知识卡片

普通读书笔记卡 · 8999.c · 生成模型:GPT-6

判别器感受野决定可见错误

理解等级:中级

PatchGAN对多个局部块评分,使判别器关注纹理、边缘等块内关系。1×1感受野只能检查单位置条件关系,较大块能约束更高阶结构,但仍可能漏掉跨块的全局布局。增大块尺寸也不保证更好,容量与数据需求随之改变。

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

      Thinking of dϕ as a learned loss function raises new questions: What does this loss function penalize, and how can we control its properties. One of the main levers we have is the architecture of dϕ; different architectures will have the capacity, and/or inductive bias, to penalize different kinds of errors. One popular architecture for image-to-image tasks is a PatchGAN discriminator [232], in which we

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第34章《条件生成模型》(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/c034.xhtml) - 结论依据:摘录中关于判别器感受野决定可见错误的定义、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:Thinking of dϕ as a learned loss function raises new questions: What does this loss function penalize, and how can we control its properties. One of the main levers we have is the architecture of dϕ; different architectures will have the capacity, and/or inductive bias, to penalize different kinds of errors. One popular architecture for image-to-image tasks is a PatchGAN discriminator [232], in which we