How to skip parameter updating

autograd
1

Hello.

I am creating an architecture which uses two networks(N1, N2).
N1 is already trained, so I want to use the output of N1 as input of N2.
However, I don’t want to update the parameters on N1 in the process.(I want to train only N2)

How could it be possible?

2

You can just give the parameters of N2 to the optimizer that you choose.

The code can look like

optimizer = torch.optim.Adam(N2.parameters(), blah blah)
input = Variable(sth)
N1_out = N1(input)
N2_out = N2(N1_out)
loss = compute_loss(N2_out)
...