知识卡片
梯度裁剪限制范数不修复目标
范数超过阈值时整体缩放梯度可限制某次更新幅度,但不能解决错误标签、无效目标或所有非有限值。它与学习率、损失缩放和优化器行为要分别记录,不把开启裁剪当作稳定训练保证。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
- · 微调
原书出处 · 1 处
- · 查看原书依据
max_grad_norm This is used for gradient clipping, which is a solution for the exploding gradients issue that is sometimes encountered during training. The max_grad_norm value is the threshold for gradient clipping. If the L2 gradient norm is above the threshold, then it will be rescaled to max_grad_norm. For more details on gradient clipping, see “Understanding Gradient Clipping (and How It Can Fix Exploding Gradients Problem)”.
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:摘录中关于梯度裁剪限制范数不修复目标的定义、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:max_grad_norm This is used for gradient clipping, which is a solution for the exploding gradients issue that is sometimes encountered during training. The max_grad_norm value is the threshold for gradient clipping. If the L2 gradient norm is above the threshold, then it will be rescaled to max_grad_norm. For more details on gradient clipping, see “Understanding Gradient Clipping (and How It Can Fix Exploding Gradients Problem)”.