(CVPR 2018) MobileNetV2:Inverted Residuals and Linear Bottlenecks

Keyword [MobileNetV2]

Sandler M, Howard A, Zhu M, et al. Mobilenetv2: Inverted residuals and linear bottlenecks[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018: 4510-4520.

1. Overview

In this paper, it proposes MobileNetV2
1) Inverted Residuals (expand→depthwise→compress)
2) Linear Bottlenecks (remove last ReLu)

2. Details

2.1. Depthwise Separable Conv (DSC)

1) Conv. $k \times k \times c_{in} \times c_{out}$
2) DSC. $k \times k \times c_{in} + 1 \times 1 \times c_{in} \times c_{out}$
Almost $k^2$ times.

2.2. Inverted Residuals and Linear Bottlenecks

1) expansion rates between 5 and 10. set $t=6$ in this paper
2) remove ReLu behind final $1\times 1$ Conv

2.3. MobileNetV2

3. Comparison