Vanilla GAN

Vanilla GAN is a PyTorch implementation of a Generative Adversarial Network trained on the MNIST digit dataset. The project provides lightweight scripts for data preparation and model training along with a configuration file for customizing hyper‑parameters.

Features

• Configurable Training – main parameters such as number of epochs, batch size and learning rate are defined in config.yaml:

  NUM_EPOCHS: 50
  NOISE_DIMENSION: 50
  BATCH_SIZE: 128
  TRAIN_ON_GPU: True
  PRINT_STATS_AFTER_BATCH: 50
  OPTIMIZER_LR: 0.0002
  OPTIMIZER_BETAS: [0.5, 0.999]
  GENERATOR_OUTPUT_IMAGE_SHAPE: 784

• Automatic Dataset Download – the prepare_dataset function downloads MNIST and returns a DataLoader with normalized tensors.
• Simple Network Architecture – generator and discriminator are built from fully connected layers.
• Training Utilities – images and model checkpoints are saved for each run under data/runs/UNIQUE_RUN_ID.
• Developer Tools & CI – linting with ruff and black, end‐to‐end smoke test in GitHub Actions

This repository is the PyTorch implementation of my blog post on the original GAN paper. The Easiest Guide to Understanding the Original GAN Paper, Actually!

Installation

1. Install the core dependencies:

   pip install -r requirements.txt

2. (Optional) install development tools and pre‑commit hooks:

   pip install -r requirements-dev.txt
   pre-commit install
   pre-commit run --all-files

3. Install the package in editable mode:

   pip install -e .

Training

To start training with the default configuration run:

bash scripts/run_train.sh

This script calls python -m vanilla_gan.train which handles dataset preparation, model initialization and epoch loops. Hyper‑parameters may be adjusted by editing config.yaml.

Model Progress

Epoch0 Batch0

Epoch0 Batch200

Epoch1 Batch100

Epoch2 Batch350

Epoch4 Batch200

Epoch13 Batch450

Epoch26 Batch150

Repository Structure

• src/vanilla_gan/ – core modules with data loading, model definitions and training logic.
• data/ – where datasets and runs would be generated
• scripts/ – helper script for launching training.
• notebooks/ – Jupyter notebook used for early experiments.

Requirements

• Python >= 3.10
• PyTorch
• torchvision
• matplotlib

Contributing

Contributions are welcome! Feel free to open issues or submit pull requests.

Acknowledgements

• Building a Vanilla GAN with Pytorch

License

This project is released under the MIT License.