I’m training some vae style models with the recently added conditional planar flows transforms and finding that I get nan as training proceeds. I’m mostly curious if this behavior is or is not so crazy, specifically in relation to the planar flows as they are.
Based on how they’re constructed, and the definition of the determinant of the Jacobian, it seems like they shouldn’t have any sort of numerical instability as they are…
not sure if the version matters, but this is working off a slighlty older version of the official release version of the planar flows.