知识卡片

普通读书笔记卡 · 8986.b · 生成模型:GPT-6

逐步取众数不同于祖先抽样

理解等级:中级

祖先抽样按p(x1)、p(x2|x1)等依次随机取值,保留对应联合分布的随机性。每步改取最大概率值会形成贪心解码,既不是分布采样,也不保证联合最可能图像。可用均匀随机数落入累计概率区间实现类别抽样。

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

      Autogressive models give us an explicit density function, equation (32.39). To sample from this density we use ancestral sampling, which is the process described previously, of sampling the first pixel from p(x1), then, conditioned on this pixel, sampling the second from p(x2 | x1) an so forth. Since each of these densities is a categorical distribution, sampling is easy: one option is to partition a

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第32章《生成模型》(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/c032.xhtml) - 结论依据:摘录中关于逐步取众数不同于祖先抽样的定义、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:Autogressive models give us an explicit density function, equation (32.39). To sample from this density we use ancestral sampling, which is the process described previously, of sampling the first pixel from p(x1), then, conditioned on this pixel, sampling the second from p(x2 | x1) an so forth. Since each of these densities is a categorical distribution, sampling is easy: one option is to partition a