ModelPS : an open-source and collaborative model edit platform with interactive transfer learning

Abstract

Many applications are utilizing deep neural networks (DNNs) to provide intelligent services. However, due to frequent changes in the business requirements, developers often have to modify these DNNs, resulting in a massive cost in terms of monetary and time. To alleviate the issue, we propose a novel system, termed ModelPS (Model Photoshop), offering an interactive workspace for DNN editing.

In our system, we first design and implement a no-code web interface that is intuitive and tailored for model improvement with different transfer learning methods (e.g., knowledge distillation). Second, we propose an intelligent module named Model Genie that assists the user to heuristically find the most favored settings under certain deployment scenarios and model performance constraints. Third, ModelPS provides a collaborative workspace that enables a team to review and share DNNs. Finally, we propose a seamless connection between improved models and efficient Machine Learning Services (MLaaS) with an automatic model deployment platform -- MLModelCI. Our system is released on GitHub\footnote{\url{https://github.com/cap-ntu/ML-Model-CI}}.