← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 8988.d · 生成模型:GPT-6

判别器不是数据密度函数

理解等级:中级

GAN判别器学习在给定真假混合采样方案下区分两类的分数,其值依赖当前生成分布与训练比例,不是对图像空间归一的数据密度。把它类比能量可帮助理解反馈方向,但不能据此直接用其数值当作真实图像似然。

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

      Notice that dϕ has a form similar to an energy function, but, unlike energy functions, dϕ is not, in general, interpretable as an unnormalized probability density. Nonetheless, we can roughly think of a GAN as a type of energy-based generative model where we train another network gθ to directly sample from the energy function rather than relying on MCMC to sample from the energy function.

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第32章《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/c032.xhtml) - 结论依据:摘录中关于判别器不是数据密度函数的定义、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:Notice that dϕ has a form similar to an energy function, but, unlike energy functions, dϕ is not, in general, interpretable as an unnormalized probability density. Nonetheless, we can roughly think of a GAN as a type of energy-based generative model where we train another network gθ to directly sample from the energy function rather than relying on MCMC to sample from the energy function.