如何加快Levenshtein距离的计算 [英] How to speed up Levenshtein distance calculation

问题描述

我正在尝试进行仿真以测试随机之间平均 Levenshtein距离 二进制字符串.

I am trying to run a simulation to test the average Levenshtein distance between random binary strings.

我的程序在python中，但是我正在使用 C扩展.与此相关且花费大量时间的函数是计算两个字符串之间的Levenshtein距离的方法.

My program is in python but I am using this C extension. The function that is relevant and takes most of the time computes the Levenshtein distance between two strings and is this.

lev_edit_distance(size_t len1, const lev_byte *string1,
                  size_t len2, const lev_byte *string2,
                  int xcost)
{
  size_t i;
  size_t *row;  /* we only need to keep one row of costs */
  size_t *end;
  size_t half;

  /* strip common prefix */
  while (len1 > 0 && len2 > 0 && *string1 == *string2) {
    len1--;
    len2--;
    string1++;
    string2++;
  }

  /* strip common suffix */
  while (len1 > 0 && len2 > 0 && string1[len1-1] == string2[len2-1]) {
    len1--;
    len2--;
  }

  /* catch trivial cases */
  if (len1 == 0)
    return len2;
  if (len2 == 0)
    return len1;

  /* make the inner cycle (i.e. string2) the longer one */
  if (len1 > len2) {
    size_t nx = len1;
    const lev_byte *sx = string1;
    len1 = len2;
    len2 = nx;
    string1 = string2;
    string2 = sx;
  }
  /* check len1 == 1 separately */
  if (len1 == 1) {
    if (xcost)
      return len2 + 1 - 2*(memchr(string2, *string1, len2) != NULL);
    else
      return len2 - (memchr(string2, *string1, len2) != NULL);
  }
  len1++;
  len2++;
  half = len1 >> 1;
  /* initalize first row */
  row = (size_t*)malloc(len2*sizeof(size_t));
  if (!row)
    return (size_t)(-1);
  end = row + len2 - 1;
  for (i = 0; i < len2 - (xcost ? 0 : half); i++)
    row[i] = i;

  /* go through the matrix and compute the costs.  yes, this is an extremely
   * obfuscated version, but also extremely memory-conservative and relatively
   * fast.  */
  if (xcost) {
    for (i = 1; i < len1; i++) {
      size_t *p = row + 1;
      const lev_byte char1 = string1[i - 1];
      const lev_byte *char2p = string2;
      size_t D = i;
      size_t x = i;
      while (p <= end) {
        if (char1 == *(char2p++))
          x = --D;
        else
          x++;
        D = *p;
        D++;
        if (x > D)
          x = D;
        *(p++) = x;
      }
    }
  }
  else {
    /* in this case we don't have to scan two corner triangles (of size len1/2)
     * in the matrix because no best path can go throught them. note this
     * breaks when len1 == len2 == 2 so the memchr() special case above is
     * necessary */
    row[0] = len1 - half - 1;
    for (i = 1; i < len1; i++) {
      size_t *p;
      const lev_byte char1 = string1[i - 1];
      const lev_byte *char2p;
      size_t D, x;
      /* skip the upper triangle */
      if (i >= len1 - half) {
        size_t offset = i - (len1 - half);
        size_t c3;

        char2p = string2 + offset;
        p = row + offset;
        c3 = *(p++) + (char1 != *(char2p++));
        x = *p;
        x++;
        D = x;
        if (x > c3)
          x = c3;
        *(p++) = x;
      }
      else {
        p = row + 1;
        char2p = string2;
        D = x = i;
      }
      /* skip the lower triangle */
      if (i <= half + 1)
        end = row + len2 + i - half - 2;
      /* main */
      while (p <= end) {
        size_t c3 = --D + (char1 != *(char2p++));
        x++;
        if (x > c3)
          x = c3;
        D = *p;
        D++;
        if (x > D)
          x = D;
        *(p++) = x;
      }
      /* lower triangle sentinel */
      if (i <= half) {
        size_t c3 = --D + (char1 != *char2p);
        x++;
        if (x > c3)
          x = c3;
        *p = x;
      }
    }
  }

  i = *end;
  free(row);
  return i;
}

可以加快速度吗?

我将在AMD FX(tm)-8350八核处理器上的32位ubuntu中运行代码.

I will be running the code in 32 bit ubuntu on an AMD FX(tm)-8350 Eight-Core Processor.

这是调用它的python代码.

Here is the python code that calls it.

from Levenshtein import distance
import random
for i in xrange(16):
    sum = 0
    for j in xrange(1000):
        str1 = bin(random.getrandbits(2**i))[2:].zfill(2**i)
        str2 = bin(random.getrandbits(2**i))[2:].zfill(2**i)
        sum += distance(str1,str2)
    print i,sum/(1000*2**i)

推荐答案

也许可以并行运行.在开始时生成一个庞大的随机变量列表，然后在循环中，一次生成线程(8个线程)以每个线程处理列表中的一个块并将其最终结果添加到sum变量中.或一次生成8个列表，然后一次生成8个.

You could run this parallel maybe. Generate one giant list of randoms at the start, then in your loop, spawn threads (8 threads) at a time to each process one chunk of the list and add its final result to the sum variable. Or generate a list of 8 at once and do 8 at a time.

openmp建议的问题是由于大量数据依赖关系，该算法并行性很差"-Wikipedia

The problem with the openmp suggestion is "This algorithm parallelizes poorly, due to a large number of data dependencies" - Wikipedia

from threading import Thread

sum = 0

def calc_distance(offset) :
    sum += distance(randoms[offset][0], randoms[offset][1]) #use whatever addressing scheme is best

threads = []
for i in xrange(8) :
    t = new Thread(target=calc_distance, args=(i))
    t.start()
    threads.append(t)

稍后....

for t in threads :
     t.join()

我认为，如果有levenshtein距离内核可用(或可编码)，此方法也将在以后很好地移植到opencl.

i think this method would port nicely to opencl later as well if levenshtein distance kernel was available (or codable).

这只是记忆中的一则快速帖子，因此可能有一些问题需要解决.

This is just a quick post from memory so there are probably some kinks to work out.

这篇关于如何加快Levenshtein距离的计算的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！