TinyGAN

Tiny, readable PyTorch Lightning GAN template — DCGAN-style generator + spectral-norm discriminator. A compact starter to design and scale up GAN models.

What’s inside

• tiny_celeba_gan_spectral_norm.py — single-file training script
• Checkpointing for generator & critic (state_dict()), optional optimizer files
• Hinge loss, manual optimization, TensorBoard logging, saves on epoch end + SIGINT

Demo

Training example: 100_epochs_training_results.gif (included).

Quick start

1. Download CelebA-HQ and put it at the repo root as celeba_hq (Kaggle): https://www.kaggle.com/lamsimon/celebahq
2. Run:

python tiny_celeba_gan_spectral_norm.py --image_size 64 --batch_size 128 --max_epochs 50 --save_dir ./outputs --tb_logdir ./logs

3. View samples in ./outputs and open TensorBoard:

tensorboard --logdir logs

Notes

• Default checkpoint paths: drive/MyDrive/model/*.pth — override with --gen_checkpoint / --crit_checkpoint.
• Use --load_optimizers / --save_optimizers to restore/save optimizer state dicts (separate files).
• Script attempts to load the generator checkpoint at startup; edit if you prefer to start fresh.

License

MIT — free to use and extend.