协调 np.fromiter 和 Python 中的多维数组

Question

我喜欢使用 numpy 中的 np.fromiter，因为它是一种构建 np.array 对象的资源惰性方法。不过好像不支持多维数组，多维数组也挺好用的

import numpy as np

def fun(i):
    """ A function returning 4 values of the same type.
    """
    return tuple(4*i + j for j in range(4))

# Trying to create a 2-dimensional array from it:
a = np.fromiter((fun(i) for i in range(5)), '4i', 5) # fails

# This function only seems to work for 1D array, trying then:
a = np.fromiter((fun(i) for i in range(5)),
        [('', 'i'), ('', 'i'), ('', 'i'), ('', 'i')], 5) # painful

# .. `a` now looks like a 2D array but it is not:
a.transpose() # doesn't work as expected
a[0, 1] # too many indices (of course)
a[:, 1] # don't even think about it

如何让 a 成为一个多维数组，同时保持这种基于生成器的惰性构造？

Answer 1

就其本身而言，np.fromiter only supports constructing 1D arrays, and as such, it expects an iterable that will yield individual values rather than tuples/lists/sequences etc. One way to work around this limitation would be to use itertools.chain.from_iterable 懒惰地 'unpack' 将生成器表达式的输出转换为单个一维值序列：

import numpy as np
from itertools import chain

def fun(i):
    return tuple(4*i + j for j in range(4))

a = np.fromiter(chain.from_iterable(fun(i) for i in range(5)), 'i', 5 * 4)
a.shape = 5, 4

print(repr(a))
# array([[ 0,  1,  2,  3],
#        [ 4,  5,  6,  7],
#        [ 8,  9, 10, 11],
#        [12, 13, 14, 15],
#        [16, 17, 18, 19]], dtype=int32)

协调 np.fromiter 和 Python 中的多维数组

Reconcile np.fromiter and multidimensional arrays in Python

python

arrays

numpy

lazy-evaluation

multidimensional-array