Sample multivariate normal with per-example standard deviation

Hi, I want to do something similar to this:

mu = torch.zeros(5, 2)
sd = torch.ones(5)
torch.normal(mu, sd)

But I get RuntimeError: inconsistent tensor size.

I noticed in the 1d case it works:

mu = torch.zeros(5, 1)
sd = torch.ones(5)
torch.normal(mu, sd)

Sorry for the spam, sd = torch.ones(5, 2) works, so we could do sd.repeat(1, 2) if sd is one-dimensional.

Or[sd, sd], 1) if sd is a Variable, since repeat is not supported by Variable.

Actually, if you want the mean/std to be the same for all samples, you can just pass a number to torch.normal.

Good call. The use case I had was a network outputting gaussian parameters mu in N x D and log_sd in N x 1, so the example above was a bit off.

After doing some more searching, I think that using expand_as might be the most efficient? To summarize:

mu = Variable(torch.zeros(5, 2))
sd = Variable(torch.rand(5, 1))
torch.normal(mu, sd.expand_as(mu))
1 Like

Yes, expand will be much better! There will be no memory copy in this case.

1 Like