Approach #1
We could use some broadcasting
here to get all the indices of all those sliding windows in one go and thus with indexing achieve a vectorized solution
. This is inspired by Efficient Implementation of im2col and col2im
.
Here’s the implementation –
def im2col_sliding_broadcasting(A, BSZ, stepsize=1):
# Parameters
M,N = A.shape
col_extent = N - BSZ[1] + 1
row_extent = M - BSZ[0] + 1
# Get Starting block indices
start_idx = np.arange(BSZ[0])[:,None]*N + np.arange(BSZ[1])
# Get offsetted indices across the height and width of input array
offset_idx = np.arange(row_extent)[:,None]*N + np.arange(col_extent)
# Get all actual indices & index into input array for final output
return np.take (A,start_idx.ravel()[:,None] + offset_idx.ravel()[::stepsize])
Approach #2
Using newly gained knowledge of NumPy array strides
that lets us create such sliding windows, we would have another efficient solution –
def im2col_sliding_strided(A, BSZ, stepsize=1):
# Parameters
m,n = A.shape
s0, s1 = A.strides
nrows = m-BSZ[0]+1
ncols = n-BSZ[1]+1
shp = BSZ[0],BSZ[1],nrows,ncols
strd = s0,s1,s0,s1
out_view = np.lib.stride_tricks.as_strided(A, shape=shp, strides=strd)
return out_view.reshape(BSZ[0]*BSZ[1],-1)[:,::stepsize]
Approach #3
The strided method listed in the previous approach has been incorporated into scikit-image
module for a less messier, like so –
from skimage.util import view_as_windows as viewW
def im2col_sliding_strided_v2(A, BSZ, stepsize=1):
return viewW(A, (BSZ[0],BSZ[1])).reshape(-1,BSZ[0]*BSZ[1]).T[:,::stepsize]
Sample runs –
In [106]: a # Input array
Out[106]:
array([[ 0, 1, 2, 3, 4],
[ 5, 6, 7, 8, 9],
[10, 11, 12, 13, 14],
[15, 16, 17, 18, 19]])
In [107]: im2col_sliding_broadcasting(a, (2,3))
Out[107]:
array([[ 0, 1, 2, 5, 6, 7, 10, 11, 12],
[ 1, 2, 3, 6, 7, 8, 11, 12, 13],
[ 2, 3, 4, 7, 8, 9, 12, 13, 14],
[ 5, 6, 7, 10, 11, 12, 15, 16, 17],
[ 6, 7, 8, 11, 12, 13, 16, 17, 18],
[ 7, 8, 9, 12, 13, 14, 17, 18, 19]])
In [108]: im2col_sliding_broadcasting(a, (2,3), stepsize=2)
Out[108]:
array([[ 0, 2, 6, 10, 12],
[ 1, 3, 7, 11, 13],
[ 2, 4, 8, 12, 14],
[ 5, 7, 11, 15, 17],
[ 6, 8, 12, 16, 18],
[ 7, 9, 13, 17, 19]])
Runtime test
In [183]: a = np.random.randint(0,255,(1024,1024))
In [184]: %timeit im2col_sliding(img, (8,8), skip=1)
...: %timeit im2col_sliding_broadcasting(img, (8,8), stepsize=1)
...: %timeit im2col_sliding_strided(img, (8,8), stepsize=1)
...: %timeit im2col_sliding_strided_v2(img, (8,8), stepsize=1)
...:
1 loops, best of 3: 1.29 s per loop
1 loops, best of 3: 226 ms per loop
10 loops, best of 3: 84.5 ms per loop
10 loops, best of 3: 111 ms per loop
In [185]: %timeit im2col_sliding(img, (8,8), skip=4)
...: %timeit im2col_sliding_broadcasting(img, (8,8), stepsize=4)
...: %timeit im2col_sliding_strided(img, (8,8), stepsize=4)
...: %timeit im2col_sliding_strided_v2(img, (8,8), stepsize=4)
...:
1 loops, best of 3: 1.31 s per loop
10 loops, best of 3: 104 ms per loop
10 loops, best of 3: 84.4 ms per loop
10 loops, best of 3: 109 ms per loop
Around 16x
speedup there with the strided method over the original loopy version!