Autogluon: AutoGluon issues on Mac OSX

[jwmueller]

Running AutoGluon on Mac OSX, you may see Segmentation fault: 11 errors whose stack trace points to libmxnet.so. If you get such errors, please use Linux instead. You can also try building MXNet from source with appropriate configuration options if you really want to use Mac OSX (which is not fully supported yet). The error occurs because the current pip installed MXNet does not work properly with AutoGluon on Mac. We are now working with the MXNet team to update the MXNet version to resolve this issue.

In particular, TabularPrediction tasks with hyperparameter_tune = True are known to produce this error, as are ImageClassification tasks. However, TabularPrediction tasks with hyperparameter_tune = False should work on Mac OSX.

Instructions for building MXNet from source on Mac: https://mxnet.apache.org/get_started/osx_setup

[larroy]

I will figure out what's our CD story in mac for mxnet.

[larroy]

Pleas reassign

[Innixma]

I have tested after latest HPO fixes and it appears that hyperparameter_tune=True is now stable on MacOS (Including for Neural Network). I have not built mxnet from source.

mxnet version: 1.5.1.post0

[tlienart]

FWIW, I'm on macOS, built MxNet 1.6 from source (with MKL but I think that's irrelevant), and used AG 6b247bfea9d504381cc512e36ba1909e6c54c0c3 I was able to run tabular-indepth without issue (with HPO). Maybe this issue can be closed?

[Innixma]

I'll add a clarifying statement to this issue: If you build mxnet from source, it works correctly. Issues only happen if mxnet is not built from source but instead pip installed. Even if you pip installed mxnet, it still works sometimes, but fails seemingly at random.

Also: Do not use Intel MKL in Tabular, it slows down NN by 2-3x

[zhreshold]

I got a data point recently that on mac the autogluon unittests will segfault 11 with

• the pip installed mxnet
• mxnet built from source with cmake

However, the unittests pass with

• mxnet built from source with Makefile