Unpaired image-to-image translation

A fastai/PyTorch package for unpaired image-to-image translation currently with CycleGAN implementation.

This is a package for training and testing unpaired image-to-image translation models. It currently only includes the CycleGAN, DualGAN, and GANILLA models, but other models will be implemented in the future.

This package uses fastai to accelerate deep learning experimentation. Additionally, nbdev was used to develop the package and produce documentation based on a series of notebooks.

Install

To install, use pip:

pip install git+https://github.com/tmabraham/UPIT.git

The package uses torch 1.7.1, torchvision 0.8.2, and fastai 2.3.0 (and its dependencies). It also requires nbdev 1.1.13 if you would like to add features to the package. Finally, for creating a web app model interface, gradio 1.1.6 is used.

How to use

Training a CycleGAN model is easy with UPIT! Given the paths of the images from the two domains trainA_path and trainB_path, you can do the following:

from upit.data.unpaired import *
from upit.models.cyclegan import *
from upit.train.cyclegan import *

dls = get_dls(trainA_path, trainB_path)
cycle_gan = CycleGAN(3,3,64)
learn = cycle_learner(dls, cycle_gan,opt_func=partial(Adam,mom=0.5,sqr_mom=0.999))
learn.fit_flat_lin(100,100,2e-4)

The GANILLA model is only a different generator model architecture (that’s meant to strike a better balance between style and content), so the same cycle_learner class can be used.

from upit.models.ganilla import *

ganilla = GANILLA(3,3,64)
learn = cycle_learner(dls, ganilla,opt_func=partial(Adam,mom=0.5,sqr_mom=0.999))
learn.fit_flat_lin(100,100,2e-4)

Finally, we provide separate functions/classes for DualGAN model and training:

from upit.models.dualgan import *
from upit.train.dualgan import *

dual_gan = DualGAN(3,64,3)
learn = dual_learner(dls, dual_gan, opt_func=RMSProp)
learn.fit_flat_lin(100,100,2e-4)

Additionally, we provide metrics for quantitative evaluation of the models, as well as experiment tracking with Weights and Biases. Check the documentation for more information!

Citing UPIT

If you use UPIT in your research please use the following BibTeX entry:

@Misc{UPIT,
    author =       {Tanishq Mathew Abraham},
    title =        {UPIT - A fastai/PyTorch package for unpaired image-to-image translation.},
    howpublished = {Github},
    year =         {2021},
    url =          {https://github.com/tmabraham/UPIT}
}