pytorch: .dot(), .mm(), .matmul(), .add(), .add_(), .mul(), .mul_(), .zero_()

1. torch.dot()函数，专门计算两个一维张量的点积；不支持广播规则

import torch

x1 = torch.tensor([1, 2, 3])  # shape: (3, )

y1 = torch.tensor([2, 3, 4])  # shape: (3, )

z1 = x1.dot(y1)  # 也可写成torch.dot(x1, y1)
print(z1)  # tensor(20)

2. torch.mm()函数，专门计算两个二维张量矩阵的乘法；不支持广播规则

# 将x1, y1扩展一个维度使其为二维张量矩阵
x2 = x1.unsqueeze(dim=0)  # shape: (1, 3)
y2 = y1.unsqueeze(dim=0)  # shape: (1, 3)

# 对y2进行转置，使得符合矩阵乘法规则
y2 = y2.T  # shape: (3, 1)

z2 = torch.mm(x2, y2)  # 根据矩阵乘法规则，得到1*1的二维张量矩阵
print(z2)  # tensor([[20]])

3. torch.matmul()函数，可用于任意维度的张量乘法；当处理一维张量，类似于torch.dot()，当处理二维张量，类似于torch.mm()；

# 创建两个随机张量
x3 = torch.randint(low=1, high=10, size=(2, 3, 2))
print(x3)
"""
tensor([[[9, 6],
         [6, 5],
         [8, 1]],
        [[1, 1],
         [1, 4],
         [1, 8]]])
"""
y3 = torch.randint(low=1, high=10, size=(2, 3))
print(y3)
"""
tensor([[1, 4, 6],
        [3, 3, 6]])
"""
z3 = torch.matmul(x3, y3)
print(z3)
"""
tensor([[[27, 54, 90],
         [21, 39, 66],
         [11, 35, 54]],
        [[ 4,  7, 12],
         [13, 16, 30],
         [25, 28, 54]]])
"""
print(z3.shape)  # torch.Size([2, 3, 3])

4. torch.add()函数，用于张量之间的逐元素相加，支持广播规则；

x4 = torch.tensor([1, 2, 3])  # shape: (3, )
y4 = torch.tensor([[1, 1, 1], [2, 2, 2]])  # shape: (2, 3)

z4 = x4.add(y4)
print(z4)
# tensor([[2, 3, 4],
#         [3, 4, 5]])
print(x4)  # tensor([1, 2, 3])

5. torch.add_()函数，是torch.add()函数的in_place版本；in_place操作在函数作用后即将结果重新赋值给变量。比如x.add_(y)，x+y的结果会存储到原来的x中。pytorch里面所有带"_"的操作，都是in-place的。

x4 = torch.tensor([1, 2, 3])  # shape: (3, )
y4 = torch.tensor([[1, 1, 1], [2, 2, 2]])  # shape: (2, 3)

z5 = y4.add_(x4)  # 注意这里是y4作为主元，因为我们要保证in_place之后的张量shape满足广播规则
print(y4)
# tensor([[2, 3, 4],
#         [3, 4, 5]])
print(y4 == z5)
# tensor([[True, True, True],
#         [True, True, True]])

6. torch.mul()函数，用于张量之间的逐元素相乘，支持广播规则；

x4 = torch.tensor([1, 2, 3])  # shape: (3, )
y4 = torch.tensor([[1, 1, 1], [2, 2, 2]])  # shape: (2, 3)

z6 = torch.mul(x4, y4)
print(z6)
# tensor([[1, 2, 3],
#         [2, 4, 6]])
print(x4)
# tensor([1, 2, 3])
print(y4)
# tensor([[1, 1, 1],
#         [2, 2, 2]])

7. torch.mul_()函数，是torch.mul()函数的in_place版本；

x4 = torch.tensor([1, 2, 3])  # shape: (3, )
y4 = torch.tensor([[1, 1, 1], [2, 2, 2]])  # shape: (2, 3)

z7 = y4.mul_(x4)
print(y4)
# tensor([[1, 2, 3],
#         [2, 4, 6]])
print(y4 == z7)
# tensor([[True, True, True],
#         [True, True, True]])

8. torch.zero_()函数，将张量元素清0，shape不变，也是in_place操作；

y4 = torch.tensor([[1, 1, 1], [2, 2, 2]])  # shape: (2, 3)

z8 = y4.zero_()
print(y4)
# tensor([[0, 0, 0],
#         [0, 0, 0]])
print(y4 == z8)
# tensor([[True, True, True],
#         [True, True, True]])

参考博客：PyTorch: .add()和.add_()，.mul()和.mul_()，.exp()和.exp_()_exp() in-place操作-CSDN博客