How to save the tensorflow's word2vec in text/

问题:

I have trained a word2vec model in tensorflow. But when I save the session, it only outputted model.ckpt.data / .index / .meta files.

I was thinking of implementing KNN method in retrieving nearest words. I saw answers of using gensim, but how can I save my tensorflow word2vec model into .txt first?

回答1:

Simply evaluate the embeddings matrix into a numpy array and write it to the file along with resolved words. Sample code:

vocabulary_size = 50000
embedding_size = 128

# Assume your word to index map
word_to_idx = { ... }
# Assume your embeddings variable
embeddings = tf.Variable(tf.random_uniform([vocabulary_size, embedding_size],0,1))

with tf.Session() as sess:
  embeddings_val = sess.run(embeddings)
  with open('embeddings.txt', 'w') as file_:
    for i in range(vocabulary_size):
      embed = embeddings_val[i, :]
      word = word_to_idx[i]
      file_.write('%s %s\n' % (word, ' '.join(map(str, embed))))

回答2:

I just had the same problem and tried Maxim's solution.

You need to replace the line:

word = word_to_idx[i]

with

word = idx_to_word[i]

You can simply reverse the word_to_idx-dictionary with the following code:

idx_to_word = dict(zip(word_to_idx.values(), word_to_idx.keys()))

Except for that, his solution works fine.