How can you parallelize a regex search of one long string? [duplicate] - python

This question already has answers here:
How can I tell if a string repeats itself in Python?
(13 answers)
Closed 7 years ago.
I'm testing the output of a simulation to see if it enters a loop at some point, so I need to know if the output repeats itself. For example, there may be 400 digits, followed by a 400000 digit cycle. The output consists only of digits from 0-9. I have the following regex function that I'm using to match repetitions in a single long string:
def repetitions(s):
r = re.compile(r"(.+?)\1+")
for match in r.finditer(s):
if len(match.group(1)) > 1 and len(match.group(0))/len(match.group(1)) > 4:
yield (match.group(1), len(match.group(0))/len(match.group(1)))
This function works fantastically, but it takes far too long. My most recent test was 4 million digits, and it took 4.5 hours to search. It found no repetitions, so I now need to increase the search space. The code only concerns itself with subsequences that repeat themselves more than 4 times because I'm considering 5 repetitions to give a set that can be checked manually: the simulation will generate subsequences that will repeat hundreds of times. I'm running on a four core machine, and the digits to be checked are generated in real time. How can I increase the speed of the search?

Based on information given by nhahtdh in one of the other answers, some things have come to light.
First, the problem you are posing is called finding "tandem repeats" or "squares".
Second, the algorithm given in http://csiflabs.cs.ucdavis.edu/~gusfield/lineartime.pdf finds z tandem repeats in O(n log n + z) time and is "optimal" in the sense that there can be that many answers. You may be able to use parallelize the tandem searches, but I'd first do timings with the simple-minded approach and divide by 4 to see if that is in the speed range you expect.
Also, in order to use this approach you are going to need O(n) space to store this suffix tree. So if you have on the order of 400,000 digits, you are going to need on the order of 400,000 time to build and 400,000 bytes to and store this suffix tree.
I am not totally what is meant by searching in "real time", I usually think of it as a hard limit on how long an operation can take. If that's the case, then that's not going to happen here. This algorithm needs to read in the entire input string and processes that before you start to get results. In that sense, it is what's called an "off-line" algorithm,.
http://web.cs.ucdavis.edu/~gusfield/strmat.html has C code that you can download. (In tar file strmat.tar.gz look for repeats_tandem.c and repeats_tandem.h).
In light of the above, if that algorithm isn't sufficiently fast or space efficient, I'd look for ways to change or narrow the problem. Maybe you only need a fixed number of answers (e.g. up to 5)? If the cycles are a result of executing statements in a program, given that programming languages (other than assembler) don't have arbitrary "goto" statements, it's possible that this can narrow the kinds of cycles that can occur and somehow by make use of that structure might offer a way to speed things up.

When one algorithm is too slow, switch algorithms.
If you are looking for repeating strings, you might consider using a suffix tree scheme: https://en.wikipedia.org/wiki/Suffix_tree
This will find common substrings in for you in linear time.
EDIT: #nhahtdh inb a comment below has referenced a paper that tells you how to pick out all z tandem repeats very quickly. If somebody upvotes
my answer, #nhahtdh should logically get some of the credit.
I haven't tried it, but I'd guess that you might be able to parallelize the construction of the suffix tree itself.

I'm sure there's room for optimization, but test this algorithm on shorter strings to see how it compares to your current solution:
def partial_repeat(string):
l = len(string)
for i in range(2, l//2+1):
s = string[0:i]
multi = l//i-1
factor = l//(i-1)
ls = len(s)
if s*(multi) == string[:ls*(multi)] and len(string)-len(string[:ls*factor]) <= ls and s*2 in string:
return s
>>> test_string
'abc1231231231231'
>>> results = {x for x in (partial_repeat(test_string[i:]) for i in range(len(test_string))) if x}
>>> sorted(sorted(results, key=test_string.index), key=test_string.count, reverse=True)[0]
'123'
In this test string, it's unclear whether the non-repeating initial characters are 'abc' or 'abc1', so the repeating string could be either '123' or '231'. The above sorts each found substring by its earliest appearance in the test string, sorts again (sorted() is a stable sort) by the highest frequency, and takes the top result.
With standard loops and min() instead of comprehensions and sorted():
>>> g = {partial_repeat(test_string[i:]) for i in range(len(test_string))}
>>> results = set()
>>> for x in g:
... if x and (not results or test_string.count(x) >= min(map(test_string.count, results))):
... results.add(x)
...
>>> min(results, key=test_string.index)
'123'
I tested these solutions with the test string 'abc123123a' multiplied by (n for n in range(100, 10101, 500) to get some timing data. I entered these data into Excel and used its FORECAST() function to estimate the processing time of a 4-million character string at 430 seconds, or about seven minutes.

Related

Code complexity Explanation | Powerset Generation

I am trying to understand difference/similarity in complexities when writing the code to generate powerset in 2 ways:
def powerset(s, i, cur):
if i==len(s):
print(cur)
return
powerset(s, i+1, cur+s[i]) # string addition is also possibly O(n^2) here?
powerset(s, i+1, cur)
powerset("abc", 0, "")
Output:
['abc', 'ab', 'ac', 'a', 'bc', 'b', 'c', '']
This is going into recursion, with 2 choices at each step (adding s[i] or not), creating 2 branches. Leading to 2^n and adding to the array/printing is another O(n), leading to O(n*2^n)
Also thinking about it in the terms of branches^depth = O(2^n)
The space complexity for this will be: O(n)? Considering max depth of the tree to go up to n by the above logic.
And with this:
s = "abc"
res = [""]
for i in s:
res += [j+i for j in res]
I get the same output.
But here, I see 2 for loops and the additional complexity for creating the strings -- which is possibly O(n^2) in Python. Leading to possible O(N^4) as opposed to O(n*2^n) in the solution above.
Space complexity here seems to me to be O(n) since we are reserving space for just the output. But no additional space, so overall: O(1)
Is my understanding for these solutions in time and space correct? I was under the impression that computing powerset is O(2^n). But I figured maybe its a more optimized solution? (even though the second solution seems more naive).
https://stackoverflow.com/questions/34008010/is-the-time-complexity-of-iterative-string-append-actually-on2-or-on
Here, they suggest using arrays to avoid the `O(n^2)` complexity of string concatenation.
It's clear that both of those solutions involving creating 2len(s) strings, which are all the subsets of s. The amount of time that takes is O(2len(s) * T) where T is the time to create and process each string. The processing time is impossible to know without seeing the code which consumes the subsets; if the processing consisting only of printing, then it's presumably O(len(s)) since printing has to be done character by character.
I think the core of your question is how much time constructing the string takes.
It's often said that constructing a string by repeated concatenation is O(n²), since the entire string needs to be copied in order to perform each concatenation. And that's basically true (although Python has some tricks up its sleeve which can optimise in certain cases). But that's assuming that you're constructing the string from scratch, tossing away the intermediate results as soon as they're no longer necessary.
In this case, however, the intermediate results are also part of the desired result. And you only copy the prefixes implicitly, when you append the next character. That means that the additional cost of producing each subset is O(n), not O(n²), because the cost of producing the immediate prefix was already accounted for. That's true in both of the solutions you provide.

Compare strings in a list to another list of strings: pairwise string comparison vs check existence in set [duplicate]

This question already has answers here:
Can hash tables really be O(1)?
(10 answers)
Closed last year.
I'm comparing a list of strings words_to_lookup to a list of strings word_list and for every match I'll do something.
I wonder whether direct string comparison or checking existence in a set will be more time efficient.
String comparison is O(n)
for w in word_list:
for s in words_to_lookup:
if s == w:
# do something
Checking existence in a set is O(1) but getting the hash of the string (probably) takes O(n).
w_set = set(word_list)
for s in words_to_lookup:
if s in w_set:
# do something
So is string comparison an efficient approach? If the set approach is faster, how?
EDIT:
I was not thinking clearly when I first posted the question. Thank you for the comments. I find it hard to convert my real problem to a concise one suited for online discussions. Now I made the edit, the answer is obvious that the first approach is O(n^2) and the second approach is O(n).
My real question should have been this: Can hash tables really be O(1)?.
And the answer is:
The hash function however does not have to be O(m) - it can be O(1). Unlike a cryptographic hash, a hash function for use in a dictionary does not have to look at every bit in the input in order to calculate the hash. Implementations are free to look at only a fixed number of bits.
If you only need to search once, your first approach is more efficient.
One advantage of constructing a set is the following: if you need to search against the same set many times, you only need to build the set once.
In other words, suppose you have N words in the dictionary (dictionary_list) and you have a list of M words that you want to look up (words_to_lookup). If you go with the set approach, the complexity is O(N+M). If you don't build a set, the complexity is O(N*M) because you may have to go over the whole dictionary of N words for each of the M words that you are looking up.
For this problem, the following code is the more efficient approach.
w_set = set(dictionary_list)
for w in words_to_lookup:
if w in w_set:
# do something
EDIT
Ok. Now I see what you mean. In that case, the set version is definitely better. Note, you can also do:
for s in words_to_lookup:
if s in word_list:
# do something
That's the same thing as your set way, but the running time of the "in" operator will be worse.
list - Average: O(n)
set/dict - Average: O(1), Worst: O(n)
So the set way is probably best.

Giant set of strings leads to memory error - alternative?

I'm trying to create a really huge list of incrementing 9 digits numbers (worst case). My plan is to have something like this:
['000000001', '000000002' , ..............,'999999999']
I already wrote the code. However, as soon as I run the code, my console prints "Memory Error" message.
Here is my current code:
HUGE_LIST = [''.join(i) for i in product('012345678', repeat = 9)
I know this might not be the best code to produce the list. Thus, can someone help me find a better way to solve this memory issue?
I'm planning to use HUGE_LIST for comparison with user input.
Example: a user enters '12345678' as input, then I want my code to assert that input with the HUGE_LIST.
The best way to solve an issue like this is to avoid a memory-intensive algorithm entirely. In this case, since your goal is to test whether a particular string is in the list, just write a function that checks whether the string satisfies the criteria to be in the list. For example, if your list contains all sequences of 9 digits, then your function just has to check whether a given input is a sequence of 9 digits.
def check(string):
return len(string) == 9 and all(c.isdigit() for c in string)
(in practice, give it a better name than check). Or if you want all sequences of 9 digits in which none of them is a 9, as your current code defining HUGE_LIST suggests, you could write
def check(string):
return len(string) == 9 and all(c.isdigit() and c != '9' for c in string)
Or so on.
If you can't write an algorithm to decide whether a string (or whatever) is in the list or not, the next best thing is to make a generator that will produce the values one at a time. If you already have a list comprehension, like
HUGE_LIST = [<something> for <variable> in <expression>]
then you can turn that into a generator by replacing the square brackets with parentheses:
HUGE_GENERATOR = (<something> for <variable> in <expression>)
Then you can test for membership using string in HUGE_GENERATOR. Note that after doing so, HUGE_GENERATOR will be (at least partially) consumed, so you can't use it for another membership test; you will have to recreate it if you want to test again.

How to change numbers in a number

I'm currently trying to learn python.
Suppose there was a a number n = 12345.
How would one go about changing every digit starting from the first spot and iterating it between (1-9) and every other spot after (0-9).
I'm sadly currently learning python so I apologize for all the syntax error that might follow.
Here's my last few attempts/idea for skeleton of the code.
define the function
turn n into string
start with a for loop that for i in n range(0,9) for i[1]
else range(10)
Basically how does one fix a number while changing the others?
Please don't give solution just hints I enjoy the thinking process.
For example if n =29 the program could check
19,39,49,59,69,79,89,99
and
21,22,23,24,25,26,27,28
Although you are new, the process seems far easy than you think.
You want to make that change to every digit of the number (let's say n=7382). But you cannot iterate over numbers (not even changing specific digits of it as you want to): only over iterables (like lists). A string is an iterable. If you get the way to do a int-str conversion, you could iterate over every number and then print the new number.
But how do you change only the digit you are iterating to? Again, the way is repeating the conversion (saving it into a var before the loop would make great DRY) and getting a substring that gets all numbers except the one you are. There are two ways of doing this:
You search for that specific value and get its index (bad).
You enumerate the loop (good).
Why 2 is good? Because you have the real position of the actual number being change (think that doing an index in 75487 with 7 as the actual one changing would not work well when you get to the last one). Search for a way to iterate over items in a loop to get its actual index.
The easiest way to get a substring in Python is slicing. You slice two times: one to get all numbers before the actual one, and other to get all after it. Then you just join those two str with the actual variable number and you did it.
I hope I didn't put it easy for you, but is hard for a simple task as that.

python: search in file

I have two text files that are both about 1M lines.
Let's call them f1 and f2.
For each line in f1, I need to find the index of line in f2, where the line in f1 is a substring of the line in f2. Since I need to do it for all lines of f1, using nested for loop is too time-costly, and I was wondering if there is a workaround that could significantly reduce the time.
Thank you in advance for the help.
There certainly are better ways than using two for loops :D That would give you an O(n^2) runtime. Something very useful for finding substrings is called a rolling hash. It is a way to use previous information to speed up finding substrings. It goes something like this:
Say I have string, f1 = "cat" and a long string f2 = "There once was a cat named felix". What you do is define a "hash" based on the letters of your f1 string. The specifics on this can be found online in various sources but for this example lets simplify things and say that letters are assigned to numbers starting at 0 going up to 25 and we'll multiply each letter's value to form a decimal number with the number of digits equaling the lenght of the string:
hash("cat") = 10^2 * 2 + 10^1 * 0 + 10^0 * 19
= some value (in python the "hash" values of letters
are not 0 through 25 but given by using the ord cast:
ord("a") will give you 97)
Now this next part is super cool. We designate windows of the size of our f1 string, so size 3, and hash the f2 string in the same way we did with f1. You start with the first three letters. The hash doesn't match so we move on. If the hash matched, we make sure it's the same string (sometimes hashes equal each other but are not the same string because of the way we assign hashes but that's ok).
COOL PART** Instead of simply shifting the window and rehashing the 2 through 4th letters of f2, we "roll" the window and don't recalculate the entire hash (which if f1 is really long would be a waste of time) since the only letters changing are the first and last! The trick is to subtract the first hash value (in our example would be ord("t")*10^2), then multiply the entire number remaining by ten (because we moved everything to the left), and add the new hash letter, ord("r") * 10^0. Check for a match again and move on. If we match, return the index.
Why are we doing this: if you have a long enough f1 string, you reduce the runtime down to O(n*len(n)) so asymptotically linear!
Now, the actual implementation takes time and could get messy but there are plenty of sources online for this kind of answer. My algorithms class has great course notes online which help understand the theory a little better and there are tons of links with python implementations. Hope this helps!

Categories

Resources