More detail about model structure?

[vietnamican]

Hello, I'm reimplement HAttMatting. But I confuse about Pyramidal Features Distillation module. I have some question:

1. Is Maxpool layer global max pooling layer? And which dimension does it get along with? Because in Fig.2, the channel dimension of input and channel dimention of max pooling is not equal (max pooling layer is so much 'thick').
2. What's the role of shared MLP? Is it calculate attention between channels?
3. Can you give me the shape of layers in this module?

Thanks