MobileNetV2:Inverted Residuals and Linear Bottlenecks

論文閱讀筆記

論文連結(CVPR2018)https://arxiv.org/abs/1801.04381

For example, the depthwise layer may work on a tensor with 144 channels, which the projection layer will then shrink down to only 24 channels. This kind of layer is also called a bottleneck layer because it reduces the amount of data that flows through the network. (This is where the “bottleneck residual block” gets its name from: the output of each block is a bottleneck.)

So the input and the output of the block are low-dimensional tensors, while the filtering step that happens inside block is done on a high-dimensional tensor.

Reference

1. https://machinethink.net/blog/googles-mobile-net-architecture-on-iphone/
2. http://machinethink.net/blog/mobilenet-v2/