← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8948.b · 生成模型:GPT-6

多尺度正弦描述坐标

理解等级:中级

书中用多个尺度的正弦分别表示x、y坐标,再拼入词元。不同频率提供不同变化尺度,但这里的纯正弦方案不同于常见成对正余弦编码,不能直接保证任意坐标范围内唯一、无歧义或具有理想外推。

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

      Sometimes, however, we may wish to retain positional knowledge. For example, knowing that a token is a representation of the top region of an image can help us identify that the token is likely to represent sky. Positional encoding concatenates a code representing position within the signal onto each token. If the signal is an image, then the positional code should represent the x- and

    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:摘录中关于多尺度正弦描述坐标的定义、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:Sometimes, however, we may wish to retain positional knowledge. For example, knowing that a token is a representation of the top region of an image can help us identify that the token is likely to represent sky. Positional encoding concatenates a code representing position within the signal onto each token. If the signal is an image, then the positional code should represent the x- and