根据下面的代码和输出,不难发现在 loss 执行完 backward 操作之后
只有那些 requires_grad 和 is_leaf 的属性同时为 True 的 Tensor 才会有 grad 属性。
不满足上面的条件也想要保存 grad,见这篇博文 Pytorch关于requires_grad_(True)的理解 。
而根据 这篇文章
Tesor 在下面几种情况下,才会是 leaf
# Creating the graph
x = torch.tensor([1.,2.,3.], requires_grad = True)
y = torch.tensor([4.,5.,6.], requires_grad = True)
z = x+y
loss = (z*z).sum()
loss.backward() #Computes the gradient
print("x requires_grad",x.requires_grad)
print('x is_leaf',x.is_leaf)
print('x grad_data',x.grad.data)
print("\ny requires_grad",y.requires_grad)
print('y is_leaf',y.is_leaf)
print('y grad_data',y.grad.data)
print("\nz requires_grad",z.requires_grad)
print('z is_leaf',z.is_leaf)
print("\nloss requires_grad",loss.requires_grad)
print('loss is_leaf',loss.is_leaf)
print('loss grad_data',loss.grad.data)
print('z grad_data',z.grad.data)
输出:
x requires_grad True
x is_leaf True
x grad_data tensor([10., 14., 18.])
y requires_grad True
y is_leaf True
y grad_data tensor([10., 14., 18.])
z requires_grad True
z is_leaf False
loss requires_grad True
loss is_leaf False
File "xx.py", line 27, in <module>
print('loss grad_data',loss.grad.data) #Prints '3' which is dz/dx
AttributeError: 'NoneType' object has no attribute 'data'
z grad_data 也是一样的报错,就不贴上来了