Udacity Deep Learning Nanodegree (nd101) — Course 5: Building Generative Adversarial Networks, final project.

Train a Deep Convolutional GAN from scratch on CelebA faces (64×64).

Standard DCGAN (Radford et al. 2015):

• Generator: z(100) → 4×4×512 → 8×8 → 16×16 → 32×32 → 64×64×3 (Tanh)
• Discriminator: 64×64×3 → 32×32 → 16×16 → 8×8 → 4×4 → logit (LeakyReLU + BatchNorm)
• Loss: BCE with logits
• Optimizer: Adam(lr=2e-4, β₁=0.5)
• Init: Normal(0, 0.02)

Generated faces are recognizable (hair, skin tone, basic facial structure) — more epochs on the full 162K dataset sharpens features further.

• Python 3.14 / PyTorch 2.10 / MPS backend
• HuggingFace datasets for CelebA loading (tpremoli/CelebA-attrs)

pip install torch torchvision datasets matplotlib
BATCH=64 EPOCHS=3 SUBSET=5000 python run_gan.py

• project.ipynb / project.html — submission notebook with loaded generator + sample images
• run_gan.py — full training script
• samples/samples_ep0N.png — face grids per epoch
• loss_curve.png — G vs D loss over training
• checkpoints/ — generator.pt + discriminator.pt (ignored; regenerate with run_gan.py)

Part of Leo's Udacity Masters in AI journey. See also:

• landmark-classifier — Course 3 CNN
• text-translation-sentiment — Course 4 Transformers