← 返回刚才阅读的位置

知识卡片

结构图卡 · 8994.f · 生成模型:GPT-6

变分编码训练的数据路线

理解等级:中级

结构图:

flowchart TD
 A[观测x] --> B[编码均值与方差]
 C[独立标准高斯噪声] --> D[重参数化抽取z]
 B --> D
 D --> E[解码观测分布参数]
 E --> F[观测的条件对数似然]
 B --> G[近似后验对先验的KL]
 F --> H[最大化重建期望减KL]
 G --> H
 H --> I[更新编码器与解码器]

生成新样本时改从先验抽z,不需先输入一幅待重建图像。

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

      You may have noticed that in the previous sections we made use of both an encoder fψ (which parameterizes qψ(Z | x)) and a decoder gθ (which parameterizes pθ(X | z)); it looks like we are using the two pieces of an autoencoder but what's the exact connection?

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第33章《生成建模与表示学习》(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/c033.xhtml) - 结论依据:摘录中关于变分编码训练的数据路线的定义、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:You may have noticed that in the previous sections we made use of both an encoder fψ (which parameterizes qψ(Z | x)) and a decoder gθ (which parameterizes pθ(X | z)); it looks like we are using the two pieces of an autoencoder but what's the exact connection?