TensorFlow increases the Keras user-friendliness through a new handling of attention layers and enters into a cooperation with Intel, AWS and others.
The TensorFlow machine learning framework developed by Google has reached version 2.10. The TensorFlow team is bringing various new features to the integrated deep learning library Keras to increase usability and has been collaborating with various companies to provide official TensorFlow builds.
In addition, the release raises the library TensorFlow Decision Forests (TF-DF) to version 1.0. This milestone release aims to highlight the stability and maturity of the library, which consists of a collection of algorithms for training, deploying and interpreting decision forest models and can be used on Linux. The TensorFlow team is still working on the use of TF-DF under macOS and Windows.
Increased usability of Keras
TensorFlow 2.10 is intended to make it easier to use the Deep Learning Library Keras when it comes to developing transformers and initializers. For example, the release expands and unifies masking support for Keras Attention Layer. These include tf.keras.layers.Attention, tf.keras.layers.AdditiveAttention and tf.keras.layers.MultiHeadAttention. Causal Attention and Implicit Masking should now work together to simplify the implementation of a Transformer-style model, since masking is often a difficult step. As part of the new Causal Attention, all three layers mentioned can use_causal_mask-argument for call use. These layers can now handle implicit masking as well.
Other Keras innovations include its initializers, which become stateless and deterministic with this release, as well as a new utility tf.keras.utils.audio_dataset_from_directory for creating audio classification records from directories of .wav files.
A TensorFlow blog post describes the highlights in detail and the GitHub repository lists all changes in version 2.10.
Cooperation with Intel, AWS & Co.
Starting with the current release, the TensorFlow team is collaborating with Intel, AWS, ARM, and software engineering organization Linaro to create official TensorFlow builds. That means in practice who means pip TensorFlow installed on Windows Native and Linux AArch24 hosts will receive an official TensorFlow build. This is characterized by the fact that experts from Google and the collaborating platforms have checked it with regard to functionality and performance standards. In the future, the TensorFlow team expects to work with other cooperation partners.
For the majority of users, installing and uninstalling TensorFlow should be done using pip install respectively pip uninstall work as usual. However, under certain circumstances, an additional pip uninstall-Step may be needed on Windows Native and Linux AArch64 systems.
TensorFlow Lite generally available on Google Play Services
Another announcement concerns TensorFlow Lite, version 1.0 of an inference runtime that was released in 2019 and is optimized for mobile devices. Now the Google Play Services API for TensorFlow Lite is generally available on Android devices, aiming to improve machine learning capabilities in mobile apps. For example, the size of an application should now be reduced by up to 5 MB compared to the static bundling of TensorFlow Lite with an app, and the applications receive regular automatic performance updates in the background.
