使用numpy构建两个数组的所有组合的数组

Question 1

I’m trying to run over the parameters space of a 6 parameter function to study it’s numerical behavior before trying to do anything complex with it so I’m searching for a efficient way to do this.

My function takes float values given a 6-dim numpy array as input. What I tried to do initially was this:

First I created a function that takes 2 arrays and generate an array with all combinations of values from the two arrays

from numpy import *
def comb(a,b):
    c = []
    for i in a:
        for j in b:
            c.append(r_[i,j])
    return c

Then I used reduce() to apply that to m copies of the same array:

def combs(a,m):
    return reduce(comb,[a]*m)

And then I evaluate my function like this:

values = combs(np.arange(0,1,0.1),6)
for val in values:
    print F(val)

This works but it’s waaaay too slow. I know the space of parameters is huge, but this shouldn’t be so slow. I have only sampled 10⁶ (a million) points in this example and it took more than 15 seconds just to create the array values.

Do you know any more efficient way of doing this with numpy?

I can modify the way the function F takes it’s arguments if it’s necessary.

Question 2

In newer version of numpy (>1.8.x), numpy.meshgrid() provides a much faster implementation:

@pv’s solution

In [113]:

%timeit cartesian(([1, 2, 3], [4, 5], [6, 7]))
10000 loops, best of 3: 135 µs per loop
In [114]:

cartesian(([1, 2, 3], [4, 5], [6, 7]))

Out[114]:
array([[1, 4, 6],
       [1, 4, 7],
       [1, 5, 6],
       [1, 5, 7],
       [2, 4, 6],
       [2, 4, 7],
       [2, 5, 6],
       [2, 5, 7],
       [3, 4, 6],
       [3, 4, 7],
       [3, 5, 6],
       [3, 5, 7]])

numpy.meshgrid() use to be 2D only, now it is capable of ND. In this case, 3D:

In [115]:

%timeit np.array(np.meshgrid([1, 2, 3], [4, 5], [6, 7])).T.reshape(-1,3)
10000 loops, best of 3: 74.1 µs per loop
In [116]:

np.array(np.meshgrid([1, 2, 3], [4, 5], [6, 7])).T.reshape(-1,3)

Out[116]:
array([[1, 4, 6],
       [1, 5, 6],
       [2, 4, 6],
       [2, 5, 6],
       [3, 4, 6],
       [3, 5, 6],
       [1, 4, 7],
       [1, 5, 7],
       [2, 4, 7],
       [2, 5, 7],
       [3, 4, 7],
       [3, 5, 7]])

Note that the order of the final resultant is slightly different.

Question 3

Here’s a pure-numpy implementation. It’s about 5× faster than using itertools.


import numpy as np

def cartesian(arrays, out=None):
    """
    Generate a cartesian product of input arrays.

    Parameters
    ----------
    arrays : list of array-like
        1-D arrays to form the cartesian product of.
    out : ndarray
        Array to place the cartesian product in.

    Returns
    -------
    out : ndarray
        2-D array of shape (M, len(arrays)) containing cartesian products
        formed of input arrays.

    Examples
    --------
    >>> cartesian(([1, 2, 3], [4, 5], [6, 7]))
    array([[1, 4, 6],
           [1, 4, 7],
           [1, 5, 6],
           [1, 5, 7],
           [2, 4, 6],
           [2, 4, 7],
           [2, 5, 6],
           [2, 5, 7],
           [3, 4, 6],
           [3, 4, 7],
           [3, 5, 6],
           [3, 5, 7]])

    """

    arrays = [np.asarray(x) for x in arrays]
    dtype = arrays[0].dtype

    n = np.prod([x.size for x in arrays])
    if out is None:
        out = np.zeros([n, len(arrays)], dtype=dtype)

    m = n / arrays[0].size
    out[:,0] = np.repeat(arrays[0], m)
    if arrays[1:]:
        cartesian(arrays[1:], out=out[0:m, 1:])
        for j in xrange(1, arrays[0].size):
            out[j*m:(j+1)*m, 1:] = out[0:m, 1:]
    return out

Question 4

itertools.combinations is in general the fastest way to get combinations from a Python container (if you do in fact want combinations, i.e., arrangements WITHOUT repetitions and independent of order; that’s not what your code appears to be doing, but I can’t tell whether that’s because your code is buggy or because you’re using the wrong terminology).

If you want something different than combinations perhaps other iterators in itertools, product or permutations, might serve you better. For example, it looks like your code is roughly the same as:

for val in itertools.product(np.arange(0, 1, 0.1), repeat=6):
    print F(val)

All of these iterators yield tuples, not lists or numpy arrays, so if your F is picky about getting specifically a numpy array you’ll have to accept the extra overhead of constructing or clearing and re-filling one at each step.

Question 5

You can do something like this

import numpy as np

def cartesian_coord(*arrays):
    grid = np.meshgrid(*arrays)        
    coord_list = [entry.ravel() for entry in grid]
    points = np.vstack(coord_list).T
    return points

a = np.arange(4)  # fake data
print(cartesian_coord(*6*[a])

which gives

array([[0, 0, 0, 0, 0, 0],
   [0, 0, 0, 0, 0, 1],
   [0, 0, 0, 0, 0, 2],
   ..., 
   [3, 3, 3, 3, 3, 1],
   [3, 3, 3, 3, 3, 2],
   [3, 3, 3, 3, 3, 3]])

Question 6

The following numpy implementation should be approx. 2x the speed of the given answer:

def cartesian2(arrays):
    arrays = [np.asarray(a) for a in arrays]
    shape = (len(x) for x in arrays)

    ix = np.indices(shape, dtype=int)
    ix = ix.reshape(len(arrays), -1).T

    for n, arr in enumerate(arrays):
        ix[:, n] = arrays[n][ix[:, n]]

    return ix

Question 7

It looks like you want a grid to evaluate your function, in which case you can use numpy.ogrid (open) or numpy.mgrid (fleshed out):

import numpy
my_grid = numpy.mgrid[[slice(0,1,0.1)]*6]

Question 8

you can use np.array(itertools.product(a, b))

Question 9

Here’s yet another way, using pure NumPy, no recursion, no list comprehension, and no explicit for loops. It’s about 20% slower than the original answer, and it’s based on np.meshgrid.

def cartesian(*arrays):
    mesh = np.meshgrid(*arrays)  # standard numpy meshgrid
    dim = len(mesh)  # number of dimensions
    elements = mesh[0].size  # number of elements, any index will do
    flat = np.concatenate(mesh).ravel()  # flatten the whole meshgrid
    reshape = np.reshape(flat, (dim, elements)).T  # reshape and transpose
    return reshape

For example,

x = np.arange(3)
a = cartesian(x, x, x, x, x)
print(a)

gives

[[0 0 0 0 0]
 [0 0 0 0 1]
 [0 0 0 0 2]
 ..., 
 [2 2 2 2 0]
 [2 2 2 2 1]
 [2 2 2 2 2]]

Question 10

For a pure numpy implementation of Cartesian product of 1D arrays (or flat python lists), just use meshgrid(), roll the axes with transpose(), and reshape to the desired ouput:

 def cartprod(*arrays):
     N = len(arrays)
     return transpose(meshgrid(*arrays, indexing='ij'), 
                      roll(arange(N + 1), -1)).reshape(-1, N)

Note this has the convention of last axis changing fastest (“C style” or “row-major”).

In [88]: cartprod([1,2,3], [4,8], [100, 200, 300, 400], [-5, -4])
Out[88]: 
array([[  1,   4, 100,  -5],
       [  1,   4, 100,  -4],
       [  1,   4, 200,  -5],
       [  1,   4, 200,  -4],
       [  1,   4, 300,  -5],
       [  1,   4, 300,  -4],
       [  1,   4, 400,  -5],
       [  1,   4, 400,  -4],
       [  1,   8, 100,  -5],
       [  1,   8, 100,  -4],
       [  1,   8, 200,  -5],
       [  1,   8, 200,  -4],
       [  1,   8, 300,  -5],
       [  1,   8, 300,  -4],
       [  1,   8, 400,  -5],
       [  1,   8, 400,  -4],
       [  2,   4, 100,  -5],
       [  2,   4, 100,  -4],
       [  2,   4, 200,  -5],
       [  2,   4, 200,  -4],
       [  2,   4, 300,  -5],
       [  2,   4, 300,  -4],
       [  2,   4, 400,  -5],
       [  2,   4, 400,  -4],
       [  2,   8, 100,  -5],
       [  2,   8, 100,  -4],
       [  2,   8, 200,  -5],
       [  2,   8, 200,  -4],
       [  2,   8, 300,  -5],
       [  2,   8, 300,  -4],
       [  2,   8, 400,  -5],
       [  2,   8, 400,  -4],
       [  3,   4, 100,  -5],
       [  3,   4, 100,  -4],
       [  3,   4, 200,  -5],
       [  3,   4, 200,  -4],
       [  3,   4, 300,  -5],
       [  3,   4, 300,  -4],
       [  3,   4, 400,  -5],
       [  3,   4, 400,  -4],
       [  3,   8, 100,  -5],
       [  3,   8, 100,  -4],
       [  3,   8, 200,  -5],
       [  3,   8, 200,  -4],
       [  3,   8, 300,  -5],
       [  3,   8, 300,  -4],
       [  3,   8, 400,  -5],
       [  3,   8, 400,  -4]])

If you want to change the first axis fastest (“FORTRAN style” or “column-major”), just change the order parameter of reshape() like this: reshape((-1, N), order='F')

Question 11

Pandas merge offers a naive, fast solution to the problem:

# given the lists
x, y, z = [1, 2, 3], [4, 5], [6, 7]

# get dfs with same, constant index 
x = pd.DataFrame({'x': x}, index=np.repeat(0, len(x))
y = pd.DataFrame({'y': y}, index=np.repeat(0, len(y))
z = pd.DataFrame({'z': z}, index=np.repeat(0, len(z))

# get all permutations stored in a new df
df = pd.merge(x, pd.merge(y, z, left_index=True, righ_index=True),
              left_index=True, right_index=True)

使用numpy构建两个数组的所有组合的数组

问题：使用numpy构建两个数组的所有组合的数组

回答 0

回答 1

回答 2

回答 3

回答 4

回答 5

回答 6

回答 7

回答 8

回答 9

排行榜展示

Python 情人节超强技能导出微信聊天记录生成词云

你不得不知道的python超级文献批量搜索下载工具

7行代码 Python热力图可视化分析缺失数据处理

Python 流程图 — 一键转化代码为流程图

Python 优化—算出每条语句执行时间

你的10W块放哪里能赚最多钱？

文章展示

如何在matplotlib中设置纵横比？

ValueError：不支持的泡菜协议：3，python2泡菜无法加载python 3泡菜转储的文件？

如何将熊猫系列或索引转换为Numpy数组？

Python-唯一词典列表

Django使用get_user_model与settings.AUTH_USER_MODEL

Python3的“函数注释”有什么好的用处

使用numpy构建两个数组的所有组合的数组

问题：使用numpy构建两个数组的所有组合的数组

回答 0

回答 1

回答 2

回答 3

回答 4

回答 5

回答 6

回答 7

回答 8

回答 9

相关文章

排行榜展示

文章展示