Open jczh98 opened 4 years ago
@neverfelly Thank you for your pointing. I agree with your interpretation of CX. I don't remember why the inversion was placed in the line... Anyway, according to https://gist.github.com/yunjey/3105146c736f9c1055463c33b4c989da, I'm sure that it was my fault.
On the current master branch, it has been fixed. If you confirm it or have the other questions, please post here.
Thanks for your great contribution to the contextual loss package. Also, I found a problem with contextual bilateral loss, the original implementation uses MSE distance instead of L1 distance.
CoBi is still under developing because of the OOM during L2 distance computation. I could test it only with L1 so I selected L1 for the computation of spatial loss to keep consistency in the test.
The intrinsic solution is to avoid the OOM but I don't have any idea for memory-efficient computation now.
Origin paper mention that CX(X, X) = 1 while CX(X,Y) = 1/N if X is far from Y. So contextual loss is zero when two images are equal and large when two images are dissimilarity. I found that your code which computes equation (1) uses 1 - CX and I'm confused with this and its result. I wish your help, thanks.