自动微分 - 回到 PyTorch 笔记

import torch

x = torch.tensor([
    [2., 3., ],
    [4., 5., ],
], requires_grad=True)
print('==1==')
print(x.requires_grad)
print(x)
print(x.grad)

print('\n==2==')
y = x + 2
print(y)

print('\n==3==')
z = y * y * 3
print(z)

print('\n==4==')
out = z.mean()  # 求算数平均值
print(out)

print('\n==5==')
print(x)
out.backward(torch.tensor(1.))
print(x)
print(x.grad)

==1==
True
tensor([[2., 3.],
        [4., 5.]], requires_grad=True)
None

==2==
tensor([[4., 5.],
        [6., 7.]], grad_fn=<AddBackward0>)

==3==
tensor([[ 48.,  75.],
        [108., 147.]], grad_fn=<MulBackward0>)

==4==
tensor(94.5000, grad_fn=<MeanBackward0>)

==5==
tensor([[2., 3.],
        [4., 5.]], requires_grad=True)
tensor([[2., 3.],
        [4., 5.]], requires_grad=True)
tensor([[ 6.0000,  7.5000],
        [ 9.0000, 10.5000]])

import torch

x = y = torch.randn(2, requires_grad=True)
print(x)
print(x.data)

print('==')
while y.data.norm() < 1000:  # 张量的标量成分的归一化
    y = y * 2
print(y)

print('==')
v = torch.tensor([1., 2.], dtype=torch.float)
print(v)
y.backward(v)
print(x.grad)  # TODO 这里比较困惑，x 与 y 的关系

tensor([ 0.9871, -0.9948], requires_grad=True)
tensor([ 0.9871, -0.9948])
==
tensor([ 1010.7546, -1018.6384], grad_fn=<MulBackward0>)
==
tensor([1., 2.])
tensor([1024., 2048.])