Group Convolution

Inception Module

  • 多尺度特征提取再融合(来自Inception-V1)
    Inception Module Naive
    Inception Module

Residual Block

  • 残差结构(来自ResNet,先降维再升维)
    Residual Block

Element-wise Addition

  • 更激进的密集连接机制(来自ResNet)
    ResNet网络的短路连接机制(其中+代表的是元素级相加操作)

BottleNeck

  • 瓶颈结构(来自ResNet)
    BottleNeck

Channel-wise Concatenation

  • 实现特征重用,提升效率(来自DenseNet)
    DenseNet网络的密集连接机制

Inverted Residual Block

  • 倒残差结构(来自MobileNet-V2,先升维再降维)
    Input Operator Output
    H×W×tK 1×1 conv2d, ReLU6 H×W×tK
    H×W×tK 3×3 dwise s=s, ReLU6 H/s×W/s×tK
    H/s×W/s×tK linear 1×1 conv2d H/s×W/s×K’

Inverted Residual Block

Squeeze-and-Excitation-Block

  • 压缩-扩张模块(来自SeNet)
    Squeeze-and-Excitation-Block

未完…