TensorFlow Serving using TensorFlow 2.0.0
Major Features and Improvements
- Some Tensorflow Text ops have been added to ModelServer (specifically constrained_sequence_op, sentence_breaking_ops, unicode_script_tokenizer, whitespace_tokenizer, wordpiece_tokenizer)
Breaking Changes
- As previously announced123, Contrib ops will not be packaged with Tensorflow, and therefore will not be available in Tensorflow Serving. If serving with Tensorflow Serving >1.15, please ensure your models do not contain any tf.contrib ops. If you are critically dependent on custom ops, please review this guide for instructions to statically build ops into the model server.
- After being deprecated for multiple years, as a part of tf.contrib deprecation, SessionBundle API will be removed starting from Tensorflow Serving 2.0 - if currently using SessionBundle, please migrate to SavedModel APIs.
Bug Fixes and Other Changes
- Add a section in the documentation for testing custom op manually. (commit: 1b65af1)
- Add ops delegate library to enable running TF ops. (commit: 1411235)
- Add command line tool to load TF Lite model for manual testing/debugging. (commit: 0b0254d)
- Fixes broken relative docs links (commit: 1281314)
- Cleaning up BUILD visibility for tf_pyclif_proto_library intermediate targets. (commit: 81ed5ef)
- Remove unused load statements from BUILD files (commit: d0e01a3)
- Manual tests for model server and including tf.Text in serving build. (commit: 142d0ad)
- Remove tensorflow/contrib/session_bundle as dependency for Tensorflow Serving. (commit: 1bdd349)
Thanks to our Contributors
This release contains contributions from many people at Google, as well as:
chaox