GANexLib

A Python library for expanding image datasets using Deep Convolutional Generative Adversarial Networks (DCGAN).

Overview

GANexLib provides an easy-to-use interface for generating synthetic images to augment your existing image datasets.

Installation

pip install GANexLib

Requirements

• Python >= 3.8
• TensorFlow >= 2.8.0
• NumPy >= 1.19.0
• Pillow >= 8.0.0
• tqdm >= 4.60.0
• joblib >= 1.0.0

Quick Start

from GANexLib import expand_dataset

# Generate 100 new images based on images in the specified directory
expand_dataset("/path/to/your/images", 100)

Usage

Basic Usage

from GANexLib import expand_dataset

# Generate new images
expand_dataset(
    absolute_path="/path/to/your/images",
    image_amount=100
)

Advanced Usage

from GANexLib import expand_dataset

# Custom training parameters
expand_dataset(
    absolute_path="/path/to/your/images",
    image_amount=100,
    epochs=30,              # Number of training epochs (default: 20)
    batch_size=64,          # Batch size for training (default: 32)
    callback=lambda msg: print(f"Training complete: {msg}")
)

Parameters

• absolute_path (str, required): Absolute path to directory containing your training images
• image_amount (int, required): Number of new images to generate
• callback (callable, optional): Function to call when training completes
• epochs (int, optional): Number of training epochs (default: 20)
• batch_size (int, optional): Batch size for training (default: 32)

Supported Image Formats

• JPEG (.jpg, .jpeg)
• PNG (.png)
• BMP (.bmp)
• WebP (.webp)

How It Works

1. Data Loading: Loads all valid images from the specified directory
2. Preprocessing: Resizes images to 64x64 and normalizes pixel values
3. Training: Trains a DCGAN model on your dataset
4. Generation: Creates new synthetic images that match your data distribution
5. Saving: Saves generated images as generated_img_0.png, generated_img_1.png, etc.

Example

# Example: Expand a dataset of cat images
from GANexLib import expand_dataset

def training_complete(message):
    print(f"✓ {message}")
    print("Check your image folder for new generated images!")

expand_dataset(
    absolute_path="/datasets/cats",
    image_amount=200,
    epochs=25,
    callback=training_complete
)

Tips for Best Results

1. Dataset Size: Use at least 1000 images for better quality results
2. Image Consistency: Training works best with similar-styled images
3. Training Time: More epochs generally produce better results (try 30-50)
4. GPU: Training is much faster with GPU support (CUDA-enabled TensorFlow)

Limitations

• Generated images are fixed at 64x64 resolution
• Requires TensorFlow installation (can be large)
• Training time depends on dataset size and hardware

Citation

If you use GANexLib in your research, please cite the original DCGAN paper:

Radford, A., Metz, L., & Chintala, S. (2015). 
Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks. 
arXiv preprint arXiv:1511.06434.

Acknowledgments

This library implements the DCGAN architecture as described in the 2014 paper by Radford et al.