RuntimeWarning: Cannot provide views on a non-contiguous input array without copying

Question

when using skimage I get the following error:

win = skimage.util.view_as_windows(x, windowSize, windowShift)

C:\Program Files\Anaconda2\lib\site-packages\skimage\util\shape.py:247: RuntimeWarning: Cannot provide views on a non-contiguous input array without copying.
  warn(RuntimeWarning("Cannot provide views on a non-contiguous input "

as far I understood this is because x is a non contiguous array.

I think I solved the problem adding in my code np.ascontiguousarray as below:

win = skimage.util.view_as_windows(np.ascontiguousarray(x), windowSize, windowShift)

Is this the right thing to do? Note: I do it all the time I call this function from skimage...does it have any particular implication?

How do you create this `x`? I doubt if the `skimage` developers expect you to see this warning every time you use this util. Non contiguous arrays aren't common. — hpaulj, Apr 12 '17 at 10:25
I can use np.ascontigousarray() when I load x the first time... is it OK? — gabboshow, Apr 12 '17 at 11:03
It is somewhat strange that you do not get a contiguous array when loading. This might indicate a problem with loading the data, so better double-check if `x` is correct. That aside, converting to contiguous right after loading is exactly where you should most likely do it. The worst thing that can happen is that you run out of memory with large images because `np.ascontigousarray` needs to copy the data. — MB-F, Apr 12 '17 at 11:27
Matlab files are order `F` contiguous. Look at the `.flags`. So yes a one time `copy` or `ascontiguousarray` before further use in your code makes sense. — hpaulj, Apr 12 '17 at 13:48

hpaulj · Accepted Answer · 2017-04-12T15:08:12.360

In [44]: from scipy.io import loadmat
In [45]: d = loadmat('test7.mat')
In [46]: d
Out[46]: 
{'__globals__': [],
 '__header__': b'MATLAB 5.0 MAT-file, written by Octave 4.0.0, 2016-09-01 15:43:02 UTC',
 '__version__': '1.0',
 'x': array([[ 1.,  2.,  3.],
        [ 4.,  5.,  6.]])}

In [48]: np.info(d['x'])
class:  ndarray
shape:  (2, 3)
strides:  (8, 16)
itemsize:  8
aligned:  True
contiguous:  False
fortran:  True
data pointer: 0xabfa13d8
byteorder:  little
byteswap:  False
type: float64
In [49]:

or the FLAGS attribute:

In [52]: x.flags
Out[52]: 
  C_CONTIGUOUS : False
  F_CONTIGUOUS : True
  OWNDATA : False
  WRITEABLE : True
  ALIGNED : True
  UPDATEIFCOPY : False
In [54]: d['x'].flags['C_CONTIGUOUS']
Out[54]: False
In [55]: d['x'].flags['F_CONTIGUOUS']
Out[55]: True

np.ascontiguous just does

array(a, dtype, copy=False, order='C', ndmin=1)

It only does a copy (of the databuffer) if needed to get the right order. See the np.array docs for more details. x.copy() would make a copy regardless.

An ascontiguous call for all loadmat arrays makes sense if you are going to use them in skimage code that expects C contiguous arrays. view_as_windows is probably doing some sort of a striding tricks to make a (sliding) window.

RuntimeWarning: Cannot provide views on a non-contiguous input array without copying

1 Answers1

Linked