用torch实现求loss中包含一阶梯度项的二阶梯度求导

如果我的loss中包含有一阶梯度项，该如何用torch进行求导。
 

from torch import nn
import torch.nn.functional as F
import torch


class Classifier(nn.Module):
    def __init__(self, dim_in=5):
        super(Classifier, self).__init__()
        self.fc = nn.Linear(dim_in, 1)
        self.relu = nn.ReLU()

    def forward(self, x):
        x = self.fc(x)
        return self.relu(x)

    def loss(self, x, y):
        return torch.mean((x-y)**2)


if __name__ == '__main__':
    h_hat = torch.rand(13, 5).requires_grad_(True)
    y = torch.rand(13, 1).requires_grad_(True)
    d = Classifier()
    opt = torch.optim.Adam(d.parameters())
    loss = d.loss(d(h_hat), y)
    opt.zero_grad()
    #求一阶梯度
    loss.backward(create_graph=True)#二阶导数设置为True
    grad = d.fc.weight.grad
    print("grad", grad)
    L_grad = torch.norm(grad, p=2)
    #新的loss中包含一阶梯度项
    loss = loss + L_grad
    #opt.zero_grad() #这里不能放zero_grad()，会将一阶梯度清0，导致两次求出的梯度相同
    #求二阶梯度
    loss.backward() # 
    print("L_grad", d.fc.weight.grad)
    opt.step()



#Terminal output:
#grad tensor([[ 0.1285, -0.0961,  0.1789,  0.1829,  0.0649]], grad_fn=<CopyBackwards>)
#L_grad tensor([[1.2037, 0.4403, 1.4736, 1.4062, 1.1052]], grad_fn=<CopyBackwards>)
#可以看到这两个梯度是不一样的