the value of 'm‘ in “Batched Adversarial Attack” of your paper

[chenpan0103]

Excuse me, what's the value of 'm‘ in “Batched Adversarial Attack” of your paper? Could you give a detailed guidance to reproduce the result of Table in your paper?

[cdluminate]

By default -m is not specified. In that way, the algorithm will go through the whole dataset. The step-wise, detailed guidance is written in README.md -- please be specific about the part which you don't understand. Meanwhile, this code base is connected with two papers, please be specific about the name of the paper.

[cdluminate]

In the paper you can find that the algorithm will go through the whole dataset for reporting the number. -m is mostly used for debugging, and can be directly omitted.

[chenpan0103]

ok,thanks a lot. I want to reproduce the result of 《Enhancing Adversarial Robustness for Deep Metric Learning》. However, it is relatively different from the results of the paper. Is it normal?

[chenpan0103]

I follow the step of README.md, I think there's no wrong step. And the result is trained on CUB dataset

[cdluminate]

This difference is normal. What you see here is within the error bar. Due to different initialization and other factors (such as the number of GPUs in DDP mode), the performance differs slightly. If you want to see a higher ERS, just try to limit the GPU number to 1 or 2 (if I remember correctly. If not, it should be the reverse way -- more GPUs -- there will be a slight trade-off between R1 and ERS when changing the GPU number. This is a common phenomenon in parallel training), and try some more initialization.

[chenpan0103]

ok ,I will try it. As to the R1/R2/mAP/NMI in your paper, are they the result of training end?

[cdluminate]

Yes, they are reported at the training end status. Because in adversarial training, these standard benign metrics may look like a U-shape curve or directly a descending curve ... That's a part of adversarial training sacrificing the benign performance.

[chenpan0103]

Get it, thank you very much!

[chenpan0103]

Excuse me, which is the mAP in 《Enhancing Adversarial Robustness for Deep Metric Learning》, mAP or mAP@R?

[cdluminate]

It's simply the original mAP. If mAP@R is used, it should have been explicitly justified.