TensorFlow 2.0 is almost here. Late last year, the TensorFlow team announced high level concepts that would drive TensorFlow 2.0. This week, the team announced more details, including an anticipated preview release of early 2019.
TensorFlow 2.0 will use Keras, a machine learning API standard, as the central API used to build and train models. Kera includes multiple model-building APIs (Sequential, Functional, and Subclassing). Although TensorFlow 2.0 in not yet available, developers can already start using Keras to build, train, and validate models. This will help prepare developers for the release of TensorFlow 2.0.
TensorFlow 2.0 will also improve support across platforms and components. To achieve this, 2.0 will include standardized exchange formats and API alignment. Once a model is trained, it can be directly executed in an application or served through a deployment library (TensorFlow Serving, TensorFlow Lite, or TensorFlow.js).
2.0 has a particular focus on moving ideas from concept to code and model to publication with new experimentation features. The Keras Functional API and the Model Subclassing API allow complex topology creation. Further, new extensions (Ragged Tensors, TensorFlow Probability, and Tensor2Tensor) enable new options for experimentation.
As is expected, developers will face some changes from 1.x to 2.0. Major changes include removal of queue runners (in favor of tf.data), removal of graph collections, variable treatment changes, renaming and moving of API symbols. Stay tuned to the TensorFlow blog for more announcements. 2.0 will be released in Preview first.