知识卡片
图像分块后投影成词元
将RGB图像切成K×K小块,每块展平成3K²维向量,再用共享线性映射投影到d维,得到N×d词元矩阵。非重叠分块可用核宽和步幅均为K的卷积实现;分块大小影响词元数、细节保留和后续注意力成本。
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 to working with tokens is to tokenize the raw input data. Once we have done this, all subsequent layers will operate over tokens, until the output layer, which will make some decision or prediction as a function of the final set of tokens. How can we tokenize an input image? Well, how did we “neuronize” an image for processing in a vanilla
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:摘录中关于图像分块后投影成词元的定义、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 to working with tokens is to tokenize the raw input data. Once we have done this, all subsequent layers will operate over tokens, until the output layer, which will make some decision or prediction as a function of the final set of tokens. How can we tokenize an input image? Well, how did we “neuronize” an image for processing in a vanilla