This repository implements a Deep Convolutional Generative Adversarial Network (DCGAN) to generate realistic handwritten digit images based on the MNIST dataset. The project showcases the power of GANs in image synthesis through a step-by-step pipeline, resulting in high-quality digit generation.
A Generative Adversarial Network (GAN) consists of two neural networks:
- Generator: Creates fake images from random noise.
- Discriminator: Distinguishes between real images and generated images.
This project uses a Deep Convolutional GAN (DCGAN) architecture, optimized for generating 28x28 grayscale handwritten digits inspired by the MNIST dataset. The results are saved and visualized, including an animated GIF of generated images during training.
- Generates realistic handwritten digit images.
- Uses the MNIST dataset for training.
- Includes a training workflow with checkpointing.
- Visualizes training progress through an animated GIF.
-
Clone the repository:
git clone https://github.com/yourusername/DCGAN-MNIST-Digit-Image-Generator.git cd DCGAN-MNIST-Digit-Image-Generator -
Install required dependencies:
pip install -r requirements.txt
-
Run the script to start training:
python generate_hand_written_digit_images_dcgan.py
- Data Preparation: The MNIST dataset is loaded and preprocessed.
- Model Creation:
- Generator creates images from random noise using transposed convolutions.
- Discriminator evaluates whether an image is real or fake.
- Loss Definition:
- Discriminator loss ensures the model distinguishes real from fake images.
- Generator loss ensures generated images can fool the discriminator.
- Training Loop:
- Both models are trained iteratively using a defined training loop.
- Checkpoints are saved periodically.
- Visualization: An animated GIF,
dcgan.gif, showcases training progress.
This GIF demonstrates the improvement in the quality of generated digit images over 100 training epochs.
- Modify the training parameters such as epochs, noise dimension, or batch size in the script for experimentation.
- Use the saved checkpoints to restore and generate images without retraining.
Contributions are welcome! If you have ideas to improve the architecture, code efficiency, or visualization, feel free to create a pull request.