知识卡片
注意力图提供诊断而非完整归因
目标源权重集中可提示词对齐模式,跨头平均和层选取又会影响图。权重大不证明该词是输出的唯一原因,也不保证翻译已正确;诊断应配错例、对照和真实源目标方向。
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
Let’s focus our attention on, well … attention. In your model, you defined a CustomDecoder that saves the average attention weights for each decoder layer on each forward pass. You have the attention weights from the translation. Now, write a function to visualize self-attention for each decoder layer, using matplotlib.
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:摘录中关于注意力图提供诊断而非完整归因的定义、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:Let’s focus our attention on, well … attention. In your model, you defined a CustomDecoder that saves the average attention weights for each decoder layer on each forward pass. You have the attention weights from the translation. Now, write a function to visualize self-attention for each decoder layer, using matplotlib.