def calc_gradient_penalty(self, netD, real_data, fake_data): alpha = torch.rand(1, 1) alpha = alpha.expand(real_data.size()) alpha = alpha.cuda() interpolates = alpha * real_data + ((1 - alpha) * fake_data) interpolates = interpolates.cuda() interpolates = Variable(interpolates, requires_grad=True) disc_interpolates, s = netD.forward(interpolates) gradients1 = autograd.grad(outputs=disc_interpolates, inputs=interpolates, grad_outputs=torch.ones(disc_interpolates.size()).cuda(), create_graph=True, retain_graph=True, only_inputs=True)[0] print(gradients1) print(gradients1.shape) gradient_penalty = (((gradients1.norm(2, dim=1) - 1) ** 2).mean() * self.LAMBDA) return gradient_penalty，已知上述代码中，s为超像素池化后的分数，如何将s加入到patchgan 的梯度惩罚中，输入图像的梯度是哪个，怎么计算，给出改进后的代码

可以将超像素池化的分数s与判别器的输出disc_interpolates相加，作为最终的判别结果。即将以下代码：

disc_interpolates, s = netD.forward(interpolates)

改为：

disc_interpolates = netD.forward(interpolates) + s

而输入图像的梯度应该是interpolates的梯度，即将以下代码：

gradients1 = autograd.grad(outputs=disc_interpolates, inputs=interpolates, grad_outputs=torch.ones(disc_interpolates.size()).cuda(), create_graph=True, retain_graph=True, only_inputs=True)[0]

改为：

gradients1 = autograd.grad(outputs=disc_interpolates, inputs=interpolates, grad_outputs=torch.ones(disc_interpolates.size()).cuda(), create_graph=True, retain_graph=True, only_inputs=True, allow_unused=True)[0]

这样可以允许s没有梯度信息时，梯度惩罚仍能正常计算。最终改进后的代码如下：

def calc_gradient_penalty(self, netD, real_data, fake_data):
    alpha = torch.rand(1, 1)
    alpha = alpha.expand(real_data.size())
    alpha = alpha.cuda()

    interpolates = alpha * real_data + ((1 - alpha) * fake_data)
    interpolates = interpolates.cuda()
    interpolates = Variable(interpolates, requires_grad=True)

    disc_interpolates = netD.forward(interpolates) + s

    gradients1 = autograd.grad(outputs=disc_interpolates, inputs=interpolates, grad_outputs=torch.ones(disc_interpolates.size()).cuda(), create_graph=True, retain_graph=True, only_inputs=True, allow_unused=True)[0]
    gradient_penalty = (((gradients1.norm(2, dim=1) - 1) ** 2).mean() * self.LAMBDA)

    return gradient_penalty