Info
- Title: Xception: Deep Learning with Depthwise Seperable Convolutions
- Author: F. Chollet
- Arxiv: 1610.02357
- Date: Oct. 2016
Highlights & Drawbacks
Replaced 1×1 convolution and 3×3 convolution in Inception unit with Depth-wise seperable convolution
Motivation & Design
The article points out that the assumption behind the Inception unit is that the correlation between the channel and the space can be fully decoupled, similarly the convolution structure in the length and height directions (the 3 × 3 convolution in Inception-v3 is 1 × 3 and 3 × 1 convolution replacement).
Further, Xception is based on a stronger assumption: the correlation between channels and cross-space is completely decoupled. This is also the concept modeled by Depthwise Separable Convolution. A simple Inception Module:
is equal to:
Push # of channel to extreme, we obtain Depthwise Separable Convolution:
NetScope Visualization and source code: awesome_cnn.
