← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 9220.g · 生成模型:GPT-6

自动求导不替代损失与图的设计

理解等级:中级

框架能按已构建的可微计算图求导,但无法替用户保证损失目标正确、离散操作可导或计算路径未断开。理解δ递推有助于定位梯度缺失和索引错误,不能把“自动”理解为任意程序都有有效训练梯度。

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

      Fortunately, calculating the partial derivatives is not very hard in practice since there are tools that automatically calculate the gradient along a computational chain in various programming languages (e.g., autograd, PyTorch, among others). These tools are at the core of modern machine learning frameworks and allow you to construct arbitrary network architectures without worrying about how to actually calculate the gradients. Yet it is difficult

    参考来源

    - 位置:《Introduction to Autonomous Robots Mechanisms, Sensors, Actuators, and Algorithms (Nikolaus Correll, Bradley Hayes etc.)》附录D《附录D:反向传播》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Introduction to Autonomous Robots Mechanisms, Sensors, Actuators, and Algorithms (Nikolaus Correll, Bradley Hayes etc.)/_epub-src/OEBPS/xhtml/appendixD.xhtml) - 结论依据:摘录中关于自动求导不替代损失与图的设计的定义、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:Fortunately, calculating the partial derivatives is not very hard in practice since there are tools that automatically calculate the gradient along a computational chain in various programming languages (e.g., autograd, PyTorch, among others). These tools are at the core of modern machine learning frameworks and allow you to construct arbitrary network architectures without worrying about how to actually calculate the gradients. Yet it is difficult