lsgos / uncertainty-adversarial-paper Goto Github PK

Hi,
thank you for making the code needed to reproduce the experiments from your paper publicly available. Currently, I'm trying to reproduce the ROC curves in figure 14 from the paper. Apparently the default parameters of this code version produce very different results, some examples attached below.

Could you please share the necessary parameter configuration for

• number of examples
• number of MC passes
• exact configuration for the attack parameters
• seed for the random number generator to select the same images

Training the model worked fine. The validation accuracy after 40 epochs was about 94%. However, it would also be nice if you could share your model parameters to rule out any doubts in this regard.

I just added a title to the plot to identify the related attack configuration. I haven't changed any of the code's logic.

Hi,

some of the attack parameter dictionaries contain the key "eps_iter" two times. The dictionaries with

• "method": "bim", "eps": 5 and
• "method": "mim", "eps": 5 do so.

Cheers,
Johannes