Could you please tell me the gpu edition used in your CAA search? - Githubissues

vtddggg / CAA

The implementation of our paper: Composite Adversarial Attacks (AAAI2021)

https://arxiv.org/abs/2012.05434

Apache License 2.0

30 stars 3 forks source link

Could you please tell me the gpu edition used in your CAA search? #3

Open yyl-github-1896 opened 3 years ago

yyl-github-1896 commented 3 years ago

Hi, I found there is no NAGA-II search code relased, so I reimplemented by myself. I followed the direction in your paper, used pymoo, set the search paramater as in your paper, and used 3 2080Ti GPU. But I found that my search time consumption on CIFAR-10 (more than 10 days on 3GPU) is much more than that the search time reported in your paper (3GPU/d). Did you mean by 3GPU/d that you searched on 3GPU and finished in 1 day? Is that because your gpu is much faster than mine? Could you please tell me the gpu edition you used in your CAA search?

vtddggg commented 3 years ago

The speed is determined by many factors.

2080Ti has only 11G memory but P100/V100 has 16G. It means the batch-size can be smaller if you use 2080Ti. I use V100 in the experiments.
If the installed torch is accelerated by cudnn.
the hyper-params used in NAGA-II implementation.

Also there will be other factors which is not mentioned above.

zzx528 commented 3 years ago

We had a lot of problems reproducing the paper, can you provide NAGA-II search code?

vtddggg commented 3 years ago

Sorry, there is no plan to open the search code till now.

You can refer to pymoo and MixPath to implement the NAGA-II search. If you have any other problems during reimplementing the CAA, feel free to put questions here.

yyl-github-1896 commented 3 years ago

We had a lot of problems reproducing the paper, can you provide NAGA-II search code?

My reimplementation is very far from perfect, but if you like, I can share my code through e-mail (please tell me your e-mail address), and I will be happy to talk about these problems.

Rainwind1995 commented 3 years ago

We had a lot of problems reproducing the paper, can you provide NAGA-II search code?

My reimplementation is very far from perfect, but if you like, I can share my code through e-mail (please tell me your e-mail address), and I will be happy to talk about these problems.

Can you send me a code for reference?

zzx528 commented 3 years ago

We had a lot of problems reproducing the paper, can you provide NAGA-II search code?

My reimplementation is very far from perfect, but if you like, I can share my code through e-mail (please tell me your e-mail address), and I will be happy to talk about these problems.

OK，thank you ! my e-mail 2432035318@qq.com

yyl-github-1896 commented 3 years ago

We had a lot of problems reproducing the paper, can you provide NAGA-II search code?

My reimplementation is very far from perfect, but if you like, I can share my code through e-mail (please tell me your e-mail address), and I will be happy to talk about these problems.

Can you send me a code for reference?

Please tell me your e-mail, I will send to you soon

Rainwind1995 commented 3 years ago

We had a lot of problems reproducing the paper, can you provide NAGA-II search code?

My reimplementation is very far from perfect, but if you like, I can share my code through e-mail (please tell me your e-mail address), and I will be happy to talk about these problems.

Can you provide your email address? I'd like to ask you some questions about the NSGA2 code you provided

yyl-github-1896 commented 3 years ago

We had a lot of problems reproducing the paper, can you provide NAGA-II search code?

My reimplementation is very far from perfect, but if you like, I can share my code through e-mail (please tell me your e-mail address), and I will be happy to talk about these problems.

Can you provide your email address? I'd like to ask you some questions about the NSGA2 code you provided

1634751580@qq.com

zzx528 commented 3 years ago

Sorry, there is no plan to open the search code till now.

You can refer to pymoo and MixPath to implement the NAGA-II search. If you have any other problems during reimplementing the CAA, feel free to put questions here.

Now, I'm having some problems with my experiment, why is res.F and res.X always None, no search results. Can you tell me what caused this?

shudong-zhang commented 2 years ago

We had a lot of problems reproducing the paper, can you provide NAGA-II search code?

My reimplementation is very far from perfect, but if you like, I can share my code through e-mail (please tell me your e-mail address), and I will be happy to talk about these problems.

Can you provide your email address? I'd like to ask you some questions about the NSGA2 code you provided

1634751580@qq.com

你好，可以提供下复现代码吗？sdong_zhang@163.com

Asber777 commented 2 years ago

Hi, I found there is no NAGA-II search code relased, so I reimplemented by myself. I followed the direction in your paper, used pymoo, set the search paramater as in your paper, and used 3 2080Ti GPU. But I found that my search time consumption on CIFAR-10 (more than 10 days on 3GPU) is much more than that the search time reported in your paper (3GPU/d). Did you mean by 3GPU/d that you searched on 3GPU and finished in 1 day? Is that because your gpu is much faster than mine? Could you please tell me the gpu edition you used in your CAA search?

请问可以要一份复现代码吗? 我的邮箱是xq.chen@siat.ac.cn 十分感谢

magic-liu2021 commented 2 years ago

Hi, I found there is no NAGA-II search code relased, so I reimplemented by myself. I followed the direction in your paper, used pymoo, set the search paramater as in your paper, and used 3 2080Ti GPU. But I found that my search time consumption on CIFAR-10 (more than 10 days on 3GPU) is much more than that the search time reported in your paper (3GPU/d). Did you mean by 3GPU/d that you searched on 3GPU and finished in 1 day? Is that because your gpu is much faster than mine? Could you please tell me the gpu edition you used in your CAA search?

Hello, I would also like to learn your reproduction code? Could you send me a copy, 76960446@qq.com. Thank you very much!

MyGoddess1 commented 8 months ago

Hi, I found there is no NAGA-II search code relased, so I reimplemented by myself. I followed the direction in your paper, used pymoo, set the search paramater as in your paper, and used 3 2080Ti GPU. But I found that my search time consumption on CIFAR-10 (more than 10 days on 3GPU) is much more than that the search time reported in your paper (3GPU/d). Did you mean by 3GPU/d that you searched on 3GPU and finished in 1 day? Is that because your gpu is much faster than mine? Could you please tell me the gpu edition you used in your CAA search?

Hello,i would also like to learn your reproceduction code about CCA,i am also a adversarial menthod learner.Could you send me a copy?ying_liu999@163.com.Thank you very much.