有没有办法将ndy中的2d数组切成较小的2d数组？
示例
[[1,2,3,4], -> [[1,2] [3,4]
[5,6,7,8]] [5,6] [7,8]]
所以我基本上想要将2x4阵列减少为2个2x2阵列。寻找用于图像的通用解决方案。
答案 0 :(得分：63)
您应该可以使用reshape和swapaxes的某种组合将数组分解为“块”：
def blockshaped(arr, nrows, ncols):
"""
Return an array of shape (n, nrows, ncols) where
n * nrows * ncols = arr.size
If arr is a 2D array, the returned array should look like n subblocks with
each subblock preserving the "physical" layout of arr.
"""
h, w = arr.shape
assert h % nrows == 0, "{} rows is not evenly divisble by {}".format(h, nrows)
assert w % ncols == 0, "{} cols is not evenly divisble by {}".format(w, ncols)
return (arr.reshape(h//nrows, nrows, -1, ncols)
.swapaxes(1,2)
.reshape(-1, nrows, ncols))
转c
c = np.arange(24).reshape((4,6))
print(c)
# [[ 0 1 2 3 4 5]
# [ 6 7 8 9 10 11]
# [12 13 14 15 16 17]
# [18 19 20 21 22 23]]
进入
print(blockshaped(c, 2, 3))
# [[[ 0 1 2]
# [ 6 7 8]]
# [[ 3 4 5]
# [ 9 10 11]]
# [[12 13 14]
# [18 19 20]]
# [[15 16 17]
# [21 22 23]]]
我发布了inverse function, unblockshaped, here和N维概括here。这种推广可以更深入地了解这种算法背后的原因。
请注意，还有superbatfish's
blockwise_view。它安排了
不同格式的块（使用更多轴），但它具有（1）的优点
总是返回一个视图和（2）能够处理任何数组
尺寸。
答案 1 :(得分：5)
还有一些其他答案似乎已经非常适合您的特定情况，但是您的问题引起了我对可以使用内存高效解决方案的可能性的兴趣，该解决方案可以使用numpy支持的最大维数，我最终下午大部分时间都有可能的方法。 （方法本身相对简单，只是我还没有使用numpy支持的大多数真正奇特的功能，所以大部分时间花在研究上看看numpy有哪些可用，以及它可以做多少，所以我没有不得不这样做。）
def blockgen(array, bpa):
"""Creates a generator that yields multidimensional blocks from the given
array(_like); bpa is an array_like consisting of the number of blocks per axis
(minimum of 1, must be a divisor of the corresponding axis size of array). As
the blocks are selected using normal numpy slicing, they will be views rather
than copies; this is good for very large multidimensional arrays that are being
blocked, and for very large blocks, but it also means that the result must be
copied if it is to be modified (unless modifying the original data as well is
intended)."""
bpa = np.asarray(bpa) # in case bpa wasn't already an ndarray
# parameter checking
if array.ndim != bpa.size: # bpa doesn't match array dimensionality
raise ValueError("Size of bpa must be equal to the array dimensionality.")
if (bpa.dtype != np.int # bpa must be all integers
or (bpa < 1).any() # all values in bpa must be >= 1
or (array.shape % bpa).any()): # % != 0 means not evenly divisible
raise ValueError("bpa ({0}) must consist of nonzero positive integers "
"that evenly divide the corresponding array axis "
"size".format(bpa))
# generate block edge indices
rgen = (np.r_[:array.shape[i]+1:array.shape[i]//blk_n]
for i, blk_n in enumerate(bpa))
# build slice sequences for each axis (unfortunately broadcasting
# can't be used to make the items easy to operate over
c = [[np.s_[i:j] for i, j in zip(r[:-1], r[1:])] for r in rgen]
# Now to get the blocks; this is slightly less efficient than it could be
# because numpy doesn't like jagged arrays and I didn't feel like writing
# a ufunc for it.
for idxs in np.ndindex(*bpa):
blockbounds = tuple(c[j][idxs[j]] for j in range(bpa.size))
yield array[blockbounds]
答案 2 :(得分：4)
在我看来，这是numpy.split或某些变体的任务。
e.g。
a = np.arange(30).reshape([5,6]) #a.shape = (5,6)
a1 = np.split(a,3,axis=1)
#'a1' is a list of 3 arrays of shape (5,2)
a2 = np.split(a, [2,4])
#'a2' is a list of three arrays of shape (2,5), (2,5), (1,5)
如果您有NxN图像，则可以创建例如2个NxN / 2个子图像的列表，然后沿另一个轴划分它们。
答案 3 :(得分：2)
如果你想要一个解决方案，它也处理矩阵的情况 没有平分，你可以用这个：
<td>
<input class="form-control box-size clearForm" type="text" name="orderChargeDescription" placeholder="Charges" />
</td>
<td>
<input type="radio" name="orderWaChargeAddSub" value="ADD" checked>C
<input type="radio" name="orderWaChargeAddSub" value="SUB">U
</td>
<td>
<button type="submit" name="orderDispUp" class="btn btn-warning button-height-lp"><span class="glyphicon glyphicon-refresh" aria-hidden="true"></span></button>
<button type="submit" name="orderDispDn" class="btn btn-danger button-height-lp"><span class="glyphicon glyphicon-remove" aria-hidden="true"></span></button>
</td>
答案 4 :(得分：1)
现在它可以在大二维数组完全切成相同大小的子阵列时起作用。
代码低于切片
a ->array([[ 0, 1, 2, 3, 4, 5],
[ 6, 7, 8, 9, 10, 11],
[12, 13, 14, 15, 16, 17],
[18, 19, 20, 21, 22, 23]])
进入这个
block_array->
array([[[ 0, 1, 2],
[ 6, 7, 8]],
[[ 3, 4, 5],
[ 9, 10, 11]],
[[12, 13, 14],
[18, 19, 20]],
[[15, 16, 17],
[21, 22, 23]]])
p ang q确定块大小
<强>代码
a = arange(24)
a = a.reshape((4,6))
m = a.shape[0] #image row size
n = a.shape[1] #image column size
p = 2 #block row size
q = 3 #block column size
block_array = []
previous_row = 0
for row_block in range(blocks_per_row):
previous_row = row_block * p
previous_column = 0
for column_block in range(blocks_per_column):
previous_column = column_block * q
block = a[previous_row:previous_row+p,previous_column:previous_column+q]
block_array.append(block)
block_array = array(block_array)
答案 5 :(得分：1)
你问practically the same as this one。您可以将{one}与np.ndindex()和reshape()：
def cutter(a, r, c):
lenr = a.shape[0]/r
lenc = a.shape[1]/c
np.array([a[i*r:(i+1)*r,j*c:(j+1)*c] for (i,j) in np.ndindex(lenr,lenc)]).reshape(lenr,lenc,r,c)
创建您想要的结果：
a = np.arange(1,9).reshape(2,1)
#array([[1, 2, 3, 4],
# [5, 6, 7, 8]])
cutter( a, 1, 2 )
#array([[[[1, 2]],
# [[3, 4]]],
# [[[5, 6]],
# [[7, 8]]]])
答案 6 :(得分：0)
这是一个基于unutbu答案的解决方案，用于处理矩阵无法平分的情况。在这种情况下，它将在使用某些插值之前调整矩阵的大小。你需要OpenCV。请注意，我必须交换ncols和nrows以使其有效，但没有想到原因。
import numpy as np
import cv2
import math
def blockshaped(arr, r_nbrs, c_nbrs, interp=cv2.INTER_LINEAR):
"""
arr a 2D array, typically an image
r_nbrs numbers of rows
r_cols numbers of cols
"""
arr_h, arr_w = arr.shape
size_w = int( math.floor(arr_w // c_nbrs) * c_nbrs )
size_h = int( math.floor(arr_h // r_nbrs) * r_nbrs )
if size_w != arr_w or size_h != arr_h:
arr = cv2.resize(arr, (size_w, size_h), interpolation=interp)
nrows = int(size_w // r_nbrs)
ncols = int(size_h // c_nbrs)
return (arr.reshape(r_nbrs, ncols, -1, nrows)
.swapaxes(1,2)
.reshape(-1, ncols, nrows))
答案 7 :(得分：0)
TheMeaningfulEngineer的答案进行了一些细微的增强，该答案处理了无法将大2d数组 完美地切成相等大小的子数组的情况
def blockfy(a, p, q):
'''
Divides array a into subarrays of size p-by-q
p: block row size
q: block column size
'''
m = a.shape[0] #image row size
n = a.shape[1] #image column size
# pad array with NaNs so it can be divided by p row-wise and by q column-wise
bpr = ((m-1)//p + 1) #blocks per row
bpc = ((n-1)//q + 1) #blocks per column
M = p * bpr
N = q * bpc
A = np.nan* np.ones([M,N])
A[:a.shape[0],:a.shape[1]] = a
block_list = []
previous_row = 0
for row_block in range(bpc):
previous_row = row_block * p
previous_column = 0
for column_block in range(bpr):
previous_column = column_block * q
block = A[previous_row:previous_row+p, previous_column:previous_column+q]
# remove nan columns and nan rows
nan_cols = np.all(np.isnan(block), axis=0)
block = block[:, ~nan_cols]
nan_rows = np.all(np.isnan(block), axis=1)
block = block[~nan_rows, :]
## append
if block.size:
block_list.append(block)
return block_list
示例：
a = np.arange(25)
a = a.reshape((5,5))
out = blockfy(a, 2, 3)
a->
array([[ 0, 1, 2, 3, 4],
[ 5, 6, 7, 8, 9],
[10, 11, 12, 13, 14],
[15, 16, 17, 18, 19],
[20, 21, 22, 23, 24]])
out[0] ->
array([[0., 1., 2.],
[5., 6., 7.]])
out[1]->
array([[3., 4.],
[8., 9.]])
out[-1]->
array([[23., 24.]])
答案 8 :(得分：0)
a = np.random.randint(1, 9, size=(9,9))
out = [np.hsplit(x, 3) for x in np.vsplit(a,3)]
print(a)
print(out)
收益
[[7 6 2 4 4 2 5 2 3]
[2 3 7 6 8 8 2 6 2]
[4 1 3 1 3 8 1 3 7]
[6 1 1 5 7 2 1 5 8]
[8 8 7 6 6 1 8 8 4]
[6 1 8 2 1 4 5 1 8]
[7 3 4 2 5 6 1 2 7]
[4 6 7 5 8 2 8 2 8]
[6 6 5 5 6 1 2 6 4]]
[[array([[7, 6, 2],
[2, 3, 7],
[4, 1, 3]]), array([[4, 4, 2],
[6, 8, 8],
[1, 3, 8]]), array([[5, 2, 3],
[2, 6, 2],
[1, 3, 7]])], [array([[6, 1, 1],
[8, 8, 7],
[6, 1, 8]]), array([[5, 7, 2],
[6, 6, 1],
[2, 1, 4]]), array([[1, 5, 8],
[8, 8, 4],
[5, 1, 8]])], [array([[7, 3, 4],
[4, 6, 7],
[6, 6, 5]]), array([[2, 5, 6],
[5, 8, 2],
[5, 6, 1]]), array([[1, 2, 7],
[8, 2, 8],
[2, 6, 4]])]]
答案 9 :(得分：0)
我发布了解决方案。请注意，此代码实际上并没有创建原始数组的副本，因此它适用于大数据。此外，如果无法均匀分割数组，也不会崩溃（但是您可以通过删除ceil并检查v_slices和h_slices是否在不休息的情况下被分割而轻松地添加条件）。 / p>
import numpy as np
from math import ceil
a = np.arange(9).reshape(3, 3)
p, q = 2, 2
width, height = a.shape
v_slices = ceil(width / p)
h_slices = ceil(height / q)
for h in range(h_slices):
for v in range(v_slices):
block = a[h * p : h * p + p, v * q : v * q + q]
# do something with a block
此代码更改（或更准确地说，使您可以直接访问数组的一部分）：
[[0 1 2]
[3 4 5]
[6 7 8]]
对此：
[[0 1]
[3 4]]
[[2]
[5]]
[[6 7]]
[[8]]
如果需要实际的副本，Aenaon code是您要寻找的。
如果您确定可以将大数组平均分割，则可以使用numpy splitting工具。
答案 10 :(得分：0)
要添加到@Aenaon答案及其blockfy函数中，如果您正在使用 COLOR IMAGES / 3D ARRAY （色彩图像/ 3D阵列），这是我的管道，可为3通道输入创建224 x 224的作物
def blockfy(a, p, q):
'''
Divides array a into subarrays of size p-by-q
p: block row size
q: block column size
'''
m = a.shape[0] #image row size
n = a.shape[1] #image column size
# pad array with NaNs so it can be divided by p row-wise and by q column-wise
bpr = ((m-1)//p + 1) #blocks per row
bpc = ((n-1)//q + 1) #blocks per column
M = p * bpr
N = q * bpc
A = np.nan* np.ones([M,N])
A[:a.shape[0],:a.shape[1]] = a
block_list = []
previous_row = 0
for row_block in range(bpc):
previous_row = row_block * p
previous_column = 0
for column_block in range(bpr):
previous_column = column_block * q
block = A[previous_row:previous_row+p, previous_column:previous_column+q]
# remove nan columns and nan rows
nan_cols = np.all(np.isnan(block), axis=0)
block = block[:, ~nan_cols]
nan_rows = np.all(np.isnan(block), axis=1)
block = block[~nan_rows, :]
## append
if block.size:
block_list.append(block)
return block_list
然后在上方扩展到
for file in os.listdir(path_to_crop): ### list files in your folder
img = io.imread(path_to_crop + file, as_gray=False) ### open image
r = blockfy(img[:,:,0],224,224) ### crop blocks of 224 x 224 for red channel
g = blockfy(img[:,:,1],224,224) ### crop blocks of 224 x 224 for green channel
b = blockfy(img[:,:,2],224,224) ### crop blocks of 224 x 224 for blue channel
for x in range(0,len(r)):
img = np.array((r[x],g[x],b[x])) ### combine each channel into one patch by patch
img = img.astype(np.uint8) ### cast back to proper integers
img_swap = img.swapaxes(0, 2) ### need to swap axes due to the way things were proceesed
img_swap_2 = img_swap.swapaxes(0, 1) ### do it again
Image.fromarray(img_swap_2).save(path_save_crop+str(x)+"bounding" + file,
format = 'jpeg',
subsampling=0,
quality=100) ### save patch with new name etc