知识卡片
逐词元投影对应一点卷积
QKV投影和逐词元MLP中的共享线性层,可在相应张量布局下表示为1×1卷积。它们只混合当前位置的特征通道。注意力的跨位置权重随输入改变,因此不能因这些局部映射可写成卷积,就把整个自注意力等同固定卷积核。
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
The first step in working with transformers is to tokenize the input. The most basic way to do this is to chop up the input image into non-overlapping patches of size K × K, then convert these patches to vectors via a linear projection. You might already have noticed that this operation can be written as convolution; after all we said the whole idea of
参考来源
- 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第26章《Transformer》(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/c026.xhtml)
- 结论依据:摘录中关于逐词元投影对应一点卷积的定义、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:The first step in working with transformers is to tokenize the input. The most basic way to do this is to chop up the input image into non-overlapping patches of size K × K, then convert these patches to vectors via a linear projection. You might already have noticed that this operation can be written as convolution; after all we said the whole idea of