TensorFlow Lite Update: Version 2.16.1 now supported
We are happy to announce that we have upgraded to Tensorflow Lite v2.16.1. We encourage users with existing projects to deploy and test their models using this latest version.
This update brings several enhancements and bug fixes, particularly aimed at improving performance on mobile and embedded devices.
While we don’t anticipate any issues as the update has been tested extensively, the significant version jump may present some edge cases.
-
Performance Enhancements:
- Improved Quantization: Enhancements in quantization techniques, including better support for per-channel quantization and dynamic range quantization, which helps in reducing model sizes while maintaining accuracy.
- Optimized Delegates: Expanded and optimized support for delegates, including GPU, NNAPI, and XNNPACK, allowing more efficient model execution on various hardware platforms.
-
Interpreter and Converter Updates:
- StableHLO Operations: Added support for various stableHLO operations like gather, add, multiply, and others, making the interpreter more powerful and flexible.
- C API Enhancements: Several new functions were added to the C API, improving control over tensor allocations and buffer handling, as well as better support for managing variable tensors.
-
Bug Fixes and Security Improvements:
- Numerous bug fixes across versions addressed issues in model conversion processes, interpreter stability, and hardware-specific optimizations, enhancing the overall reliability and security of TensorFlow Lite.
-
New Features:
- Support for New Platforms: Added support for new hardware platforms and operating systems, including enhanced compatibility with the latest Android and iOS versions.
- Experimental Features: Introduced experimental support for features like float16 auto-mixed precision on X86 CPUs, leveraging new instruction sets for better performance.
-
Deprecations and Breaking Changes:
- Some APIs and functions have been deprecated or removed, which may require users to update their existing codebases to maintain compatibility with the latest versions.
For a detailed list of all changes, please refer to the full TensorFlow Lite changelog from: v2.7.0…v2.16.1
Best
Eoin