GAN does not learn when using symmetric outputs from generator to disciminator

Question

I'm currently trying to implement the paper Generative modeling for protein structures and I have succesfully been able to train a model following Pytorch's DCGAN Tutorial which has a similar model structure to the paper. The two implementations differ when it comes to output of the generator.

In the tutorial's model, the generator simply passes a normal output matrix to the discriminator. This works fine when I implement the paper's model (ommiting the symmetry and clamping) but the paper specifies:

During training, we enforce that G(z) be positive by clamping output values above zero and symmetric

when I put this into my training loop I receive a loss graph that indicates that the generator isn't learning.

Here is my training loop:

# Training Loop
# Lists to keep track of progress
img_list = []
G_losses = []
D_losses = []
iters = 0

print("Starting Training Loop...")
# For each epoch
for epoch in range(num_epochs):
    # For each batch in the dataloader
    for i, data in enumerate(dataloader, 0):
        ############################
        # (1) Update D network: maximize log(D(x)) + log(1 - D(G(z)))
        ###########################
        ## Train with all-real batch
        netD.zero_grad()
        # Format batch
        # Unsqueezed dim one to convert [128, 64, 64] to [128, 1, 64, 64] to conform to D architecture 
        real_cpu = (data.unsqueeze(dim=1).type(torch.FloatTensor)).to(device)
        b_size = real_cpu.size(0)
        label = torch.full((b_size,), real_label, device=device)
        # Forward pass real batch through D
        output = netD(real_cpu).view(-1)
        # Calculate loss on all-real batch
        errD_real = criterion(output, label)
        # Calculate gradients for D in backward pass
        errD_real.backward()
        D_x = output.mean().item()

        ## Train with all-fake batch
        # Generate batch of latent vectors
        noise = torch.randn(b_size, nz, 1, 1, device=device)
        # Generate fake image batch with G
        fake = netG(noise)
        label.fill_(fake_label)
        # Make Symmetric
        sym_fake = (fake.detach().clamp(min=0) + fake.detach().clamp(min=0).permute(0, 1, 3, 2)) / 2
        # Classify all fake batch with D
        output = netD(sym_fake).view(-1)
        # Calculate D's loss on the all-fake batch
        errD_fake = criterion(output, label)
        # Calculate the gradients for this batch
        errD_fake.backward()
        D_G_z1 = output.mean().item()
        # Add the gradients from the all-real and all-fake batches
        errD = errD_real + errD_fake
        # Update D
        optimizerD.step()
        #adjust_optim(optimizerD, iters)
        ############################
        # (2) Update G network: maximize log(D(G(z)))
        ###########################
        netG.zero_grad()
        label.fill_(real_label)  # fake labels are real for generator cost
        # Since we just updated D, perform another forward pass of all-fake batch through D
        output = netD(fake.detach()).view(-1)
        # Calculate G's loss based on this output
        errG = criterion(output, label)
        # Calculate gradients for G
        errG.backward()
        D_G_z2 = output.mean().item()
        # Update G
        optimizerG.step()
        adjust_optim(optimizerG, iters)

        # Output training stats
        if i % 50 == 0:
            print('[%d/%d][%d/%d]\tLoss_D: %.4f\tLoss_G: %.4f\tD(x): %.4f\tD(G(z)): %.4f / %.4f'
                  % (epoch, num_epochs, i, len(dataloader),
                     errD.item(), errG.item(), D_x, D_G_z1, D_G_z2))

        # Save Losses for plotting later
        G_losses.append(errG.item())
        D_losses.append(errD.item())

        # Check how the generator is doing by saving G's output on fixed_noise
        if (iters % 500 == 0) or ((epoch == num_epochs-1) and (i == len(dataloader)-1)):
            with torch.no_grad():
                fake = netG(fixed_noise).detach().cpu()
            img_list.append(vutils.make_grid(fake, padding=2, normalize=True))

        iters += 1

Here is the training loss.

Here is my expected loss.

I make the output symmetric with the following line

sym_fake = (fake.detach().clamp(min=0) + fake.detach().clamp(min=0).permute(0, 1, 3, 2)) / 2

and I pass it to the discriminator on the lines that call sym_fake

Question

Is my implementation in pytorch wrong or is there something I'm missing? I don't understand why the paper makes the matrix symmetric and clamps if the network is capable of generating images without the need for symmetry and clamping.

Answer 1

It could be because after the criterion for netG obtains an output that was detached from the parameters of netG thus the optimizer is not / can't be updating the parameters for netG.