← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 9000.a · 生成模型:GPT-6

重叠图块不是独立新样本

理解等级:中级

局部判别可复用卷积参数并提供许多块级训练信号,但同图重叠块高度相关,不能按独立图像样本数计算统计优势。图中70×70等效果只说明特定设置下的权衡,最大块退化的原因是作者解释,非一般最优尺寸规律。

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

      The PatchGAN strategy has two advantages over just scoring the whole image as real or synthetic with a classifier architecture: (1) it can be easier to model if a patch is real or synthetic than to model if an entire image is real or synthetic ( needs fewer parameters), (2) there are more patches in the training data than their are images. These two properities

    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:摘录中关于重叠图块不是独立新样本的定义、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 PatchGAN strategy has two advantages over just scoring the whole image as real or synthetic with a classifier architecture: (1) it can be easier to model if a patch is real or synthetic than to model if an entire image is real or synthetic ( needs fewer parameters), (2) there are more patches in the training data than their are images. These two properities