New features for ML development

TensorFlow 2.2.0 arrives with breaking changes and drops Python 2 support

Maika Möbus
© Shutterstock / Illus_man

TensorFlow 2.2.0 has been released, nearly four months after v2.1.0. The TensorFlow team has been keeping busy: In the latest version of the machine learning platform, they have added lots of new features and breaking changes, and have also fixed several bugs.

The open source machine learning library TensorFlow is now available in version 2.2.0. It adds several major features as well as bug fixes and breaking changes. Furthermore, Python 2 support is being dropped in this release as the language version reached its end of life in January 2020.

SEE ALSO: IBM open sources toolkit for AI development with Jupyter Notebook

Let’s dive right in and see what else is new.

Major features in TensorFlow 2.2.0

TensorFlow 2.2.0 has several major updates on board, and one of them is that you can now export C++ functions to Python via pybind11. This change has taken place in line with the deprecation efforts of SWIG, which had previously been used for this. Another new feature is a new Profiler for TensorFlow 2 for CPU/GPU/TPU, and the scalar type for string tensors  std::string has been replaced by the now ABI-stable tensorflow::tstring.

Among other areas, tf.keras has received a number of updates. For example, it is now possible to use custom training logic with by overriding Model.train_step, and the SavedModel format now comes with support for all Keras built-in layers, including metrics, preprocessing layers, and stateful RNN layers.

Breaking changes, bug fixes & more

tf.keras not only comes with new features, but also breaking changes such as updating the Huber loss function to be consistent with other Keras losses. Other breaking changes include the deprecation of XLA_CPU and XLA_GPU devices, and AutoGraph will no longer convert functions passed to tf.py_function, tf.py_func and tf.numpy_function.

SEE ALSO: “Julia is comparable to Python for simple machine learning tasks and better for complex ones”

Additionally, this release adds several bug fixes and other updates regarding, tf.lite, tf.keras, and more. And, as the TensorFlow team points out, TensorFlow 2.2.0 requires gast version 0.3.3.

See all the changes in the release notes.

Maika Möbus
Maika Möbus has been an editor for Software & Support Media since January 2019. She studied Sociology at Goethe University Frankfurt and Johannes Gutenberg University Mainz.

Inline Feedbacks
View all comments