← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8928.a · 生成模型:GPT-6

空洞卷积扩大核跨度

理解等级:中级

d倍空洞卷积把核权重间距扩大,K个采样点的有效跨度变为1+(K−1)d,而权重数保持不变。它改变的是取样支撑,并不必然降低输出分辨率。相邻权重间留下的空洞使空间覆盖稀疏,因此扩大跨度不能等同于获得同样宽度的任意稠密核。

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

      Dilated convolution is similar to strided convolution but spaces out the filter itself rather than spacing out where the filter is applied to the image: (24.9) Here we dilate by factor dk in both spatial dimensions but we could choose a different dilation in each dimension. Or, we could even dilate in the channel dimension, if we were using a multichannel convolution, but this is

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第24章《卷积神经网络》(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/c024b.xhtml) - 结论依据:摘录中关于空洞卷积扩大核跨度的定义、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:Dilated convolution is similar to strided convolution but spaces out the filter itself rather than spacing out where the filter is applied to the image: (24.9) Here we dilate by factor dk in both spatial dimensions but we could choose a different dilation in each dimension. Or, we could even dilate in the channel dimension, if we were using a multichannel convolution, but this is