(Un?)expected nans in loss function #2

If you fit a matrix Y with inverse variance weights W and any element in W is 0.0 then the loss will always be nan. It makes sense why that happens, but I think for the uninitiated user it might make sense to do a safe-divide by internally clipping W to have a minimum of some keyword argument epsilon (default 1e-12 or something), or by adding that epsilon to all W for internal operations.