DL-4 CNN basic

Convolution layer 계산

• output size $(N-F+2P) / S + 1$

• number of parameters $K(F^2N + 1)$

N : input D

F : filter D

P : padding

S : stride

K : Number of filters

Pooling 의 경우

• output size $(N-F) /S+ 1$

• number of parameters $0$ (no learning happens)

1x1 convolution 왜 함?

• dimension reduction
• 깊이는 깊어지면서도 reduce number of parameters
  • bottleneck architecture