Question 1: Short answer
mediumAI-checkedGenerative Adversarial NetworksWith one-sided label smoothing, the discriminator's loss on real images uses target 0.9 instead of 1:
labels = torch.ones(batch_size) * 0.9
loss = nn.BCEWithLogitsLoss()(D_out.squeeze(), labels)For a single real image, which logit D_out minimizes this loss? Give the value
(natural log).