Fritz Obermeyer

Hi @kamathhrishi no rush, just FYI we switched from travis-ci to github actions so you'll need to merge the dev branch into this for ci to pass.

Note you can already do this by wrapping `loss_fn` in a tracing hook: ```py sgpr = ... optimizer = ... loss_fn = pyro.infer.Trace_ELBO().differentiable_loss # Wrap loss_fn in a tracing hook....

This is probably better done upstream in PyTorch, but you can add a patch in Pyro if you want. Note we intentionally left this unimplemented because it leads to exponential...

@ngoodman, @neerajprad can you provide any simple example models? I thought I had implemented a model with a sparse [coo](https://en.wikipedia.org/wiki/Sparse_matrix#Coordinate_list_(COO))-style plate, but the only examples I can find use a...

Here's a start: I'd recommend jumping into a debugger and examining the invalid values. The error means that values are falling outside of the distributions support, typically either NAN values...

@fehiepsi sure, and let me know if you want me to review any code.

Hi @AlexanderAivazidis thanks for pointing out this bug. I wonder if the issue is with `.expand()`? Could you try removing the `.expand()` (instead expanding the arguments before creating the distribution)?...

@karalets do you recall your return value assumptions on `RelaxedBernoulliStraightThrough`?

Hi @nipunbatra, regarding 1. I like the idea of adding a `render_params` argument to `render_model` and possibly printing their constraints if `render_distributions == True`. @fehiepsi might also like this feature...

@karm216 this looks great, we'd love PR contributing this feature! Note there are rigorous tests for `pyro.infer.inspect`, so we'll need to (1) update a bunch of those tests and (2)...