1. 정리

요약

• SE (Squeeze and Excitation)
  • input : [H, W, F]
  • squeezed
    • pooling : [1, 1, F]
    • fc, relu : [1, 1, 1/4F]
    • fc, sigmoid : [1, 1, F]
  • output : [H, W, F]
    • input * squeezed
• (b) : MBConv3 (k5x5)
  • depth를 3배 expand, dwconv 적용, 원래 사이즈로 축소.
  • process
    • Conv1x1, BN, Relu : F -> 3F
    • DWConv5x5, BN, Relu : 3F -> 3F
    • SE
    • Conv1x1, BN, Relu : 3F -> F
    • Residual
• (c) : MBConv6 (k3x3)
• (d) : SepConv (k3x3)
  • dwconv3x3, bn, relu
  • conv1x1, bn

질문

• mobile inverted bottleneck에서 relu를 생략하는 이유?
• sepconv 에서 relu를 생략하는 이유
• SENet 리뷰

내용

2. Related Work

• 모바일용 cnn 구조
  • squeezenet
    • using lower cost (1x1)-convolutions
    • reducing filter size
  • mobilenet v2
    • inverted residual
    • linear bottleneck
  • shufflenet
    • group convolution
    • channel shuffle
  • densenet
    • learns to connect group convolutions across layer