Optimum Selection of DNN Model and Framework for Edge Inference

Abstract

This paper describes a methodology to select the optimum combination of deep neuralnetwork and software framework for visual inference on embedded systems. As a first step, benchmarkingis required. In particular, we have benchmarked six popular network models running on four deep learningframeworks implemented on a low-cost embedded platform. Three key performance metrics have beenmeasured and compared with the resulting 24 combinations: accuracy, throughput, and power consumption.Then, application-level specifications come into play. We propose a figure of merit enabling the evaluationof each network/framework pair in terms of relative importance of the aforementioned metrics for a targetedapplication. We prove through numerical analysis and meaningful graphical representations that only areduced subset of the combinations must actually be considered for real deployment. Our approach can beextended to other networks, frameworks, and performance parameters, thus supporting system-level designdecisions in the ever-changing ecosystem of embedded deep learning technology.