← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8661.a · 生成模型:GPT-6

预热与衰减控制不同训练阶段

理解等级:中级

预热从较小学习率逐步增至目标值,衰减在后段降低步幅,重启则按指定周期重新提高。线性、余弦与常数路线各有条件;应按优化器更新步等明确定义时间轴,不从示意图推断普遍优劣。

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

      Learning schedules Toward the end of the training process, it is a good idea to lower the learning rate because you do not want to overshoot when you are so close to convergence. In a similar vein, you would like to prevent your model from learning too much from the first few batches of examples. In either case, we would like to be able to

    参考来源

    - 位置:《Designing Large Language Model Applications》第06章《微调》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch06.html) - 结论依据:摘录中关于预热与衰减控制不同训练阶段的定义、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:Learning schedules Toward the end of the training process, it is a good idea to lower the learning rate because you do not want to overshoot when you are so close to convergence. In a similar vein, you would like to prevent your model from learning too much from the first few batches of examples. In either case, we would like to be able to