Unbundle tensorflow from spark-tensorflow-distributor

spark/spark-tensorflow-distributor/README.md

@@ -12,6 +12,8 @@ pip install spark-tensorflow-distributor
 ```
 
 The installation does not install PySpark because for most users, PySpark is already installed.
+In addition, `tensorflow` not installed so that users may choose between regular and CPU-only
+installation via `pip install tensorflow` and `pip install tensorflow-cpu`.
 If you do not have PySpark installed, you can install it directly:
 
 ```bash

spark/spark-tensorflow-distributor/requirements.txt

@@ -6,13 +6,11 @@
 ## Tensorflow Requirements ##
 # Tensorflow is required to run this code but depends on specific configurations. Install from:
 # https://www.tensorflow.org/get_started/os_setup#overview
+# Either `tensorflow` or `tensorflow-cpu` should be installed via pypi depending on if you want GPU support
 
 # NOTE: If you want to use the GPU version, you will also need Nvidia's CUDA toolkit and cuDNN:
 # https://developer.nvidia.com/cuda-downloads
 # https://developer.nvidia.com/cudnn
-# Note that if you want to use the GPU version, you have to `pip uninstall tensorflow`
-# and `pip install tensorflow-gpu`, since both cannot coexist. 
-tensorflow>=2.1.0
 
 # We'll use pytest to run our tests; this isn't really necessary to run the code, but it is to run
 # the tests.  With this here, you can run the tests with `py.test` from the base directory.

spark/spark-tensorflow-distributor/setup.py

@@ -6,7 +6,7 @@
 
 setuptools.setup(
   name="spark_tensorflow_distributor",
-  version="0.0.3",
+  version="1.0.0",
   author="sarthfrey",
   author_email="[email protected]",
   description="This package helps users do distributed training with TensorFlow on their Spark clusters.",
@@ -24,8 +24,5 @@
     "Topic :: Software Development :: Libraries :: Python Modules",
     "Topic :: Software Development :: Version Control :: Git",
   ],
-  install_requires=[
-    "tensorflow>=2.1.0",
-  ],
   python_requires='>=3.6',
 )