TensorFlow 2.9 Released
TensorFlow version 2.9 has been released, bringing a number of improvements and changes, including new models and optimizers for Keras, Intel processor optimizations, experimental model parallelism utility, and more.
Created on May 18|Last edited on May 19
Comment
Almost exactly a month ago, the first release candidate version of TensorFlow 2.9 was released. Today we've been given the announcement that TensorFlow version 2.9 is complete and fully released. This update comes with a number of improvements to look forward to.
What's changed in TensorFlow 2.9?
Be sure to check out the changelog on GitHub for the full detailed list of changes in TensorFlow 2.9: https://github.com/tensorflow/tensorflow/releases/tag/v2.9.0
New Features
The Keras module is seeing a lot of new things getting added to it with this update. Various sizes of ResNet model architectures have been added under tf.keras.applications.resnet_rs. A reworked optimizer is in the works, offering more control over your optimizers, available at tf.keras.optimizers.experimental.Optimizer. Though, it will be dropping the experimental in the future as it's planned to fully replace the current optimizer code, sending it to legacy.
CPU users get to see a performance boost with oneDNN, a package of optimizations for Intel processors. The feature has been in TensorFlow since version 2.5 as experimental, but now it's being enabled by default. oneDNN grants performance improvements, however some rounding errors may appear, so if something is going wrong with your models, you still have the option to disable it.
An experimental extension to TensorFlow for handling model parallelism has been added, called DTensor. It's still very much experimental, so don't be upset if things go
Additional changes include new operations support for various datatypes in TFLite, an experimental extension for model parallelism called DTensor, eager mode now running ops as a tf.function, and more.
Removals
There's been a few code removals in this version you should be aware of in case you're still using them.
BoostedTrees was deprecated in version 2.8 and is being completely removed in version 2.9 - users should switch to TensorFlow Decision Forests instead.
The tf.keras.mixed_precision.experimental API has been removed, though the non-experimental tf.keras.mixed_precision API has been available since version 2.4 - if you're still using the experimental API for some reason, it's time to switch over.
Additional Changes
Numerous bug fixes and security improvements are also seen in the update. For a full list of changes, take a loot at the changelog on GitHub here: https://github.com/tensorflow/tensorflow/releases/tag/v2.9.0
Find out more
Add a comment
Tags: ML News
Iterate on AI agents and models faster. Try Weights & Biases today.