带深度学习部件的智能化软件DevOps关键技术研究

带深度学习模块的智能化软件在近十年的飞速发展中取得了巨大的成效。面向各类任务的深度神经网络模型作为其软件核心组件，已广泛应用于各行各业之中。当深度神经网络模型脱离了传统的实验室场景，转而步入真实的应用场景时，将面临很多和传统软件开发、部署和运维过程中面对的类似的软件工程问题。深度学习模型这类智能化软件天生带有开发过程快速迭代、部署运行环境差异大、以及脆弱的鲁棒性催生质量保障需求等特点。DevOps作为日渐兴起的软件开发过程模型，其核心思想在构建开发团队和运维团队的紧密协作，使得软件得以有效的进行持续演化。面对这一类新兴的带深度学习模块的智能化软件，本研究将模型及其训练得到的参数当作一种“新型的代码”，从软件工程的视角着眼智能化软件的DevOps思想，同时结合深度学习乃至机器学习的技术基石研究其中的软件工程问题，对于推动智能化软件的软件工程化有着深远的意义。

Intelligent software with deep learning models has achieved outstanding results in the rapid development of the past decade. As the core component of intelligent software, deep neural network models for various tasks have been widely used in many industry applications. When the deep neural network model deviates from the traditional laboratory scenario and enters the scenario of real-world application, it faces many software engineering problems similar to those encountered in the process of conventional software development, deployment, and operation and maintenance. Deep learning model based intelligent software is inherently characterized by rapid iteration of the development process, large differences in deployment and operating environments, and fragile robustness that give rise to quality assurance requirements. DevOps is an emerging software development process model. Its core idea is to build a close collaboration between the development team and the operation and maintenance team so that the software can effectively continue to evolve. Faced with this kind of emerging intelligent software with deep learning models, this research treats the model and its training parameters as a "new type of code." It focuses on the DevOps idea of intelligent software from the perspective of software engineering. Combining the technical cornerstones of deep learning and even machine learning to study the software engineering issues is of far-reaching significance for promoting the software engineering of intelligent software.