Now pass some random images to train discriminator.

Let's look at the code for it and then explore the important features:

fake = netG(noisev)
output = netD(fake.detach())
errD_fake = criterion(output, labelv)
errD_fake.backward()
optimizerD.step()

The first line in this code passes a vector with a size of 100, and the generator network (netG) generates an image. We pass on the image to the discriminator for it to identify whether the image is real or fake. We do not want the generator to get trained, as the discriminator is getting trained. So, we remove the fake image from its graph by calling the detach method on its variable. Once all the gradients are calculated, we call the optimizer to train the discriminator.