[Docker] update onnx to 1.6 and torch to 1.4

[masahi]

v1.4 was out last week and we are still at 1.2.

[tqchen]

Right now the CI update was blocked by #4764, but we could build a torch update separately, depending on how quickly we can resolve the blocker.

[masahi]

There is no hurry, we can wait until mxnet issues are fixed.

[leandron]

Hi @masahi, I think it is also relevant to see whether these change I sent a few weeks ago (#4632), can be reverted by updating torch to 1.4.

[masahi]

@leandron yes, torchvision 0.5 seems to be compatible with PIL 7.0 (fixed in #1501). My installation was already 7.0.

In [5]: import PIL                                                                                                                                

In [6]: import torchvision                                                                                                                        

In [7]: PIL.__version__                                                                                                                           
Out[7]: '7.0.0'

In [8]: torchvision.__version__                                                                                                                   
Out[8]: '0.5.0'

does this answer your question?

[leandron]

Yes @masahi. Then, if you want, you can remove those <7 in Pillow installations, in the context of this PR, so it is a self-contained torch update? I'm happy to review it for you.

[masahi]

ok reverted

[masahi]

hmm no idea why CI is failing. @leandron do you know what is going with sphinx issues?

[leandron]

Sorry @masahi , I don't know how that error is related to the Pillow dependency and couldn't find any related issue on sphinx.

[masahi]

ok CI is fixed.

@tqchen since the update of MXNet seems to be having issues, and we want to test the PyTorch frontend on the latest version, we want the PyTorch update to be built first.

[masahi]

@jwfromm have you tried pytorch 1.4 + onnx 1.6? So far I have no problem with this combo. I want to update onnx CI as well if possible.

[jwfromm]

I don't think I've tried that version combo. If it's working for you then I'd say let's update!

[masahi]

@jwfromm I posted what I think is the solution to "onnx segfault problem" at onnx/onnx#2394 (comment). I encourage you to try it with pytorch 1.3 or 1.4.

[jwfromm]

You're right that importing onnx first resolves the issue. I tried it out with a few different pytorch and onnx version combinations and it runs fine with them all. Your explanation of the issue makes perfect sense as well. Funny that it was something so simple all along.

[masahi]

@tqchen we are ready to upgrade!

[tqchen]

OK, will report back in one to two days

[tqchen]

updates

• [CI][DOCKER] Update ci-gpu to v0.60 #4827
• [CI][DOCKER] Update ci-gpu torch1.4 and onnx1.6 #4826

[masahi]

@tqchen can we send torch 1.4 and onnx 1.6 dependent changes now that #4827 and #4826 were merged? It seems only ci-gpu was updated, I wonder what happen if I run my test on ci-cpu.

Also it seems this PR can be closed.

[tqchen]

Only ci-cpu is updated atm. Let me know if we also need to update ci-cpu

[masahi]

does ci-cpu run on CI to test PRs? If so I cannot send torch 1.4 and onnx 1.6 dependent changes.

[tqchen]

I believe it runs unit-tests but not integration(frontend tests)