Pyro: Some tests fail under pytorch 0.3

[zou3519]

Hi,
The folks over at pytorch are working on cutting a new 0.3 release. We'd like to make the transition as smooth as possible so we're testing various community repos that use pytorch with the release.

I built pyro from source and ran the tests with make test, replacing test with a number of the targets in the Makefile, and at least three are failing right now. Since make exits if any of the tests fail, more than two could be failing.

The outputs of these tests are listed in the gist below. These failures could be problems with pytorch / I also might not be running them correctly, so I'm looking into them as well.

https://gist.github.com/zou3519/e9cc880defd9741d0317c2ab9204661d

cc @soumith

[fritzo]

Thanks, @zou3519 ! I'm looking into this now by testing PyTorch's v0.3.0 branch + Pyro 0.1.1 tag.

[fritzo]

@zou3519 I'm still diagnosing a failure of test_subsample_gradient that you found. To reproduce:

pytest -v tests/infer/test_gradient.py::test_subsample_gradient

[fritzo]

It looks like the only issues with PyTorch v0.3.0 are slight changes to randomness that cause some of Pyro's flaky tests to fail.

BTW I noticed PyTorch v0.3.0 is about 2x faster than v0.2.0 on some tests. Great work!

[fritzo]

Tests still pass on PyTorch v0.3.0 f20fb2c + Pyro 0.1.2 daa7476.

[martinjankowiak]

@fritzo can we close this?

[fritzo]

Yes, we fixed all tests.