s2anet icon indicating copy to clipboard operation
s2anet copied to clipboard

Published 20 hours ago • verified publisher icon csuhan

s2anet支持使用其他分类损失吗？我尝试了eql_loss(用于处理长尾现象），但是训练时，报错如下：如果使用别的损失，需要注意哪些地方吗，希望您可以给出一些建议。

Open xc-chengdu opened this issue 3 years ago • 3 comments

baocuo gengai2

xc-chengdu avatar May 03 '22 03:05 xc-chengdu

Maybe you can add 'EQLv2' to this line so that we do not use another sampler.

https://github.com/csuhan/s2anet/blob/d2870963815b10c8f7796577e30b7c3b14e47ac6/mmdet/models/anchor_heads_rotated/s2anet_head.py#L64

csuhan avatar May 03 '22 10:05 csuhan

Thanks for your suggestion, I will try it.

xc-chengdu avatar May 04 '22 03:05 xc-chengdu

Hello, I would like to ask again, have you tried using the non-cuda version of focal loss? Why is there such a big difference between the two trained models?

xc-chengdu avatar May 04 '22 10:05 xc-chengdu