aamir-mustafa / pcl-adversarial-defense

Adversarial Defense by Restricting the Hidden Space of Deep Neural Networks, in ICCV 2019
59 stars 14 forks source link

Proxity and Conproxity losses share different prototype centers in code #9

Closed krishnakanthnakka closed 4 years ago

krishnakanthnakka commented 4 years ago

As per Eq 5 of the paper, the prototypes centers for proximity loss and con-proximity loss should be the same. However, from the code, we can see different sets of prototypes for both losses. Can you please share your thoughts.

Please correct me I'm wrong.

aamir-mustafa commented 4 years ago

Hi,

Since the optimization step is made after computing the centres for both proximity and con_proximity losses, we have asssumed in our implementation that the centres chosen whould be closer (if not exact) in each case.

Hope that helps.