WGAN-GP 关键源码

#https://github.com/yunjey/stargan/blob/master/solver.py
def gradient_penalty(self, y, x):
    """Compute gradient penalty: (L2_norm(dy/dx) - 1)**2."""
    weight = torch.ones(y.size()).to(self.device)
    dydx = torch.autograd.grad(outputs=y,
                               inputs=x,
                               grad_outputs=weight,
                               retain_graph=True,
                               create_graph=True,
                               only_inputs=True)[0]

    dydx = dydx.view(dydx.size(0), -1)
    dydx_l2norm = torch.sqrt(torch.sum(dydx**2, dim=1))
    return torch.mean((dydx_l2norm-1)**2)
# Compute loss for gradient penalty.
alpha = torch.rand(x_real.size(0), 1, 1, 1).to(self.device)
x_hat = (alpha * x_real.data + (1 - alpha) * x_fake.data).requires_grad_(True)
out_src, _ = self.D(x_hat)
d_loss_gp = self.gradient_penalty(out_src, x_hat)