WGANGP

critic_loss(m, x_true, x_generated, batch_size, λ)

WGAN-GP relaxed critic loss with lagrange multiplier λ

lipschitz1_gradient_loss(m, x_true, x_generated, batch_size)

Estimates 𝐄ₓ(‖∇ₓD(x)‖₂ - 1)², where x is sampled uniformly on lines between points from the data distribution and the generators distribution

step_critic!(opt, m, x_true, x_generated, batch_size; λ = 10.0f0)

A single optimisation step for the critic, with λ gradient penalty factor.

A single optimisation step for the generator