(torch)如何构造一种forward 和backward 不同的layer

在BNN中，往往训练阶段较为复杂，而inference phase is simple.

So how can we build a layer with complex training progress and simple inference algorithm?

n1 = nn.Linear(4,5)
n1.train()
print(n1.__dict__)
n1.eval()
print(n1.__dict__)

　　

The different phase has its corresponding __dict__