Changes

• Upgraded TensorFlow Lite version from 2.9 to 2.13.

Changes

• Added the ImageSearcher iOS API. It’s available via the TensorFlowLiteVision pod on CocoaPods.
• Upgraded TensorFlow Lite version from 2.9 to 2.10.

Fixes

• Fixed the “image_utils has no attribute 'ImageDataFree'” error occurred when using the Task Library Python API.

Changes

• Added a new task BertCLUAnnotator. It’s available on Java, Python and C++.
• Added these Natural Language Processing tasks’ Python APIs: NLClassifier, BertNLClassifier, BertQuestionAnswerer.
• Added the AudioClassifier iOS API. It’s available via the new TensorFlowLiteAudio pod on CocoaPods.

Fixes

• Fixed “TypeError: create_from_options(): incompatible function arguments.” error when using the Task Library Python API without the tensorflow pip package installed.
• Fixed the num_thread option not being applied on Task Library iOS API.

Changes

• Migrated a number of Python APIs to use Dataclasses instead of Protobuf for custom configuration
• Renamed some Swift methods to better aligned with Swift naming conventions
• Updated TFLite dependency to 2.9

Fixes

• libportaudio is now only required for audio tasks. The Python library can be installed without this dependency.

Major features

• Added support for on-device image and text search in Task Library. See the documentation of ImageSearcher and TextSearcher.
• Added Python API for all Task Library Vision / NLP / Audio tasks. See the Run inference in Python section of each task in Task Library’s documentation for more details.
• Supported building the TensorFlow Lite Support Pypi package on Raspberry Pi and Coral.
• Added Object-C API for Task Library Vision tasks. Note that the Object-C NLP API already exists. See the Run inference in Swift section of each task in Task Library’s documentation for more details.
• Added support for CoreML delegate in Task Library C++ API.

Fixes

• Fixed an offset bug in Java TensorAudio when loading short[] audio buffers.

Major features

• Add scripts to build TFLite Support Python wheels for ARM systems, i.e. Raspberry Pi.
• Add method to export the "final" metadata string in MetadataWriter.
• Throw RuntimeExceptions instead of AssertionError on native errors in Task Java Library.

Major features

• Task
  • ODML Image support in Java
  • C headers in AAR
  • Acceleration support in Java and C++ API
  • Audio Embedder in C++ API
• Support
  • ODML Image proxy support in Java
• Metadata
  • Generic metadata writer

Fixes

• Task
  • No extra TFLite runtime in dependency tree (binary size decreased)
• Support
  • No TFLite runtime dependency

This release only targets on fixing a ByteBuffer incompatibility issue in the TFLite Metadata Java library. The updated metadata library can be found on Maven Central Repository.

Fixes

• On some devices, using TFLite Metadata library throws NoSuchMethodError at runtime. The error message looks like: "java.lang.NoSuchMethodError: No virtual method position(I)Ljava/nio/ByteBuffer; in class Ljava/nio/ByteBuffer; or its super classes"

Release changes

• Java / Android artifacts are now pushed to Maven Central and Sonatype OSSRH Snapshot
• Python artifacts now support Python 3.9

Major Features

Task Library

• Add a new task: Audio classification in a new prebuilt artifact (maven: org.tensorflow:tensorflow-lite-task-audio)
• Support YUV image and Android media.Image as inputs in vision Tasks
• Support configuring max_seq_length in BertNLClassifier

Support Library

• Support YUV, grayscale and Android media.Image in TensorImage
• Add Audio data supports: TensorAudio and AudioBuffer

Metadata

• Add Metadata Writers for common tasks in the pip package

Fixes

Task

• Fix a JNI local ref overflow issue on lower version Android devices
• Fix a memory leak issue in NL Classification

Metadata

• Prevent using Flatbuffers 2.0 which is not compatible with TFLite Support

Codegen

• Fix a class name generation error

Release 0.1.0

Major features

It's the first release of the TFLite Support toolkit. We provides the following components:

• TFLite Task Library (source, JCenter): enables integrating popular ML use cases in a few lines of code.
  • Image Classifier
  • Object Detector
  • Image Segmenter
  • NL Classifier
  • Bert NL Classifier
  • Bert Question Answerer
• TFLite Support Library (source, JCenter): simplifies pre-processing/post-processing code for models.
  • Utilities for image pre/post processing
• TFLite Metadata Library (source, PyPI, JCenter, MavenCentral): reads and writes TFLite Metadata in models.
  • Utilities for extracting and populating TFLite Metadata
• TFLite Support Codegen (source, PyPI): generates model wrapper automatically based on Metadata.
• TFLite Custom Ops (source): helps deploying cutting-edge models on devices.

Compatibility

Verified TF version

TF 2.4.0

Verified TFLite Model Maker version

TFLite Model Maker 0.2.4

The final RC for the formal release 0.1.0.