← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 7011.h · 生成模型:DeepSeek-V4.1-Flash

全周期监控的心态转变

理解等级:高级

技术挑战虽多,但抵触全周期监控的组织与文化因素 arguably 最相关:建模者通常想着部署前的建模 KPI 表现,而不是部署后的问题检测,而建模 KPI 未必直接连到业务 KPI,这显然给全周期监控带来麻烦——部署前后都重要,且像成功的软件一样,部署后阶段往往比你以为的更长久。急于开发部署的团队常对严谨交付框架不耐烦,觉得它们妨碍自己按最方便的方式组织训练与服务;这在某种程度上确实如此,但换来的是一套生产中的监控与管理保证,否则只能靠临时拼凑、难以达成。作者由此定下本章总体目标:推荐全周期监控方法,尤其给出一组除「模型本就旨在改善的具体业务指标」之外的默认监控项(那些指标已被充分理解)。既然 TensorBoard、Weights & Biases 这类开发期专用工具通常不能自然迁移到生产,眼下许多做法难免要自己补,那么最该做的是准备一个合理、灵活的方案,在整个生命周期维持尽可能广而有用的模型行为图景,并按自身情境调整。

这张卡片看不懂?

已有反馈 · 0 条

    还没有人提交反馈。

    在整书笔记中的位置

    原书出处 · 1 处

    • Monitoring and Observability for Models · 查看原书依据

      Given that, the overall goal for this chapter is to recommend a whole-lifecycle approach to monitoring, and in particular, suggest a default set of things to monitor other than the specific business metrics the model is intended to improve, since they are already well understood. Best Practices for ML Model Monitoring Let’s start off with a few framing assumptions: for the purposes of this chapter, model development is general

    参考来源

    - 位置:《Reliable Machine Learning》第09章《Monitoring and Observability for Models》(源文件:extracted/Reliable Machine Learning/_epub-src/OEBPS/ch09.xhtml) - 结论依据:原文明确阐述该主题,并给出定义、实践或反例。 - 原始内容:Given that, the overall goal for this chapter is to recommend a whole-lifecycle approach to monitoring, and in particular, suggest a default set of things to monitor other than the specific business metrics the model is intended to improve, since they are already well understood. Best Practices for ML Model Monitoring Let’s start off with a few framing assumptions: for the purposes of this chapter, model development is general