When computing KL divergences between two multivariate Gaussians, there are warnings if one of them has zero determinant. The warning(actually it's also a minor bug since one of the KL divergences would be evaluated as -inf and later truncated to 0) is eliminated by directly returning KL divergences as np.inf in that special case.
When computing KL divergences between two multivariate Gaussians, there are warnings if one of them has zero determinant. The warning(actually it's also a minor bug since one of the KL divergences would be evaluated as
-inf
and later truncated to 0) is eliminated by directly returning KL divergences asnp.inf
in that special case.