Implantação de modelos de aprendizado de máquina no formato onnx utilizando diferentes frameworks

Abstract

This paper presents a systematic analysis of the Open Neural Network Exchange (ONNX) and ONNXRuntime, along with the main frameworks for training and serving machine learning models. The general objective is to evaluate the performance and interoperability of ONNX Runtime on multiple platforms, both in batch and real-time inference, comparing it with other inference solutions. In addition, specific objectives are presented, including a literature review on ONNXandONNXRuntime, the development of a pipeline for distributed execution of tested algorithms, the evaluation of ONNX Runtime performance, the analysis of ONNX interoperability with different machine learning frameworks, and the documentation of the results of the systematic analysis and experiments. The study also includes a comparative analysis of fundamental characteristics among related works, such as inference engines, platforms used, application domains, evaluated metrics, models used, and hardware accelerators. The obtained results provide insights and conclusions on the performance and interoperability of ONNX Runtime, contributing to the understanding and improvement of deploying machine learning models in different environments and usage scenarios.