← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 9000.d · 生成模型:GPT-6

循环一致鼓励保留可逆信息

理解等级:中级

同时学习g:X→Y与f:Y→X,用||x−f(g(x))||1和||y−g(f(y))||1约束往返重建。它鼓励输入信息在转换中保留,减少只生成任意目标域图像的解。两方向还各需域判别约束,权重决定重建与外观匹配的取舍。

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

      So, a regular GAN objective, applied to mapping from X to Y, solves the unpaired translation problem, but it does not distinguish between the correct mapping and the many other mappings that achieve the same output distribution. To break the symmetry we need to add additional constraints or inductive biases. One that works especially well is cycle-consistency, which was introduced in this context by CycleGAN

    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:摘录中关于循环一致鼓励保留可逆信息的定义、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:So, a regular GAN objective, applied to mapping from X to Y, solves the unpaired translation problem, but it does not distinguish between the correct mapping and the many other mappings that achieve the same output distribution. To break the symmetry we need to add additional constraints or inductive biases. One that works especially well is cycle-consistency, which was introduced in this context by CycleGAN