MobileNetV2: Inverted Residuals and Linear Bottlenecks (MobileNets v2)

April 2019

tl;dr: Added linear bottleneck layer to MobileNets module. This adds to the computation per module but requires fewer layers to achieve the same performance, thus saving computation.

Overall impression

The authors provided some intuition behind MobileNetsV2, but the paper is really hard to comprehend. Good thing is that the implementation of MobileNetsV2 builds upon and requires only slight change from MobileNetsV1.

Key ideas

Technical details

V1 alt text

V2 alt text

<img src=https://machinethink.net/images/mobilenet-v2/ExpandProject@2x.png alt=”alt text” height=”100”>