Make it safer to use factor statements in guides

[fritzo]

Blocking #2948

It has always been dangerous to use pyro.factor statements in guides #2829, because the guide doesn't know whether the factor arose from reparameterized or non-reparameterized sampling. This PR adds a has_rsample kwarg to pyro.factor and the underlying Unit distribution. The has_rsample need not be specified for pyro.factor statements in models, but is now required when using pyro.factor statements in guides. While this does cause some backwards incompatibility, I believe the likely small breakage will be worth introducing, so as to avoid silent errors (I've just spent half a week hunting down a bug in my own code due to this).

Tested

• added smoke tests
• added tests of value in Fix AutoGaussian and tests #2948

[martinjankowiak]

lgtm.

"Defaults to False when used in models" how so?

"added tests of value in Fix AutoGaussian and tests #2948" can you pleae elaborate?

[fritzo]

Thanks for reviewing!

"Defaults to False when used in models" how so?

By default dist.Unit.has_rsample = False, but in models this fact is not used in any inference algorithms.

"added tests of value in Fix AutoGaussian and tests #2948" can you pleae elaborate?

The strong high-precision tests of AutoGaussian in #2948 rely on the new behavior of pyro.factor(..., has_rsample=True). Since the new behavior is tested in that PR which is scheduled to merge soon, I have added only simple smoke tests in this PR.

[martinjankowiak]

re: The strong high-precision tests

can you please point to a specific test?

[fritzo]

can you please point to a specific test?

• tests/infer_test_autoguide.py::test_exact
• tests/infer_test_autoguide.py::test_exact_batch
• tests/infer_test_autoguide.py::test_exact_tree
• tests/infer/autoguide/test_gaussian.py -k structure(9 tests in total) which use check_backends_agree()