LowRankMultivariateNormal example raises ValueError

1

I am trying to make use of LowRankMultivariateNormal but I have trouble to correctly specify cov_factor and cov_diag. Unfortunately, the example in the documentation also raises a ValueError for me (pytorch version 1.3.1):

m = LowRankMultivariateNormal(torch.zeros(2), torch.tensor([1, 0]), torch.tensor([1, 1]))
# ValueError: cov_factor must be at least two-dimensional, with optional leading batch dimensions

I haven’t been able to figure out the right amount of (un-) squeezing to make it work. Can someone help?

2

Thanks for reporting this issue!
The example doesn’t seem to work indeed.
Could you use these test use cases as an example code, while this issue is being fixed?

CC @fehiepsi

1 Like
3

Oops… the cov_factor in the example is missing a trailing dimension (which has size “rank”). Thanks for pointing it out! I’ll fix it soon. In the meantime, you can simply unsqueeze cov_factor to say that the distribution has rank 1. In general, cov_factor will have shape event_shape + rank

3 Likes
4

Ah great, thank you! This works for me (I had to change the integers for the tensor creation of cov_factor to floats as otherwise one gets a RuntimeError: "cholesky_cpu" not implemented for 'Long')

LowRankMultivariateNormal(torch.zeros(2), torch.tensor([1.0, 0.0]).unsqueeze(-1), torch.tensor([1, 1]))
1 Like
5

Dear smonays

May I ask: if we have a low-rank cov-matrix at hand, what is the correct way to find out cov_factor and cov_diag, in order to use this LowRankMultivariateNormal() function ??

The example here is a very simplified case I guess.

Thanks