Thanks for your great work. I found the inference time in table 14. However, you could see the actual time cost to run one batch of 64 images on cifar-rand-Linf-rn50.sh is so long(autoattack apgd-dlr about 118754.8seconds). I run the experiment with 4 A100 GPUs but the time shows on table14 with V100 is about 10seconds per image. Have you faced the long waiting time when you run the experiment? Is there any mistake?
slurm-56811695.txt