6

Just came across this:

Vectorized way of calculating row-wise dot product two matrices with Scipy

This numpy.einsum is really awesome but its a little confusing to use. Suppose I have:

import numpy as np
a = np.array([[1,2,3], [3,4,5]])
b = np.array([[0,1,2], [1,1,7]])

How would i use the "ij" in einsum to get a "cross dot product" between a and b?

Using the example basically I would like to compute dot product of

[1,2,3] and [0,1,2]

[1,2,3] and [1,2,7]

[3,4,5] and [0,1,2]

[3,4,5] and [1,1,7]

and end up with [[8,26],[14,42]]

I know if I use 

np.einsum("ij,ij->i",a,b)

I would just end up with [8, 42] which means I am missing the "cross" elements

bFig8
  • 417
  • 2
  • 7
  • 20

1 Answers1

8

Your result is still 2 dimensional, so you need two indices. What you need is a matrix multiplication with the second array transposed, so instead of normal ij,jk->ik, you transpose the second matrix by ij,kj->ik:

np.einsum('ij,kj->ik', a, b)

#array([[ 8, 24],
#       [14, 42]])

which is equivalent to:

np.dot(a, b.T)

#array([[ 8, 24],
#       [14, 42]])

import numpy as np
a = np.array([[1,2,3], [3,4,5]])
b = np.array([[0,1,2], [1,1,7]])
Psidom
  • 209,562
  • 33
  • 339
  • 356
  • 1
    Yeah this works! Any chance you can point me to where I can learn this "ijk" indexing for einsum? Preferably with examples? The manual https://docs.scipy.org/doc/numpy/reference/generated/numpy.einsum.html isn't much to work with. – bFig8 Aug 26 '17 at 15:30
  • 3
    You can read [this](https://stackoverflow.com/questions/26089893/understanding-numpys-einsum), or [this](http://ajcr.net/Basic-guide-to-einsum/) and some explanation about [einstein notation](https://en.wikipedia.org/wiki/Einstein_notation) may help as well. – Psidom Aug 26 '17 at 15:36