Design of Neural Network Model Converting Framework based on NNEF

Abstract

There are various engines that process artificial neural networks, such as Caffe, PyTorch, and Darknet. These engines can be operated on PCs, servers, as well as on-board devices. To cope with this kind of the diversity in these computing environments and neural network engines, compatibility among these engines is emerging as an important issue. This paper describes a neural network model converting framework to provide compatibility among these neural network engines. In this paper, we adopted a pivot model that can be used as common format among the engines. This pivot model is NNEF that was proposed by Khronos Group. With NNEF, we designed Neural Network Model Converting Framework. This framework provides user interface that configures the conversion modules between NNEF format and neural network engine's model saving formats. The framework has Manager that enables conversion and optimization of the neural network models. We also made a prototype for this framework to verify its functionalities. The prototype was tested with a neural network that recognizes hand-written numbers on Darknet and PyTorch engines.