Tensorflow pluggable device

[Apurer]

This is support for pluggable device using experimental TensorFlow feature.
Motivated by the desire to use the M1 Chip GPU cores.
Thanks to gists of @lnshi, @wangjia184 and my own trial and error.

[Apurer]

Resolves issue 381.

[Apurer]

In order for the experimental feature to work properly on an Apple M1 chip, it is necessary to satisfy these dependencies:

1. libmetal_plugin.dylib library in lib path:
   • Setup TensorFlow metal
   • Search for location of libmetal_plugin.dylib by running: pip show tensorflow-metal
   • Copy displayed path after Location:
     • It could be /miniconda/lib/python3.9/site-packages
   • Add /tensorflow-plugins to the returned path and set env variable
     • Example export DYLD_LIBRARY_PATH=~/miniconda/lib/python3.9/site-packages/tensorflow-plugins:$DYLD_LIBRARY_PATH
2. libtensorflow in lib path:
   • Install libtensorflow using brew brew install libtensorflow
   • Check path of installed package brew --cellar libtensorflow
     • Example path /opt/homebrew/Cellar/libtensorflow
   • Add libtensorflow to the lib path
     • Example export DYLD_LIBRARY_PATH=/opt/homebrew/Cellar/libtensorflow/2.10.0/lib:$DYLD_LIBRARY_PATH
   • Create soft symlink for file libtensorflow.so.{version}
     • Example ln -s /opt/homebrew/Cellar/libtensorflow/2.10.0/lib/libtensorflow.so.2.10.0 /opt/homebrew/Cellar/libtensorflow/2.10.0/lib/libtensorflow.dylib

[adamcrume]

Please ensure that ./test-all succeeds.

[Apurer]

Can you rerun the checks @adamcrume? I don't remember what was wrong :)

[Apurer]

Closes #402

[Apurer]

I tested code by running workflow in my repository.

[Apurer]

I believe I resolved all the issues. Can you please trigger the workflow run, @adamcrume?

[Apurer]

I am happy to see all the checks pass :)

[adamcrume]

There was another pull request for this (#402), but it didn't have the macOS test. Could you rebase? It would still be nice to use this PR for the test.

[adamcrume]

This doesn't render properly (https://github.com/Apurer/rust/tree/feature/experimental/tensorflow-sys). It also needs some explanation. Do users need to follow that link and do anything specific?

[Apurer]

I'll fix the rendering issue. Yes, users need to click on the provided link to set up the TensorFlow plugin on Apple computers. I considered including the steps listed on developer.apple.com, but I was unsure if the steps would be updated.

[Apurer]

There was another pull request for this (#402), but it didn't have the macOS test. Could you rebase? It would still be nice to use this PR for the test.

I will rebase my pull request and possibly refactor.