The TensorFlow team gives an outlook on TensorFlow 2.0. The new version should be easier to use and offer new features. The transition to TensorFlow 2.0 will be supported with updates and tools.
TensorFlow 2.0 should be easier to use, according to an official blog post . This is to be realized through the integration of Keras as a central high-level API for building and training models in TensorFlow. Thanks to various features of Keras, the team behind TensorFlow wants to make using TensorFlow 2.0 even easier for newcomers.
Some of Keras’ lighter-featured features include included APIs for building models. The deployment libraries of TensorFlow for different platforms are also to be mentioned here, as they allow the use of trained models regardless of the system used and partly also of the programming language used. The team wants to facilitate the unproblematic exchange of models between the offshoots through the standardization of SavedModel . The areas of prototyping and debugging should also be simplified as a result of the implementation of Eager Execution. For the research TensorFlow 2.0 should also bring new features. Among them are several new enhancements such as Ragged Tensors, TensorFlow Probability or Tensor2Tensor for experiments, but also a combination of high and low level APIs is planned.
Easy transition to TensorFlow 2.0
During a 12-month transition period, additional security patches are being released for the latest 1.x version of TensorFlow. In addition, the transition is facilitated by a conversion tool. This tool, according to the announcement, updates Python code from TensorFlow 1.x to use TensorFlow 2.0 compliant APIs and marks codes for which this is not automatically possible.
The release of TensorFlow 2.0 as a public preview is scheduled for an early date in 2019. More information about the features mentioned here and other features of the planned new version can be found in the blog post of the TensorFlow team . All details about Keras can be found in the official documentation .