← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8988.c · 生成模型:GPT-6

交替更新形成变化中的目标

理解等级:中级

固定生成器时更新判别器识别当前伪造,固定判别器时更新生成器减少可识别差异,然后交替。生成器改变后旧判别器通常不再最优,但增加判别步数或持续训练不保证收敛、样本质量单调提升或覆盖全部模式,需实际监控。

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

      Now, because the optimal discriminator depends on the current behavior of gθ, as soon as we change gθ, updating it to better fool , then no longer is the optimal discriminator and we need to again solve problem in equation (32.50). To optimize a GAN, we simply alternate between taking one gradient step on the objective in equation (32.51) and then K gradient steps on

    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:摘录中关于交替更新形成变化中的目标的定义、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:Now, because the optimal discriminator depends on the current behavior of gθ, as soon as we change gθ, updating it to better fool , then no longer is the optimal discriminator and we need to again solve problem in equation (32.50). To optimize a GAN, we simply alternate between taking one gradient step on the objective in equation (32.51) and then K gradient steps on