I understand that there’s a lot of work done in providing unbiased gradient estimators for SVI in numpyro. If my model() (I.e. my simulation or forward model) is fully differentiable with respect to its inputs (I’ve checked this explicitly in Jax), can I leverage this? Is this automatically done in numpyro? I’ve looked for docs on this, but haven’t been able to find anything. In my specific use-case, my guide is a BNAF and my likelihood is a deterministic simulation. I’m performing SVI to search the input space conditioned on a desired output of the simulation.
if you’re using SVI and your guide is reparameterizable (e.g. AutoNormal, BNAF, …) and your model log density is differentiable w.r.t. to all latent variables then yes the differentiability will be leveraged automatically