TensorFlow Releases TensorFlow v2.9 With New Features

This Article Is Based On The Research Article 'What's new in TensorFlow 2.9?'. All Credit For This Research Goes To The Researchers of This Project 👏👏👏

Please Don't Forget To Join Our ML Subreddit

TensorFlow has announced the release of version 2.9 just three months after the release of version 2.8. OneDNN, a novel model distribution API, and DTensor, an API for smooth data and model parallelism migration, are the key highlights of this release.

OneDNN

The oneDNN performance package was added to TensorFlow to improve Intel CPUs’ performance. The experimental support for oneDNN in TensorFlow has been available since version 2.5, delivering a four-fold increase in speed. Linux x86 packages and CPUs with neural-network-focused hardware capabilities like AVX512 VNNI, AVX512 BF16, AMX, and others found on Intel Cascade Lake and newer CPUs, oneDNN optimizations will be turned on by default.

Dtensor

Dtensor is a new API for disseminating models and is one of the most notable features of this edition. DTensorflow allows shifting from data parallelism to single program multiple data (SPMD) based model parallelism, including spatial partitioning. Model inputs that are too massive for a single device can now be trained using new tools available to developers. A model code can be utilized on CPU, GPU, or TPU, regardless of the device, because it is a device-agnostic API. This job likewise gets rid of the coordinator and instead uses the task’s local devices to control them all. Model scaling can be accomplished without affecting startup time.

Model code was previously built around replicas; however, with DTensor, the model code is written from a global perspective, and per replica, code is given by the DTensor runtime. Additionally, the Windows Subsystem for Linux will allow developers to run a Linux environment on Windows without the requirement for a virtual machine or dual-boot setup.

TensorFlow 2.8 added an API that allowed for deterministic operations. Using the same inputs and the same outputs repeatedly are called determinism. The determinism performance has been increased in version 2.9. The Keras Optimizer API has been updated to include an experimental version of the TensorFlow Optimizer API, which provides a catalog of both built-in and custom optimizers.

Source: https://blog.tensorflow.org/2022/05/whats-new-in-tensorflow-29.html