Experiments for blog post 'Reassessing adversarial training with fixed data augmentation'

Setup

Install fork of robustness package:

git clone https://github.com/iamgroot42/robustness.git
cd robustness
python setup.py install

Training models

1. Model without data-augmentation

   python -m robustness.main --dataset cifar --data data/ --arch resnet50 --out-dir models/ --exp-name standard_noaug --data-aug 0 --adv-train 0

2. Model with data-augmentation (faulty worker-init)

   python -m robustness.main --dataset cifar --data data/ --arch resnet50 --out-dir models/ --exp-name standard_aug --data-aug 1 --adv-train 0

3. Model with data-augmentation (fixed worker-init)

   python -m robustness.main --dataset cifar --data data/ --arch resnet50 --out-dir models/ --exp-name standard_fixed_aug --data-aug 1 --adv-train 0 --better_init 1

Evaluating models

    python -m robustness.main --dataset cifar --data data/ --arch resnet50 --out-dir evals/ --attack-steps 20 --constraint inf --eps 8/255 --attack-lr 20/5100 --resume models/robust_fixed_aug/checkpoint.pt.best --eval-only 1 --adv-eval 1