This repository contains a Tensorflow implementation and demonstration of CycleGAN technique. Note that this project is still work in progress and requires some polishing.
The technique performs unpaired image to image translation using conditional GAN's. It can be trained to translate instances from one domain to another without a one-to-one mapping between the source and target domain.
CycleGAN can be used to solve many interesting problems such as photo-enhancement, image colorization, style transfer, etc. The capabilities are demonstrated in two tasks:
- translation of handwritten digits
- style transfer based on paintings by Leonid Afremov
TODO
- Linux, macOS or Windows
- Python 3 (tested version 3.6)
- CPU or NVIDIA GPU + CUDA CuDNN
- Clone this repository
- Setup Python environment (preferably version 3.6) with pip command available for this project
- Check requirements.txt with the list of all dependencies. Select appropriate Tensorflow dependency given your configuration.
- Install all dependencies by executing:
pip install -r requirements.txtinto the Python environment.
- Start a Jupyter server within your Python environment by executing a command:
jupyter notebook. Make sure that this command is executed relative to the project directory, so the project's script can be imported with no changes to the code. - Check the log of a Jupyter server in the console. There should be a server's address. Use it to interact with the server.
- Open digit_transformation or style_transfer notebook on the server and run all cells. The example should run as it is (it will automatically download the data set for you)
TODO
- https://medium.com/analytics-vidhya/transforming-the-world-into-paintings-with-cyclegan-6748c0b85632
- https://machinelearningmastery.com/how-to-develop-cyclegan-models-from-scratch-with-keras/
- https://medium.com/datadriveninvestor/style-transferring-of-image-using-cyclegan-3cc7aff4fe61
- https://www.tensorflow.org/tutorials/generative/cyclegan
- https://arxiv.org/pdf/1703.10593.pdf
- https://towardsdatascience.com/style-transfer-with-gans-on-hd-images-88e8efcf3716
- https://www.tensorflow.org/tutorials/generative/pix2pix
- https://gluon.mxnet.io/chapter14_generative-adversarial-networks/pixel2pixel.html
- https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix