How big is the difference in training time and memory consumption between using and not using adversarial training in the search network process？

[wen-alan]

Excuse me, I want to know the difference between the training time and the memory consumption of the search network process with and without adversarial training? I want to know how much additional time and memory is needed to obtain robnet. I will be grateful for your answers to my questions。

[gmh14]

I am sorry I have no experience in running the search algorithm w/o adversarial training with the current pipeline, but I can provide the training cost w/ adversarial training for your reference: for CIFAR10, the adversarial training of the super-net needs ~17.5h with 32 1080Ti for 300 epochs.

[wen-alan]

Thank you for your reply. I initially think that the time consumed by adversarial training should be several times that of searching the network directly, so I want to find out if this is the case.

[gmh14]

You are correct. The adversarial training itself costs times of additional memory consumption and the super-net structure even exacerbates the situation.