Reduce memory usage for edit distances? about comparator HOT 4 OPEN

Hi Olivier,

I'd be in favor of this. Keeping the distance matrix in memory is only useful if one is interested in recording the optimal sequence of edit operations. But here, we're only interested in the distance (the last entry in the matrix).

One thing to be mindful of, is that the Levenshtein distance is related to other edit-based distances in the package through class inheritance. The inheritance diagram is as follows: DamerauLevenshtein -> OSA -> Levenshtein -> LCS.

Since all of these classes currently rely on the full distance matrix, they would also be impacted by your proposal. I believe it's possible to implement LCS and Levenshtein by keeping 2 rows in memory, and OSA by keeping 3 rows in memory. For DamerauLevenshtein, it may be necessary to keep all rows in memory (need to check this).

Would you be happy to submit a pull request?

from comparator.

You might be interested in this: https://web.archive.org/web/20180612143641/https://bitbucket.org/clearer/iosifovich/

from comparator.

@ngmarchant thanks for the link! I'll look into it for further optimization, and then see if I can make a pull request.

from comparator.

For reference, here's the optimization I could get by reducing the memory usage a bit more:

def levenshtein(s, t):
    m = len(s)
    n = len(t)
    buffer = np.arange(max(n, m) + 1)

    p = m
    for j in range(1, n+1):
        temp = j-1
        p = j
        for i in range(1, m+1):
            p = min(p + 1, buffer[i] + 1, temp + (s[i-1] != t[j-1]))
            temp = buffer[i]
            buffer[i] = p
    return p

from comparator.