DCGAN MNIST Generator

This repository contains a Deep Convolutional GAN (DCGAN) trained on the MNIST dataset. The model generates handwritten-like digit images from random noise.

Model Architecture

The model implementation is based on the paper Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks.

Generator architecture: 5 transposed convolutional layers with batch normalization
Latent space dimension: 100
Output: 64x64 grayscale images

Demo App

The included Gradio app allows you to generate new MNIST-like images using the pre-trained model.

Running Locally

Install dependencies:

pip install -r requirements.txt

Run the app:

python app.py

Features

Generate multiple images at once
Set a random seed for reproducible outputs
Visualize the generated images in a grid

Training Details

This model was trained for 25 epochs on the MNIST dataset using PyTorch. For optimal results, the model checkpoint from epoch 21 is used for inference, as it produced the most realistic images without mode collapse.

Acknowledgments

Original DCGAN implementation based on PyTorch examples
Training was tracked using Weights & Biases