This article may rely excessively on sources too closely associated with the subject, potentially preventing the article from being verifiable and neutral. (February 2019) |
Neural Network Exchange Format (NNEF) is an artificial neural network data exchange format developed by the Khronos Group. It is intended to reduce machine learning deployment fragmentation by enabling a rich mix of neural network training tools and inference engines to be used by applications across a diverse range of devices and platforms.[2][3]
 | |
| Developer(s) | Khronos Group |
|---|---|
| Stable release | 1.0.5
/ February 16, 2022[1] |
| Operating system | Cross-platform |
| Platform | Cross-platform |
| Type | API |
| Website | www |
History
editNNEF was proposed in 2015 by member companies of the Khronos Group as a device and implementation independent transfer format capable of describing any artificial neural net in terms of its structure, operations and data.
The first version of the standard was launched in provisional form in December 2017, and was ratified as an official Khronos standard in August 2018.
Objectives
editThe goal of NNEF is to enable data scientists and engineers to easily transfer trained networks from their chosen training framework into a wide variety of inference engines. NNEF encapsulates a complete description of the structure, operations and parameters of a trained neural network, independent of the training tools used to produce it and the inference engine used to execute it.
Governance and Availability
editNNEF is maintained by the Khronos Group under its Open Governance Principles[4] as follows:
- Any company is invited and able to join Khronos to contribute to and influence the development of its specifications;
- Finalized specifications are publicly and freely distributed at zero cost from the Khronos web-site;
- Any company can implement a Khronos specification and participating implementers can obtain a trademark license for conformant implementations and pay zero royalties to Khronos participants; and
- Developers may freely use implementations of Khronos specifications.
The NNEF specification is available on the Khronos NNEF registry and tools are available on Github
Versions
editIndustry Participation
editThe following Khronos members have participated in the NNEF working group:
- AIMotive.
- Advanced Micro Devices.
- Arm Holdings, Ltd.
- Axell
- Axis Communications.
- Cadence
- Ceva
- Codeplay
- Digital Media Professionals
- ETRI
- Huawei
- Intel Corp.
- Imagination technologies
- LG
- Los Alamos National Lab
- LunarG
- Mediatek
- Mentor Graphics
- NXP
- On Semiconductor
- Qualcomm
- The Qt Company
- Renesas
- Samsung
- Silicon Studio
- Socionext
- Sony
- Synopsys
- Texas Instruments
- Think Silicon
- Verisilicon
- Xilinx
Tools
editThe NNEF tools project on GitHub contains the following open source tools:
- File format Parser
- Bidirectional converters between NNEF and ONNX, Caffe, Caffe2, TensorFlow (python), TensorFlow (protobuf)
- Model zoo: reference collection of models converted to NNEF
See also
editReferences
edit- ^ "Releases".
- ^ "NNEF - Neural Network Exchange Format (NNEF)". The Khronos Group. 2016-10-04. Retrieved 2019-02-07.
- ^ Seo, B.; Shin, M.; Mo, Y. J.; Kim, J. (January 2018). "Top-down parsing for Neural Network Exchange Format (NNEF) in TensorFlow-based deep learning computation". 2018 International Conference on Information Networking (ICOIN). pp. 522–524. doi:10.1109/ICOIN.2018.8343173. ISBN 978-1-5386-2290-2. S2CID 5053900.
- ^ Khronos IP Framework
- ^ v1.0p Khronos PR
- ^ "The Khronos Group launches new standard for deploying trained neural networks". SD Times. 2018-08-13. Retrieved 2019-02-11.
- ^ "Khronos NNEF Registry - The Khronos Group Inc". www.khronos.org. Retrieved 2019-08-15.