torch.bmm()和torch.matmul()函数的用法和区别，矩阵相乘
torch.bmm()和torch.matmul()都是矩阵乘法的运算函数，区别是，torch.matmul更强大。
两者都可以支持3维的矩阵运算，实际是第一维只是找下标，后面2维才是矩阵，然后对应做矩阵乘法。
import torch
x = torch.rand(2,3,2)#2个3x2大小的矩阵
y = torch.rand(2,2,4)#2个2x4大小的矩阵
#矩阵乘法的结果为2个3x4大小的矩阵
print('矩阵x:',x)
print('矩阵y:',y)
print('matmul运算：')
print(torch.matmul(x,y))
print('bmm运算：')
print(torch.bmm(x,y))
运算结果
矩阵x: tensor([[[0.9071, 0.4671],
[0.4678, 0.9219],
[0.2115, 0.1356]],
[[0.9514, 0.4708],
[0.6294, 0.0667],
[0.2318, 0.7602]]])
矩阵y: tensor([[[0.3470, 0.9198, 0.8931, 0.3055],
[0.2137, 0.0132, 0.4909, 0.8819]],
[[0.4307, 0.1535, 0.2024, 0.9597],
[0.5891, 0.7382, 0.1799, 0.8932]]])
matmul运算：
tensor([[[0.4146, 0.8405, 1.0394, 0.6891],
[0.3594, 0.4424, 0.8703, 0.9559],
[0.1024, 0.1963, 0.2554, 0.1842]],
[[0.6871, 0.4936, 0.2772, 1.3335],
[0.3104, 0.1459, 0.1394, 0.6637],
[0.5477, 0.5968, 0.1837, 0.9016]]])
bmm运算：
tensor([[[0.4146, 0.8405, 1.0394, 0.6891],
[0.3594, 0.4424, 0.8703, 0.9559],
[0.1024, 0.1963, 0.2554, 0.1842]],
[[0.6871, 0.4936, 0.2772, 1.3335],
[0.3104, 0.1459, 0.1394, 0.6637],
[0.5477, 0.5968, 0.1837, 0.9016]]])
区别来了，matmul可支持单独的矩阵运算，但是bmm不支持，如下：
import torch
x = torch.rand(3,2)#单独的3x2大小的矩阵
y = torch.rand(2,4)#单独的2x4大小的矩阵
print('矩阵x:',x)
print('矩阵y:',y)
print('matmul运算：')
print(torch.matmul(x,y))
print('bmm运算：')
print(torch.bmm(x,y))
运算结果
矩阵x: tensor([[0.6754, 0.4208],
[0.7033, 0.2135],
[0.2919, 0.3624]])
矩阵y: tensor([[0.5826, 0.6710, 0.3636, 0.3826],
[0.8267, 0.8261, 0.6235, 0.8551]])
matmul运算：
tensor([[0.7414, 0.8009, 0.5080, 0.6183],
[0.5863, 0.6483, 0.3888, 0.4517],
[0.4696, 0.4952, 0.3321, 0.4215]])
bmm运算：
Traceback (most recent call last):
File "E:/储物柜/study/研究生/深度学习/works/SiameseNetwork.py", line 11, in
print(torch.bmm(x,y))
RuntimeError: Expected 3-dimensional tensor, but got 2-dimensional tensor for argument #1 'batch1' (while checking arguments for bmm)
后面报错的原因是bmm只支持3维矩阵的运算。