本节介绍有关张量Tensor的基本操作

Tensor相当于numpy中的ndarrays

• 创建空Tensor和全零Tensor，torch.zeros(d0,d1)类似于numpy.zeros(size)，torch.empty(d0,d1)类似于numpy.empty(size)

x1 = torch.empty(2,2)
x2 = np.empty((2,2))
x2 = torch.zeros(2,2)
x4 = np.zeros((2,2))
print(x1)
print(x2)
print(x3)
print(x4)

tensor([[ 1.4013e-45,  0.0000e+00],
        [-4.2944e+22,  6.7683e-43]])
tensor([[0., 0.],
        [0., 0.]])
[[0. 0. 0.]
 [0. 0. 0.]
 [0. 0. 0.]
 [0. 0. 0.]
 [0. 0. 0.]]
[[0. 0.]
 [0. 0.]]

• 创建随机Tensor，torch.rand(d0,d1)相当于numpy.random.rand(d0,d1)，rand表示随机变量服从uniform distribution，torch.randn(d0,d1)相当于numpy.random.randn(d0,d1)，randn表示随机变量服从 normal distribution

y1 = torch.rand(2,2)
y2 = np.random.rand(2,2)
y3 = np.random.random((2,2))
print(y1)
print(y2)
print(y3)

tensor([[0.4407, 0.1455],
        [0.0214, 0.6033]])
[[0.91633254 0.74086994]
 [0.11203967 0.78700098]]
[[0.89562162 0.63706194]
 [0.07474063 0.94183217]]

• 手动创建Tensor，torch.tensor()类似于numpy.array()

z1 = torch.tensor(np.random.rand(2,2))
z2 = np.array([2,2])
print(z1)
print(z2)

tensor([[0.8592, 0.2296],
        [0.3234, 0.0014]], dtype=torch.float64)
[2 2]

• 创建全一Tensor，torch.ones(d0,d1)类似于numpy.ones(size)

c1 = torch.ones(2,2)
c2 = np.ones((2,2))
print(c1)
print(c2)

tensor([[1., 1.],
        [1., 1.]])
[[1. 1.]
 [1. 1.]]

• 创建独热码Tensor，torch.eye(d0,d1)类似于np.eye(d0,d1)

c1 = torch.ones(2,2)
c2 = np.ones((2,2))
print(c1)
print(c2)

tensor([[1., 1.],
        [1., 1.]])
[[1. 1.]
 [1. 1.]]

• 获得Tensor的size，tensor.size()返回tensor的shape

r = torch.eye(2,2)
print(r.size())

• 加法，使用+与numpy类似

a1 = torch.eye(2,2)
a2 = torch.rand(2,2)
print(a1+a2)
b1 = np.eye(2,2)
b2 = np.random.rand(2,2)
print(b1+b2)

tensor([[1.1682, 0.4159],
        [0.5044, 1.4019]])
[[1.7929664  0.96477472]
 [0.3380899  1.35091993]]

• 加法，使用torch.add()与numpy.add()类似

a1 = torch.eye(2,2)
a2 = torch.rand(2,2)
print(torch.add(a1,a2))
b1 = np.eye(2,2)
b2 = np.random.rand(2,2)
print(np.add(b1,b2))

tensor([[1.5244, 0.8070],
        [0.2586, 1.1021]])
[[1.52558655 0.85622143]
 [0.92030175 1.18823413]]

• 加法，tensor.add(a,b,out=result)，将加法的结果保存在预先开辟好的result张量中

result = torch.empty(3,3)
torch.add(a1,a2, out=result)
print(result)
result = torch.add(a1,a2)
print(result)

tensor([[1.8654, 0.8892, 0.9849],
        [0.4269, 1.3964, 0.7995],
        [0.2235, 0.3375, 1.3471]])
tensor([[1.8654, 0.8892, 0.9849],
        [0.4269, 1.3964, 0.7995],
        [0.2235, 0.3375, 1.3471]])

• 加法，in-place原地替换的做法tensor.add_()，注：所有的in-place的做法都有一个_

a1 = torch.eye(3,3)
a2 = torch.eye(3,3)
a1.add_(a2)
print(a1)

tensor([[2., 0., 0.],
        [0., 2., 0.],
        [0., 0., 2.]])

• tensor改变形状，tensor.reshape(size)与ndarrays.reshape(size)类似，但tensor.reshape(size)不是in-place的做法，ndarryas.reshape(size)也不是in-place的做法

a = torch.eye(4,4)
b = a.reshape(2,8)
print(a)
print(b)

c = np.eye(4,4)
d = c.reshape(2,8)
print(c)
print(d)

tensor([[1., 0., 0., 0.],
        [0., 1., 0., 0.],
        [0., 0., 1., 0.],
        [0., 0., 0., 1.]])
tensor([[1., 0., 0., 0., 0., 1., 0., 0.],
        [0., 0., 1., 0., 0., 0., 0., 1.]])
[[1. 0. 0. 0.]
 [0. 1. 0. 0.]
 [0. 0. 1. 0.]
 [0. 0. 0. 1.]]
[[1. 0. 0. 0. 0. 1. 0. 0.]
 [0. 0. 1. 0. 0. 0. 0. 1.]]

• 使用tensor.item()来获得 scalar tensor的值

a = torch.randn(3,3)
print(a)
print(a[0,0])
print(a[0,0].item())

tensor([[ 0.1261, -0.7185,  0.3167],
        [ 0.7252,  0.9447,  1.6690],
        [ 0.4250, -0.3057,  0.7201]])
tensor(0.1261)
0.126137375831604