NVlabs / DeepInversion

Official PyTorch implementation of Dreaming to Distill: Data-free Knowledge Transfer via DeepInversion (CVPR 2020)
Other
488 stars 79 forks source link

Question about R_compete term #26

Open minhquoc0712 opened 1 year ago

minhquoc0712 commented 1 year ago

Thank you for your interesting research,

I have a question about the R{compete} term. The paper states as follows about R{compete}:

"During optimization, this new term leads to new images the student cannot easily classify whereas the teacher can."

Since the Jensen-Shanon divergence is symmetry in terms of p(x^{hat}) and q(x^{hat}), the opposite might be true: The images are generated such that the teacher cannot classify them, but the student can.

Is it correct? And does it contradicts the purpose of the R_{compete} term?