v2.2.5

• Allow use with tensorflow-cpu package.
• Dependency updates
• Minor fixes.

v2.2.4

• Added option to defer releasing temporary socket/port to user map_function for cases where user code may not bind to the assigned port soon enough to avoid other processes binding to the same port, e.g. extensive pre-processing before invoking TF APIs.
• Updated screwdriver.cd build template.
• Trigger documentation publish after PyPI push.

v2.2.3

• Added ability to disable spark barrier execution in TFParallel
• Updated with spark 3 + scala 2.12 dependencies
• Fixed documentation build

v2.2.2

• Migrated build from travis-ci to screwdriver.cd

v2.2.1

• Added support for port ranges in TFOS_SERVER_PORT environment variable.
• Updated mnist/keras/mnist_tf.py example with workaround for tensorflow datasets issue.
• Added more detailed error message for missing executor_id.
• Added unit tests for gpu allocation variants.

v2.2.0

• Added support for Spark 3.0 GPU resources
• Updated to support Spark 2.4.5
• Fixed dataset ordering in mnist_inference.py examples (thanks to @qsbao)
• Added optional environment variables to configure TF server/grpc ports and TensorBoard ports on executors
• Fixed bug with TFNode.start_cluster_server in backwards-compatibility code for TF1.x
• Fixed file conflict issue with compat.export_saved_model in TF2.1
• Removed support for Python 2.x

v2.1.3

• Detect TF version w/o importing to avoid runtime initialization before GPU allocation.

v2.1.2

• Use tf.config.list_physical_devices() to avoid TF runtime initialization.

v2.1.1

• added compat.is_gpu_available() method to use:
  • tf.config.list_logical_devices('GPU') (for TF2.1)
  • tf.test.is_cuda_available() (for earlier versions of TF).
• added ability to launch TensorBoard on chief:0 or master:0 nodes (for small clusters without worker nodes).

v2.1.0

• Added compat module to manage minor API changes in TensorFlow.
• Added compatibility for TF2.1.0rc0 (exporting saved_models and configuring auto-shard policy)
• Re-introduced compatibility for TF1.x (except support for InputMode.TENSORFLOW in the ML Pipeline API).
• Added TFParallel class for parallelized single-node inferencing via Spark executors.
• Updated examples for TF API changes.
• Updated to use module-level loggers.