知识卡片

普通读书笔记卡 · 9098.c · 生成模型:GPT-6

跨注意力由文本查询图像

理解等级:中级

一种图像描述模型让文本token产生Q,图像token产生K和V,对图像位置分配权重后聚合图像值向量。输出保留文本位置个数;纯跨注意力这一支聚合的是图像V,文本自身信息可经残差或独立自注意力再融合,不能混为同一V集合。

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

      There is one special layer to point out, the cross-attention layer (cross-attn). This layer addresses the question of how to insert image tokens into a text model. The answer is quite simple: just let the text tokens attend to the image tokens. The image tokens will emit their value vectors which will be mixed with the text token value vectors according to the attention weights

    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由文本查询图像的定义、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 is one special layer to point out, the cross-attention layer (cross-attn). This layer addresses the question of how to insert image tokens into a text model. The answer is quite simple: just let the text tokens attend to the image tokens. The image tokens will emit their value vectors which will be mixed with the text token value vectors according to the attention weights