Reputation: 614
Sorry for the badly explained title. I am trying to parallelise a part of my code and got stuck on a dot product. I am looking for an efficient way of doing what the code below does, I'm sure there is a simple linear algebra solution but I'm very stuck:
puy = np.arange(8).reshape(2,4)
puy2 = np.arange(12).reshape(3,4)
print puy, '\n'
print puy2.T
zz = np.zeros([4,2,3])
for i in range(4):
zz[i,:,:] = np.dot(np.array([puy[:,i]]).T,
np.array([puy2.T[i,:]]))
Upvotes: 6
Views: 314
Reputation: 221764
Here's another way with broadcasting
-
(puy[None,...]*puy2[:,None,:]).T
Upvotes: 3
Reputation: 353604
One way would be to use np.einsum
, which allows you to specify what you want to happen to the indices:
>>> np.einsum('ik,jk->kij', puy, puy2)
array([[[ 0, 0, 0],
[ 0, 16, 32]],
[[ 1, 5, 9],
[ 5, 25, 45]],
[[ 4, 12, 20],
[12, 36, 60]],
[[ 9, 21, 33],
[21, 49, 77]]])
>>> np.allclose(np.einsum('ik,jk->kij', puy, puy2), zz)
True
Upvotes: 6