消除Python和Numpy构造中的for循环

时间:2012-01-02 13:03:01

标签: python for-loop numpy

我正在寻找Python和/或Numpy矢量化的方法,以消除以下内容的for循环使用:

for i in list_range_values:
    v[list_list_values[i]] += list_comp_values[i]

其中:

  • list_range_values是一个整数值的Python列表,例如。 [1,3,5],范围(0,R-1,1)

  • list_comp_values是数值的Python列表,例如。 [0.7,9.8,1.2,5,10,11.7,6,0.2]使len(list_comp_values)= R

  • v是长度为V的numpy向量,使得R可以是<,=,>比V

  • list_list_values是一个列表的Python列表(每个列表包含不同数量的整数值,例如[[3,6,7],[5,7,11,25,99],[8,45] ],[4,7,8],[0,1],[21,31,41],[9,11,22,33,44],[17,19]])从范围(0, V-1,1)和len(list_list_values)= R

例如

for i in list_range_values (= [1, 3, 5]):
    i=1: v[[5, 7, 11, 25, 99]] += list_comp_values[1] (= 9.8)
    i=3: v[[4, 7, 8]] += list_comp_values[3] (= 5)
    i=5: v[[21, 31, 41]] += list_comp_values[5] (= 11.7)

是否有可用的方法可以消除for循环?

Cython,Scipy / Weave / Blitz和C模块是替代解决方案,但是想要确定是否首先存在Numpy矢量化答案。

3 个答案:

答案 0 :(得分:6)

虽然它经常导致大规模的加速消除for循环并利用numpy内置插件/矢量化。我只想指出情况并非总是如此。定时简单的for循环与更多涉及的矢量化,不会给你一个大的加速,而且更冗长。需要考虑的事情:

from timeit import Timer

setstr="""import numpy as np
import itertools
import random

Nlists = 1000
total_lists = 5000
outsz = 100
maxsublistsz = 100


# create random list of lists
list_range_values = random.sample(xrange(total_lists),Nlists)
list_list_values = [random.sample(xrange(outsz),np.random.randint(1,maxsublistsz)) for k in xrange(total_lists)]

list_comp_values = 10*np.random.uniform(size=(total_lists,))

v = np.zeros((outsz,))

def indices(start, end):
    lens = end - start
    np.cumsum(lens, out=lens)
    i = np.ones(lens[-1], dtype=int)
    i[0] = start[0]
    i[lens[:-1]] += start[1:]
    i[lens[:-1]] -= end[:-1]
    np.cumsum(i, out=i)
    return i

def sum_by_group(values, groups):
    order = np.argsort(groups)
    groups = groups[order]
    values = values[order]
    values.cumsum(out=values)
    index = np.ones(len(groups), 'bool')
    index[:-1] = groups[1:] != groups[:-1]
    values = values[index]
    groups = groups[index]
    values[1:] = np.diff(values)
    return values, groups


"""

method1="""
list_list_lens = np.array(map(len, list_list_values))
comp_vals_expanded = np.repeat(list_comp_values, list_list_lens)

list_vals_flat = np.fromiter(itertools.chain.from_iterable(list_list_values),dtype=int)
list_list_starts = np.concatenate(([0], np.cumsum(list_list_lens)[:-1]))

toadd = indices(list_list_starts[list_range_values],(list_list_starts + list_list_lens)[list_range_values])

v[list_vals_flat[toadd]] += comp_vals_expanded[toadd]
"""

method2="""
for k in list_range_values:
    v[list_list_values[k]] += list_comp_values[k]

"""

method3="""
llv = [list_list_values[i] for i in list_range_values]
lcv = [list_comp_values[i] for i in list_range_values]
counts = map(len, llv)
indices = np.concatenate(llv)
values = np.repeat(lcv, counts)

totals, indices_unique = sum_by_group(values, indices)
v[indices_unique] += totals
"""


t1 = Timer(method1,setup=setstr).timeit(100)
print t1

t2 = Timer(method2,setup=setstr).timeit(100)
print t2

t3 = Timer(method3,setup=setstr).timeit(100)
print t3

对于列表中的大量元素:

方法1 :(不用于循环-jterrace)1.43秒

方法2 :( for loop)4.62秒

方法3 :(不用于循环 - bago)2.99秒

对于少量列表(将Nlists更改为10),for循环明显快于jterrace的解决方案:

方法1 :(不用于循环-jterrace)1.05秒

方法2 :( for循环)0.045秒

方法3 :(不用于循环 - bago)0.041秒

这不是要敲@jterrace或@ bago的解决方案,这是非常优雅的。相反,它要指出的是,通常简单的for循环不能很好地执行。

答案 1 :(得分:2)

使用您的示例输入:

>>> list_list_values = [[3, 6, 7], [5, 7, 11, 25, 99], [8, 45], [4, 7, 8], 
                        [0, 1], [21, 31, 41], [9, 11, 22, 33, 44], [17, 19]]
>>> list_comp_values = [0.7, 9.8, 1.2, 5, 10, 11.7, 6, 0.2]
>>> list_range_values = [1, 3, 5]

首先,一些发电机恶作剧:

>>> indices_weights = ((list_list_values[i], list_comp_values[i]) 
                       for i in list_range_values)
>>> flat_indices_weights = ((i, weight) for indices, weight in indices_weights 
                             for i in indices)

现在我们将数据传递给numpy。我无法弄清楚如何从迭代器生成rec.array,所以我不得不将上面的生成器转换为列表。也许有办法避免......

>>> i_w = numpy.rec.array(list(flat_indices_weights),       
                          dtype=[('i', int), ('weight', float)])
>>> numpy.histogram(i_w['i'], bins=range(0, 100), weights=i_w['weight'])
(array([  0. ,   0. ,   0. ,   0. ,   5. ,   9.8,   0. ,  14.8,   5. ,
         0. ,   0. ,   9.8,   0. ,   0. ,   0. ,   0. ,   0. ,   0. ,
         0. ,   0. ,   0. ,  11.7,   0. ,   0. ,   0. ,   9.8,   0. ,
         0. ,   0. ,   0. ,   0. ,  11.7,   0. ,   0. ,   0. ,   0. ,
         0. ,   0. ,   0. ,   0. ,   0. ,  11.7,   0. ,   0. ,   0. ,
         0. ,   0. ,   0. ,   0. ,   0. ,   0. ,   0. ,   0. ,   0. ,
         0. ,   0. ,   0. ,   0. ,   0. ,   0. ,   0. ,   0. ,   0. ,
         0. ,   0. ,   0. ,   0. ,   0. ,   0. ,   0. ,   0. ,   0. ,
         0. ,   0. ,   0. ,   0. ,   0. ,   0. ,   0. ,   0. ,   0. ,
         0. ,   0. ,   0. ,   0. ,   0. ,   0. ,   0. ,   0. ,   0. ,
         0. ,   0. ,   0. ,   0. ,   0. ,   0. ,   0. ,   0. ,   9.8]), 
 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, 25, 26, 27, 28, 29, 30, 31, 32, 33,
       34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50,
       51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67,
       68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84,
       85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99]))

我有一点时间用自己的几个来跟进JoshAdel的测试。到目前为止,最快的解决方案是使用Bago的设置,但用sum_by_group函数取代histogram函数。以下是我获得的数字(已更新)

Method1(jterrace):2.65

方法2(for循环):2.25

Method3(Bago):1.14

Method4(直方图):2.82

方法5(3/4组合):1.07

请注意,如此处所实现的,第一种方法根据我的测试给出了错误的结果。我没有时间弄清问题是什么。我的测试代码如下;它只是轻轻地调整了JoshAdel的原始代码,但为了方便我在这里发布了它。 (更新后包括Bago的评论,并且有点说不清楚。)

from timeit import Timer

setstr="""import numpy as np
import itertools
import random

Nlists = 1000
total_lists = 5000
outsz = 100
maxsublistsz = 100

# create random list of lists
list_range_values = random.sample(xrange(total_lists),Nlists)
list_list_values = [random.sample(xrange(outsz),np.random.randint(1,maxsublistsz)) for k in xrange(total_lists)]

list_comp_values = list(10*np.random.uniform(size=(total_lists,)))

v = np.zeros((outsz,))

def indices(start, end):
    lens = end - start
    np.cumsum(lens, out=lens)
    i = np.ones(lens[-1], dtype=int)
    i[0] = start[0]
    i[lens[:-1]] += start[1:]
    i[lens[:-1]] -= end[:-1]
    np.cumsum(i, out=i)
    return i

def sum_by_group(values, groups):
    order = np.argsort(groups)
    groups = groups[order]
    values = values[order]
    values.cumsum(out=values)
    index = np.ones(len(groups), 'bool')
    index[:-1] = groups[1:] != groups[:-1]
    values = values[index]
    groups = groups[index]
    values[1:] = np.diff(values)
    return values, groups


"""

setstr_test = setstr + "\nprint_v = True\n"

method1="""
list_list_lens = np.array(map(len, list_list_values))
comp_vals_expanded = np.repeat(list_comp_values, list_list_lens)

list_vals_flat = np.fromiter(itertools.chain.from_iterable(list_list_values),dtype=int)
list_list_starts = np.concatenate(([0], np.cumsum(list_list_lens)[:-1]))

toadd = indices(list_list_starts[list_range_values],(list_list_starts + list_list_lens)[list_range_values])

v[list_vals_flat[toadd]] += comp_vals_expanded[toadd]
"""

method2="""
for k in list_range_values:
    v[list_list_values[k]] += list_comp_values[k]
"""

method3="""
llv = [np.fromiter(list_list_values[i], 'int') for i in list_range_values]
lcv = [list_comp_values[i] for i in list_range_values]
counts = map(len, llv)
indices = np.concatenate(llv)
values = np.repeat(lcv, counts)

totals, indices_unique = sum_by_group(values, indices)
v[indices_unique] += totals
"""

method4="""
indices_weights = ((list_list_values[i], list_comp_values[i]) for i in list_range_values)
flat_indices_weights = ((i, weight) for indices, weight in indices_weights for i in indices)
i_w = np.rec.array(list(flat_indices_weights), dtype=[('i', 'i'), ('weight', 'd')])
v += np.histogram(i_w['i'], bins=range(0, outsz + 1), weights=i_w['weight'], new=True)[0]
"""

method5="""
llv = [np.fromiter(list_list_values[i], 'int') for i in list_range_values]
lcv = [list_comp_values[i] for i in list_range_values]
counts = map(len, llv)
indices = np.concatenate(llv)
values = np.repeat(lcv, counts)

v += np.histogram(indices, bins=range(0, outsz + 1), weights=values, new=True)[0]
"""


t1 = Timer(method1,setup=setstr).timeit(100)
print t1

t2 = Timer(method2,setup=setstr).timeit(100)
print t2

t3 = Timer(method3,setup=setstr).timeit(100)
print t3

t4 = Timer(method4,setup=setstr).timeit(100)
print t4

t5 = Timer(method5,setup=setstr).timeit(100)
print t5

exec(setstr_test + method1 + "\nprint v\n")
exec("\nv = np.zeros((outsz,))\n" + method2 + "\nprint v\n")
exec("\nv = np.zeros((outsz,))\n" + method3 + "\nprint v\n")
exec("\nv = np.zeros((outsz,))\n" + method4 + "\nprint v\n")
exec("\nv = np.zeros((outsz,))\n" + method5 + "\nprint v\n")

答案 2 :(得分:1)

首先,设置你给出的变量:

import numpy as np
list_range_values = [1, 3, 5]
list_list_values = [[3, 6, 7], [5, 7, 11, 25, 99], [8, 45],
                    [4, 7, 8], [0, 1], [21, 31, 41]]
list_comp_values = [0.7, 9.8, 1.2, 5, 10, 11.7]
v = np.arange(100, dtype=float)

接下来,list_list_valueslist_comp_values需要展平,因此它们是连续的:

list_list_lens = np.array(map(len, list_list_values))
comp_vals_expanded = np.repeat(list_comp_values, list_list_lens)
import itertools
list_vals_flat = np.fromiter(itertools.chain.from_iterable(list_list_values),
                             dtype=int)

然后,需要每个子阵列的起始索引:

list_list_starts = np.concatenate(([0], np.cumsum(list_list_lens)[:-1]))

现在我们有了起始值和结束值,我们可以使用indices function from this question来获取一个选择器索引数组:

def indices(start, end):
    lens = end - start
    np.cumsum(lens, out=lens)
    i = np.ones(lens[-1], dtype=int)
    i[0] = start[0]
    i[lens[:-1]] += start[1:]
    i[lens[:-1]] -= end[:-1]
    np.cumsum(i, out=i)
    return i

toadd = indices(list_list_starts[list_range_values],
                (list_list_starts + list_list_lens)[list_range_values])

现在我们已经完成了所有这些魔术,可以像这样添加数组:

v[list_vals_flat[toadd]] += comp_vals_expanded[toadd]