Question

Say I have the array:

 >>> import numpy as np
 >>> myarr = np.array([[1],[2],[3],[2]])

and I want to use logical indexing to return the sub-array where myarr is not equal to 2.

 >>> subarr = myarr[myarr != 2]
 >>> print subarr
 [1 3]

Thus, myarr.shape = (4,1) but subarr.shape = (2,).

Why is the size of dimension 2 being changed when the logical indexing doesn’t affect dimension 2? I could easily just reshape the output, but if I need to use logical indexing often in my code, I don’t want to have to always account for extra dimensions (i.e. dimensions that are obviously not relevant to the logical slice I am doing) that may have changed.

This (N,1) –> (K,) example is especially bothersome, since any logical slice of a something-by-one array must also be a something-by-one array.

I feel like there must be a cleaner way to have the NumPy slice operations leave trivial array dimensions alone. Any suggestions?

Answer 1

In numpy indexing with a boolean array is the same as indexing with array.nonzero(). Because array.nonzero() returns two sets of indices for a 2darray, you are in fact indexing on both the first and the second dimension:

>>> myarr = np.array([[1],[2],[3],[2]])
>>> (myarr != 2).nonzero()
(array([0, 2]), array([0, 0]))
>>> myarr[(array([0, 2]), array([0, 0]))]
array([1, 3])

What you really want is:

>>> myarr[array([0, 2])]
array([[1],
       [3]])

Which you can get by:

>>> myarr[myarr[:,0] != 2]
array([[1],
       [3]])

or:

>>> myarr[(myarr != 2).nonzero()[0]]
array([[1],
       [3]])