PyTorch的张量tensor和自动求导autograd详解

import torch

t = torch.tensor([[5, 6], [7, 8]])  # 转化为tensor
t1 = torch.empty(5, 3)  # 创建随机5×3的tensor
t2 = torch.rand(5, 3)
t3 = torch.zeros(5, 3, dtype=torch.float16)  # 创建全为0的5×3数组
t4 = torch.ones(5, 3)	# 创建全为1的
t5 = torch.ones_like(t3)	# 创建形状和t3一样、全为1的数组
print(t5.size())  # 输出：torch.Size([5, 3])

t = torch.tensor([[1, 2, 3],
                  [4, 5, 6]])
print(t[0, :2])  # 选取首行前两个元素，输出tensor([1, 2])
print(t.view(-1, 2))	# 改变形状
'''
tensor([[1, 2],
        [3, 4],
        [5, 6]])
'''
t = torch.unsqueeze(t, dim=0)  # 给张量第一维增加一个维度
print(t)
'''
tensor([[[1, 2, 3],
         [4, 5, 6]]])
'''
print(t.permute(2, 1, 0))  # 交换维度位置
'''
tensor([[1, 3, 5],
        [2, 4, 6]])
'''
print(t.expand(2, 2, 3))    # 将(2,3)的tensor扩维成为(2,2,3)
'''
tensor([[[1, 2, 3],
         [4, 5, 6]],
        [[1, 2, 3],
         [4, 5, 6]]])
'''

x = torch.tensor([[1, 2, 3],
                  [4, 5, 6]])
y = torch.tensor([[7, 8, 9],
                  [10, 11, 12]])
print(x + y)  # 用运算符
print(torch.add(x, y))  # 用add()函数
z = torch.empty(2, 3)
torch.add(x, y, out=z)  # 指明输出对象
print(z)
y.add_(x)  # 在原数组上操作
print(y)

t = torch.ones(5)
print(t)
n1 = t.numpy()  # 转化为numpy
n2 = t.clone().numpy()
t.add_(1)  # 对原来的tensor进行操作
print(n1)
print(n2)

n = np.ones(5)	
t2 = torch.from_numpy(n)	# 从numpy创建tensor
'''
tensor([1., 1., 1., 1., 1.])	原来的t
[2. 2. 2. 2. 2.]				操作t后对应的n1也随之改变
[1. 1. 1. 1. 1.]				但是克隆产生的n2不会改变
'''

if torch.cuda.is_available():
    device = torch.device("cuda")  # a CUDA device object
    y = torch.ones_like(x, device=device)  # 直接在GPU上创建tensor
    x = x.to(device)  # 或者使用.to("cuda")，将tensor移入GPU
    z = x + y
    print(z)
    print(z.to("cpu", torch.double))  # 将tensor移出GPU，也能在移动时改变dtype
'''
tensor([2., 2., 2., 2., 2.], device='cuda:0')
tensor([2., 2., 2., 2., 2.], dtype=torch.float64)
'''

# 定义张量
x = torch.ones(5, requires_grad=True)
w1 = torch.tensor(2.0, requires_grad=True)
w2 = torch.tensor(3.0, requires_grad=True)
b = torch.tensor(4.0, requires_grad=False)

# 执行正向操作
l1 = x * w1
l2 = l1 + b
l3 = l2 * w2
y = l3.mean()

# 反向传播
y.backward()

print(l1.data, l1.grad, l1.grad_fn)
# tensor([2., 2., 2., 2., 2.]) None <MulBackward0 object at 0x0000024D8E921BE0>
print(l2.data, l2.grad, l2.grad_fn)
# tensor([6., 6., 6., 6., 6.]) None <AddBackward0 object at 0x000001B960FC0F98>
print(y)
# tensor(18., grad_fn=<MeanBackward0>)
print(w1.grad, w2.grad)
# tensor(3.) tensor(6.)
print(x.grad)
# tensor([1.2000, 1.2000, 1.2000, 1.2000, 1.2000])

...
l1.retain_grad()
l2.retain_grad()
y.backward()

print(l1.data, l1.grad, l1.grad_fn)
# tensor([2., 2., 2., 2., 2.]) tensor([0.6000, 0.6000, 0.6000, 0.6000, 0.6000]) <MulBackward0 object at 0x000001DF52100F60>
print(l2.data, l2.grad, l2.grad_fn)
# tensor([6., 6., 6., 6., 6.]) tensor([0.6000, 0.6000, 0.6000, 0.6000, 0.6000]) <AddBackward0 object at 0x000001DF52111048>

x = torch.ones(5, requires_grad=True)
print((x ** 2).requires_grad)   # 输出True，记录grad

with torch.no_grad():
    print((x ** 2).requires_grad)   # 输出False，不记录