如何在没有嵌套循环的情况下从二维阵列切片三维阵列



我有四个矩阵,部分由彼此创建:

  • A是表示灰度图像堆栈的3D矩阵,其形状为(n, h, w)
  • B是也表示图像堆栈的3D矩阵,其中每个切片是根据A中的相应切片单独计算的,并且也是形状为(n, h, w)
  • C是2D矩阵,包含z方向上最大值为B的索引,形状为(h, w)
  • D是2D矩阵,其中来自a的值从某个切片复制,该切片由位置(x, y)处的C中的值指示

用循环实现的最小示例如下:

import numpy as np
A = np.random.randint(0, 255, size=(3, 4, 5))
B = np.random.randint(0, 255, size=(3, 4, 5))
C = np.argmax(B, axis=0)
D = np.zeros(C.shape, dtype=int)
for y in range(C.shape[0]):
for x in range(C.shape[1]):
D[y, x] = A[C[y, x], y, x]

> A
array([[[ 24,  84, 155,   8, 147],
[ 25,   4,  49, 195,  57],
[ 93,  76, 233,  83, 177],
[ 70, 211, 201, 132, 239]],
[[177, 144, 247, 251, 207],
[242, 148,  28,  40, 105],
[181,  28, 132,  94, 196],
[166, 121,  72,  14,  14]],
[[ 55, 254, 140, 142,  14],
[112,  28,  85, 112, 145],
[ 16,  72,  16, 248, 179],
[160, 235, 225,  14, 211]]])
> B
array([[[246,  14,  55, 163, 161],
[  3, 152, 128, 104, 203],
[ 43, 145,  59, 169, 242],
[106, 169,  31, 222, 240]],
[[ 41,  26, 239,  25,  65],
[ 47, 252, 205, 210, 138],
[194,  64, 135, 127, 101],
[ 63, 208, 179, 137,  59]],
[[112, 156, 183,  23, 253],
[ 35,   6, 233,  42, 100],
[ 66, 119, 102, 217,  64],
[ 82,  67, 135,   6,   8]]])
> C
array([[0, 2, 1, 0, 2],
[1, 1, 2, 1, 0],
[1, 0, 1, 2, 0],
[0, 1, 1, 0, 0]])
> D
array([[ 24, 254, 247,   8,  14],
[242, 148,  85,  40,  57],
[181,  76, 132, 248, 177],
[ 70, 121,  72, 132, 239]])

我的问题是:如何用CA进行切片,有效地消除嵌套的for循环?我最初的想法是将C扩展到一个3D布尔掩码,其中只有位置[c, y, x]设置为True,然后简单地将其与A元素相乘,并在z轴上求和。但我想不出一个没有循环的Python式实现(如果我知道如何编写的话,我可能就不需要再创建布尔掩码了(。

我发现最接近已经实现的函数是np.chose((,但它只需要32个元素就可以实现C.

这里的标准方法是将np.take_along_axis()np.expand_dims()结合使用(核心思想也在np.argmax()文档中介绍(:

np.take_along_axis(A, np.expand_dims(C, axis=0), axis=0).squeeze()

将所提出的方法与显式循环和基于np.ogrid()的方法进行比较,可以得到:

import numpy as np

def take_by_axis_loop_fix(arr, pos):
result = np.zeros(pos.shape, dtype=int)
for i in range(pos.shape[0]):
for j in range(pos.shape[1]):
result[i, j] = arr[pos[i, j], i, j]
return result

def take_by_axis_ogrid_fix(arr, pos):
i, j = np.ogrid[:pos.shape[0], :pos.shape[1]]
return arr[pos[i, j], i, j]

def take_by_axis_np(arr, pos, axis=0):
return np.take_along_axis(arr, np.expand_dims(pos, axis=axis), axis=axis).squeeze()

def take_by_axis_ogrid(arr, pos, axis=0):
ij = tuple(np.ogrid[tuple(slice(None, d, None) for d in pos.shape)])
ij = ij[:axis] + (pos[ij],) + ij[axis:]
return arr[ij]
A_ = np.random.randint(0, 255, size=(300, 400, 500))
B_ = np.random.randint(0, 255, size=(300, 400, 500))
C_ = np.argmax(B_, axis=0)
funcs = take_by_axis_loop_fix, take_by_axis_ogrid_fix, take_by_axis_ogrid, take_by_axis_np
for func in funcs:
print(func.__name__, np.all(func(A_, C_) == take_by_axis_loop_fix(A_, C_)))
%timeit func(A_, C_)
print()
# take_by_axis_loop_fix True
# 10 loops, best of 3: 114 ms per loop
# take_by_axis_ogrid_fix True
# 100 loops, best of 3: 5.94 ms per loop
# take_by_axis_ogrid True
# 100 loops, best of 3: 5.54 ms per loop
# take_by_axis_np True
# 100 loops, best of 3: 3.34 ms per loop

表明这是迄今为止提出的最有效的方法。还要注意,与_fix方法相反,基于np.take_along_axis()take_by_axis_ogrid()方法对于具有更高维度的输入基本上不变。

特别地,take_by_axis_ogrid()take_by_axis_ogrid_fix()axis不可知版本,本质上是第n个答案。

y, x = np.ogrid[:C.shape[0],:C.shape[1]]
D = A[C[y, x], y, x]

最新更新