This is an unoffical and simple implementation of NIPS2019 paper "Cross-channel Communication Networks"
If I'm right, it seems cross-channle module is another version of Squeeze-and-Excitation (SE) block, because the attention is also a single scalar.
I hope I'm wrong about this.
The office implementation can be find at https://github.com/jwyang/C3Net.pytorch