Why are you using diagonal matrix in the MultivariateNormalDiag?

[Gurkirat Singh]

In the class we can use loc as 1d array and also scale as 1d array of same shape. If it not passed you can use tf.ones(loc.shape). Why taking matrix?

loc = [lx,  ly,   lz, ......]

           |.     |.    |

std = [sx, sy, sz, .....]

[Gurkirat Singh]

I mean instead of Diag we can use this MultivariateNormal

[Gurkirat Singh]

This is to generalise that covariance matrix must have the following properties:

1. It is symmetric
2. It is positive (semi-)definite

Not all covar matrix are valid.