Problem with batch sampling from Multivariate Normal Distribution

Misc.
1

Suppose batch size is 8 and vector length is 50. Basically, this is what happens:

a = torch.randn(8, 50)
b = torch.randn(8, 50)
with pyro.plate("c"):
    d = pyro.sample('d', Normal(a, b).to_event(1))
e = Normal(a, b).to_event(1).log_prob(d)

We get e = tensor([nan] * 8).

I can see why Pyro does not know how to calculate the log probability for a converted distribution, but this is the same with that in the semi-supervised tutorial. So if I want to sample from MultivariateNormal(a[i], diag(b[i])) for each i=1,...,8 in the mini-batch, what’s a correct and efficient way of dong it?

2

scale b has to be positive.

3

Hi JP, thanks for replying. That helps!