GAN for MNIST Synthesis
Generative Adversarial Network trained on MNIST and deployed as a live demo on HuggingFace Spaces.
PythonPyTorchGradioHuggingFace SpacesDocker
What This Is
A GAN that generates handwritten digits, trained on MNIST and deployed as a live interactive demo on HuggingFace Spaces. Training the model was the straightforward part. Building the pipeline around it — Gradio app, Docker container, CI/CD to HuggingFace — was where most of the work went.
How It Works
- Generator (1.49M params) and Discriminator (1.46M params) trained on MNIST with LeakyReLU and Batch Normalization for stable training
- Gradio web app with structured logging, type hints, and input validation
- Automatic CUDA, MPS, or CPU selection, deterministic seeding, and input validation
- Error images for failed requests and a headless Matplotlib backend for server use
- Deployed to HuggingFace Spaces with CI/CD integration
Results
- Live demo running on HuggingFace Spaces where anyone can generate digits
- Containerized with Docker for portable deployment
- Trained for 200 epochs, reaching about 0.981 generator loss and 1.213 discriminator loss
- Exported weights of about 17 MB, with the training notebook and setup instructions in the repository