Build Addons for Windows (CPU)

[seanpmorgan]

Fixes #173

• Compile custom ops on MSVC2017
• Fix packaging of whl
• Travis build

TODO (Future Issues):

• Travis Testing (Win/MacOS) Add nightly tests for windows/macos #787
• Parsetime needs to link a DLL with ABSL symbols exported (currently links with .pyd) Build ParseTime on windows  #782
• Windows build in documentation (waiting for 0.7 release)
• Build GPU custom ops for windows Build GPU custom Ops for windows #784

[seanpmorgan]

@tensorflow/sig-addons-maintainers ready for review.

Built with travis and pushed to pypi test:
https://travis-ci.org/seanpmorgan/addons/builds/626967246
https://test.pypi.org/project/tensorflow-addons/0.7.0.dev0/#files

No py27 package since theres no binary for TF2.1.0rc1 but it should be there for the full release.

Tested on windows VM:

pip install tensorflow==2.1.0rc1
pip install -i https://test.pypi.org/simple/ tensorflow-addons==0.7.0.dev0 --no-deps

>>> import tensorflow as tf
>>> import tensorflow_addons as tfa
>>>
>>> print(tf.__version__)
2.1.0-rc1
>>> print(tfa.__version__)
0.7.0-dev
>>> tf.debugging.set_log_device_placement(True)
>>>
>>> x = tf.constant([[0.5, 1.2, -0.3]])
2019-12-18 19:29:22.789977: W tensorflow/stream_executor/platform/default/dso_loader.cc:55] Could not load dynamic library 'nvcuda.dll'; dlerror: nvcuda.dll not found
>>> layer = tfa.layers.GeLU()
>>> print(layer(x))
2019-12-18 19:29:35.200914: I tensorflow/core/common_runtime/eager/execute.cc:573] Executing op __inference_gelu_7 in device /job:localhost/replica:0/task:0/device:CPU:0
x: (_Arg): /job:localhost/replica:0/task:0/device:CPU:0
2019-12-18 19:29:35.209845: I tensorflow/core/common_runtime/placer.cc:54] x: (_Arg): /job:localhost/replica:0/task:0/device:CPU:0
Addons>Gelu: (Addons>Gelu): /job:localhost/replica:0/task:0/device:CPU:0
2019-12-18 19:29:35.217653: I tensorflow/core/common_runtime/placer.cc:54] Addons>Gelu: (Addons>Gelu): /job:localhost/replica:0/task:0/device:CPU:0
Identity: (Identity): /job:localhost/replica:0/task:0/device:CPU:0
2019-12-18 19:29:35.224838: I tensorflow/core/common_runtime/placer.cc:54] Identity: (Identity): /job:localhost/replica:0/task:0/device:CPU:0
identity_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:CPU:0
2019-12-18 19:29:35.235247: I tensorflow/core/common_runtime/placer.cc:54] identity_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:CPU:0
tf.Tensor([[ 0.345714    1.0617028  -0.11462908]], shape=(1, 3), dtype=float32)

[WindQAQ]

Thank you for the effort on build stuff!

[WindQAQ]

Why do we get rid of the workspace name here?

[seanpmorgan]

Yeah, so I tried for a while to get this to work. Windows seemed unwilling to copy the files during build_pip_pkg if the directory was specified using workspace. Possibly related to windows sym-links but I'm not really sure. It's worth figuring out, but I've had about as much of windows build as I can stand for a while :P

The workspace name is something I added during CUDA10.1 PR and isn't being done for custom-op:
https://github.com/tensorflow/custom-op/blob/master/WORKSPACE

[WindQAQ]

cc @fsx950223

[WindQAQ]

Thanks! Let's get this merged and deal with other issues later :D