I've implemented the J-S divergence simply by taking
m(x) = (mu(x) + log_sigma(x)) / 2
and then evaluated the kl divergence by replacing the log_sigma term in the original K-L divergence term with m(x), which obviously isn't right. Could someone help me out with a way to implement it?
I've implemented the J-S divergence simply by taking
m(x) = (mu(x) + log_sigma(x)) / 2and then evaluated the kl divergence by replacing the log_sigma term in the original K-L divergence term with m(x), which obviously isn't right. Could someone help me out with a way to implement it?