Top-down parsing for Neural Network Exchange Format (NNEF) in TensorFlow-based deep learning computation

Abstract

As the demand for neural network technology grows in recent years, there is an increasing interest in the standardization related to the technology, and research on standardization independent file format for data exchange between deep learning data and learning systems is underway. In this paper, we introduce a more reasonable standardized method called the NNEF. Neural Network Exchange Format (NNEF) is one of the standardized methods of neural network. Neural network graph defined by NNEF is possible to exchange various neural network configuration platform. Accordingly, NNEF provides the simple process of using constant format to create a network and train network, and it has a significant impact on configuring neural network used in cross-platform. Tensorflow framework is the most common and widely adopted today. The Tensorflow provide independent frameworks for neural network configuration. In TensorFlow frameworks, artificial neural networks are represented by computational graphs. In addition, represented data (tensor) are shared among nodes. Therefore, a standardized method is desired because the method of representing a computational graph in each framework is all different. This paper introduces the Neural Network Exchange Format (NNEF) at first and then presents the purpose of such an exchange format, i.e., enabling neural network training in deep learning frameworks to be executed in a standardization manner. The aim of this approach is to describe neural network computations in a platform independent manner, while enabling the possibility for inference engines to highly optimize the run-time execution.