In general, PyTorch has 3 different types of matrix multiplication and each of the types has its own characteristics and the way to use it. They are torch.matmul, torch.mm, and torch.bmm. Then, let’s look at them one by one in detail!.
파이토치는 3종 행렬 곱셈이 기본적으로 가지고 있으며 각각 종은 다른 특징과 방식도 갖는다. 그것들은 torch.matmul, torch.mm, torch.bmm. 그렇다면, 각각 행렬 곱셈의 기능에 대해 상세하게 알아보자!.
torch.matmul
The torch.matmul has an ability to perform the matrix multiplication in several ways. This torch.matmul can automatically broadcast the tensor size to fit the the other tensor sizes without explicitly define the compatible tensor sizes during multiplication. Additionally, torch.matmul is also able to perform matrix with scalar multiplication, vanilla matrix multiplications, and batched matrix multiplication.
torch.matmul은 몇 몇 행렬 곱셈의 기능을 갖는다. 텐서의 크기가 맞을 수 있도록 이 torch.matmul이 자동적으로 텐서를 broadcast할 수 있으며 텐서의 크기를 스스로 정의하는 것이 필요없다. 게다가, torch.matmul도 행렬과 스칼라 곱셈, 평범한 행렬 곱셈, 그리고 배치 행렬 곱셈이다.
torch.matmul 구현 코드
# 1D (3) tensor x 1D (3) tensor
a = torch.rand(3)
b = torch.rand(3)
print(torch.matmul(a,b).size())
# The output is 0D ([]) tensor (scalar)
# 2D (3,3) tensor x 1D (3) broadcasted tensor
a = torch.rand(3,3)
b = torch.rand(3)
print(torch.matmul(a,b).size())
# The output is 1D ([3]) tensor (vector)
# 2D (3,3) tensor x 2D (3x3) tensor
a = torch.rand(3,3)
b = torch.rand(3,3)
print(torch.matmul(a,b).size())
# The output is 2D ([3,3]) tensor (matrix)
# 3D (3,1,2) tensor (batched matrix) x 1D (2) broadcasted tensor
a = torch.rand(3,1,2)
b = torch.rand(2)
print(torch.matmul(a,b).size())
# The output is 2D ([3,1]) tensor
# 3D (3,1,2) tensor (batched matrix) x 3D (3,2,3) batched matrix
a = torch.rand(3,1,2)
b = torch.rand(3,2,3)
print(torch.matmul(a,b).size())
# The output is 3D ([3,1,3]) tensor
torch.mm
torch.mm function is only used for multiplying 2D tensor or we can say to perform a vanilla 2D matrix multiplication. Different with torch.matmul the torch.mm function doesn’t have tensor broadcasting ability.
torch.mm은 2D 텐서를 곱셈하기 위하거나 평범한 2D 행렬 곱셈을 사용한다. torch.matmul과 torch.mm이 다른 점은 텐서 broadcasting 기능을 갖지 않는다.
torch.mm 구현 코드
# Vanilla matrix multiplication (tensor 2D) (2,3) x (3,10)
a = torch.rand(2,3)
b = torch.rand(3,10)
print(torch.mm(a,b).size())
# The matrix output size is (2,10) => (n, m) * (m, a) = (n, a)
torch.bmm
The last one is torch.bmm, this function is only for doing batched tensor multiplication. However, compared to the torch.matmul, the torch.bmm also doesn’t have ability to do tensor broadcasting. Therefore, torch.bmm is merely a special function for doing batched tensor multiplication.
마지막으로, torch.bmm이다, 이 기능은 배치 텐서 곱셈을 할 수 밖에 없다. 그렇지만, torch.matmul에 비해 torch.bmm도 텐서 broadcasting 기능을 가지지 않는다. 따라서 torch.bmm은 특별한 배치 텐서를 하기 위한 기능일 뿐이다.
torch.bmm 구현 코드
# Batched tensor multiplication
a = torch.rand(3,1,2)
b = torch.rand(3,2,4)
print(torch.bmm(a,b).size())
# The output size is (3,1,4)
Written with StackEdit ❤️.