Attention apply personal understanding to images

The attention mechanism should not be directly in the original backbone network （backbone） Add , So as not to destroy the original pre training parameters , It is best to backbone Then add .
When adding , Pay attention to the input channels The number of .