PyTorch Code for "Adversarial Interpolation Training"

Usage

The training environment can be setup as follows:

virtualenv -p python3 /YOUR_PATH_TO_VENVS/adv_interp_env
source /YOUR_PATH_TO_VENVS/adv_interp_env/bin/activate

git clone URL_TO_REPO
cd REPO
python setup.py install

Specify the path for saving checkpoints in adv_interp_train.sh, and then run

sh ./adv_interp_train.sh

Specify the corresponding model path and attack method in eval.sh and then run

sh ./eval.sh

A model trained on CIFAR10 using Adversarial Interpolation Training is here. Download it to ./pre_trained_adv_interp_models/ and then run

sh ./eval_pretrain.sh