pytorch中torch.mul、torch.mm/torch.bmm、torch.matmul的区别

预备知识：矩阵的各种乘积

三者对比

torch.mul: 两个输入按元素相乘，内积

分两种情况，默认其中一个是tensor

1. 另一个也是tensor，那么两者就要符合broadcasedt的规则
2. 另一个是标量，那就用标量的数值乘对应tensor就好

---

torch.mm: 特指二维矩阵乘法，外积

If input is (n×m) tensor, mat2 is a (m×p) tensor, out will be a (n×p) tensor.

bonus: torch.bmm就是加了一维batch_size, 也是一样的，只能说3维的两个向量，而且必须符合boardcasted规则

---

torch.matmul: 两个tensor的矩阵乘积，与mm的区别就是不限制是不是二维，符合broadcasted规则的都可以

例如
if input is a (j×1×n×n) tensor and other is a(k×n×n) tensor, out will be a (j×k×n×n) tensor.

代码演示

import torch


num=10
mat1=torch.randn(2,3)
mat2=torch.randn(3,3)
mat3=torch.randn(3,2,3)
#
print("torch.mul的结果1\n",torch.mul(mat1,num))
print("torch.mul的结果2\n",torch.mul(mat1,mat1))
print("torch.mm的结果[二维矩阵相乘]\n",torch.mm(mat1,mat2))
print("torch.matmul的结果[二维矩阵相乘]\n",torch.matmul(mat3,mat2))

'''
torch.mul的结果1
 tensor([[ 3.9774,  8.8436, -0.6948],
        [20.7512,  6.7674,  8.6104]])
torch.mul的结果2
 tensor([[0.1582, 0.7821, 0.0048],
        [4.3061, 0.4580, 0.7414]])
torch.mm的结果[二维矩阵相乘]
 tensor([[-0.4850, -1.0539, -0.9389],
        [-0.4043, -3.4724,  0.1518]])
torch.matmul的结果[二维矩阵相乘]
 tensor([[[-1.8273, -2.6934, -0.2099],
         [-0.7700,  0.9479,  2.2710]],

        [[-0.4027,  1.9117,  0.0667],
         [ 1.1908,  1.0069, -2.3486]],

        [[-1.1636,  0.1499,  3.8286],
         [ 1.0558, -1.7520,  0.8731]]])
         '''

注意：
*，@是两个运算符，他们分别映射到函数torch.mul和torch.matmul()