知识卡片
为生成器学习近似逆映射
GAN 通常没有随附编码器。可采样 z,生成 g(z),再用配对 (g(z),z) 训练编码器恢复潜变量。得到的是生成分布上的近似逆;生成器非单射时逆可能不唯一,对真实图像的编码也不由这个训练目标自动保证。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
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
GANs do not necessarily come with encoders fψ, but for applications that need one, like we have here, you can simply train an encoder via supervised learning on pairs {gθ(z), z}, solving arg minψ 𝔼z[L(fψ(gθ(z)), z)]. See [107] for discussion of this method and more advanced techniques.
参考来源
- 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第36章《鲁棒性与通用性训练》(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/c036.xhtml)
- 结论依据:摘录中关于为生成器学习近似逆映射的定义、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:GANs do not necessarily come with encoders fψ, but for applications that need one, like we have here, you can simply train an encoder via supervised learning on pairs {gθ(z), z}, solving arg minψ 𝔼z[L(fψ(gθ(z)), z)]. See [107] for discussion of this method and more advanced techniques.