← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 9261.b · 生成模型:GPT-6

训练步数与数据遍历轮数不同

理解等级:中级

一次优化更新通常称一步,一个epoch通常表示训练数据的一次遍历;小批训练每轮可含许多步。早停应明确观察量、改善阈值和耐心,而非把每次迭代都等同完整轮次。

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

    • · 查看原书依据

      Repeat: Steps 2to 5 are repeated for a certain number of iterations (training iterations are commonly called epochs) that are predefined by the data scientist. Other stopping criteria can also be applied. For example, after several iterations, the loss may only decrease by a nominal amount and not contribute to the training process. Thus, one stopping criterion could be to stop the training process before

    参考来源

    - 位置:《Machine Learning Fundamentals: Concepts, Models, and Applications》第09章《训练机器学习模型》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Machine Learning Fundamentals: Concepts, Models, and Applications/_epub-src/OEBPS/xhtml/23_chapter09.xhtml) - 结论依据:摘录中关于训练步数与数据遍历轮数不同的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。 - 原始内容:Repeat: Steps 2to 5 are repeated for a certain number of iterations (training iterations are commonly called epochs) that are predefined by the data scientist. Other stopping criteria can also be applied. For example, after several iterations, the loss may only decrease by a nominal amount and not contribute to the training process. Thus, one stopping criterion could be to stop the training process before