PaddlePaddle / PaddleClas

A treasure chest for visual classification and recognition powered by PaddlePaddle
Apache License 2.0
5.47k stars 1.17k forks source link

多标签 weight loss 的疑惑 #2951

Closed thsno02 closed 10 months ago

thsno02 commented 1 year ago

Rethinking of Pedestrian Attribute Recognition: Realistic Datasets with Efficient Method 这篇论文中提到了三种 weight loss 的方法,

image

实验结果如下:

image

PaddleClas 在多标签中仅包含了 https://github.com/PaddlePaddle/PaddleClas/blob/9984080a3d7ca52dd9044afaae1df1bed676fd1c/ppcls/loss/multilabelloss.py#L6-L14 WF1 这一种,请问这里是有什么考量吗?比如基于实验,WF2 和 WF3 的表现不如 WF1?还是说目前没有官方支持?

REF: https://arxiv.org/abs/2107.03576

thsno02 commented 1 year ago

对多标签感兴趣的朋友可以试试这两个 weight function,在我的任务上,有显著提升。

cuicheng01 commented 1 year ago

@thsno02 目前没有支持,欢迎提PR给PaddleClas~