Python 列表理解与生成器表达式';结果如何?

Python 列表理解与生成器表达式';结果如何?,python,list,list-comprehension,timeit,generator-expression,Python,List,List Comprehension,Timeit,Generator Expression,我在回答这个问题时,我更喜欢这里的生成器表达式,并使用了这个表达式,我认为它会更快,因为生成器不需要首先创建整个列表: >>> lis=[['a','b','c'],['d','e','f']] >>> 'd' in (y for x in lis for y in x) True Levon在他的书中使用了列表理解 但当我计算时间时,这些LC的结果比发电机快: ~$ python -m timeit -s "lis=[['a','b','c'],['d',

我在回答这个问题时,我更喜欢这里的生成器表达式,并使用了这个表达式,我认为它会更快,因为生成器不需要首先创建整个列表:

>>> lis=[['a','b','c'],['d','e','f']]
>>> 'd' in (y for x in lis for y in x)
True
Levon在他的书中使用了列表理解

但当我计算时间时,这些LC的结果比发电机快:

~$ python -m timeit -s "lis=[['a','b','c'],['d','e','f']]" "'d' in (y for x in lis for y in x)"
    100000 loops, best of 3: 2.36 usec per loop
~$ python -m timeit -s "lis=[['a','b','c'],['d','e','f']]" "'d' in [y for x in lis for y in x]"
    100000 loops, best of 3: 1.51 usec per loop
然后我增加了列表的大小,并再次计时:

lis=[['a','b','c'],['d','e','f'],[1,2,3],[4,5,6],[7,8,9],[10,11,12],[13,14,15],[16,17,18]]
这次搜索
'd'
生成器的速度比LC快,但当我搜索中间元素(11)和最后一个元素时,LC再次击败生成器表达式,我不明白为什么

~$ python -m timeit -s "lis=[['a','b','c'],['d','e','f'],[1,2,3],[4,5,6],[7,8,9],[10,11,12],[13,14,15],[16,17,18]]" "'d' in (y for x in lis for y in x)"
    100000 loops, best of 3: 2.96 usec per loop

~$ python -m timeit -s "lis=[['a','b','c'],['d','e','f'],[1,2,3],[4,5,6],[7,8,9],[10,11,12],[13,14,15],[16,17,18]]" "'d' in [y for x in lis for y in x]"
    100000 loops, best of 3: 7.4 usec per loop

~$ python -m timeit -s "lis=[['a','b','c'],['d','e','f'],[1,2,3],[4,5,6],[7,8,9],[10,11,12],[13,14,15],[16,17,18]]" "11 in [y for x in lis for y in x]"
100000 loops, best of 3: 5.61 usec per loop

~$ python -m timeit -s "lis=[['a','b','c'],['d','e','f'],[1,2,3],[4,5,6],[7,8,9],[10,11,12],[13,14,15],[16,17,18]]" "11 in (y for x in lis for y in x)"
100000 loops, best of 3: 9.76 usec per loop

~$ python -m timeit -s "lis=[['a','b','c'],['d','e','f'],[1,2,3],[4,5,6],[7,8,9],[10,11,12],[13,14,15],[16,17,18]]" "18 in (y for x in lis for y in x)"
100000 loops, best of 3: 8.94 usec per loop

~$ python -m timeit -s "lis=[['a','b','c'],['d','e','f'],[1,2,3],[4,5,6],[7,8,9],[10,11,12],[13,14,15],[16,17,18]]" "18 in [y for x in lis for y in x]"
100000 loops, best of 3: 7.13 usec per loop

与流行的观点相反,列表理解对于中等范围来说是非常好的。迭代器协议意味着调用
迭代器。uuu next_uuu()
,而Python中的函数调用——老实说——代价非常昂贵


当然,在某个时候,生成器内存/cpu的权衡将开始奏效,但对于小集合,列表理解是非常有效的。

完全取决于数据

发电机有一个固定的设置时间,必须在调用多少项上摊销;列表理解最初更快,但随着较大数据集使用更多内存,其速度会大大降低

回想一下,随着cPython列表的扩展,列表的大小将按照的增长模式进行调整。对于更大的列表理解,Python可能会分配比数据大小多4倍的内存。一旦你击中虚拟机——真的很慢!但是,如上所述,对于小数据集,列表理解比生成器更快

考虑案例1,一个2x26列表:

LoL=[[c1,c2] for c1,c2 in zip(string.ascii_lowercase,string.ascii_uppercase)]  

def lc_d(item='d'):
    return item in [i for sub in LoL for i in sub]

def ge_d(item='d'):
    return item in (y for x in LoL for y in x)    

def any_lc_d(item='d'):
    return any(item in x for x in LoL)    

def any_gc_d(item='d'):
    return any([item in x for x in LoL])     

def lc_z(item='z'):
    return item in [i for sub in LoL for i in sub]

def ge_z(item='z'):
    return item in (y for x in LoL for y in x)    

def any_lc_z(item='z'):
    return any(item in x for x in LoL)    

def any_gc_z(item='z'):
    return any([item in x for x in LoL])               

cmpthese.cmpthese([lc_d,ge_d,any_gc_d,any_gc_z,any_lc_d,any_lc_z, lc_z, ge_z])   
这些时间安排的结果:

         rate/sec   ge_z   lc_z   lc_d any_lc_z any_gc_z any_gc_d   ge_d any_lc_d
ge_z      124,652     -- -10.1% -16.6%   -44.3%   -46.5%   -48.5% -76.9%   -80.7%
lc_z      138,678  11.3%     --  -7.2%   -38.0%   -40.4%   -42.7% -74.3%   -78.6%
lc_d      149,407  19.9%   7.7%     --   -33.3%   -35.8%   -38.2% -72.3%   -76.9%
any_lc_z  223,845  79.6%  61.4%  49.8%       --    -3.9%    -7.5% -58.5%   -65.4%
any_gc_z  232,847  86.8%  67.9%  55.8%     4.0%       --    -3.7% -56.9%   -64.0%
any_gc_d  241,890  94.1%  74.4%  61.9%     8.1%     3.9%       -- -55.2%   -62.6%
ge_d      539,654 332.9% 289.1% 261.2%   141.1%   131.8%   123.1%     --   -16.6%
any_lc_d  647,089 419.1% 366.6% 333.1%   189.1%   177.9%   167.5%  19.9%       --

现在考虑<强>案例2 <强>,这显示了LC和GEN之间的巨大差异。在这种情况下,我们在100×97×97列表中寻找一种元素:列表结构:

LoL=[[str(a),str(b),str(c)] 
       for a in range(100) for b in range(97) for c in range(97)]

def lc_10(item='10'):
    return item in [i for sub in LoL for i in sub]

def ge_10(item='10'):
    return item in (y for x in LoL for y in x)    

def any_lc_10(item='10'):
    return any([item in x for x in LoL])    

def any_gc_10(item='10'):
    return any(item in x for x in LoL)     

def lc_99(item='99'):
    return item in [i for sub in LoL for i in sub]

def ge_99(item='99'):
    return item in (y for x in LoL for y in x)    

def any_lc_99(item='99'):
    return any(item in x for x in LoL)    

def any_gc_99(item='99'):
    return any([item in x for x in LoL])      

cmpthese.cmpthese([lc_10,ge_10,any_lc_10,any_gc_10,lc_99,ge_99,any_lc_99,any_gc_99],c=10,micro=True)   
这些时间的结果:

          rate/sec  usec/pass       ge_99      lc_99      lc_10  any_lc_99  any_gc_99  any_lc_10   ge_10 any_gc_10
ge_99            3 354545.903          --     -20.6%     -30.6%     -60.8%     -61.7%     -63.5% -100.0%   -100.0%
lc_99            4 281678.295       25.9%         --     -12.6%     -50.6%     -51.8%     -54.1% -100.0%   -100.0%
lc_10            4 246073.484       44.1%      14.5%         --     -43.5%     -44.8%     -47.4% -100.0%   -100.0%
any_lc_99        7 139067.292      154.9%     102.5%      76.9%         --      -2.4%      -7.0% -100.0%   -100.0%
any_gc_99        7 135748.100      161.2%     107.5%      81.3%       2.4%         --      -4.7% -100.0%   -100.0%
any_lc_10        8 129331.803      174.1%     117.8%      90.3%       7.5%       5.0%         -- -100.0%   -100.0%
ge_10      175,494      5.698  6221964.0% 4943182.0% 4318339.3% 2440446.0% 2382196.2% 2269594.1%      --    -38.5%
any_gc_10  285,327      3.505 10116044.9% 8036936.7% 7021036.1% 3967862.6% 3873157.1% 3690083.0%   62.6%        --
正如您所看到的——这取决于函数调用的开销,这是一种折衷…

扩展到的答案,生成器表达式通常比列表理解慢。在这种情况下,中的
的短路行为会抵消在相当早的时间发现项目时的慢度,但在其他情况下,该模式保持不变

为了进行更详细的分析,我在分析器中运行了一个简单的脚本。以下是脚本:

lis=[['a','b','c'],['d','e','f'],[1,2,3],[4,5,6],
     [7,8,9],[10,11,12],[13,14,15],[16,17,18]]

def ge_d():
    return 'd' in (y for x in lis for y in x)
def lc_d():
    return 'd' in [y for x in lis for y in x]

def ge_11():
    return 11 in (y for x in lis for y in x)
def lc_11():
    return 11 in [y for x in lis for y in x]

def ge_18():
    return 18 in (y for x in lis for y in x)
def lc_18():
    return 18 in [y for x in lis for y in x]

for i in xrange(100000):
    ge_d()
    lc_d()
    ge_11()
    lc_11()
    ge_18()
    lc_18()
以下是相关结果,重新排序以使模式更清晰

         5400002 function calls in 2.830 seconds

   Ordered by: standard name

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
   100000    0.158    0.000    0.251    0.000 fop.py:3(ge_d)
   500000    0.092    0.000    0.092    0.000 fop.py:4(<genexpr>)
   100000    0.285    0.000    0.285    0.000 fop.py:5(lc_d)

   100000    0.356    0.000    0.634    0.000 fop.py:8(ge_11)
  1800000    0.278    0.000    0.278    0.000 fop.py:9(<genexpr>)
   100000    0.333    0.000    0.333    0.000 fop.py:10(lc_11)

   100000    0.435    0.000    0.806    0.000 fop.py:13(ge_18)
  2500000    0.371    0.000    0.371    0.000 fop.py:14(<genexpr>)
   100000    0.344    0.000    0.344    0.000 fop.py:15(lc_18)
我不确定剩余的时间是什么原因造成的;即使没有额外的函数调用,生成器表达式似乎也会慢一些。我想这证实了“创建生成器理解比列表理解有更多的本地开销”的断言。但无论如何,这非常清楚地表明,生成器表达式的速度较慢主要是因为调用了
next

我要补充一点,当短路不起作用时,列表理解速度更快,即使对于非常大的列表也是如此。例如:

>>> counter = itertools.count()
>>> lol = [[counter.next(), counter.next(), counter.next()] 
           for _ in range(1000000)]
>>> 2999999 in (i for sublist in lol for i in sublist)
True
>>> 3000000 in (i for sublist in lol for i in sublist)
False
>>> %timeit 2999999 in [i for sublist in lol for i in sublist]
1 loops, best of 3: 312 ms per loop
>>> %timeit 2999999 in (i for sublist in lol for i in sublist)
1 loops, best of 3: 351 ms per loop
>>> %timeit any([2999999 in sublist for sublist in lol])
10 loops, best of 3: 161 ms per loop
>>> %timeit any(2999999 in sublist for sublist in lol)
10 loops, best of 3: 163 ms per loop
>>> %timeit for i in [2999999 in sublist for sublist in lol]: pass
1 loops, best of 3: 171 ms per loop
>>> %timeit for i in (2999999 in sublist for sublist in lol): pass
1 loops, best of 3: 183 ms per loop
正如您所看到的,当短路不相关时,即使对于百万项长的列表,列表理解也始终更快。显然,在这些标度下,实际使用
中的
,发电机会因为短路而更快。但是对于其他类型的迭代任务,在项目数量上是真正线性的,列表理解总是快得多。如果您需要在一个列表上执行多个测试,这一点尤其正确;您可以非常快速地迭代已构建的列表:

在这种情况下,列表理解速度要快一个数量级


当然,这只会在内存耗尽之前保持不变。这就引出了我的最后一点。使用发电机有两个主要原因:利用短路和节省内存。对于非常大的sequences/iterables,生成器是显而易见的方法,因为它们可以节省内存。但如果短路不是一个选项,你几乎永远不会选择发电机超过列表的速度。您选择它们是为了节省内存,这始终是一种折衷。

+1我也会关注答案:)可能是因为缓存。。。和发电机。。。可能需要更多的调用(下一步、屈服、保存状态等)。而且生成器的内存效率也很高。这是肯定的。为什么不干脆
any(x中的d代表lis中的x)
?@gnibbler
any()
比生成器表达式本身慢,请看你上面提到的问题。生成器也不会耗尽整个列表,但它很慢。@Ashwini Chaudhary:与具有少量数据的LC相比,生成器只慢。查看案例2中的比较速度,发电机或使用发电机的
any
结构踢LC的屁股。生成器速度很快,但它们必须克服较长的安装时间。@senderle:非常好的观点,我在文章的编辑中提到了它们。谢谢,这对我现在来说更有意义。此外,创建生成器理解比列表理解有更多的本机开销
>>> .634 - .278 - .333
0.023
>>> .806 - .371 - .344
0.091
>>> counter = itertools.count()
>>> lol = [[counter.next(), counter.next(), counter.next()] 
           for _ in range(1000000)]
>>> 2999999 in (i for sublist in lol for i in sublist)
True
>>> 3000000 in (i for sublist in lol for i in sublist)
False
>>> %timeit 2999999 in [i for sublist in lol for i in sublist]
1 loops, best of 3: 312 ms per loop
>>> %timeit 2999999 in (i for sublist in lol for i in sublist)
1 loops, best of 3: 351 ms per loop
>>> %timeit any([2999999 in sublist for sublist in lol])
10 loops, best of 3: 161 ms per loop
>>> %timeit any(2999999 in sublist for sublist in lol)
10 loops, best of 3: 163 ms per loop
>>> %timeit for i in [2999999 in sublist for sublist in lol]: pass
1 loops, best of 3: 171 ms per loop
>>> %timeit for i in (2999999 in sublist for sublist in lol): pass
1 loops, best of 3: 183 ms per loop
>>> incache = [2999999 in sublist for sublist in lol]
>>> get_list = lambda: incache
>>> get_gen = lambda: (2999999 in sublist for sublist in lol)
>>> %timeit for i in get_list(): pass
100 loops, best of 3: 18.6 ms per loop
>>> %timeit for i in get_gen(): pass
1 loops, best of 3: 187 ms per loop