Asymmetric Loss

[zhongqiu1245]

Describe the feature
An effective Loss to balance the postive samples and negtive samples.

Related resources
offical paper: https://arxiv.org/abs/2009.14119
offical code: https://github.com/Alibaba-MIIL/ASL
open_mmlab code(mmcls): https://github.com/open-mmlab/mmclassification/blob/master/mmcls/models/losses/asymmetric_loss.py

Thank you!

[MengzhangLI]

Hi, it's noted but we do not have clear time schedule due to limited human resources.

PR of you is welcomed!

[MengzhangLI]

I just scaned related pr. It won't take too much time to transfer to mmsegmentation.

If time available, I would like to add it. But it would be better if you could make a pr and we review it because we want more community members getting involved. If you want to do that, feel free to contact me.

Best,

[zhongqiu1245]

Sorry for the late reply.
I haven't create pr yet, so I am learning it.
If time available, I will create pr about Asymmetric Loss.

[MengzhangLI]

Here are some steps:

(1) Installation of MMSegmentation.
Please refer to get_started.md for installation and dataset_prepare.md for dataset preparation.

(2) Making a pull request.
Usefule links: Contributing to OpenMMLab and Chinese article from zhihu.

(3) Support the Asymmetric Loss.
Usefule links: doc of customized loss function and PR about supporting focal loss.

[zhongqiu1245]

@MengzhangLI thx！