alirezakazemipour / cycle-gan-pytorch

PyTorch implementation of the Cycle GAN paper.

Cycle-GAN-PyTorch

PyTorch implementation of the Cycle GAN paper.

Cycle GAN: An approach for learning to translate an image from a source domain X to a target domain Y in the absence of paired examples.

• This repo covers only the Cycle GAN architecture with 9 residual blocks that is suitable for images with 256 x 256 size.

Demo

Original image	Fake image	Reconstructed image

Dataset tested

• Horse to Zebra & Vice Versa
• Apple to Orange & Vice Versa

Dependencies

• imageio == 2.9.0
• numpy == 1.19.2
• opencv_contrib_python == 4.4.0.44
• torch == 1.6.0
• torchsummary == 1.5.1
• tqdm == 4.50.0

Installation

pip3 install -r requirements.txt

Usage

The training requires a suitable (not necessarily a 1080 Ti or a 2080 RTX Nvidia gpu 😁) gpu-enabled machine. Google Colab provides what is enough to train such an algorithm but if you a more powerful free online gpu provider, take a look at: paperspace.com.

• To run the code:

python3 main.py

• **If the weights with name CycleGan.pth is available in the directory, then automatically the training procedure would be continued otherwise, it starts from scratch. **.
• If you want to test the algorithm, turn TRAIN_FLAG flag to False. Modify the name of your testing instances accordingly in the main.py.

Reference

1. Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks, Zhu et al., 2017

Acknowledgement

Most of the materials is inspired by the original implmentation and the TensorFlow's great tutorial especially about pre-processing images.