def augment_pretrained_with_random_initialization(args):
words = OrderedDict()
# words in pretrained word embedding
for word in open(args.pretrained_vocab):
words[word.strip()] = 1
# words in specials e.g. PAD, START, END
for word in args.specials:
words[word] = 1
# words found in training data
for word, freq in read_model_defs(args.new_words).items():
if freq >= args.freq_cut:
words[word.encode("utf-8")] = freq
new_pretrained_vocab = os.path.join(args.out, "new_words.txt")
print >> sys.stderr, "writing to", new_pretrained_vocab
with open(new_pretrained_vocab, "w") as f:
for word, freq in words.items():
f.write("{} {}\n".format(word, freq))
embeddings = read_pretrained_embeddings(args.pretrained)
assert embeddings.shape[0] <= len(words), "pretrained size: {}, read words: {}".format(embeddings.shape[0], len(words))
new_embeddings = 0.02 * np.random.random_sample(
(len(words), embeddings.shape[1])).astype('f') - 0.01
for i in range(embeddings.shape[0]):
new_embeddings[i] = embeddings[i]
new_pretrained = os.path.join(args.out, "new_embeddings.txt")
print >> sys.stderr, "writing to", new_pretrained
np.savetxt(new_pretrained, new_embeddings)
print >> sys.stderr, "vocabulary size", len(embeddings), "-->", len(new_embeddings)
def extract_subset_of_pretrained_embeddings(args):
embeddings = read_pretrained_embeddings(args.pretrained)
emb_words = [word.strip().decode("utf-8") for word in open(args.pretrained_vocab)]
subset = read_model_defs(args.new_words).keys()
new_pretrained = os.path.join(args.out, "extracted_embeddings.vector")
new_vocab = os.path.join(args.out, "extracted_embeddings.words")
print >> sys.stderr, "writing to", new_pretrained
with open(new_vocab, "w") as v:
with open(new_pretrained, "w") as f:
for i, word in enumerate(emb_words):
if word in subset:
f.write(" ".join([str(u) for u in embeddings[i]]) + "\n")
v.write(word.encode("utf-8") + "\n")
def load_pretrained_embeddings(self, path):
self.emb_word.W.data = read_pretrained_embeddings(path)