Enable ONNX test in circle CI

[jiafatom]

The PR disable onnx test since it is failing since 11/14/2020. I did a test locally and find that onnx need update to 1.8.0. So This PR aims to enable onnx test by pulling latest onnx==1.8.0.
The root cause is that onnx_shape_inference is enabled on 11/13/2020, but the shape inference PR (committed 9/30/2020) does not handle invalid dynamic axes input well. In torch vision CI, there are some invalid axes, so we fixed it in this PR, and enable onnx test. In a separate pytorch PR, we add exception handling for invalid axes input, then it will give meaningful failure msg rather than seg fault here.

[jiafatom]

@fmassa I try to enable onnx test, it works in my local machine, but fails in the Ci build here, it shows
free(): invalid next size (fast)
which is a memory issue. Do you have any ideas on why this happen? thanks!

[fmassa]

Hi @jiafatom

Thanks for the PR!

Do you have any ideas on why this happen?

I don't know what could be the issue. This looks like something inside ORT might be problematic?

[codecov]

Codecov Report

Merging #3144 (8a1b8b9) into master (25e2085) will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master    #3144   +/-   ##
=======================================
  Coverage   72.75%   72.75%           
=======================================
  Files          99       99           
  Lines        8979     8979           
  Branches     1431     1431           
=======================================
  Hits         6533     6533           
  Misses       1999     1999           
  Partials      447      447           

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 25e2085...8a1b8b9. Read the comment docs.

[jiafatom]

@fmassa Just fixed the issue, and test_onnx passes in CircleCI, so we are enabling test_onnx now. Could you please sign off? Thank you!

[datumbox]

Awesome work and investigation @jiafatom. Thanks for the PR.

[datumbox]

@jiafatom I think the test broke again, possibly due to upstream changes. Could you have a look on your side and confirm?

[jiafatom]

@datumbox
Could you please show me the link of test broken? I am checking https://travis-ci.org/github/pytorch/vision/builds
The last time the master branch runs is 12/8/2020, that is 3 days ago, before this commits checked in.

The latest job failure is https://travis-ci.org/github/pytorch/vision/jobs/748808277
but this is a v0.8.2 branch, not related to this PR.

v0.8.2 did pip install -q --user -i https://test.pypi.org/simple/ ort-nightly==1.4.0.dev202008122
I cannot find this ort-nightly in pytorch/vision master code.
This ort-nightly cannot be fetched, the possible explanation is that onnxruntime 1.6.0 released on 12/10/2020 (yesterday),
so some old ort-nightly is removed.

Try to use the released version of onnxruntime.
If you want to use nightly build, try pip install --user -i https://test.pypi.org/simple/ ort-nightly==1.5.2.dev202012031 instead.
Please let me know if you have other questions. Thanks.

[fmassa]

@jiafatom check #3123 for example, where the only failing job is an ONNX test job.

[jiafatom]

@datumbox @fmassa
There is a PR merged in pytorch at 12:30 pm 12/10/2020 (yesterday). That PR changes pytorch-onnx exporter API for a special case when the last input is a dictionary. So we need fix our unit test here.
The fix is here, please sign off, thank you!

cc @spandantiwari

[spandantiwari]

Hi @fmassa - thank you for your help with this. @jiafatom is working with our team. This latest PR #3160 should fix the tests. It was broken because of a commit that went into PyTorch recently (just yesterday). It seems like there was a bit of difference in test running in PyTorch CI which is why this wasn't immediately detected for the PyTorch commit. We are going to update the version of tests running in PyTorch CI so that it matches vision tests with priority.
Thanks again for your help.

[datumbox]

@spandantiwari @jiafatom I confirm that the PR #3160 fixed the issue. Thanks a lot for your speedy response and support. We just merged the fix into master. Have great weekend.

[fmassa]

Thanks @jiafatom and @spandantiwari for the help!