net = your_network().cuda()
def train():
 ...
 outputs = net(inputs)
    loss = criterion(outputs, targets)
    loss.backward()
 for name, parms in net.named_parameters(): 
  print('-->name:', name, '-->grad_requirs:',parms.requires_grad, \
   ' -->grad_value:',parms.grad)
 ...

import torch
from torch.autograd import Variable
x = torch.Tensor([[1.,2.,3.],[4.,5.,6.]])  #grad_fn是None
x = Variable(x, requires_grad=True)
y = x + 2
z = y*y*3
out = z.mean()
#x->y->z->out
print(x)
print(y)
print(z)
print(out)
#结果:
tensor([[1., 2., 3.],
        [4., 5., 6.]], requires_grad=True)
tensor([[3., 4., 5.],
        [6., 7., 8.]], grad_fn=<AddBackward>)
tensor([[ 27.,  48.,  75.],
        [108., 147., 192.]], grad_fn=<MulBackward>)
tensor(99.5000, grad_fn=<MeanBackward1>)

out.backward()
print(x.grad)
#结果:
tensor([[3., 4., 5.],
        [6., 7., 8.]])
#如果是z关于x求导就必须指定gradient参数：
gradients = torch.Tensor([[2.,1.,1.],[1.,1.,1.]])
z.backward(gradient=gradients)
#若z不是一个标量，那么就先构造一个标量的值：L = torch.sum(z*gradient)，再关于L对各个leaf Variable计算梯度
#对x关于L求梯度
x.grad
#结果:
tensor([[36., 24., 30.],
        [36., 42., 48.]])

z.backward()
print(x.grad) 
#报错:RuntimeError: grad can be implicitly created only for scalar outputs只能为标量创建隐式变量
    
x1 = Variable(torch.Tensor([[1.,2.,3.],[4.,5.,6.]])) 
x2 = Variable(torch.arange(4).view(2,2).type(torch.float), requires_grad=True)
c = x2.mm(x1)
c.backward(torch.ones_like(c))
# c.backward()
#RuntimeError: grad can be implicitly created only for scalar outputs
print(x2.grad)