[RELAY][FRONTEND] Initial OneFlow frontend support.

[hhhfccz]

Initial OneFlow frontend support, also add tutorials and tests.
tested models:

• alexnet
• mobilenetv2
• resnet50
• SRGAN
• yolov3

[hhhfccz]

@areusch@comaniac@Huyuwei

[BBuf]

Hello, TVM developers. We have recently done a work to add a new deep learning framework OneFlow to the TVM front-end, hope this work can be received.

@tqchen

[comaniac]

Thanks for the PR. For large scope new features, we recommend to file an RFC first to propose the motivation, approach and interfaces to the community for discussions. Please see the guideline here: https://github.com/apache/tvm-rfcs

[junrushao]

Yeah we find it super exciting and happy to discuss more with the RFC process

[masahi]

Have you reproduced that issue locally using docker?

Does oneflow depend on libibverbs?

[BBuf]

Have you reproduced that issue locally using docker?

Does oneflow depend on libibverbs?

It is not depend on libibverbs 。

[BBuf]

The tests related to the oneflow front-end in ci have all passed (indicating that oneflow can be imported normally), but it only hangs when the html corresponding to the document is compiled, and no specific reason is given. But I can run this from_oneflow.py normally in local. If it is convenient, can you help to confirm the source of this error? I compared the format of from_oneflow.py and from_onnx.py line by line, and found no problem. Thanks! @masahi

[masahi]

I think CI is not running any oneflow tests. You need to update https://github.com/apache/tvm/blob/main/tests/scripts/task_python_frontend.sh

It is not depend on libibverbs

Can you do more investigation on this. It may depend on it indirectly via other deps. Remember that we are using the oneflow CPU build.

[BBuf]

I think CI is not running any oneflow tests. You need to update https://github.com/apache/tvm/blob/main/tests/scripts/task_python_frontend.sh

It is not depend on libibverbs

Can you do more investigation on this. It may depend on it indirectly via other deps. Remember that we are using the oneflow CPU build.

Ok, thanks.

[areusch]

@BBuf is this ready for another look?

[BBuf]

@BBuf is this ready for another look?

Yes, but the problem here is that CI will have a segment fault bug when build html, but I can't reproduce it here, this is my pain point. So, can i get some help? thanks!

[areusch]

@BBuf can you try using our new ci.py tool (still kinda in dev, but supports docs)? it should reproduce locally. if not, then i guess we have a more annoying problem, probably specific to a GPU used in CI (also--do you have an nvidia GPU available locally?).

i think tests/scripts/ci.py docs should repro the CI locally, there are a few other options you might try there.

[BBuf]

if not, then i guess we have a more annoying problem, probably specific to a GPU used in CI (also--do you have an nvidia GPU available locally?).

Thank you very much. I run tests/scripts/ci.py docs locally, but I can't reproduce the build html error, it's normally.

but in ci, some strange bug appeared, link：

@areusch@masahi

[masahi]

Maybe some problem is happening during

generating gallery for how_to/compile_models... [100%] from_oneflow.py

How about removing from_oneflow.py for now and get the rest merged, and worry about the tutorial later?

[BBuf]

Maybe some problem is happening during

generating gallery for how_to/compile_models... [100%] from_oneflow.py

How about removing from_oneflow.py for now and get the rest merged, and worry about the tutorial later?

Ok, I'll have a try.

[masahi]

@BBuf@hhhfccz sorry it took so long, but this is finally merged!

[BBuf]

Thanks. I will rewrite the tutorial in my next PR.