Release v0.11.0-rc0

Major changes and bug fixes

• Added support for a LRU based caching system to remove functions (and their associated tensors) that are no longer used. DL models that has dynamic shapes (thus requiring frequent recompilation and cache miss - such as GNMT, RFCN etc.) will benefit from this.
• To configure the depth of the cache use the following environment variable: NGRAPH_TF_FUNCTION_CACHE_ITEM_DEPTH=<n> where replace <n> with the appropriate value (default 16).
• Added support for generating JUNIT XML output for TensorFlow python tests.
• Added support for installing PlaidML backend using optional plaidml while specifying the wheel such as: pip install ngraph_tensorflow_bridge-0.11.0-rc0-py2.py3-none-macosx_10_7_x86_64.whl[plaidml]

Dependency information

TensorFlow v1.12.0
nGraph v0.14.0-rc.1

Release v0.10.0

Major changes and bug fixes

• Added more quantization ops to support ResNet50 quantized inference workload
• Added BackendManager::ReleaseBackend() to allow for cleanup of backend resources
• Added support for PlaidML backend configuration
• Improved diagnostics
  • Added feature for reading binary input files to support layer wise computation verification
  • Added feature for loading checkpoints to support layer wise computation verification

Dependency information

TensorFlow v1.12.0
nGraph v0.12.0

Release v0.10.0-rc1

Major changes and bug fixes

• Ops implemented/enhanced: Conv3D, Any, All, MaxPool3D, DepthToSpace, Sqrt, Squeeze
• Added an example C++ file and a Makefile to illustrate how to use TensorFlow C++ library with nGraph. See examples/tf_cpp_examples for more information.
• Diagnostics: More information available in the generated log - edge contraction, more placement statistics. Added script to pre-process the pbtxt files that replaces node names starting with ‘_’
• Added build options for using PlaidML when the SDK is installed

Dependency information

TensorFlow v1.12.0
nGraph v0.12.0-rc.1

Release v0.10.0-rc0

Major changes and bug fixes

• Fixed ResNet50 training slowness
• Fixed compilation issues on Mac - now Mac build works
• Expanded the unit test coverage
  • Now running Resnet50 training and inference tests from test_ngtf.py
• Enhanced the build_ngtf.py to support the following options:
  • --verbose_build
  • --target_archtecture
  • --use_prebuilt_binaries
• Fixed a race condition that would hang execution when an execution thread throws exception
• Added code to capture the nGraph serialized graphs automatically when a backend call fails

Known issues

• SSD-VGG16 and Wide and Deep inference is not working due to some changes in the CPU backend.
• RFCN and few other inference models are running slowly due to some changes in the CPU backend.

Dependency information

TensorFlow v1.12.0
nGraph v0.12.0-rc.0

Release v0.9.0

Major changes and bug fixes

• Fixed MacOS build that was crashing for some ops
• Added proper parsing of backend names so that individual backends can process the passed in options
• Added support for running TensorFlow python tests using nGraph backends
• Added new Python scripts to build and run unit tests (build_ngtf.py and test_ngtf.py)
• Fixed various bugs

Dependency information

TensorFlow v1.12.0
nGraph v0.11.0

Release v0.8.0

Major changes and bug fixes

• Renamed the Python module to ngraph_bridge and the Python whl to ngraph-tensorflow-bridge
• Provided optimizations for additional DL networks.
• Added support for selection of various nGraph backends and fixed relevant Python API functions
• Support for running TensorFlow Python unit tests using nGraph
• Fixed bugs in graph processing that were affecting proper execution of several models
• Build system updated to allow use of pre-existing nGraph and TensorFlow binaries to speed up the building from source significantly.

Dependency information

This release uses the following dependencies:
TensorFlow v1.12.0
nGraph v0.10.1

Release 0.7.0

Major changes and bug fixes

• Added more ops support including a few quantization ops
• Added support for using multiple backends in the unit testing framework
• Fixed a bug that was crashing nGraph when XLA is enabled
• Added abseil C++ library as a third party dependency
• Added ability to build the nbench tool
• Added a python test runner for testing nGraph using TensorFlow python unit tests
• Added support for building on CentOS
• Added more support for layer wise result comparison for debugging convergence issues

Dependency information

This release uses the following dependencies:
TensorFlow v1.12.0-rc2
nGraph v0.9.1

Release v0.7.0-rc1

• Upgraded to nGraph release v0.9.1

Release v0.7.0-rc0

• Added more ops support including a few quantization ops
• Added support for multiple backends in the unit testing framework
• Fixed a bug that was crashing nGraph when XLA is enabled
• Added abseil C++ library as a third party dependency
• Added ability to specify the build of the nbench tool
• Added a python test runner for testing nGraph using TensorFlow python unit tests
• Added support for building on CentOS

Versions used: TensorFlow v1.12.0 nGraph: v0.9.0

v0.6.1

Release v0.6.1

Changes

• Bug fixes
• Documentation update