Can we convert event dims to batch dims, e.g., using negative .to_event(-n)?


I know that using .to_event(n) converts batch dimensions to event dimensions. But I haven’t seen any documentation about how to convert event dimensions to batch dimensions.

I tried using .to_event() with a negative number and that seemed to reduce the event_dim of the distribution, but I’m a little weary that it might mess up something later (in inference maybe) because I can’t find any tutorials or documentation that uses .to_event() with negative numbers.

So, I’m wondering is it valid to use .to_event(-n) to convert event dims back to batch dims or could that mess something up in inference? Thanks!

i believe it’s ok. if you git grep the codebase you’ll find a few examples