NumPy 'as_strided' for strided sliding window over RGBA image (3D array)

Question

I've done research and looked over several questions on SO to figure out the proper set-up of strides and shape, but it's been giving me problems.

I have an image array of shape (250, 250, 4) (example) and need to use as_strided to create a sliding window of size (50, 50) across all channels.

Assuming:

x = np.random.random((250, 250, 4)) * 255
image = x.astype(np.uint8)
strided = np.lib.stride_tricks.as_strided(image, shape=(?, ?, ?, ?, ?), strides=(?, ?, ?, ?, ?))

Meaning strided[0, 0, 0] would return the R channel of shape (50,50) equivalent to image[0:50, 0:50, 0], and strided[1, 1, 2] would return the B channel equivalent to image[50:100, 50:100, 2]

I'm needing to use as_strided because I am given a memory view of a massive image (multiple GB in size), the above is just an example to illustrate the problem.

Answer 1

A couple of possibilities with sliding_window:

In [9]: arr = np.ones((250,250,4),int)
In [10]: x = np.lib.stride_tricks.sliding_window_view(arr,(50,50),(0,1))
In [11]: x.shape
Out[11]: (201, 201, 4, 50, 50)
In [12]: x = np.lib.stride_tricks.sliding_window_view(arr,(50,50,4))
In [13]: x.shape
Out[13]: (201, 201, 1, 50, 50, 4)

In the first case, strides is: (8000, 32, 8, 8000, 32)

In the second, (8000, 32, 8, 8000, 32, 8)