Open LeeJZh opened 5 years ago
[x] 108页最下方margin"通常可以通过选择合适的卷积大小以及步长来使得(n − m + 2p)/s + 1为整数。"发生了文本重合.
[x] 117页margin文本溢出
[x] 124页最下margin文本重合
[x] 等宽卷积在本章中第一次, 第二次出现在残差网络一节122页, 第三次出现在习题中. 之前并没有对该名词做出解释, 疑似与109页等长卷积(equal-width convolution)发生了混淆.此外width应译为宽? 总之,这部分 narrow, wide, equal-width convolution的说法值得商榷. 既然其本质与padding及stride有关, 不妨从这两点出发加以诠释.
[ ] 123页5.5节其他卷积方式没有提及group convolution, depth-wise convolution及graph convolution等较为重要的卷积方式.
[ ] zero padding以外的填充方式没有涉及, 如reflection padding.
谢谢建议,部分已修正。
建议关闭issue #116
[x] 108页最下方margin"通常可以通过选择合适的卷积大小以及步长来使得(n − m + 2p)/s + 1为整数。"发生了文本重合.
[x] 117页margin文本溢出
[x] 124页最下margin文本重合
[x] 等宽卷积在本章中第一次, 第二次出现在残差网络一节122页, 第三次出现在习题中. 之前并没有对该名词做出解释, 疑似与109页等长卷积(equal-width convolution)发生了混淆.此外width应译为宽? 总之,这部分 narrow, wide, equal-width convolution的说法值得商榷. 既然其本质与padding及stride有关, 不妨从这两点出发加以诠释.
[ ] 123页5.5节其他卷积方式没有提及group convolution, depth-wise convolution及graph convolution等较为重要的卷积方式.
[ ] zero padding以外的填充方式没有涉及, 如reflection padding.