How to vectorise a list of matrix vector multiplications using pytorch/numpy

Question

For example, I have a list of N B x H tensor(i.e. a N x B x H tensor) and a list of N vectors (i.e. N x B tensor). And I want multiply each B x H tensor in the list with corresponding B dimensional tensor, resulting a N x H tensor.

I know how to use a single for-loop with PyTorch to implement the computation, but is there any vectorised implantation? (i.e. no for-loop, just using PyTorch/numpy operations)

Seconded. I have an idea in my head of what you're trying but it would be much more reassuring if I have some defined input and output to test against. — roganjosh

benjaminplanche benjaminplanche · Accepted Answer · 2018-06-23T16:31:01

You could achieve this with torch.bmm() and some torch.squeeze()/torch.unsqueeze().

I am personally rather fond of the more generictorch.einsum() (which I find more readable):

import torch
import numpy as np

A = torch.from_numpy(np.array([[[1, 10, 100], [2, 20, 200], [3, 30, 300]],
                               [[4, 40, 400], [5, 50, 500], [6, 60, 600]]]))
B = torch.from_numpy(np.array([[ 1,  2,  3],
                               [-1, -2, -3]]))

AB = torch.einsum("nbh,nb->nh", (A, B))
print(AB)
# tensor([[   14,   140,  1400],
#         [  -32,  -320, -3200]])

How to vectorise a list of matrix vector multiplications using pytorch/numpy

1 Answers