y commented 2 years ago

Hi! @Jack-Hu-2001 I believe the following modifications will be helpful to create a PR.

Add configs/gflv2/ https://github.com/shinya7y/UniverseNet/tree/master/configs/gflv2
Update mmdet/models/dense_heads/gfl_head.py https://github.com/shinya7y/UniverseNet/commit/10ad833bf0c9e341be0a99cf4e116824a509921f#diff-46684aaa9343774704db6d90236b2513f19b3a71f8c7732a75ec52dbeaac956a The update of L224-L227 is not needed unless supporting SEPC.
Update mmdet/models/losses/gfocal_loss.py https://github.com/shinya7y/UniverseNet/commit/4e8c51e72d6029c4e090b63fa0f98bd34dd7237e#diff-8e73774c02458bbeded443614ff1071129eee1653d62cef269c90733be380358

In my opinion, GFLv2 is so similar to GFLv1 that we don't need new files in mmdet/models.

jackhu-bme commented 2 years ago

Tha's so nice of you! I simply replaced the new files in mmdet2.16.0 with old files in 2.6.0 yesterday.Surely your update is prefect and impatiable with both versions. Thanks a lot! I' ll start using and testing right now! If other problems got arisen, I'll discuss them later here. It' my great honor and luck to use such a well maintained open-source program!!!