← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 9000 · 生成模型:GPT-6

平均块分数不等于平均块损失

理解等级:中级

若各块输出真假概率,先平均概率再取对数,与对各块对数损失求平均通常不等价。实现PatchGAN需明确输出logit或概率、损失归约轴、有效块数及边界处理。两类Softmax或单logit均可表达二元判断,不必固定两个输出通道。

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

      where k × k is the patch size. Notice that this operation is equivalent to the sliding window action of CNN, so can simply be implemented as a CNN that outputs a 2 × N × M feature map of real-versus-synthetic scores. Naturally, patches can be sampled at different strides and resolutions, depending on the exact architecture of the CNN .

    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:摘录中关于平均块分数不等于平均块损失的定义、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:where k × k is the patch size. Notice that this operation is equivalent to the sliding window action of CNN, so can simply be implemented as a CNN that outputs a 2 × N × M feature map of real-versus-synthetic scores. Naturally, patches can be sampled at different strides and resolutions, depending on the exact architecture of the CNN .