The PSA module that is defined here is only used in the basic block of HRNet and not the bottleneck block. The paper says this though:
For any baseline networks with the bottleneck or basic residual blocks, such as ResNet and HRnet, we add PSAs after the first 3x3 convolution in every residual blocks, respectively.
The PSA module that is defined here is only used in the basic block of HRNet and not the bottleneck block. The paper says this though: