知识卡片
点积缩放按键宽而非序列长
点积的数值尺度可随键查询维数增加而变大,使softmax过于集中;除以√d_k用于调节。它不是按文本词数缩放,也不保证全部层没有梯度消失或爆炸;稳定训练还依残差、归一化与优化设置。
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
原书出处 · 1 处
- · 查看原书依据
There are two ways to implement the linear algebra of an attention algorithm: additive attention and dot-product attention. A scaled version of dot-production attention is most effective in transformers. For dot-product attention, the scalar products between the query vectors Q and the key vectors K are scaled down based on how many dimensions there are in the model. This makes the dot product more numerically
参考来源
- 位置:《Natural Language Processing in Action, Second Edition》第09章《可堆叠的Transformer》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Natural Language Processing in Action, Second Edition/_epub-src/OEBPS/Text/chapter-9.html)
- 结论依据:摘录中关于点积缩放按键宽而非序列长的定义、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:There are two ways to implement the linear algebra of an attention algorithm: additive attention and dot-product attention. A scaled version of dot-production attention is most effective in transformers. For dot-product attention, the scalar products between the query vectors Q and the key vectors K are scaled down based on how many dimensions there are in the model. This makes the dot product more numerically