I'm just trying to make the support for these functions faster. Feel free to close this if it is more convenient to implement all ONNX opset version bump together in a later time rather than in separate PRs like this.

These seem pretty low risk, even if we don't bump the opset version.

@bddppq It should be easy to add in tree tests for these now, is that right?

@ezyang yes in test/onnx/test_operators.py

But note caffe2 currently does not suppory asin, acos, atan and tan yet. Should be easy to add though.

@ezyang these are new ops added in the newest opset, so the pytorch export opset has to be bumped

I think if we add test, we have to bump up the op set version. Otherwise the test will not pass. If needed, I can bump up the version and take care of the other ops changed as specified in ONNX/Changelog.md.

But, is it safe to do this now, since ONNX opset version 7 is not stable yet? (In the future, if some changes happens on ONNX, then pytorch exporter might fail automatically (or maybe even worse, silently), and new changes are required to make unit tests pass. This might introduce additional maintenance burden.)

Alright, sounds good, let's wait. (BTW, this is one of the reasons why I originally advocated for having all opset versions in ONNX be stable.)

@zasdfgbnm is this good to merge now?

@zou3519
I don't think it it good to merge now, because we still have in _onnx_opset_version = 6 in https://github.com/pytorch/pytorch/blob/master/torch/onnx/symbolic.py#L93

But I do think it is time to bump up this version to 7, because ONNX has just released 1.2 with opset version 7. The opset version 7 should be considered stable now, since by @linkerzhang:

Operator Changes
Latest operator set version is 8 for ONNX domain and 1 for ONNX_ML domain.

@bddppq Are we ready for opset version bump?

How long is the expected time that it will be bumped to version 7?

Opset version has been bumped to 7 https://github.com/pytorch/pytorch/blob/master/torch/onnx/symbolic.py#L146

@pytorchbot test this please

@pytorchbot retest this please

@zasdfgbnm could you rebase to master and add test cases in tests/onnx/test_pytorch_onnx_caffe2.py? (test_operators only check the export successfully, it's better to also test the real computation).

Need to update the expect files: https://ci.pytorch.org/jenkins/job/caffe2-builds/job/onnx-py2-gcc5-ubuntu16.04-test/2003/console

@bddppq All done. Should be working now. I have some trouble pip install caffe2 and python setup_caffe2.py install, so I can not test it on my computer. And the CI system seems to have some trouble right now...

17:35:10 >       test_func('acos')
...
17:35:10 E               AssertionError: 
17:35:10 E               Not equal to tolerance rtol=0.001, atol=1e-07
17:35:10 E               
17:35:10 E               (mismatch 100.0%)
17:35:10 E                x: array([nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
17:35:10 E                      nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
17:35:10 E                      nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,...
17:35:10 E                y: array([nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
17:35:10 E                      nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
17:35:10 E                      nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,...

Need to choose better test input data :-)

@pytorchbot retest this please

@bddppq all fixed