Label smoothing in training option

[ptran1203]

Hi, I could not find any questions about label smoothing, so I wonder if there is label smoothing option in the training script?
I think it would be useful, as the authors (from this paper) demonstrated the performance boost.

[glenn-jocher]

@ptran1203 we do have a class label smoothing option, though we couldn't get better performance with it. You can experiment with it here, and let us know if it helps you!

I would also start with very small eps, i.e. 0.01 maybe for your experiments.

yolov5/utils/loss.py

Lines 99 to 101 in fab5085

	# Class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3
	self.cp, self.cn = smooth_BCE(eps=0.0)

[ptran1203]

@glenn-jocher Thank you, I see, I will do some experiments with this.
And maybe it is useful when adding this option into hyper params config. If there is no problem, can you consider a pull request?

[ptran1203]

@glenn-jocher , I saw hyp params already attached in the model and it can be easily obtained via h.get('label_smoothing_eps', 0.0)

[glenn-jocher]

Yes PR is good idea!

On Mon, Mar 1, 2021 at 7:44 PM Phat Tran ***@***.***> wrote: @glenn-jocher <https://github.com/glenn-jocher> , I saw hyp params already attached in the model and it can be easily obtained via h.get('label_smoothing_eps', 0.0) — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#2332 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AGMXEGLXLQNEPCOAG32XWL3TBRNJTANCNFSM4YN3N5MQ> .

-- <https://www.ultralytics.com/> *Glenn Jocher*Founder & CEO, Ultralytics LLC +1 301 237 6695 <https://www.facebook.com/ultralytics> <https://www.twitter.com/ultralytics> <https://www.youtube.com/ultralytics> <https://www.github.com/ultralytics> <https://www.linkedin.com/company/ultralytics> <https://www.instagram.com/ultralytics> <https://contact.ultralytics.com/>

[Robotatron]

@ptran1203 Did you obtain better mAP results with label smoothing?