删除特定索引(我从矩阵中删除了16和21)

import numpy as np
mat = np.arange(12,26)
a = [4,9]
del_map = np.delete(mat, a)
del_map.reshape(3,4)

输出:

array([[12, 13, 14, 15],
      [17, 18, 19, 20],
      [22, 23, 24, 25]])

删除特定索引(我从矩阵中删除了16和21)

import numpy as np
mat = np.arange(12,26)
a = [4,9]
del_map = np.delete(mat, a)
del_map.reshape(3,4)

输出:

array([[12, 13, 14, 15],
      [17, 18, 19, 20],
      [22, 23, 24, 25]])

如果不知道索引，就不能使用logical_and

x = 10*np.random.randn(1,100)
low = 5
high = 27
x[0,np.logical_and(x[0,:]>low,x[0,:]<high)]

如果没有想要删除的元素的索引，可以使用numpy提供的in1d函数。

如果一个一维数组的元素也存在于另一个数组中，则该函数返回True。要删除元素，只需对该函数返回的值求负即可。

注意，这个方法保持原始数组的顺序。

In [1]: import numpy as np

        a = np.array([1, 2, 3, 4, 5, 6, 7, 8, 9])
        rm = np.array([3, 4, 7])
        # np.in1d return true if the element of `a` is in `rm`
        idx = np.in1d(a, rm)
        idx

Out[1]: array([False, False,  True,  True, False, False,  True, False, False])

In [2]: # Since we want the opposite of what `in1d` gives us, 
        # you just have to negate the returned value
        a[~idx]

Out[2]: array([1, 2, 5, 6, 8, 9])

有一个numpy内置函数可以帮助实现这一点。

import numpy as np
>>> a = np.array([1, 2, 3, 4, 5, 6, 7, 8, 9])
>>> b = np.array([3,4,7])
>>> c = np.setdiff1d(a,b)
>>> c
array([1, 2, 5, 6, 8, 9])

我不是一个麻木的人，我试了一下:

>>> import numpy as np
>>> import itertools
>>> 
>>> a = np.array([1,2,3,4,5,6,7,8,9])
>>> index=[2,3,6]
>>> a = np.array(list(itertools.compress(a, [i not in index for i in range(len(a))])))
>>> a
array([1, 2, 5, 6, 8, 9])

根据我的测试，这优于numpy.delete()。我不知道为什么会这样，也许是因为初始数组的大小较小?

python -m timeit -s "import numpy as np" -s "import itertools" -s "a = np.array([1,2,3,4,5,6,7,8,9])" -s "index=[2,3,6]" "a = np.array(list(itertools.compress(a, [i not in index for i in range(len(a))])))"
100000 loops, best of 3: 12.9 usec per loop

python -m timeit -s "import numpy as np" -s "a = np.array([1,2,3,4,5,6,7,8,9])" -s "index=[2,3,6]" "np.delete(a, index)"
10000 loops, best of 3: 108 usec per loop

这是一个相当显著的差异(与我预期的方向相反)，有人知道为什么会这样吗?

更奇怪的是，传递numpy.delete()一个列表的性能比遍历列表并给它单个索引的性能更差。

python -m timeit -s "import numpy as np" -s "a = np.array([1,2,3,4,5,6,7,8,9])" -s "index=[2,3,6]" "for i in index:" "    np.delete(a, i)"
10000 loops, best of 3: 33.8 usec per loop

编辑:这似乎与数组的大小有关。对于大型数组，numpy.delete()要快得多。

python -m timeit -s "import numpy as np" -s "import itertools" -s "a = np.array(list(range(10000)))" -s "index=[i for i in range(10000) if i % 2 == 0]" "a = np.array(list(itertools.compress(a, [i not in index for i in range(len(a))])))"
10 loops, best of 3: 200 msec per loop

python -m timeit -s "import numpy as np" -s "a = np.array(list(range(10000)))" -s "index=[i for i in range(10000) if i % 2 == 0]" "np.delete(a, index)"
1000 loops, best of 3: 1.68 msec per loop

显然，这一切都是相当无关紧要的，因为您应该始终保持清晰，避免重复工作，但我发现它有点有趣，所以我想我就把它留在这里。