Loading pre-trained embeddings

emeddings.py

# -*- coding: utf-8 -*-

import numpy as np

import logging

logger = logging.getLogger(__name__)


def load_glove(stream, words=None):
    """
    Loads GloVe word embeddings.

    :param stream: An opened stream to the GloVe file.
    :param words: Words in the existing vocabulary.
    :return: dict {word: embedding}
    """
    word_to_embedding = {}

    for n, line in enumerate(stream):
        split_line = line.split(' ')
        word = split_line[0]

        if words is None or word in words:
            try:
                word_to_embedding[word] = [float(f) for f in split_line[1:]]
            except ValueError:
                logger.error('{}\t{}\t{}'.format(n, word, str(split_line)))

    return word_to_embedding


def load_word2vec(stream, words=None):
    """
    Loads word2vec word embeddings.

    :param stream: An opened stream to the GloVe file.
    :param words: Words in the existing vocabulary.
    :return: dict {word: embedding}
    """
    word_to_embedding = {}

    vec_n, vec_size = map(int, stream.readline().split())
    byte_size = vec_size * 4

    for n in range(vec_n):
        word = b''

        while True:
            c = stream.read(1)
            if c == b' ':
                break
            else:
                word += c
        word = word.decode('utf-8')
        vector = np.fromstring(stream.read(byte_size), dtype=np.float32)
        if words is None or word in words:
            word_to_embedding[word] = vector.tolist()

    return word_to_embedding

  # Sample usage:
  def main():
    if os.path.isfile(glove_path):
    from derte.io.embeddings import load_glove
    logger.info('Initialising the embeddings with GloVe vectors ..')

    word_set = {w for w, w_idx in qs_tokenizer.word_index.items()
                if w_idx < vocab_size}
    with open(glove_path, 'r') as stream:
        word_to_embedding = load_glove(stream=stream, words=word_set)

    for word in tqdm(word_to_embedding):
        word_idx, word_embedding = qs_tokenizer.word_index[word], word_to_embedding[word]
        session.run(assign_word_embedding, feed_dict={
            word_idx_ph: word_idx,
            word_embedding_ph: word_embedding
        })

    logger.info('Done!')

Forgot this - here's the assign graph:

    word_idx_ph = tf.placeholder(dtype=tf.int32, name='word_idx')
    word_embedding_ph = tf.placeholder(dtype=tf.float32, shape=[None], name='word_embedding')
    assign_word_embedding = model.embeddings[word_idx_ph, :].assign(word_embedding_ph)