What do we need to do this operation for computing image embedding in AC?

[raymondchua]

Hi,
I have some trouble trying to understand why we need to do this in the ACModel:

rl-starter-files/model.py

Line 38 in 317da04

self.image_embedding_size = ((n-1)//2-2)*((m-1)//2-2)*64

Can someone explain? I understand that 64 could be the number of output channel from the previous layer. But why do we need to do ((n-1)//2-2)?

Best,
Raymond