Listen to this story
TensorFlow today released TensorFlow 2.12.0!
Check out the GitHub repo to learn more about the update.
In the updated version, the tensorflow-gpu and tf-nightly-gpu packages have been effectively removed and replaced with packages that direct users to switch to tensorflow or tf-nightly. The name was the only difference between the two sets of packages ever since TensorFlow 2.1, so there is no loss of functionality or GPU support.
Subscribe to our Newsletter
Join our editors every weekday evening as they steer you through the most significant news of the day, introduce you to fresh perspectives, and provide unexpected moments of joy
Furthermore, the ‘tf.function’ now uses the Python inspect library directly for parsing the signature of the Python function. This factor can break certain previously ignored functions where signature is malformed including the use of ‘functools.wraps’ on a function with different signature and using ‘functools.partial’ with an invalid tf.function input. The updated function also enforces input parameter names to be valid Python identifiers and incompatible names are automatically sanitized similar to the existing SavedModel signature behavior. The function now has an experimental.Fingerprint that contains the fingerprint of the SavedModel. It also has an API for reading the fingerprint of a SavedModel.
The new version also includes TraceType which now requires an additional placeholder_value method to be defined. Instead of the value itself, ‘tf.function’ now traces with TraceTyle generated placeholder values.
Major improvements have been made to Keras as well. All saving related tasks have moved to keras.saving, a new namespace. The previous API locations (in keras.utils and keras.models) are still available but it’s recommended to update the code to point to the new API locations. The previous Keras version may have ignored the mark but now the users will get an error if they pass a mask with an incompatible shape.
For TensorFlow Lite, float LSTM will now support diagonal recurrent tensors and the users can add 16-bit float type support for built-in op fill. Also, transpose now supports 6D tensors.