robust-principles
robust-principles copied to clipboard
How do I train a model like in the RobustBench ranking?
I saw on RobustBench that there is a RaWideResNet-101-2 from here. But when I look at your Makefile, I can train a RaWRN101_2 via make experiments/RaWRN101_2/.done_train, but I cannot evaluate this exact one on Autoattack
If I want an RaWRN101_2 being able to be evaluated on Autoattack, I have to do make experiments/PGDAT_RaWRN101_2/.done_train and then make experiments/PGDAT_RaWRN101_2/.done_test_aa
Now my question is: is the PGDAT variant even the same Model like on RobustBench or is it the RaWRN101_2 without PGDAT, but how can I evaluate it then, to check if I got the same accuracy like you got?