参考：https://blog.csdn.net/omnispace/article/details/54942668

上面这篇博客讲的很好！

PS:

(1)wgan中的weight cliping后面又被升级为gradient penalty;

参考：

代码：

from torch.autograd import grad#gradient penalty , autograd wayLAMBDA_GRAD_PENALTY = 1.0alpha = torch.rand(BATCH_SIZE, 1, 1, 1).cuda()#pred_penalty是生成的分布，D_gt_v是真实分布differences = pred_penalty - D_gt_v interpolates = D_gt_v + (alpha * differences)D_interpolates = model_D(interpolates)gradients = grad(outputs=D_interpolates, inputs=interpolates, grad_outputs=torch.ones(D_interpolates.size()).cuda(), create_graph=False, retain_graph=True, only_inputs=True)[0]gradient_penalty = torch.mean(torch.sqrt(torch.sum((gradients - 1) ** 2 , dim = (1 , 2 , 3)))) * LAMBDA_GRAD_PENALTYloss_D += gradient_penalty