#SANet_CROWD_COUNTINg
implement by using pytorch
And the result are virtually same with my tensorflow implementation BIGKnight/SANet_implementation.
However, this version are more close to the paper's net structure.(except I put the BN or IN behind the relu instead of before it. and it seemed in fact the BN layer did not ameliorate the net ability)
I still can not reach the result as the paper said(in fact, far worse than that). To be honest, I somewhat doubt the results showed in the paper.