MobileDLSearch: Ontology-based Mobile Platform for Effective Sharing and Reuse of Deep Learning Models

Abstract

Recently, on-device inference using deep learning (DL) models for mobile and edge devices has attracted significant attention in ubiquitous computing due to its lower latency, better performance and increased data privacy. Adopting pre-trained DL models as the backbone for downstream tasks has become the consensus of the Artificial Intelligence (AI) community since it can remarkably accelerate the DL deployment process. However, most of the pre-trained DL models are not suitable for resource-constraint platforms. Further, there is a scarcity of platforms providing a unified way to store, query, share and reuse pre-trained DL models, especially for mobile applications. To address these limitations, this paper proposes an ontology-based platform (MobileDLSearch) that offers end-users greater flexibility to store, query, share and reuse pre-trained DL models for various mobile applications. The proposed Mo-bileDLSearch uses a standardised ontology to represent various DL models with different backends (e.g., TensorFlow, Keras and PyTorch), and provides an intuitive and interactive user interface to support search and retrieval of DL models. It also implements an automatic model converter to optimise desktop/laboratory-oriented pre-trained DL models for mobile platforms, and has an on-device real-time model integration module to benchmark the model's performance on mobile devices. The evaluation results demonstrate the usability of the proposed MobileDLSearch to help end-users quickly search, deploy and benchmark DL models for various on-device inference tasks.