I'll try to remember all the changes as best I can. It includes a new cnn model that was trained on 150 instead of 50 epochs, so the validtion accuracy starts at around 98.9% instead of 96.7% (ish -- off the top of my head).
Also updates strategic_with_carlini.py to be compatible with python3, including helper functions in model_utils and l2_attack.py. It isn't a complete fix though, as I'm still getting an error when I try to run strategic_with_carlini. I also deleted a bunch of stuff in model_setup_carlini in lib/utils/model_utils, because I don't understand why the model setup would be different. I think the best/most trustworthy way forward is to paste in some code from carlini's repo here: https://github.com/carlini/nn_robust_attacks.
also note I updated the dev_list in run_defense to be [2,4] since one of the tests was failing on the small dataset. For the actual runs, should use dev_list=np.linspace(.1,2.5,6)
I'll try to remember all the changes as best I can. It includes a new cnn model that was trained on 150 instead of 50 epochs, so the validtion accuracy starts at around 98.9% instead of 96.7% (ish -- off the top of my head).
Also updates
strategic_with_carlini.py
to be compatible with python3, including helper functions in model_utils and l2_attack.py. It isn't a complete fix though, as I'm still getting an error when I try to runstrategic_with_carlini
. I also deleted a bunch of stuff inmodel_setup_carlini
inlib/utils/model_utils
, because I don't understand why the model setup would be different. I think the best/most trustworthy way forward is to paste in some code from carlini's repo here: https://github.com/carlini/nn_robust_attacks.