Mapper networks in RadialGAN

[MargerieHD]

Question

RadialGAN: Mappers functions, is there a difference between the paper and the code ?

Further Information

In the paper, the mapper functions F are described as mapping from a target space D_i to the latent space Z. However, in the code (_train_epoch_mapper function), the loss function is the MSE between the real_X and the mapper[domain](generator[other_domain](noise). Therefore is the mapper functions mapping directly from one domain to the other?

synthcity/src/synthcity/plugins/domain_adaptation/plugin_radialgan.py

Lines 287 to 319 in 6043e1b

	def _train_epoch_mapper(
	self,
	domain: int,
	X: torch.Tensor,
	) -> float:
	batch_size = len(X)
	if batch_size == 0:
	return 0
	# Update the M network
	self.mappers[domain].optimizer.zero_grad()
	real_X = X.to(self.device)
	noise = torch.randn(batch_size, self.n_units_latent, device=self.device)
	errs = []
	for other_domain in self.domains:
	if other_domain == domain:
	continue
	fake = self.generators[other_domain](
	noise
	) # generate fake data for <other_domain>
	fake = self.mappers[domain](fake) # remap data to domain <domain>
	# Calculate M's loss based on this output
	errM = nn.MSELoss()(fake, real_X)
	errs.append(errM)
	# Calculate gradients for M
	errM = 0.1 * torch.sqrt(torch.stack(errs)).mean()
	errM.backward()

Image from the article

Here we see that F maps from the image domain to the latent space