OpenVINO™ integration with TensorFlow v2.3.0

Key Changes in this release:

• TF Version upgraded to v2.9.3 for security fixes
• OpenVINO version upgraded to v2022.3.0
• New Ops enabled in OCM with translations done through TensorFlow FrontEnd: Select, SegmentSum, ParallelDynamicStitch, DynamicPartition, Erf, Einsum, BroadcastGradientArgs, Concat, ExtractImagePatches, LogicalNot, LogicalOr , LogicalXor, Mod, RandomUniform, Roll, SelectV2, Swish
• Dynamic input shapes support added for a limited model scope. It can be enabled through the environmental variable OPENVINO_TF_ENABLE_DYNAMIC_SHAPES
• MYRIAD support removed as the OpenVINO 2022.3.0 package does not include the MYRIAD plugin anymore.
• CMake only build enabled with OpenVINO as a dependency
• Known issues in this release are listed here

In addition to the wheels hosted on PyPi for the TF version 2.9.3, wheels that work with the previous stable TF version 2.8.4 can be downloaded now from the GitHub Assets of this release below, and can be installed using pip. For example, to install the Python3.8 wheel along with TensorFlow 2.8.4.

pip install tensorflow==2.8.4 https://github.com/openvinotoolkit/openvino_tensorflow/releases/download/v2.3.0/openvino_tensorflow-2.3.0-cp38-cp38-manylinux_2_27_x86_64.whl

OpenVINO™ integration with TensorFlow v2.2.0

This release brings an overhaul in how OpenVINO™ integration with TensorFlow handles operator translations through the new TensorFlow FrontEnd API. This release also includes additional functional bug fixes from the previous release. It is based on the TensorFlow version v2.9.2 and the OpenVINO™ version v2022.2

• The TensorFlow upgrade to v2.9.2 provides bug fixes and addresses vulnerabilities over TensorFlow v2.9.1.
• [Preview] OpenVINO™ integration with TensorFlow's Backend Manager now has support to choose between various GPU backends like Intel's integrated graphics cards, Intel’s discrete graphics cards, Intel® Data Center GPU Flex Series, and Intel® Arc™ GPU for DL inferencing workloads in the intelligent cloud, edge, and media analytics workloads.
• Added a new notebook that demonstrates the performance benefits OpenVINO™ integration with TensorFlow brings for Object Detection architectures like SSD, FasterRCNN, and EfficientDet from the TensorFlow Hub.
• [Experimental] Model caching support added to reduce first inference latency on GPUs. It can be enabled by setting the environment variable OPENVINO_TF_MODEL_CACHE_DIR to the corresponding cache directory. For more details, please see USAGE.md.

For the complete list of features offered by OpenVINO™ check the release notes of the new OpenVINO™ toolkit v.2022.2

• Docker Support

  • OpenVINO™ integration with TensorFlow Runtime Dockerfiles for Ubuntu 18.04 and Ubuntu 20.04 are updated
  • OpenVINO™ integration with TensorFlow Runtime Dockerfiles with TF-Serving for Ubuntu 18.04 and Ubuntu 20.04 are updated
  • Prebuilt images are updated and can be found on Docker Hub and Azure Marketplace.

OpenVINO™ integration with TensorFlow v2.1.0

This release provides performance improvements and functional bug fixes from the previous release.

• Performance Optimizations of existing supported models

  • Updated opset support to OpenVINO™ opset8
  • Dynamic shapes support enabled for translation of multiple Ops, which resulted in larger cluster sizes and performance improvement
  • Added support for following TensorFlow Ops in OpenVINO™ integration with TensorFlow: CTCGreedyDecoder, SparseToDense, BatchMatMulV2, BatchMatMul, SquaredDifference, LessEqual, NotEqual, Cumsum

• TensorFlow version upgraded to v2.9.1. It provides bug fixes and addresses vulnerabilities over TensorFlow v2.8.0

  • OpenVINO™ integration with TensorFlow PyPi release is built against v2.9.1, and it will also work with PATCH versions like v2.9.0, and the upcoming v2.9.x releases
  • OpenVINO™ integration with TensorFlow source code is backward compatible. This means you will be able to build its source code with the past MINOR versions of TensorFlow 2.x. (validated for TensorFlow versions v2.4.4, v2.5.3, v2.6.5, v2.7.3, v2.8.2 and v2.9.1)

• Docker Support

  • OpenVINO™ integration with TensorFlow Runtime Dockerfiles for Ubuntu 18.04 and Ubuntu 20.04 are updated
  • OpenVINO™ integration with TensorFlow Runtime Dockerfiles with TF-Serving for Ubuntu 18.04 and Ubuntu 20.04 are updated
  • Prebuilt images are updated and can be found on Docker Hub and Azure Marketplace

• Enhanced GitHub documentation

OpenVINO™ integration with TensorFlow v2.0.0

This release provides functional improvements and enhanced backend support from the previous release.

• Version Upgrades

  • OpenVINO™ version upgraded to 2022.1.0. It provides functional bug fixes, and capability changes from the previous 2021.4.2 LTS release. This new release has increased the support for more deep learning models, and has improved inferencing performance

  • TensorFlow version upgraded to 2.8.0. It provides bug fixes and address vulnerabilities over TensorFlow 2.7.0

  • Increased model coverage across TensorFlow Hub on all the supported Intel backends

  • Performance Optimizations of existing supported models

• GA Support for Windows 10 - 64 bit - with improved model support and enhanced performance

• Docker Support

  • Dockerfiles for Ubuntu 18.04 and Ubuntu 20.04 OS are available. Also, Dockerfiles of TensorFlow Serving for corresponding OS's are released. These can be used to build runtime Docker images for OpenVINO™ integration with TensorFlow with CPU, GPU, VPU, and VAD-M support

  • Prebuilt images can be found on Docker Hub

  • TensorFlow Serving Docker images released for OpenVINO™ integration with TensorFlow can be used to run all the workflows supported by standalone TensorFlow Serving Docker images

• Hardware support

  • Updated platform support for AlderLake Family

• Intel® DevCloud Gen1 and Gen2 samples updated for OpenVINO™ integration with TensorFlow v2.0.0

• Enhanced GitHub documentation

OpenVINO™ integration with TensorFlow v1.1.0

1. This release provides functional improvements and enhanced backend support from the previous release. This includes
   • OpenVINO™ 2021.4.2 upgrade - Long Term Support release of OpenVINO™ with focus on stability and compatibility.
   • TensorFlow v2.7.0 upgrade – Bug fixes and address vulnerabilities over TensorFlow 2.5.1 release.
   • Increased model coverage across TensorFlow Hub on all the supported Intel backends.
   • Performance Optimizations of existing supported models on all Intel backends
2. Additional support for Windows 10 - 64 bit
   • It is a Beta Preview release
   • Performance Optimization/improvements are in progress
3. Updated platform support for AlderLake Family.
4. Python/ C++ examples and Colab notebooks are now compliant to TensorFlow 2.x.
5. Enhanced GitHub documentation.

OpenVINO™ integration with TensorFlow v1.0.1

This release provides few functional bug fixes and minor updates over the previous release

• Fixes Quantization models support
• Minor fix in Colab notebook and Build documentation
• Image Tagging utility added to examples
• Extra checks added to address vulnerabilities

OpenVINO™ integration with TensorFlow v1.0.0

1. This release provides functional improvements and enhanced backend support from the previous release. This includes
   • OpenVINO™ 2021.4.1 upgrade - Long Term Support release of OpenVINO™ with focus on stability and compatibility.
   • TensorFlow v2.5.1 upgrade – Bug fixes and address vulnerabilities over TensorFlow 2.4.1 release.
   • OpenVINO™ integration with TensorFlow samples in Intel DevCloud – User can learn and test this capability across multiple Intel hardware devices.
   • TensorFlow 2.x compliant C++ samples covering additional use-cases.
   • Enabled FP16 precision on iGPU backend.
   • Increased model (TensorFlow Hub/ OpenVINO™ Model Zoo) coverage on all the supported Intel backends.
   • Performance Optimizations of existing supported models on all Intel backends
   • Preview support of INT8 quantized models using OpenVINO™ Neural Network Compression Framework (NNCF)
   • API to export OpenVINO™ IR (Intermediate Representation) for easy transition to native OpenVINO™
2. Additional support for MacOS and Ubuntu 20.04 and Python 3.9.
3. Updated platform support for TigerLake and IceLake Family.
4. Documentation improvements
   • Added Mandarin version of GitHub documentation
   • Enhanced GitHub documentation

OpenVINO™ integration with TensorFlow

This is a Preview Release