I wanna to recurrent the model of this paper.
And when I train Discriminator,I found that the acc of Feature extractor and Classifier decline because the batchnorm’s parameters changed,but I had set the .requires_grad of them to be False.The optimizer just change the parameters of D.
I found if I use model.eval() to close
BatchNorm2d of Feature extractor and Classifier when I train Discriminator ,the acc does not decline .What’s the mistake of my code,or what’s the mistake of my understand of batchnorm?
d_optimizer = torch.optim.SGD(D.parameters(), lr=LR,momentum=0.9,weight_decay=0.0005)
` #######for discriminator###################
D.zero_grad()
pr_fm=CNN(pr)
p_label=torch.full((batch_size,), 1).to(device).float()
pr_d=D(pr_fm.detach())
pr_d=pr_d.view(pr_d.size(0)).float()
D_prloss = D_loss_func(pr_d, p_label)
D_prloss.backward()
D_PR_1 = pr_d.mean().item()
hw_fm=CNN(hw)
h_label=torch.full((batch_size,), 0).to(device).float()
hw_d=D(hw_fm.detach())
hw_d=hw_d.view(hw_d.size(0)).float()
D_hwloss = D_loss_func(hw_d, h_label)
D_hwloss.backward()
D_HW_1 = hw_d.mean().item()
D_loss=D_prloss+D_hwloss
d_optimizer.step()`