search

JDAI-CV / fast-reid

SOTA Re-identification Methods and Toolbox
Apache License 2.0
3.42k stars 837 forks source link

sbs的cfg里，没有用到cricle loss？ #591

Closed elvindp closed 2 years ago

elvindp commented 2 years ago
  1. 在sbs的说明里，提到用circle loss的trick，但cfg里好像没有设置这个loss。不加的效果更好吗？
  2. 此外，在NAIC20里，用到了cosface，请问一下，加了cosface可以提高效果吗？
  3. 用sbs r50 ibn训练DukeMTMC的结果很差，rank-1在训练68个epoch达到89.18之后就开始下降，676个epoch后只剩83.66，中间在85徘徊了很久。这是正常的现象吗？

circle softmax就是circle loss……是我傻逼了…… 但circle loss里的pairwise loss和circle softmax的区别是什么？什么时候用？

谢谢

BokyLiu commented 2 years ago

https://github.com/JDAI-CV/fast-reid/issues/225#issuecomment-673911446

L1aoXingyu commented 2 years ago

circle softmax 和 pair-wise circle 可以具体看看 circle loss 那篇 paper，主要就是分类 loss 和度量学习 loss 的区别

elvindp commented 2 years ago

circle softmax 和 pair-wise circle 可以具体看看 circle loss 那篇 paper，主要就是分类 loss 和度量学习 loss 的区别

明白了，所以， 一个输入时，是一组logits，和softmax里存放的矩阵乘 另一个是embeding pair。

JoJoliking commented 2 years ago

@elvindp @L1aoXingyu 请问以下，pair-wise circle的输入，embedding是Batch*Embedding Dim的Tensor。Target是对应什么呢？对应这个Batch个的ID号么？

github-actions[bot] commented 2 years ago

This issue is stale because it has been open for 30 days with no activity.

github-actions[bot] commented 2 years ago

This issue was closed because it has been inactive for 14 days since being marked as stale.