Gensim: word vectors encoding problems

Question

After creating word vectors in Gensim 2.2.0 from plain English text files with IMDB movie ratings:

import gensim, logging
import smart_open, os
from nltk.tokenize import RegexpTokenizer

VEC_SIZE = 300 
MIN_COUNT = 5
WORKERS = 4
data_path = './data/'
vectors_path = 'vectors.bin.gz'

class AllSentences(object):
    def __init__(self, dirname):
        self.dirname = dirname
        self.read_err_cnt = 0
        self.tokenizer = RegexpTokenizer('[\'a-zA-Z]+', discard_empty=True)

    def __iter__(self):
        for fname in os.listdir(self.dirname):
            print(fname)
            for line in open(os.path.join(self.dirname, fname)):
                words = []     
                try:
                    for word in self.tokenizer.tokenize(line):
                        words.append(word)
                    yield words
                except:
                    self.read_err_cnt += 1

sentences = AllSentences(data_path)

Training and saving model:

model = gensim.models.Word2Vec(sentences, size=VEC_SIZE, 
                               min_count=MIN_COUNT, workers=WORKERS)
word_vectors = model.wv
word_vectors.save(vectors_path)

And then trying to load it back:

vectors = KeyedVectors.load_word2vec_format(vectors_path,
                                                    binary=True,
                                                    unicode_errors='ignore')

I get 'UnicodeDecodeError: 'utf-8' codec can't decode byte 0x80 in position 0' exception (see below). I tried different combinations of 'encoding' parameters including 'ISO-8859-1' and 'Latin1'. Also different combinations of 'binary=True/False'. Nothing helps - the same exception, no matter what parameters are used. What is wrong? How to make loading vectors work?

Exception:

UnicodeDecodeError                        Traceback (most recent call last)
<ipython-input-64-f353fa49685c> in <module>()
----> 1 w2v = get_w2v_vectors()

<ipython-input-63-cbbe0a76e837> in get_w2v_vectors()
      3     vectors = KeyedVectors.load_word2vec_format(word_vectors_path,
      4                                                     binary=True,
----> 5                                                     unicode_errors='ignore')
      6 
      7                                                 #unicode_errors='ignore')

D:\usr\anaconda\lib\site-packages\gensim\models\keyedvectors.py in load_word2vec_format(cls, fname, fvocab, binary, encoding, unicode_errors, limit, datatype)
    204         logger.info("loading projection weights from %s", fname)
    205         with utils.smart_open(fname) as fin:
--> 206             header = utils.to_unicode(fin.readline(), encoding=encoding)
    207             vocab_size, vector_size = map(int, header.split())  # throws for invalid file format
    208             if limit:

D:\usr\anaconda\lib\site-packages\gensim\utils.py in any2unicode(text, encoding, errors)
    233     if isinstance(text, unicode):
    234         return text
--> 235     return unicode(text, encoding, errors=errors)
    236 to_unicode = any2unicode
    237 

UnicodeDecodeError: 'utf-8' codec can't decode byte 0x80 in position 0: invalid start byte

gojomo gojomo · Accepted Answer · 2017-08-02T15:48:40

If you save vectors using gensim's native save() method, you should load them with the native load() method.

If you want to load vectors using load_word2vec_format(), you'll need to save them with save_word2vec_format(). (You'll lose some information this way, such as the exact occurrence counts that would otherwise be inside the KeyedVectors.vocab dictionary items.)

Gensim: word vectors encoding problems

1 Answers