BLD: Package GPU Ops (#456)

* Run GPU docker image
* Run without nvidia runtime
* Update docs

Author: seanpmorgan

.travis.yml

@@ -28,7 +28,8 @@ jobs:
     - stage: build
       name: "Build on Ubuntu 16.04 for Python 2.7 3.5 3.6 3.7"
       script:
-      - bash -x -e tools/run_docker.sh -c 'tools/ci_build/builds/release_linux.sh'
+      # Build GPU kernels without runtime=nvidia as there are no GPUs/Drivers
+      - docker run -v ${PWD}:/addons -w /addons tensorflow/tensorflow:custom-op-gpu-ubuntu16 tools/ci_build/builds/release_linux.sh
       after_success:
         - twine upload -u $PYPI_USER -p $PYPI_PW wheelhouse/*.whl
 

README.md

@@ -73,6 +73,9 @@ https://bazel.build/) build system.
 git clone https://github.com/tensorflow/addons.git
 cd addons
 
+# If building GPU Ops (Requires CUDA 10.0 and CuDNN 7)
+export TF_NEED_CUDA=1
+
 # This script links project with TensorFlow dependency
 ./configure.sh
 
@@ -93,6 +96,11 @@ User experience and project maintainability are core concepts in
 TF-Addons. In order to achieve these we require that our additions
 conform to established API patterns seen in core TensorFlow.
 
+#### GPU/CPU Custom-Ops
+A major benefit of TensorFlow Addons is that there are precompiled ops. Should 
+a CUDA 10 installation not be found then the op will automatically fall back to 
+a CPU implementation.
+
 #### Proxy Maintainership
 Addons has been designed to compartmentalize subpackages and submodules so 
 that they can be maintained by users who have expertise and a vested interest