Thanks for a great package, I have been using Pyro for a few different projects now and it is really amazing! Definitely better than the other probabilistic programming packages I have tried and far better than implementing from scratch
If I may make a small suggestion/feature request, I was wondering if there are plans to reduce the need for transposes,
.reshape(-1,1) and code like
try: var1 = pyro.sample("var1", dist.Categorical(params.T)) except ValueError: var1 = pyro.sample("var1", dist.Categorical(params))
in a future version? I understand why they are necessary and it is only a minor annoyance but I find these make the code somewhat harder to read and debug, and it is not always clear without running the code where they are needed.