Closed less-and-less-bugs closed 2 years ago
This is reasonable because we use the empirical estimate of JMMD as in the original paper (Equ 9). It may be better to use other approximations, and I'm not particularly aware of improvements to this myself.
Thank you very much.
Hi, the jmmd loss will be negative given the existing implementation of JMMD in class JointMultipleKernelMaximumMeanDiscrepancy(nn.Module).
I think may be another approximation of JMMD should be better.