← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 9098.d · 生成模型:GPT-6

跨注意力归一化沿图像轴

理解等级:高级

若文本有T个位置、图像有P个token,QKᵀ产生T×P分数,softmax对每个文本位置的P个候选归一化。标准缩放因子是查询键维度√d_k;式51.4用√m应对应这一维度,不能解释为任意序列长度。

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

      where variables subscripted with t are associated with text tokens and variables subscripted with ℓ are associated with image tokens. Then, the tokens output by cross-attention can be combined with tokens output by self-attention by weighted summation: (51.6) where Tout,self is regular self-attention on the input text token Tt, and wcross and wself are optional learnable weights. The text decoder is an autoregressive model (section

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第51章《视觉与语言》(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/c051.xhtml) - 结论依据:摘录中关于跨注意力归一化沿图像轴的定义、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:where variables subscripted with t are associated with text tokens and variables subscripted with ℓ are associated with image tokens. Then, the tokens output by cross-attention can be combined with tokens output by self-attention by weighted summation: (51.6) where Tout,self is regular self-attention on the input text token Tt, and wcross and wself are optional learnable weights. The text decoder is an autoregressive model (section