TensorFlow 2.19: API Changes and Casting Improvements

TensorFlow 2.19 brings targeted updates to the popular machine learning framework, with a focus on API refinements and expanded data type support. While not a major overhaul, this release introduces changes that developers should be aware of, particularly those working with TensorFlow Lite and C++.

C++ API Changes in TensorFlow Lite Runtime

A key change in this release affects the C++ API for TensorFlow Lite Runtime (LiteRT). Two public constants, previously defined as compile-time constants, are now const references:

• tflite::Interpreter::kTensorsReservedCapacity
• tflite::Interpreter::kTensorsCapacityHeadroom

This shift aims to improve API compatibility for TensorFlow Lite in Play services while maintaining flexibility for future updates. Developers relying on these constants should review their code to ensure it accommodates this change.

Expanded bfloat16 Support in TensorFlow Lite

TensorFlow 2.19 extends bfloat16 support to the tfl.Cast operation in the runtime kernel. This addition enhances TensorFlow Lite's ability to work with this increasingly important data type, which offers a balance between the precision of float32 and the efficiency of float16.

Deprecation and API Migration

The tf.lite.Interpreter API is now deprecated and will be removed entirely in TensorFlow 2.20. Users are directed to migrate to ai_edge_litert.interpreter. This change is part of ongoing efforts to streamline and modernize the TensorFlow API. Developers should consult the migration guide to update their code accordingly.

Package Distribution Changes

TensorFlow 2.19 discontinues the release of standalone libtensorflow packages. However, users can still access these libraries by unpacking them from the PyPI package. This change may affect some deployment workflows, particularly for those who relied on separate libtensorflow downloads.

Keras Updates Move to keras.io

It's worth noting that release updates for the new multi-backend Keras (starting with Keras 3.0) will now be published on keras.io rather than bundled with TensorFlow release notes. This reflects the ongoing separation and modularization of Keras as a standalone deep learning framework.

What This Means for Developers

While TensorFlow 2.19 doesn't introduce sweeping changes, it does require attention from developers in specific areas:

1. C++ developers using TensorFlow Lite should review their use of the modified constants.
2. Projects leveraging bfloat16 in TensorFlow Lite can now take advantage of expanded casting support.
3. Any code using tf.lite.Interpreter needs to be updated before the next release.
4. Build processes that depend on standalone libtensorflow packages may need adjustment.

These updates reflect TensorFlow's ongoing evolution, balancing API stability with the need to refine and improve the framework. While the changes are targeted, they underscore the importance of staying current with TensorFlow's development and planning for future migrations.