Update Gumbel intersection to work with automatic broadcasting.

iesl / box-embeddings

Box Embeddings as Modules

https://www.iesl.cs.umass.edu/box-embeddings

Apache License 2.0

100 stars 10 forks source link

Closed dhruvdcoder closed 4 years ago

dhruvdcoder commented 4 years ago

Implemented a logsumexp2 which takes two arguments. See logaddexp for torch version 1.7+.
Updated the gumbel intersection to use logsumexp2/logaddexp. This makes it work with automatic broadcasting.
Performed random testing for numerical stability and for checking if the theoretical inequalities: max(a,b) < beta*logaddexp(a/beta, b/beta) and min(a,b) > -beta*logaddexp(-a/beta, -b/beta) hold. Currently, the use of torch.nextafter on logaddexp seem to be enough for this.

dhruvdcoder commented 4 years ago

@mboratko I discovered an issue just now!