機器學習維運（MLOps）的開源解決方案之開發與研究

機器學習在當今資料驅動的世界中扮演著重要的角色，它通過從大量資料中學習規律和模式，實現對未知情況的預測和決策。然而，建立和部署機器學習系統並非容易的事，它涉及到多個技術和管理上的挑戰，如資料處理、資料管理、模型開發、模型評估、模型部署、模型監控等。這些挑戰不僅涉及到機器學習的理論和演算法，還需要融入軟體工程、系統工程和測試工程等多個領域的知識。為了應對這些挑戰並提高機器學習系統的質量和效率，機器學習維運（MLOps）概念應運而生。現今對於機器學習維運的開發與研究有限，本研究從介紹機器學習開始，探討機器學習系統開發面臨的挑戰，再從介紹軟體開發提倡已久的開發維運（DevOps）概念，接著，說明將開發維運概念應用於機器學習領域的機器學習維運，最後，探討過去關於機器學習維運工作流程的研究。本研究提出一個機器學習維運工作流程，依照它設計機器學習系統架構，並收集相關開源解決方案，最後，使用收集到的開源軟體實作出以開源解決方案所建構的機器學習系統架構，證明研究者提出之方法可行性。

關鍵字

機器學習；機器學習維運；開發維運；深度學習；開源軟體

並列摘要

Machine learning plays a pivotal role in today's data-driven world. It learns patterns and rules from large volumes of data to predict and make decisions in unknown situations. However, building and deploying machine learning systems is not an easy task. It involves various technical and managerial challenges, such as data processing, data management, model development, model evaluation, model deployment, and model monitoring. These challenges not only involve the theories and algorithms of machine learning, but also require knowledge from multiple fields such as software engineering, systems engineering, and testing engineering. In response to these challenges and to improve the quality and efficiency of machine learning systems, the concept of Machine Learning Operations (MLOps) has emerged. The current development and research on MLOps are limited. This study starts with introducing machine learning, discusses the challenges faced in the development of machine learning systems, then introduces the long-advocated concept of Development Operations (DevOps) in software development. Next, it explains the application of the DevOps concept in the field of machine learning, leading to MLOps. Finally, it explores past research on MLOps workflows. This study proposes a workflow for MLOps, designs the architecture of machine learning systems according to it, and collects related open-source solutions. In the end, using the collected open-source software, it implements the machine learning system architecture constructed by open-source solutions, demonstrating the feasibility of the proposed method.