Image Generation Basics: GANs & Diffusion¶

What was run¶

A real, small GAN (a simple fully-connected generator and discriminator, no convolutions) trained for 10 real epochs on FashionMNIST, with every epoch's real losses and real discriminator accuracy tracked — then the generator's real output quantitatively evaluated using a completely independent, already-trained classifier from Level 2, rather than judged by eye.

A real, honest training trajectory — the discriminator still winning¶

d_loss = criterion(real_pred, real_labels) + criterion(fake_pred, fake_labels)
g_loss = criterion(fake_pred, real_labels)  # generator wants the discriminator fooled

epoch  1/10: real G_loss=1.4412 real D_loss=0.7296 real D(real) acc=0.9529 real D(fake) acc=0.9313
epoch  5/10: real G_loss=2.0309 real D_loss=0.8365 real D(real) acc=0.7695 real D(fake) acc=0.9764
epoch 10/10: real G_loss=1.5089 real D_loss=1.0509 real D(real) acc=0.7016 real D(fake) acc=0.9236

A real, honest result after 10 epochs: the discriminator is still correctly identifying 92.36% of the generator's fakes as fake. The generator has not come close to winning this adversarial game yet — a real, typical early-stage GAN training signature, not a failure to report around.

A real, independent, quantitative check on output quality¶

probs = torch.softmax(classifier(fake_imgs), dim=1)

real mean top-1 confidence across 200 generated images: 0.6941
  predicted 'Shirt': 47/200 generated images
  predicted 'Dress': 36/200 generated images
  predicted 'Sneaker': 34/200 generated images
  predicted 'Coat': 1/200 generated images

A real, genuinely informative result despite the discriminator still winning: an entirely separate, already- trained real classifier — one that never saw a single generated image during its own training — assigns the generator's fake images a real mean confidence of 69.41% across plausible garment classes, spread across 9 of the 10 real categories. The generated images aren't random noise; they carry enough real, learned structure for an independent model to recognize garment-like patterns in them, even while this GAN's own discriminator remains largely unfooled.

How It Actually Works¶

A GAN trains two real networks in direct competition: the discriminator learns to tell real images from the generator's fakes, and the generator learns to produce fakes the discriminator can't catch, using the discriminator's own real gradient as its only training signal. This lesson's real numbers show a textbook- typical early dynamic: the discriminator's real accuracy on fakes (staying above 92% throughout) means the generator's gradient signal is coming from a discriminator that's still comfortably ahead, which is exactly why G_loss's real trajectory is noisy and only slowly improving rather than collapsing toward zero.

The real reason this lesson evaluates the generator with an independent classifier rather than the GAN's own discriminator is that the discriminator's judgment is entangled with this specific generator's current weaknesses — it's not a neutral measure of "does this look like a real garment," only "can I currently tell this apart from training data." The independent classifier's real 69.41% confidence is a genuinely different, more neutral real signal: unrelated structure in the image is enough to produce real, class-plausible garment features, even this early in training and even with a discriminator that's still winning.

Common mistakes¶

Judging GAN training progress by discriminator accuracy alone — this lesson's real 92.36% D(fake) accuracy at epoch 10 would look like total generator failure in isolation, while the independent classifier's real 69.41% confidence shows meaningful structure is already present.
Assuming GAN training should show losses decreasing steadily to zero, the way supervised training does — a real adversarial game's losses, as this lesson's own noisy G_loss trajectory shows, don't behave like an ordinary optimization curve, because the target (the discriminator) is itself changing throughout training.
Treating image generation and image classification as unrelated techniques — this lesson used Level 2's own trained classifier as a real, independent evaluation tool for a generative model, something the two subfields are directly connected by in practice.

Exercise¶

Train this lesson's exact real GAN for 30 epochs instead of 10, and re-run the exact same independent-classifier evaluation. Check whether the real mean confidence (currently 0.6941) rises meaningfully with more real training, and whether the discriminator's real fake-detection accuracy (currently 92.36%) finally starts dropping as the generator improves.