对一个一般的函数利用paddle.grad是可以求高阶导数，好像每次求导都会在原本的图的基础上加入求导的部分

x = paddle.to_tensor([4])
x.stop_gradient = False
y = x*x*x
dx =  paddle.grad(y, x, grad_outputs=None, retain_graph=True, create_graph=True, only_inputs=True, allow_unused=False, no_grad_vars=None)[0]
dx2 = paddle.grad(dx, x, grad_outputs=None, retain_graph=True, create_graph=True, only_inputs=True, allow_unused=False, no_grad_vars=None)[0]
print(dx.item()," ",dx2.item())

但是对paddle.nn.Layer构造的模型无法求高阶导数，只能求一阶导数，求高阶导数会报错，无法从图上找到高阶导数与输入变量的关系，create_graph和retain_graph好像没起作用，执行到最后一行会报错

class net(paddle.nn.Layer):
    def __init__(self):
        super(net,self).__init__()
        self.linear1 = nn.Linear(1, 100)
        self.linear2 = nn.Linear(100,1)
        self.relu1   = nn.ReLU()
    def forward(self,x):
        y = self.linear1(x)
        y = self.relu1(y)
        y = self.linear2(y)
        return y
model = net()

t = paddle.to_tensor([4.0])
t.stop_gradient = False
Sx = model(t)
dSx_dT = paddle.grad(Sx, t, grad_outputs=None, retain_graph=True, create_graph=True, only_inputs=True, allow_unused=False, no_grad_vars=None)
d2Sx_dT2 = paddle.grad(dSx_dT, t, grad_outputs=None, retain_graph=True, create_graph=True, only_inputs=True, allow_unused=False, no_grad_vars=None)