CNN Cases

CNN 2017.11

1. VGG
   
   • VGG-16的参数个数
     $maxpool: \ 2 \times 2 \ pixel \ window, \ stride =2$
     $conv<receptive \ field \ size>-<number \ of \ channels>$
     $input =224 \times 224 \times 3$
     ReLU activation function is not shown for brevity.
     $$\begin{align*} conv3-64 &= (3 \times 3 \times 3) \times 64 \\ conv3-64 &= (3 \times 3 \times 64) \times 64 \\ maxpool \\ conv3-128 &= (3 \times 3 \times 64) \times 128 \\ conv3-128 &= (3 \times 3 \times 128) \times 128 \\ maxpool \\ conv3-256 &= (3 \times 3 \times 128) \times 256 \\ conv3-256 &= (3 \times 3 \times 256) \times 256 \\ conv3-256 &= (3 \times 3 \times 256) \times 256 \\ maxpool \\ conv3-512 &= (3 \times 3 \times 256) \times 512 \\ conv3-512 &= (3 \times 3 \times 512) \times 512 \\ conv3-512 &= (3 \times 3 \times 512) \times 512 \\ maxpool \\ conv3-512 &= (3 \times 3 \times 512) \times 512 \\ conv3-512 &= (3 \times 3 \times 512) \times 512 \\ conv3-512 &= (3 \times 3 \times 512) \times 512 \\ maxpool \\ there, \ the \ feature \ map \ is \ 7*7*512 \\ FC-4096 \\ FC-4096 \\ FC-1000 \\ softmax \end{align*}$$
     $$\begin{align*} the \ parameters &= \\ 7*7*512*4096+4096*4096+ 4096*1000 \\ 3*3*3*64 + 3*3*64*64*3 + 3*3*128*128*3 \\ +3*3*256*256*4 + 3*3*512*512*5 \\ &= 123633664 + 14710464 \\ &= 138344128 \end{align*}$$
     最后的结果为138millions，和论文中给出的数据相当。
     在conv layer不改变feature map的情况下,所有的maxpool layer(窗口大小和步长如上所述)总共会将图像宽高缩减到原来的$1/2^5$。