在RS.sh第19行，

Hello.
First of all, thank you for providing your code.
I have a question about your implementation of ensemble attack in vmi_di_ti_si_fgsm.py.
As you mentioned in your paper, I averaged the logits from four normally trained models in batch_grad and graph function. However, when i ran the code, error(cuda out of memory) occur, even if i reduced the batch size. Could you let me know environment for ensemble attack or provide the code for ensemble attack ?
(My GPU : Titan RTX 24Gb)
I will wait for your reply.

Thank you!

问题如上，取beta=1.5，vmi不如mi好

Thank you for sharing your nice work!

I have a question about ensemble attack.
Is there any code to run an ensemble attack? I can't find the ensemble in your codes.

您好，
请问如何使用Bit-red代码中的装饰器bit_depth_reduction进行图像处理？
能写个样例吗？十分感谢

您好！我在运行您的代码时候，程序报错最后一行提示找不到dev_data/val_rs文件，我将val_rs.csv的后缀名删除后，又报错提示找不到dev_data/val_rs.csv，我全部加上后，还是报错提示找不到dev_data/val_rs文件。
请问该如何解决

The differences in success rates are minor, but I wanted to check whether this difference is expected (and if so, why?). I chose to reproduce two arbitrary attack rows of Table 1 from the paper where the Inc-v3 is used: MI-FGSM and VNI-FGSM.

Steps taken:

1. Downloaded data and models from provided README link.
2. Ran mi_fgsm.py followed by simple_eval.py, then logged results.
3. Ran vni_fgsm.py followed by simple_eval.py, then logged results.

Results for MI-FGSM:

Results for VNI-FGSM:

In the case of MI-FGSM, my run shows that the attack sometimes underperforms reported results (by at most 1.7%). I haven't checked other rows of the table, but thought I should surface this and ask about it. Thanks!

作者你好！阅读了你的论文，想问下论文的实验细节。用的显卡是啥，用了几张卡，跑了多久？