将word2vec bin文件转换为文本

Question

从word2vec网站我可以下载GoogleNews-vectors-negative300.bin.gz..bin文件(大约3.4GB)是一种对我没用的二进制格式.Tomas Mikolov 向我们保证:"将二进制格式转换为文本格式应该相当简单(尽管这需要更多的磁盘空间).检查距离工具中的代码,读取二进制文件相当简单." 不幸的是,我不太了解C http://word2vec.googlecode.com/svn/trunk/distance.c.

据说gensim也可以做到这一点,但我发现的所有教程似乎都是关于从文本转换而不是其他方式.

有人可以建议修改C代码或gensim发出文本的说明吗？

Answer 1

我用这段代码加载二进制模型,然后将模型保存到文本文件中,

from gensim.models.keyedvectors import KeyedVectors

model = KeyedVectors.load_word2vec_format('path/to/GoogleNews-vectors-negative300.bin', binary=True)
model.save_word2vec_format('path/to/GoogleNews-vectors-negative300.txt', binary=False)

参考文献:API和nullege.

注意:

以上代码适用于gensim的新版本.对于以前的版本,我使用了以下代码:

from gensim.models import word2vec

model = word2vec.Word2Vec.load_word2vec_format('path/to/GoogleNews-vectors-negative300.bin', binary=True)
model.save_word2vec_format('path/to/GoogleNews-vectors-negative300.txt', binary=False)

Answer 2

在word2vec-toolkit邮件列表中,Thomas Mensink 以小型C程序的形式提供了一个答案,它将.bin文件转换为文本.这是distance.c文件的修改.我用下面的Thomas代码替换了原来的distance.c并重建了word2vec(make clean; make),并将编译后的距离重命名为readbin.然后./readbin vector.bin将创建vector.bin的文本版本.

//  Copyright 2013 Google Inc. All Rights Reserved.
//
//  Licensed under the Apache License, Version 2.0 (the "License");
//  you may not use this file except in compliance with the License.
//  You may obtain a copy of the License at
//
//      http://www.apache.org/licenses/LICENSE-2.0
//
//  Unless required by applicable law or agreed to in writing, software
//  distributed under the License is distributed on an "AS IS" BASIS,
//  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
//  See the License for the specific language governing permissions and
//  limitations under the License.

#include <stdio.h>
#include <string.h>
#include <math.h>
#include <malloc.h>

const long long max_size = 2000;         // max length of strings
const long long N = 40;                  // number of closest words that will be shown
const long long max_w = 50;              // max length of vocabulary entries

int main(int argc, char **argv) {
  FILE *f;
  char file_name[max_size];
  float len;
  long long words, size, a, b;
  char ch;
  float *M;
  char *vocab;
  if (argc < 2) {
    printf("Usage: ./distance <FILE>\nwhere FILE contains word projections in the BINARY FORMAT\n");
    return 0;
  }
  strcpy(file_name, argv[1]);
  f = fopen(file_name, "rb");
  if (f == NULL) {
    printf("Input file not found\n");
    return -1;
  }
  fscanf(f, "%lld", &words);
  fscanf(f, "%lld", &size);
  vocab = (char *)malloc((long long)words * max_w * sizeof(char));
  M = (float *)malloc((long long)words * (long long)size * sizeof(float));
  if (M == NULL) {
    printf("Cannot allocate memory: %lld MB    %lld  %lld\n", (long long)words * size * sizeof(float) / 1048576, words, size);
    return -1;
  }
  for (b = 0; b < words; b++) {
    fscanf(f, "%s%c", &vocab[b * max_w], &ch);
    for (a = 0; a < size; a++) fread(&M[a + b * size], sizeof(float), 1, f);
    len = 0;
    for (a = 0; a < size; a++) len += M[a + b * size] * M[a + b * size];
    len = sqrt(len);
    for (a = 0; a < size; a++) M[a + b * size] /= len;
  }
  fclose(f);
  //Code added by Thomas Mensink
  //output the vectors of the binary format in text
  printf("%lld %lld #File: %s\n",words,size,file_name);
  for (a = 0; a < words; a++){
    printf("%s ",&vocab[a * max_w]);
    for (b = 0; b< size; b++){ printf("%f ",M[a*size + b]); }
    printf("\b\b\n");
  }  

  return 0;
}

我删除了"\ b\b" printf.

顺便说一句,生成的文本文件仍然包含文本词和一些不必要的空格,我不想进行一些数值计算.我使用bash命令从每行删除了初始文本列和尾随空白.

cut --complement -d ' ' -f 1 GoogleNews-vectors-negative300.txt > GoogleNews-vectors-negative300_tuples-only.txt
sed 's/ $//' GoogleNews-vectors-negative300_tuples-only.txt

Answer 3

我正在使用 gensim 与 GoogleNews-vectors-negative300.bin 一起工作，并且binary = True在加载模型时包含了一个标志。

from gensim import word2vec

model = word2vec.Word2Vec.load_word2vec_format('Path/to/GoogleNews-vectors-negative300.bin', binary=True) 

似乎工作正常。

Answer 4

格式为IEEE 754单精度二进制浮点格式:binary32 http://en.wikipedia.org/wiki/Single-precision_floating-point_format 它们使用little-endian.

举个例子:

• 第一行是字符串格式:"3000000 300 \n"(vocabSize&vecSize,getByte until byte =='\n')

• 下一行首先包括词汇单词,然后(浮点值300*4字节,每个维度4字节):

  getByte till byte==32 (space). (60 47 115 62 32 => <\s>[space])
  

• 那么每个接下来的4个字节将代表一个浮点数

  接下来的4字节:0 0 -108 58 => 0.001129150390625.

您可以查看维基百科链接以了解具体方法,让我以此为例:

(little-endian - >逆序)00111010 10010100 00000000 00000000

• 首先是符号位=>符号= 1(否则= -1)
• next 8 bits => 117 => exp = 2 ^(117-127)
• next 23 bits => pre = 0*2 ^( - 1)+ 0*2 ^( - 2)+ 1*2 ^( - 3)+ 1*2 ^( - 5)

value = sign*exp*pre

Answer 5

您可以在word2vec中加载二进制文件,然后保存文本版本,如下所示:

from gensim.models import word2vec
 model = word2vec.Word2Vec.load_word2vec_format('Path/to/GoogleNews-vectors-negative300.bin', binary=True)
 model.save("file.txt")

`