知识卡片

普通读书笔记卡 · 9008.c · 生成模型:GPT-6

微调以旧参数作为起点

理解等级:中级

微调用预训练参数初始化新任务模型,再在目标数据上继续优化。算法 37.1 的两阶段分别使用各自数据与学习率。它改变模型参数,而非仅调用已有模型;适配幅度由优化和任务决定,“fine”并不保证参数变化必然很小。

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

      Finetuning consists of initialializing a new classifier with the parameter vector from the pretrained model, and then training it as usual. In other words, finetuning is making small (fine) adjustments (tuning) to the parameters of your model to adapt it to do something slightly different than what it did before, or even something a lot different.

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第37章《迁移学习与适配》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)/_epub-src/OPS/xhtml/c037.xhtml) - 结论依据:摘录中关于微调以旧参数作为起点的定义、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:Finetuning consists of initialializing a new classifier with the parameter vector from the pretrained model, and then training it as usual. In other words, finetuning is making small (fine) adjustments (tuning) to the parameters of your model to adapt it to do something slightly different than what it did before, or even something a lot different.