Paper Reading:MobileNetV2: Inverted Residuals and Linear Bottlenecks

MobileNetV2: Inverted Residuals and Linear Bottlenecks(1801.04381)

Motivation

相比MobileNet v1性能更好的轻量化模型,做的改进。

Architecture

  • Depthwise Separable Convolutions

    和MobileNet v1一样

  • Linear Bottlenecks

    这里通过有效的实验来证明,在特征信息更集中的缩减后的通道中,加上一个非线性**层,比如ReLU,就会有较大的信息丢失 。

Paper Reading:MobileNetV2: Inverted Residuals and Linear Bottlenecks

所以这里使用了一种新的结构,Pointwise对通道数进行增加+Relu+Depthwise+Relu+Pointwise对通道数减少,但是后面不加非线性层。

Paper Reading:MobileNetV2: Inverted Residuals and Linear Bottlenecks

Paper Reading:MobileNetV2: Inverted Residuals and Linear Bottlenecks

  • Inverted residuals

    实验证明,这样做的好处是可以减少内存的使用。

Paper Reading:MobileNetV2: Inverted Residuals and Linear Bottlenecks

Experiments

  • Classification
    Paper Reading:MobileNetV2: Inverted Residuals and Linear Bottlenecks

  • Detection

Paper Reading:MobileNetV2: Inverted Residuals and Linear Bottlenecks