def build_gan():
noise_dim = 64
gen = dcgan_gen(noise_dim=noise_dim)
disc = dcgan_disc()
gan = GAN(gen, disc)
batch_gen = MNISTBatchGenerator()
noise_gen = NoiseGenerator([(noise_dim,)])
return (gan, batch_gen, noise_gen)
theta = torch.zeros(args.p).to(device)
data, theta = generate_contaminated_data(args.eps,
args.train_size,
theta=theta,
type_cont=args.contamination,
coord_median_as_origin=False)
data = data.to(device)
theta = theta.to(device)
data_loader = torch.utils.data.DataLoader(TensorDataset(data),
batch_size=args.real_batch_size,
shuffle=True,
num_workers=0)
noise_generator = NoiseGenerator().to(device)
'''
We recommend not using coordinate-wise median as initialization.
The global minimum of Wasserstein GAN has mean square error very close to the coordinate-wise median,
thus we prefer the training starting from somewhere else in order to see the progress of training.
'''
generator = Generator(
p=args.p,
initializer=1.3 * coord_median(data_loader.dataset.tensors[0]),
# 0.5 * torch.ones(args.p),
).to(device)
sinkhorn = SinkhornIteration(lam=args.lam,
max_iter=args.sinkhorn_max_iter,
device=device,
const=args.const,