I have a 3d array
A = np.random.random((4,4,3))
and a index matrix
B = np.int_(np.random.random((4,4))*3)
How do I get a 2D array from A based on index matrix B?
In general, how to get a N-1 dimensional array from a ND array and a N-1 dimensional index array?
Lets take an example:
>>> A = np.random.randint(0,10,(3,3,2))
>>> A
array([[[0, 1],
[8, 2],
[6, 4]],
[[1, 0],
[6, 9],
[7, 7]],
[[1, 2],
[2, 2],
[9, 7]]])
Use fancy indexing to take simple indices. Note that the all indices must be of the same shape and the shape of each index will be what is returned.
>>> ind = np.arange(2)
>>> A[ind,ind,ind]
array([0, 9]) #Index (0,0,0) and (1,1,1)
>>> ind = np.arange(2).reshape(2,1)
>>> A[ind,ind,ind]
array([[0],
[9]])
So for your example we need to supply the grid for the first two dimensions:
>>> A = np.random.random((4,4,3))
>>> B = np.int_(np.random.random((4,4))*3)
>>> A
array([[[ 0.95158697, 0.37643036, 0.29175815],
[ 0.84093397, 0.53453123, 0.64183715],
[ 0.31189496, 0.06281937, 0.10008886],
[ 0.79784114, 0.26428462, 0.87899921]],
[[ 0.04498205, 0.63823379, 0.48130828],
[ 0.93302194, 0.91964805, 0.05975115],
[ 0.55686047, 0.02692168, 0.31065731],
[ 0.92822499, 0.74771321, 0.03055592]],
[[ 0.24849139, 0.42819062, 0.14640117],
[ 0.92420031, 0.87483486, 0.51313695],
[ 0.68414428, 0.86867423, 0.96176415],
[ 0.98072548, 0.16939697, 0.19117458]],
[[ 0.71009607, 0.23057644, 0.80725518],
[ 0.01932983, 0.36680718, 0.46692839],
[ 0.51729835, 0.16073775, 0.77768313],
[ 0.8591955 , 0.81561797, 0.90633695]]])
>>> B
array([[1, 2, 0, 0],
[1, 2, 0, 1],
[2, 1, 1, 1],
[1, 2, 1, 2]])
>>> x,y = np.meshgrid(np.arange(A.shape[0]),np.arange(A.shape[1]))
>>> x
array([[0, 1, 2, 3],
[0, 1, 2, 3],
[0, 1, 2, 3],
[0, 1, 2, 3]])
>>> y
array([[0, 0, 0, 0],
[1, 1, 1, 1],
[2, 2, 2, 2],
[3, 3, 3, 3]])
>>> A[x,y,B]
array([[ 0.37643036, 0.48130828, 0.24849139, 0.71009607],
[ 0.53453123, 0.05975115, 0.92420031, 0.36680718],
[ 0.10008886, 0.02692168, 0.86867423, 0.16073775],
[ 0.26428462, 0.03055592, 0.16939697, 0.90633695]])
If you prefer to use mesh as suggested by Daniel, you may also use
A[tuple( np.ogrid[:A.shape[0], :A.shape[1]] + [B] )]
to work with sparse indices. In the general case you could use
A[tuple( np.ogrid[ [slice(0, end) for end in A.shape[:-1]] ] + [B] )]
Note that this may also be used when you'd like to index by B an axis different from the last one (see for example this answer about inserting an element into a list).
Otherwise you can do it using broadcasting:
A[np.arange(A.shape[0])[:, np.newaxis], np.arange(A.shape[1])[np.newaxis, :], B]
This may be generalized too but it's a bit more complicated.
Bsupposed to be indexing? Which dimensions are you wanting to use to sliceA? e.g. What doesB[0,0] == 2indicate?