A library to easily train various existing GANs (Generative Adversarial Networks) in PyTorch.
This library targets mainly GAN users, who want to use existing GAN training techniques with their own generators/discriminators. However researchers may also find the GAN base class useful for quicker implementation of new GAN training techniques.
The focus is on simplicity and providing reasonable defaults.
You need python 3.5 or above. Then:
pip install vegans
The basic idea is that the user provides discriminator and generator networks, and the library takes care of training them in a selected GAN setting:
from vegans import WGAN
from vegans.utils import plot_losses, plot_image_samples
netD = ### Your discriminator/critic (torch.nn.Module)
netG = ### Your generator (torch.nn.Module)
dataloader = ### Your dataloader (torch.utils.data.DataLoader)
# Build a Wasserstein GAN
gan = WGAN(netG, netD, dataloader, nr_epochs=20)
# train it
gan.train()
# vizualise results
img_list, D_losses, G_losses = gan.get_training_results()
plot_losses(G_losses, D_losses)
plot_image_samples(img_list, 50)
You can currently use the following GANs:
MMGAN
: Classic minimax GAN, in its non-saturated versionWGAN
: Wasserstein GANWGANGP
: Wasserstein GAN with gradient penaltyBEGAN
: Boundary Equilibrium enforcing GAN
All of these GAN objects inherit from a GAN
base class. When building any such GAN, you must give in argument a generator and discriminator networks (some torch.nn.Module
), as well as a torch.utils.data.DataLoader
. In addition, you can specify some parameters supported by all GAN implementations:
optimizer_D
andoptimizer_G
: some PyTorch optimizers (fromtorch.optim
) for the discriminator and generator networks. By defaults those are set with default optimization parameters suggested in the original papers.nr_epochs
: the number of epochs (default: 5)nz
: size of the noise vector (input of the generator) - by defaultnz=100
.save_every
: VeGANs will store some samples produced by the generator everysave_every
iteration. Default: 500fixed_noise_size
: The number of samples to save (from fixed noise vectors)print_every
: The number of iterations between printing training progress. Default: 50
Finally, when calling train()
you can specify some parameters specific to each GAN. For example, for the Wasserstein GAN we can do:
gan = WGAN(netG, netD, dataloader)
gan.train(clip_value=0.1)
This will train a Wasserstein GAN with clipping values of 0.1
(instead of the default 0.01
).
If you are researching new GAN training algorithms, you may find it useful to inherit from the GAN
base class.
Currently the best way to learn more about how to use VeGANs is to have a look at the example notebooks. You can start with this simple example showing how to sample from a univariate Gaussian using a GAN. Alternatively, can run example scripts.
PRs and suggestions are welcome. Look here for more details on the setup.
Some of the code has been inspired by some existing GAN implementations: