知识卡片

普通读书笔记卡 · 6501.i · 生成模型:GPT-6

模型权重快照不同于完整续训

理解等级:中级

state_dict 权重足以在匹配架构中加载做推理,但精确续训通常还需优化器状态、随机状态、数据位置和训练配置。书中检查点只保存模型权重;重新调用训练函数还会新建优化器,不能称完全恢复原训练进度。

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

      Note that, next to the logging steps that print results periodically to track progress, we also save the model checkpoint (a snapshot of the model weights) periodically so we can load, evaluate, and use it later, as discussed in the next section.

    参考来源

    - 位置:《Build a Reasoning Model (From Scratch)》第06章《强化学习训练推理模型》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Build a Reasoning Model (From Scratch)/_epub-src/OEBPS/Text/chapter-6.html) - 结论依据:摘录中关于模型权重快照不同于完整续训的定义、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:Note that, next to the logging steps that print results periodically to track progress, we also save the model checkpoint (a snapshot of the model weights) periodically so we can load, evaluate, and use it later, as discussed in the next section.