Simplifying / optimizing a chain of for-loops

Question

I have a chain of for-loops that works on an original list of strings and then gradually filtering the list as it goes down the chain, e.g.:

import re

# Regex to check that a cap exist in string.
pattern1 = re.compile(r'\d.*?[A-Z].*?[a-z]')
vocab = ['dog', 'lazy', 'the', 'fly'] # Imagine it's a longer list.

def check_no_caps(s):
    return None if re.match(pattern1, s) else s

def check_nomorethan_five(s):
    return s if len(s) <= 5 else None

def check_in_vocab_plus_x(s,x):
    # s and x are both str.
    return None if s not in vocab else s+x

slist = ['the', 'dog', 'jumps', 'over', 'the', 'fly']
# filter with check_no_caps
slist = [check_no_caps(s) for s in slist]
# filter no more than 5.
slist = [check_nomorethan_five(s) for s in slist if s is not None]
# filter in vocab
slist = [check_in_vocab_plus_x(s, str(i)) for i,s in enumerate(slist) if s is not None]

The above is just an example and in reality my functions to manipulate the strings are more complicated but they do return the original string or a manipulated one.

I could use generators instead of list and do something like this:

slist = ['the', 'dog', 'jumps', 'over', 'the', 'fly']
# filter with check_no_caps and no more than 5.
slist = (s2 check_no_caps(s1) for s1 in slist 
         for s2 in check_nomorethan_five(s1) if s1)
# filter in vocab
slist = [check_in_vocab_plus_x(s, str(i)) for i,s in enumerate(slist) if s is not None]

Or in one crazy nested generator:

slist = ['the', 'dog', 'jumps', 'over', 'the', 'fly']
slist = (s3 check_no_caps(s1) for s1 in slist 
         for s2 in check_nomorethan_five(s1) if s1
         for s3 in check_in_vocab_plus_x(s2, str(i)) if s2)

There must be a better way. Is there a way to make the chain of for-loop faster?

Is there a way to do it with map, reduce and filter? Will it be faster?

Imagine that my original slist is very very large like 10s of billions. And my functions are a not as simple as the functions above, they do some computation and do around 1,000 calls per second.

Answer 1

First of all is the overall process that you make on your strings. You are taking some strings and to each of them you apply certain functions. Then you cleanup the list. Let's say for a while that all the functions you apply to strings works at a constant time (it's no true, but for now it won't matter). In your solution you iterate throgh list applying one function (that's O(N)). Then you take next function and iterate again (another O(N)), and so on. So, the obvious way to speed-up is to reduce number of loops. That's not so difficult.

The next thing to do is try to optimize your functions. E.g. you use regexp to check whether string has capital letters, but there is str.islower (Return true if all cased characters in the string are lowercase and there is at least one cased character, false otherwise).

So, this is the very first attempt to simplify and speed-up your code:

vocab = ['dog', 'lazy', 'the', 'fly'] # Imagine it's a longer list.

# note that first two functions can be combined in one
def no_caps_and_length(s):
    return s if s.islower() and len(s)<=5 else None

# this one is more complicated and cannot be merged with first two
# (not really, but as you say, some functions are rather complicated)
def check_in_vocab_plus_x(s,x):
    # s and x are both str.
    return None if s not in vocab else s+x

# now let's introduce a function that would pipe a string through all functions you need
def pipe_through_funcs(s):
    # yeah, here we have only two, but could be more
    funcs = [no_caps_and_length, check_in_vocab_plus_x]
    for func in funcs:
        if s == None: return s
        s = func(s)
    return s

slist = ['the', 'dog', 'jumps', 'over', 'the', 'fly']
# final step:
slist = filter(lambda a: a!=None, map(pipe_through_funcs, slist))

There might be one more thing that can be improved. Currently you iterate through list modifying elements and then filter it out. But if might be faster to filter and then modify. Like this:

vocab = ['dog', 'lazy', 'the', 'fly'] # Imagine it's a longer list.

# make a function that does all the checks for filtering
# you can make a big expression and return its result,
# or a sequence of ifs, or anything in-between,
# it won't affect performance,
# but make sure you put cheaper checks first
def my_filter(s):
    if len(s)>5: return False
    if not s.islower(): return False
    if s not in vocab: return False
    # maybe more checks here
    return True

# now we need modifying function
# there is a concern: if you need indices as they were in original list
# you might need to think of some way to pass them here
# as you iterate through filtered out list
def modify(s,x):
    s += x
    # maybe more actions
    return s

slist = ['the', 'dog', 'jumps', 'over', 'the', 'fly']
# final step:
slist = map(modify, filter(my_filter, slist))

Note also, that in some cases generators, maps and things can be faster, but that is not always true. I believe, that if number of items you filter out is substantial, it might be faster to use a for-loop with append. I would not vouch that it will be faster but you could just try something like this:

initial_list = ['the', 'dog', 'jumps', 'over', 'the', 'fly']
new_list = []
for s in initial_list:
    processed = pipe_through_funcs(s)
    if processed != None: new_list.append(processed)

Answer 2

If you make your transform functions unified then you could do something like this:

import random
slist = []
for i in range(0,100):
    slist.append(random.randint(0,1000))

# Unified functions which have the same function description
# x is the value
# i is the counter from enumerate
def add(x, i):
    return x + 2

def replace(x, i):
    return int(str(x).replace('2', str(i)))

# Specifying your pipelines as a list of tuples 
# Where tuple is (filter function, transformer function)
_pipeline = [
    (lambda s: True, add),
    (lambda s: s % 2 == 0, replace),
]

# Execute your pipeline
for _filter, _fn in _pipeline:
    slist = map(lambda item: _fn(*item), enumerate(filter(_filter, slist)))

The code works on both python 2 and python 3. The difference is that in Python3 everything returns a generator so it's not executed until it has to be. So effectively your going to have one iteration over your list.

print(slist)
<map object at 0x7f92b8315fd0>

However iterating over once or many times won't make much of a difference as long as it can be done in memory because regardless of the iteration method the same amount of transformation and filtering has to be executed. So to improve your code try to make your filter and transform functions as fast as possible.

For example what @Rawing mentioned to have vocab as a set instead of list will make a big difference especially with large number of items.

Answer 3

You have a bunch of checks with which you can make an iterable:

def check1(s):
    if s.islower():
        return s

def check2(s):
    if len(s) < 5:
        return s

checks = [check1, check2]

And an iterable of strings:

l = ['dog', 'Cat', 'house', 'foo']

So one questions is whether you should iterate over checks first or strings first.

def checks_first(l, checks):
    for check in checks:
        l = filter(None, map(check, l))

    return list(l)


def strings_first(l, checks):
    res = []

    for item in l:
        for check in checks:
            item = check(item)
            if item is None:
                break
        else:
            res.append(item)

    return res

You can time these two approaches with the timeit module. Note: that you may have to use a subset of the strings to get these results in a timely fashion.

import timeit

print(timeit.timeit('checks_first(l, checks)', setup='from __main__ import checks_first, checks, l', number=10))
print(timeit.timeit('strings_first(l, checks)', setup='from __main__ import strings_first, checks, l', number=10))

Which is faster depends on the ratio of number checks to the of number strings, hardware, etc. From the tests I've done they seem to run at roughly the same speed.

My guess is that the biggest time savings will be gained by optimizing some of the checks. A good starting point is to identify the checks that cost the most time. This can be done with a closure to wrap your check functions.

import functools

def time_func(func, timer_dict):

    @functools.wraps(func)
    def inner(*args, **kwargs):
        t0 = time.time()
        res = func(*args, **kwargs)
        timer_dict[func.__name__] += time.time() - t0
        return res

    return inner

To apply this to the checks:

from collections import defaultdict

timer_dict = defaultdict(lambda: 0)
checks = [time_func(check, timer_dict) for check in checks]

Then call the function(s) that apply the checks and view timer_dict for timing information.

checks_first(l, checks)
strings_first(l, checks)

print(dict(timer_dict))

# {'check1': 0.41464924812316895, 'check2': 0.2684309482574463}

Then identify the bottlenecks in the costly checks either by inspection or profiling. The latter can be done by timing lines of code with the time module or using a line profiler like this.

Optimize your algos and data structures to get rid of these bottlenecks. You can take a look at Cython for code that you need to bring to (near) C speed.

Answer 4

First off: I think your example code is not doing what you think. The result is ['the0', 'dog1', None, None, 'the4', 'fly5'], but I believe you don't want the None values.

The only reasonable answer to this is to measure your performance and identify the bottlenecks, which will probably be in your check functions and not in the external loop.

From outside the check functions the only real optimization I see is to perform the checks that will reduce your set first, so that you'll have smaller loops in the following iterations and you reduce the number of checks you perform on values you'll discard anyway. Depending on your data and the number of values that get discarded in the first checks you might see quite a jump in performance... Or not!

The only way to really know is to profile your code. You should use cProfile together with RunSnakeRun and work on your bottlenecks, otherwise you'll end up optmizing the wrong stuff.

To profile your script you can run it as follows: python -m cProfile <script-name>