CycleGAN-JAX

CycleGAN is a deep learning architecture for image-to-image translation that can be trained on unpaired datatsets. We reimplement and train it on several datasets using the JAX framework.

Curated Outputs

Installing Dependencies

TODO: script to automatically install dependencies?

conda create -n cg-jax python=3.9
conda activate cg-jax

If CUDA is available (highly recommended for training):

pip install --upgrade "jax[cuda]" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html

Otherwise:

pip install --upgrade "jax[cpu]"

Finally:

pip install flax optax matplotlib jupyter tqdm
pip install torch torchvision # For data loading

Running the Model

To train, prepare a dataset directory with subdirectories trainA, trainB, testA, testB populated accordingly, then run

python main.py --train -d <dataset dir path> -m <model checkpoints and outputs path>

To generated an image, run

python main.py --predict <A|B> -d <file path> -m <model checkpoints and outputs path>

where <A|B> is the set that the starting image belongs to.

Hyperparameters and other configuration options are gathered in train.get_default_opts.

Downloading Data

Grant execution permission to get-*.sh script, then run. For example,

chmod +x ./get-horse2zebra.sh
./get-horse2zebra.sh

The datasets we use are hosted by the original CycleGAN authors.