Design of NNEF-PyTorch Neural Network Model Converter

There are various types of inference engines based on neural networks. These engines define their own neural network representation models. Because each inference engine uses a different neural network representation model, there is a difficulty in sharing the trained data with each other. To solve...

Full description

Saved in:
Bibliographic Details
Published in2021 International Conference on Information and Communication Technology Convergence (ICTC) pp. 1710 - 1712
Main Authors Lee, Kyung Hee, Park, Jaebok, Kim, Seon-Tae, Kwak, Ji Young, Cho, Chang Sik
Format Conference Proceeding
LanguageEnglish
Published IEEE 20.10.2021
Subjects
Artificial neural networks
Data models
Engines
Information and communication technology
Interoperability
model converter
neural network
NNEF
PyTorch
Training
Online AccessGet full text

Cover

More Information
Summary:There are various types of inference engines based on neural networks. These engines define their own neural network representation models. Because each inference engine uses a different neural network representation model, there is a difficulty in sharing the trained data with each other. To solve this interoperability problem, a standard neural network model format called NNEF (Neural Network Exchange Format) has been proposed. If the result data trained through the training engine is stored in a standard neural network model format and a neural network model converter is provided for each inference engine, interoperability among the training engines and the inference engines can be increased. In this paper, we designed a NNEF-PyTorch neural network model converter for PyTorch, one of the neural network engines for neural network training and inferencing. To verity the converter, we show that experimental neural network model written in NNEF can be executed on PyTorch engine after the converter transformed the model into PyTorch model.
DOI:10.1109/ICTC52510.2021.9621003