numpy: multiply arrays rowwise
Question:
I have those arrays:
a = np.array([
[1,2],
[3,4],
[5,6],
[7,8]])
b = np.array([1,2,3,4])
and I want them to multiply like so:
[[1*1, 2*1],
[3*2, 4*2],
[5*3, 6*3],
[7*4, 8*4]]
… basically out[i] = a[i] * b[i], where a[i].shape is (2,) and b[i] then is a scalar.
What’s the trick? np.multiply seems not to work:
>>> np.multiply(a, b)
ValueError: operands could not be broadcast together with shapes (4,2) (4)
Answers:
add an axis to b:
>>> np.multiply(a, b[:, np.newaxis])
array([[ 1, 2],
[ 6, 8],
[15, 18],
[28, 32]])
>>> a * b.reshape(-1, 1)
array([[ 1, 2],
[ 6, 8],
[15, 18],
[28, 32]])
For those who don’t want to use np.newaxis or reshape, this is as simple as:
a * b[:, None]
This is because np.newaxis is actually an alias for None.
Read more here.
It looks nice, but quite naive, I think, because if you change the dimensions of a or b, the solution
np.mulitply(a, b[:, None])
doesn’t work anymore.
I’ve always had the same doubt about multiplying arrays of arbitrary size row rise, or even, more generally, n-th dimension wise.
I used to do something like
z = np.array([np.multiply(a, b) for a, b in zip(x,y)])
and that works for x or y that have dimension 1 or 2.
Does it exist with a method with “axis” argument like in other numpy methods? Such like
z = np.mulitply(x, y, axis=0)
What’s missing here is the einsum (doc) variant:
np.einsum("ij,i->ij", a, b)
This gives you full control over the indices and a and b are passed blank.
I have those arrays:
a = np.array([
[1,2],
[3,4],
[5,6],
[7,8]])
b = np.array([1,2,3,4])
and I want them to multiply like so:
[[1*1, 2*1],
[3*2, 4*2],
[5*3, 6*3],
[7*4, 8*4]]
… basically out[i] = a[i] * b[i], where a[i].shape is (2,) and b[i] then is a scalar.
What’s the trick? np.multiply seems not to work:
>>> np.multiply(a, b)
ValueError: operands could not be broadcast together with shapes (4,2) (4)
add an axis to b:
>>> np.multiply(a, b[:, np.newaxis])
array([[ 1, 2],
[ 6, 8],
[15, 18],
[28, 32]])
>>> a * b.reshape(-1, 1)
array([[ 1, 2],
[ 6, 8],
[15, 18],
[28, 32]])
For those who don’t want to use np.newaxis or reshape, this is as simple as:
a * b[:, None]
This is because np.newaxis is actually an alias for None.
Read more here.
It looks nice, but quite naive, I think, because if you change the dimensions of a or b, the solution
np.mulitply(a, b[:, None])
doesn’t work anymore.
I’ve always had the same doubt about multiplying arrays of arbitrary size row rise, or even, more generally, n-th dimension wise.
I used to do something like
z = np.array([np.multiply(a, b) for a, b in zip(x,y)])
and that works for x or y that have dimension 1 or 2.
Does it exist with a method with “axis” argument like in other numpy methods? Such like
z = np.mulitply(x, y, axis=0)
What’s missing here is the einsum (doc) variant:
np.einsum("ij,i->ij", a, b)
This gives you full control over the indices and a and b are passed blank.