sunzet/vangogh2photo_cycle_gans

CycleGAN for Image-to-Image Translation 🐴🦄🦓

This project implements a Cycle-Consistent Generative Adversarial Network (CycleGAN) for unpaired image-to-image translation tasks. Based on the original CycleGAN paper, this implementation learns to translate images between two domains without paired training examples.

Project Overview

The implementation includes:

• Custom dataset loader for unpaired image datasets
• Generator with residual blocks using instance normalization
• PatchGAN discriminator architecture
• Cycle-consistency loss for domain translation
• Adversarial loss with both MSE and BCE implementations
• Training pipeline with visualization of results

Key Features ✨

• Flexible Architecture: Supports different normalization layers and activation functions
• Comprehensive Loss Functions: Implements cycle-consistency loss and adversarial loss
• Visualization Tools: Includes functions to visualize training progress and generated images
• Efficient Training: Batch processing and GPU acceleration support
• Model Checkpointing: Automatic saving of model states during training

Installation ⚙️

1. Clone repository:

git clone https://github.com/Ansarill/cycle_gans.git
cd cycle_gans

2. Create virtual env (Python 3.10.12):

python3.10 -m venv my_env
source my_env/bin/activate

3. Install dependencies:

pip install -r requirements.txt

Usage 🚀

1. Prepare your dataset in the following structure (or follow notebook to download one of datasets) and change variables in the notebook accordingly:

datasets/
└── img2img/
    └── your_dataset/
        ├── trainA/
        ├── trainB/
        ├── testA/
        └── testB/

2. Configure the training parameters in the Jupyter notebook

3. Run the training

License 📄

This project is licensed under the Apache License, Version 2.0 - see the LICENSE file for details.

---

Inspired by The Great Impressionist Painter Vincent van Gogh
Powered by Pure PyTorch and Enthusiasm