Faster way to patchify a picture to overlapping blocks
I'm looking for a FAST (and if possible memory afficiant) way to rewrite a function I crerated as part of Visual bag of words algorithm:
def get_pic_patches(pic, l, s): # "s" stands for stride
r, c = pic.shape
i, j = [0, 0]
x_range = list(range(0, r, s ) )
y_range = list(range(0, c , s ) )
patches = []
patches_location = []
for x in x_range: # without last two since it will exceed dimensions
for y in y_range: # without last two since it will exceed dimensions
if x+ l<= r and y+l <= c:
patch = pic[x:x + l , y:y + l ]
patches_location.append([x, y]) # patch location is the upper left pixel
patches.append( patch )
return patches, patches_location
it takes a grayscale image (NOT RGB!), desired patch length and stride value, and gives back all patches as a list of numpy array.
On other qestions, I found this:
def patchify(img, patch_shape):
img = np.ascontiguousarray(img) # won't make a copy if not needed
X, Y = img.shape
x, y = patch_shape
shape = ((X-x+1), (Y-y+1), x, y) # number of patches, patch_shape
strides = img.itemsize*np.array([Y, 1, Y, 1])
return np.lib.stride_tricks.as_strided(img, shape=shape, strides=strides)
in order to get to return a list, I used it like this:
def patchify(img, patch_shape):
img = np.ascontiguousarray(img) # won't make a copy if not needed
X, Y = img.shape
x, y = patch_shape
shape = ((X-x+1), (Y-y+1), x, y) # number of patches, patch_shape
strides = img.itemsize*np.array([Y, 1, Y, 1])
patches = np.lib.stride_tricks.as_strided(img, shape=shape, strides=strides)
a,b,c,d = patches.shape
patches = patches.reshape(((a*b),c,d))
patches = patches.tolist()
return
but this was actually much slower than my original function! another problem is that is only works with stride = 1, and I want to be able to use all sorts of stride values.