I have a 2d array:
a = np.random.randint(100, size=(6, 4))
[[72 76 40 11]
[48 82 6 87]
[53 24 25 99]
[ 7 94 82 90]
[28 81 10 9]
[94 99 67 58]]
And a 1d array:
idx = np.random.randint(4, size=6)
[0, 3, 2, 1, 0, 2]
Is it possible to index the 2d array so that the result is:
a[idx]
[72, 87, 25, 94, 28, 67]
Since you have the column indices, all you need are the row indices. You can generate those with arange.
>>> a[np.arange(len(a)), idx]
array([72, 87, 25, 94, 28, 67])
Is there any way to get by this without arange? It seems counterintuitive to me that something like
a[idx.reshape(-1,1)]
or
a[:,idx]
would not produce this result.
You can also use np.diagonal if you want to avoid np.arrange.
a = np.array([[72, 76, 40, 11],
[48, 82, 6, 87],
[53, 24, 25, 99],
[ 7, 94, 82, 90],
[28, 81, 10, 9],
[94, 99, 67, 58]])
idx = np.array([0, 3, 2, 1, 0, 2])
index into each array in 2d array using idx
>>> a_idx = a[...,idx]
>>> a_idx
array([[72, 11, 40, 76, 72, 40],
[48, 87, 6, 82, 48, 6],
[53, 99, 25, 24, 53, 25],
[ 7, 90, 82, 94, 7, 82],
[28, 9, 10, 81, 28, 10],
[94, 58, 67, 99, 94, 67]])
diagonal is where the position of idx and each array in 2d array line up
>>> np.diagonal(a_idx)
array([72, 87, 25, 94, 28, 67])
An alternative answer.
(a*np.eye(6,4)[idx]).sum(axis=1)
