Chen Ma

Oh, no! I found the candidate and eval_sample returned value are depended on each other.

@HuichenLi Can you help to answer these questions?

Dear HuiChenLi, I want to add the compared experiments of QEBA in my paper **so much**. However, when I follow the code into depth, I found the passed parameter `tgt_label`...

@HuiChenLi , I will check the output label with QEBA generated adversarial example, and I will contact with you later.

你用这个选项吧 --temporal_edge_mode，具体你看看代码，这个代码几年前写的，我记得不是很清了，但整体逻辑是把人脸的各个区域变成一个个node，然后整体一个graph连接起来。

OK, I will add a part in README. BTW, I am implementing a PyTorch version of AU R-CNN now.

@Jypeng-64 我已经写好了，把邮箱发我，我发给你

The following file is the compressed zip file of pytorch version of AU R-CNN. [AU_RCNN_pytorch.zip](https://github.com/machanic/AU_R-CNN/files/13696108/AU_RCNN_pytorch.zip)

If you use GPU, you should always use NCHW

In my experiment, I found the attack successful rate is low in many pictures. Why the attack successful rate is lower than BIM(iterative FGSM) and PGD?