How to train the 55.54 AA accuracy resnet-18 on the robustbench ?

inspire-group / proxy-distributions

[ICLR 2022 official code] Robust Learning Meets Generative Models: Can Proxy Distributions Improve Adversarial Robustness?

https://openreview.net/forum?id=WVX0NNVBBkV&noteId=05ntgCksbhL

26 stars 1 forks source link

How to train the 55.54 AA accuracy resnet-18 on the robustbench ? #1

Open wooozihui opened 1 year ago

wooozihui commented 1 year ago

I run the following script as recommended and only modified the arch to resnet18， CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node=4 --master_port=6116 train.py --dataset cifar10 --syn-data-list ddpm_cifar10 --syn-data-dir dir_where_syn_data_is_stored --arch resnet18 --trainer pgd --val-method pgd --batch-size 128 --batch-size-syn 128 --exp-name name_of_experiment but get a resnet-18 with only 53.89 AA accuracy. What should be modified for training better?

VSehwag commented 1 year ago

This model is also available at RobustBench (https://robustbench.github.io/), which provides a standardized evaluation pipeline. Can you try evaluating the checkpoint available at RobustBench.