How to create view of an irregularly spaced slice of a numpy array?

Question

In the following case:

a = np.zeros(35)
# create a view
av = a[3:10]
assert np.may_share_memory(a, av)
some_ind = array([0,5,6,24])
# trying to create a second view of another slice of a 
av2 = a[some_ind] # or a[np.where(some_ind)]
assert np.may_share_memory(a, av2) # raise AssertionError

What would be the procedure such that av2 would be a view of the slice of a corresponding to the indices some_ind? I thought about creating a mask array but found in the doc that copies were made using mask arrays.

In other words, my aim, if it is possible, is to be able to create a view of a slice of an array with indices from the initial array not regularly spaced.

Not possible to get a view. You can have efficient alternative using a mask, as listed here - stackoverflow.com/a/40929672/3293881 — Divakar
– Divakar, Commented May 13, 2017 at 16:37

hpaulj · Accepted Answer · 2017-05-13 16:57:48Z

The attributes of an array consist of shape, strides and the data.

a[3:10] is a view because it can use the original data buffer, and just use a different shape (7,) and a different start point in the buffer.

a[some_ind] cannot be a view because [0,5,6,24] is not a regular pattern. It can't be expressed as shape, strides and data pointer. There for is has to have its own data copy.

In [534]: a=np.zeros(25,int)
In [535]: np.info(a)
class:  ndarray
shape:  (25,)
strides:  (4,)
itemsize:  4
aligned:  True
contiguous:  True
fortran:  True
data pointer: 0xafa5cc0
...
In [536]: np.info(a[3:10])
class:  ndarray
shape:  (7,)
strides:  (4,)
itemsize:  4
aligned:  True
contiguous:  True
fortran:  True
data pointer: 0xafa5ccc      # ccc v cc0
....
In [537]: np.info(a[[0,5,6,24]])
class:  ndarray
shape:  (4,)
strides:  (4,)
itemsize:  4
aligned:  True
contiguous:  True
fortran:  True
data pointer: 0xae9c038      # different
...

or looking at the data buffer pointer in decimal format:

In [538]: a.__array_interface__['data']
Out[538]: (184179904, False)
In [539]: a[3:10].__array_interface__['data']
Out[539]: (184179916, False)                   # 12 bytes larger
In [540]: a[[0,5,6]].__array_interface__['data']
Out[540]: (181099224, False)

Another way to put it is: the only alternative to copying elements of a, is to hang on the indexing array (or mask), and apply it each time you need those elements.

Collectives™ on Stack Overflow

How to create view of an irregularly spaced slice of a numpy array?

1 Answer 1

Comments

Your Answer

Linked

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

Comments

Your Answer

Sign up or log in

Post as a guest

Linked

Related