To train the network (and evaluate every epoch):
cd examples/undefended_pytorch_resnet
python main.py --evaluate
To run prediction and make coverage-error plot:
gsutil cp gs://unrestricted-advex/example_defenses/undefended_pytorch_resnet.pth.tar /tmp/undefended_pytorch_resnet.pth.tar
CUDA_VISIBLE_DEVICES=0 python main.py --resume=/tmp/undefended_pytorch_resnet.pth.tar --evaluate