gradients的一些注意点

Each variable has a [.grad_fn] attribute that references a Function that has created the Variable(except for Variables created by the user their grad_fn is None).

out.backward()=out.backward(torch.Tensor([1.0]))

import torch
from torch.autograd import Variable

# part 1
x = Variable(torch.ones(2, 2), requires_grad=True)
y = x + 2
print(y.grad_fn)    # <torch.autograd.function.AddConstantBackward object at 0x000001D8A156E048>
print(y.grad)       # None
z = y * y * 3
out = z.mean()
out.backward()
print(out.grad)     # None
print(y.grad)       # None
print(x.grad)       # d(out)/dx
'''
Variable containing:
 4.5000  4.5000
 4.5000  4.5000
[torch.FloatTensor of size 2x2]
'''
print(x.grad_fn)    # None
print(x.grad_output)


# part 2
x = torch.randn(3)
x = Variable(x, requires_grad=True)
y = x * 2
# print(type(y))          # <class 'torch.autograd.variable.Variable'>
# print(type(y.data))     # <class 'torch.FloatTensor'>
# print(y.data.norm())    # 4.076032686856067
while y.data.norm() < 1000:
    y = y * 2

# print(x.grad)           # None
gradients = torch.FloatTensor([0.1, 1.0, 0.0001])
# print(y)  # Variable containing: 377.3516 997.8206 11.2558 [torch.FloatTensor of size 3]
y.backward(gradients)
# y.backward()            # y.backward()=y.backward(torch.Tensor([1.0]))
# RuntimeError: grad can be implicitly created only for scalar outputs

print(x.grad)
# print(x.grad_fn)          # None
# print(x.grad_output)      # AttributeError: 'Variable' object has no attribute 'grad_output'