36

因此,在 numpy 数组中有用于获取对角线索引的内置函数,但我似乎无法弄清楚如何从右上角而不是左上角开始获取对角线。

这是从左上角开始的正常代码:

>>> import numpy as np
>>> array = np.arange(25).reshape(5,5)
>>> diagonal = np.diag_indices(5)
>>> array
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]])
>>> array[diagonal]
array([ 0,  6, 12, 18, 24])

那么如果我希望它返回,我该怎么做:

array([ 4,  8, 12, 16, 20])
4

3 回答 3

44

In [47]: np.diag(np.fliplr(array))
Out[47]: array([ 4,  8, 12, 16, 20])

或者

In [48]: np.diag(np.rot90(array))
Out[48]: array([ 4,  8, 12, 16, 20])

两者中,np.diag(np.fliplr(array))速度更快:

In [50]: %timeit np.diag(np.fliplr(array))
100000 loops, best of 3: 4.29 us per loop

In [51]: %timeit np.diag(np.rot90(array))
100000 loops, best of 3: 6.09 us per loop
于 2013-04-19T22:28:46.917 回答
3

这里有两个想法:

step = len(array) - 1

# This will make a copy
array.flat[step:-step:step]

# This will make a veiw
array.ravel()[step:-step:step]
于 2013-04-20T04:12:30.513 回答
3

这是使用 numpy 切片的简单方法。我个人觉得这对眼睛来说并不难(但承认这fliplr更具描述性!)。

为了突出这个示例对现有答案的贡献,我运行了相同的简单基准。

In [1]: import numpy as np

In [3]: X = np.random.randint(0, 10, (5, 5))

In [4]: X
Out[4]: 
array([[7, 2, 7, 3, 7],
       [8, 4, 5, 9, 6],
       [0, 2, 9, 0, 4],
       [8, 2, 1, 0, 3],
       [3, 1, 0, 7, 0]])

In [5]: Y = X[:, ::-1]

In [6]: Z1 = np.diag(Y)

In [7]: Z1
Out[7]: array([7, 9, 9, 2, 3])

现在与给出的当前最快的解决方案进行比较。

In [8]: step = len(X) - 1

In [9]: Z2 = np.take(X, np.arange(step, X.size-1, step))

In [10]: Z2
Out[10]: array([7, 9, 9, 2, 3])

In [11]: np.array_equal(Z1, Z2)
Out[11]: True

基准

In [12]: %timeit np.diag(X[:, ::-1])
1.92 µs ± 29.5 ns per loop (mean ± std. dev. of 7 runs, 1000000 loops each)

In [13]: %timeit step = len(X) - 1; np.take(X, np.arange(step, X.size-1, step))
2.21 µs ± 246 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)

初步比较表明,我的解决方案在复杂性上也是线性的,而使用第二个“步骤”解决方案则不是:

In [14]: big_X = np.random.randint(0, 10, (10000, 10000))

In [15]: %timeit np.diag(big_X[:, ::-1])
2.15 µs ± 96.3 ns per loop (mean ± std. dev. of 7 runs, 1000000 loops each)

In [16]: %timeit step = len(big_X) - 1; np.take(big_X, np.arange(step, big_X.size-1, step))
100 µs ± 1.85 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)

我通常使用这种方法来翻转图像(镜像它们),或者将(channels, height, width)opencv的格式转换为( height, width, channels)的格式。所以对于一个三维图像,它只是. 当然,您可以通过将零件放在所需的尺寸中来概括它以沿任何尺寸翻转。matplotlibflipped = image[:, :, ::-1]::-1

于 2018-07-06T10:14:56.820 回答