@fritzo - Can I just remove the pyro TransformedDistribution and Bijector classes and use the corresponding PyTorch implementations? I will modify the InverseAutoregressiveFlow class to inherit off of torch Bijector. I will probably need that for HMC too.

EDIT - I'll let it be for now, and have you and @martinjankowiak take care of that.

I agree the automated example discovery logic is getting kludgey.

Its not that bad yet. :) Just added a TODO to keep in mind that if we are messing around with that part too much, it might be worth revisiting.

Can I just remove the pyro TransformedDistribution

I am working on that in this branch https://github.com/uber/pyro/compare/torch-transformed-and-iaf
Feel free to take over the work if you like. I can open a PR if you want to discuss.

I am working on that in this branch https://github.com/uber/pyro/compare/torch-transformed-and-iaf Feel free to take over the work if you like. I can open a PR if you want to discuss.

Yeah, I will need this for HMC. Unfortunately, we can't mix RandomPrimitive and other distributions, since RandomPrimitive instances can't support properties (we need to pass in distribution parameters to those methods). As such they are not interchangeable with torch distributions.