Introduce lop-sided affine coupling (IRN, 2020)

[mgm52]

I introduce a "lop-sided" coupling layer used in IRN (Xiao et al 2020), which was an influential model in the task of invertible networks for image rescaling. In the paper, they use x1 to propagate a low-resolution image through the network, and x2 as gaussian-distributed latent data encoding the high-frequency information that x1 is missing.

The layer uses three subnets phi, rho, mu, performing:

y1 = x1 + phi(x2)
y2 = x2 * exp(rho(y1)) + mu(y1)