2 min

Tags in this article

, , , ,

The second version of Google’s popular Artificial Intelligence (AI) framework, TensorFlow, has been made widely available. TensorFlow 2.0 comes with close integration with Keras.

The new version of the framework has received several updates and changes, with the aim of increasing the ease of use, writes Venturebeat. For example, a number of redundant APIs have been removed, which is why there is a close integration and reliance on tf.keras as the central high-lel API.

The integration with Keras was already announced in March this year, when the new iteration of TensorFlow was presented. With Keras, engineers can access functions that were previously scattered across multiple APIs in one place. It also offers more options for customizing the development workflow.

The first integration with Keras, a deep learning library, already appeared at TensorFlow 1.0 in February 2017.

Eager Execution

TensorFlow 2.0 also gets a function called eager execution, which is enabled by default. This means that AI models are activated much faster than previous versions. Engineers can therefore try out different model variations with less delay between tests.

TensorFlow 2.0 is also said to deliver three times faster training performance when using mixed precision on Nvidia’s Volta and Turing GPUs.

Modules

When Google announced TensorFlow 2.0 in March this year, it also rolled out a number of associated tools. One of them is TensorFlow Privacy, which allows machine learning models to remove potentially sensitive data that they do not need to process. For this purpose, the module automatically filters input that is different from information that the algorithm normally receives.

TensorFlow Federated focuses on the growing number of mobile services that depend on AI to support core functions. The module allows apps to analyze user data directly on the user’s device.

Developers can, therefore, collect and use the resulting insights to improve the AI algorithms, without needing access to the underlying data. So, the privacy of users can be improved in this way.