Unexpected behavior with array slicing and mask

Question

It was unexpected that

x=np.empty((2,10,5))
x.shape
>>> (2, 10, 5)

x[0].shape, x[0,:,:].shape
>>> ((10, 5), (10, 5))

mask = [True,True,True,False,False]
x[0,:,mask].shape
>>> (3, 10)

I would have expected x[0,:,mask].shape to be (10,3) because the shape of x is (2,10,5) and applying the slice/mask this change was expected: (2→1, 10→10, 5→3) = (1,10,3).

It seems however transposing the result?

For another array it works as expected:

y=np.empty((2,5))
y.shape
>>> (2, 5)

y[0].shape, y[0,:].shape
>>> ((5,), (5,))

y[:,mask].shape
>>> (2, 3)

This is documented mixed basic and advanced indexing. The sliced dimension is put at the end. Same as x[0, :, [0,1,2]]. x[0][:,mask] avoids this. Lots of similar SO questions. — hpaulj
– hpaulj, Commented Apr 24 at 18:45

Bhargav · Accepted Answer · 2025-04-24 12:32:30Z

1

NumPy boolean masks don't just select elements but restructure the output array. When you use a mask on a dimension, it collapses that dimension and moves the selected elements to the first axis of the result, causing the unexpected transposition in your multidimensional array.

If you want to preserve the original dimensions, you can use numpy.compress(condition, a, axis=None, out=None)

result = np.compress(mask, x[0], axis=1) 
print(result.shape)

output
(10, 3)

answered Apr 24 at 12:32

Bhargav

4,8912 gold badges12 silver badges29 bronze badges

Sign up to request clarification or add additional context in comments.

2 Comments

hpaulj Apr 24 at 18:55

I can't test it now, but I think this is the same as x[0,:,[0,1,2]], a case of mixed basic and advanced indexing. The slice dimension is 'tacked' on to the end.

pas-calc May 22 at 8:36

np.all(x[0,:,[0,1,2]] == x[0,:,mask]) is True

joana · Accepted Answer · 2025-04-24 12:30:59Z

-1

x[0][:, mask].shape  # (10, 3)

That's happen [:, mask] performs standard boolean slicing, selecting 3 out of the 5 columns, resulting in a shape of (10, 3)

answered Apr 24 at 12:30

joana

1013 bronze badges

2 Comments

hpaulj Apr 24 at 18:56

Expand on why this works better.

pas-calc May 9 at 15:09

very interesting, this example shows that x[0,:,mask].shape == (3,10) but x[0][:, mask].shape == (10, 3). . Why is that the case?

Collectives™ on Stack Overflow

Unexpected behavior with array slicing and mask

2 Answers 2

2 Comments

2 Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

2 Answers 2

2 Comments

2 Comments

Your Answer

Sign up or log in

Post as a guest

Related