What’s new in TensorFlow 2.13 and Keras 2.13? — The TensorFlow Blog

Posted by the TensorFlow and Keras Teams

TensorFlow 2.13 and Keras 2.13 have been released! Highlights of this release include publishing Apple Silicon wheels, the new Keras V3 format being default for .keras extension files and many more!

TensorFlow Core

Apple Silicon wheels for TensorFlow

TensorFlow 2.13 is the first version to provide Apple Silicon wheels, which means when you install TensorFlow on an Apple Silicon Mac, you will be able to use the latest version of TensorFlow. The nightly builds for Apple Silicon wheels were released in March 2023 and this new support will enable more fine-grained testing, thanks to technical collaboration between Apple, MacStadium, and Google.

tf.lite

The Python TensorFlow Lite Interpreter bindings now have an option to use experimental_disable_delegate_clustering flag to turn-off delegate clustering during delegate graph partitioning phase. You can set this flag in TensorFlow Lite interpreter Python API

interpreter = new Interpreter(file_of_a_tensorflowlite_model, experimental_preserve_all_tensors=False)

The flag is set to False by default. This is an advanced feature in experimental that is designed for people who insert explicit control dependencies via with tf.control_dependencies() or need to change graph execution order.

Besides, there are several operator improvements in TensorFlow Lite in 2.13

• add operation now supports broadcasting up to 6 dimensions. This will remove explicit broadcast ops from many models. The new implementation is also much faster than the current one which calculates the entire index for both inputs the the input instead of only calculating the part that changes.
• Improve the coverage for 16×8 quantization by enabling int16x8 ops for exp, mirror_pad, space_to_batch_nd, batch_to_space_nd
• Increase the coverage of integer data types
• enabled int16 for less, greater_than, equal, bitcast, bitwise_xor, right_shift, top_k, mul, and int16 indices for gather and gather_nd
• enabled int8 for floor_div and floor_mod, bitwise_xor, bitwise_xor
• enabled 32-bit int for bitcast, bitwise_xor, right_shift

tf.data

We have improved usability and added functionality for tf.data APIs.

tf.data.Dataset.zip now supports Python-style zipping. Previously users were required to provide an extra set of parentheses when zipping datasets as in Dataset.zip((a, b, c)). With this change, users can specify the datasets to be zipped simply as Dataset.zip(a, b, c) making it more intuitive.

Additionally, tf.data.Dataset.shuffle now supports full shuffling. To specify that data should be fully shuffled, use dataset = dataset.shuffle(dataset.cardinality()). This will load the full dataset into memory so that it can be shuffled, so make sure to only use this with datasets of filenames or other small datasets.

We have also added a new tf.data.experimental.pad_to_cardinality transformation which pads a dataset with zero elements up to a specified cardinality. This is useful for avoiding partial batches while not dropping any data.

Example usage:

ds = tf.data.Dataset.from_tensor_slices({‘a’: [1, 2]})
ds = ds.apply(tf.data.experimental.pad_to_cardinality(3))
list(ds.as_numpy_iterator())
[{‘a’: 1, ‘valid’: True}, {‘a’: 2, ‘valid’: True}, {‘a’: 0, ‘valid’: False}]

This can be useful, e.g. during eval, when partial batches are undesirable but it is also important not to drop any data.

oneDNN BF16 Math Mode on CPU

oneDNN supports BF16 math mode where full FP32 tensors are implicitly down-converted to BF16 during computations for faster execution time. TensorFlow CPU users can enable this by setting the environment variable TF_SET_ONEDNN_FPMATH_MODE to BF16. This mode may negatively impact model accuracy. To go back to full FP32 math mode, unset the variable.

Keras

Keras Saving format

The new Keras V3 saving format, released in TF 2.12, is now the default for all files with the .keras extension.

You can start using it now by calling model.save(“your_model.keras”).

It provides richer Python-side model saving and reloading with numerous advantages:

A lightweight, faster format:

Human-readable: The new format is name-based, with a more detailed serialization format that makes debugging much easier. What you load is exactly what you saved, from Python’s perspective.

Safer: Unlike SavedModel, there is no reliance on loading via bytecode or pickling – a big advancement for secure ML, as pickle files can be exploited to cause arbitrary code execution at loading time.

More general: Support for non-numerical states, such as vocabularies and lookup tables, is included in the new format.

Extensible: You can add support for saving and loading exotic state elements in custom layers using save_assets(), such as a FIFOQueue – or anything else you want. You have full control of disk I/O for custom assets.

The legacy formats (h5 and Keras SavedModel) will stay supported in perpetuity. However, we recommend that you consider adopting the new Keras v3 format for saving/reloading in Python runtimes, and using model.export() for inference in all other runtimes (such as TF Serving).