search

val-iisc / NuAT

Towards Efficient and Effective Adversarial Training, NeurIPS 2021
MIT License
16 stars 1 forks source link

Some questions #4

Closed jiaxiaojunQAQ closed 2 years ago

jiaxiaojunQAQ commented 2 years ago

The used ResNet is different from the previous work, it has an extra layer ReLU? Whether the results reported in the paper are optimal or final checkpoints?

SravantiAddepalli commented 2 years ago

Hi @jiaxiaojunQAQ,

We used the ResNet-18 architecture from the following repository: https://github.com/yichuan9527/pytorch-cifar/blob/master/models/resnet.py Yes, there is an additional ReLU layer compared to the standard ResNet-18 model. However, we do not expect the results to be significantly different with the standard model as well.

We select the best model based on I-FGSM 7-step accuracy on a hold-out validation set. The code for selection of best model is also present in the same file that is used for training models.