A clean and readable Pytorch implementation of CycleGAN (https://arxiv.org/abs/1703.10593)

Code is intended to work with Python 3.6.x, it hasn't been tested with previous versions

Follow the instructions in pytorch.org for your current setup

To plot loss graphs and draw images in a nice web browser view

pip3 install visdom

First, you will need to download and setup a dataset. The easiest way is to use one of the already existing datasets on UC Berkeley's repository:

./download_dataset <dataset_name>

Valid <dataset_name> are: apple2orange, summer2winter_yosemite, horse2zebra, monet2photo, cezanne2photo, ukiyoe2photo, vangogh2photo, maps, cityscapes, facades, iphone2dslr_flower, ae_photos

Alternatively you can build your own dataset by setting up the following directory structure:

.
├── datasets                   
|   ├── <dataset_name>         # i.e. brucewayne2batman
|   |   ├── train              # Training
|   |   |   ├── A              # Contains domain A images (i.e. Bruce Wayne)
|   |   |   └── B              # Contains domain B images (i.e. Batman)
|   |   └── test               # Testing
|   |   |   ├── A              # Contains domain A images (i.e. Bruce Wayne)
|   |   |   └── B              # Contains domain B images (i.e. Batman)

python3 train.py --dataroot datasets/<dataset_name>/ --cuda

This command will start a training session using the images under the dataroot/train directory with the hyperparameters that showed best results according to CycleGAN authors. You are free to change those hyperparameters, see ./train --help for a description of those.

Both generators and discriminators weights will be saved under the output directory.

If you don't own a GPU remove the --cuda option, although I advise you to get one!

You can also view the training progress as well as live output images by running python3 -m visdom.server in another terminal and opening http://localhost:8097/ in your favourite web browser. This should generate training loss progress as shown below (default params, horse2zebra dataset):

The Generator loss is the sum of GAN loss, identity loss and cycle loss. The Discriminator loss is the sum of D_A and D_B losses. The GAN loss tells us how much our generator is able to fool the discriminator. The displayed curve is the sum of GAN_A2B and GAN_B2A. The Identity loss tells us how much our generator from domain A (resp. B) to domain B (resp. A) is able to preserve a domain B (resp. A) input. The displayed curve is the sum of identity A and identity B losses. The cycle loss compares input of cycle GAN from domain A (resp. B), and the output of the cycle A2B2A(A) (resp. B2A2B(B)). The displayed curve is the sum of the A2B2A and B2A2B cycle losses.

python3 test.py --dataroot datasets/<dataset_name>/ --cuda

This command will take the images under the dataroot/test directory, run them through the generators and save the output under the output/A and output/B directories. As with train, some parameters like the weights to load, can be tweaked, see ./test --help for more information.

Examples of the generated outputs (default params, horse2zebra dataset):

This project is licensed under the GPL v3 License - see the LICENSE.md file for details

Code is basically a cleaner and less obscured implementation of pytorch-CycleGAN-and-pix2pix. All credit goes to the authors of CycleGAN, Zhu, Jun-Yan and Park, Taesung and Isola, Phillip and Efros, Alexei A.