The Khronos Group, an industry consortium creating open standards, announced the release of Neutral Network Exchange Format [NNEF] 1.0 Provisional specification facilitating universal exchange. With the announcement, Khronos Group is very clear about aim and is ready to incorporate comments and feedbacks on NNEF GitHub repository before the specifications are finalized.
Their end goal is to empower data scientists and engineers to facilitate the exchange of trained networks from chosen training framework to a whole big variety of inference engines. The new format is versatile and standard to an extent that equipment manufacturers can depend on its deployment onto edge devices. NNEF captures the complete description of structure and parameters of the neural network, for independent training tools, and interference engine used for the execution.
The field of machine learning benefits from the vitality of the many groups working in the field, but suffers from a lack of common standards, especially as research moves closer to multiple deployed systems.
Peter McGuinness, Chairperson at NNEF work, says, “The field of machine learning benefits from the vitality of the many groups working in the field, but suffers from a lack of common standards, especially as research moves closer to multiple deployed systems.” He further adds that “Khronos anticipated this industry need and has been working for over a year on the NNEF universal standard for neural network exchange, which will act as the equivalent of a pdf for neural networks.”
Torch, Caffe, and Caffe2, TensorFlow, Theano, Chainer, MXNeT are a few of the many tools and engines across which NNEF can perform the exchange with reliability. However, NNEF 1.0 Provisional specification encompasses a bunch of use cases and multiplicity of network types and a scalable design that borrows syntactical elements from Python and aids correctness by adding formal elements. This also includes custom compound operations and posh network optimization. The entire goal is to draw the future work on this very same architecture for providing a stable platform in the field of quickly evolving machine learning.
The giant has triggered a sequence of open source projects, which includes NNEP syntax parser or validator from selected frameworks and encourages machine learning community to use NNEF for their workflows.
Additionally, NNEF is all set to work closely with popular product, OpenVX, of Khronos group for enabling incorporation of NNEF files for deep learning operations in a single graph by leveraging OpenVX 1.2 as a cross-platform interface.
An industry-wide support is apparent as the announcement of the release of NNEF 1.0, and the professionals are taking all the pride that they can during the development of the network. From hardware engineering key players like Almotive to Radeon Technologies group and from Machine learning community Arm, to wireless technology forerunners, Qualcomm technologies, everyone is on the same positive front, as the development of NNEF. As the vice presidents and chairpersons come to the support, there is an industry-wide flare of acceptance and eager wait for flexibility in handling flat and compound operations.
The new NNEF 1.0 documentation project and specification are available on the Khronos Registry. Please visit The Khronos Group official site for more information.