pankessel / adv_explanation_ref Goto Github PK

Hi, thanks for the insightful paper and making the code public!
While trying to reproduce your results (namely figure 2 in the paper), I ran into issues with multiple attribution methods:
For gradient and guided backprop, the explanation visually does not change at all.
For grad_times_input the explanation does not look like the target explanation and the image is visibly disturbed.
For LRP the method works as expected, so I feel like it's not a general issue.
In all cases I used the parameters suggested in appendix A (num_iterations, learning_rate) and otherwise did not change your implementation or default parameters.
Do you know what could be the issue or are there other parameters that need to be changed? Thanks for your help!
Guided backprop:

Grad x input:

Hello! I was trying different things with this code base, and when I ran the attack (run_attack.py) with the same target image as the original image (tiger_cat.jpg), I expected the initial total loss to be 0.0 (as MSE Loss between explanations should be 0). But, the initial total loss started with some other value and then gradually converged to 0.0. I am unable to understand this behavior.

Just a heads up: The original paper links to https://github.com/pankessel/adv_explanation_ref, which does not exist. Should be possible to fix it by changing the repo name (might not be wanted).

Best regards
Chris