upgrade to pytorch 1.2 #3182
upgrade to pytorch 1.2 #3182
Conversation
|
For the naqanet test, first things to try are checking for dropout in the test fixture, and seeing if the test is just flaky. Other than that, not so many good ideas. |
|
Hmm i'll revert that docstring test. It looks broken and takes 60 seconds... |
…e that will make it pass? :(
|
ok, this is ready for review, I guess. the one part I'm not happy with (and that you won't be happy with) is that I couldn't figure out a way to get the biaffine_dependency_parser_multilang_test to not trigger the zero gradients, so I had to add a /~https://github.com/allenai/allennlp/pull/3182/files#diff-83e1bde8a01afa0c96f68ca171f3ba8bR11 if anyone has a better way to deal with that, I'm open to it. |
| @@ -101,7 +101,7 @@ | |||
| packages=find_packages(exclude=["*.tests", "*.tests.*", | |||
| "tests.*", "tests"]), | |||
| install_requires=[ | |||
| 'torch>=0.4.1,<1.2', | |||
| 'torch>=1.2.0', | |||
There was a problem hiding this comment.
Do we strictly need features from 1.2.0 or is 1.0 sufficient?
There was a problem hiding this comment.
this PR is in response to breaking changes that happened between 1.1 and 1.2, but also going forward we'd like to use PyTorch's dataloader infrastructure, and for that to support allennlp datasets we need torch>=1.2
There was a problem hiding this comment.
Got it. Thanks for explaining.
|
@TalSchuster, I believe you wrote the biaffine_dependency_parser_multilang code. Strangely one of the tests for that code is producing zero gradients when we upgrade to PyTorch 1.2. Any insight on why that might be? We're tempted to ignore, but it's a bit unsettling. |
* first attempt at pytorch 1.2 * explicit is better than implicit * more explicit * attempt to fix flaky tests * pylint * no disable dropout * disable dropout by default * restore dropout, don't deepcopy * change batch size for biaffine_dependency_parser_multilang_test, maybe that will make it pass? :( * try batch size 10 * ignore bad gradient parameter * cleanup
as you can see, this ended up not being that much work.
there are still two tests failing for me.
the first is the new "docstring" test, which I don't entirely understand:
the second is
naqanet_test:I'll keep digging into these, but if anyone has any ideas, let me know