Releases: NervanaSystems/ngraph-tf
v0.6.0
Release 0.6.0
Major Changes
- Fixed some bugs that were preventing from running several models such as Faster RCNN, RFCN etc.
- Added support for more models
- Fixed a bug in encapsulation that was picking up requested devices and not assigned devices
- More diagnostics tools update
- More unit tests added
- Examples directory was cleaned up and new instructions provided on how to use nGraph with TensorFlow
- Documentation updated
- Fixed the Mac build
Release 0.5.0
Release 0.5.0
Major Changes
-
Added support to build nGraph-TensorFlow bridge using pre-exixisting TensorFlow installation. This
eliminates the need to download/build TensorFlow from source and makes eay to get started with nGraph. -
Removed the
NGRAPH
device - TensorFlow computation sub-graphs are automatically complied and executed by nGraph without any changes needed in the TensorFlow Python scripts (or C++ client code).
Therefore scripts no longer need to specifywith device(/NGRAPH:):
as previously required.
The only change needed run TensorFlow with nGraph is to addimport ngraph
in the TensorFlow script. -
Enables CPU backend with direct execution (DEX) which eliminates the need to download and/build LLVM which cuts down the build time significantly. This also reduces the nGraph compilation time significantly.
-
Support added for more ops that allow for more inference models and ResNet50 training.
-
Support added to use other backends (such as nGPU or SPH).
-
A new set of diagnostic scripts are added to be able to visualize and debug transformed computation graphs using
TensorBoard
. This is going to be useful to debugging new models. Also added are helpful environment variables that control various debug log levels, placement of ops on nGraph, and saving of intermediate graphs for debugging and analysis.
Updated based on TF API change (post v19.0)
v0.3.0-rc1 Updated base on TF API change.
Updated for upstreaming to TensorFlow.
This version contains incremental updates with more ops and necessary code to support building from TensorFlow tree - in preparation for upstreaming.
Updated to TensorFlow v1.9.0
This release contains support for more ops and necessary API updates for TensorFlow v1.9.0 release.
Initial release
Initial release of the nGraph bridge for TensorFlow. Supports TensorFlow v1.8.0 and focused on inference workloads. Provides very good throughput and latency for popular CNN models such as ResNet50, Inception V3 and V4 and so on. Subsequent releases will focus on training workloads and RNNs.