Google has announced the availability of TensorFlow release 1.4 which includes a number of new features and enhancements. One of the key enhancements included in this latest release is the graduation of Keras from a contribution package (tf.contrib.keras) to a core package (tf.keras).
Keras is a deep learning library written in Python that features two high-level neural networks APIs: Keras Sequential model API and Keras Functional model API. The sequential model is a linear stack of layers, and the functional model is used to define complex models such as directed acyclic graphs and multi-output models. Keras is now part of the TensorFlow core, and can be used with the Estimator API and other core TensorFlow functionality.
The Dataset API has graduated from a contribution package (tf.contrib.data) to a core package (tf.data) with this release. The Dataset API also now includes Python generators support. Google recommends that developers use this latest version of the API to build input pipelines for TensorFlow models as it offers more functionality than the older API versions. It should be noted that only the APIs for Python are covered by TensorFlow’s API stability promises, a policy for developers who need backwards compatibility for code or data. The C++, Go, and Java APIs are not covered by the stability promises policy.
For developers interested in experimenting with computer vision, Google recently released the Tensorflow Object Detection API. The API provides access to an open source framework, built on Tensorflow, for constructing, training, and deploying object detection models.
For more information about the 1.4 release of TensorFlow, visit TensorFlow.org.