Double-Attention-Network

This is the PyTorch implementation of A^2-Nets: Double Attention Networks, Y Chen et al NIPS 2018

It can be used as an additional block for building models. Right now, the output tensor has shape (B, c_n, H, W). One can re-construct the original shape (B, c, H, W) with a single line of code in PyTorch.

Layer architecture

• Two attention steps