def main(
vocab_path: str,
elmo_config_path: str,
elmo_weights_path: str,
output_dir: str,
batch_size: int,
device: int,
use_custom_oov_token: bool = False,
):
"""
Creates ELMo word representations from a vocabulary file. These
word representations are _independent_ - they are the result of running
the CNN and Highway layers of the ELMo model, but not the Bidirectional LSTM.
ELMo requires 2 additional tokens: <S> and </S>. The first token
in this file is assumed to be an unknown token.
This script produces two artifacts: A new vocabulary file
with the <S> and </S> tokens inserted and a glove formatted embedding
file containing word : vector pairs, one per line, with all values
separated by a space.
"""
# Load the vocabulary words and convert to char ids
with open(vocab_path, "r") as vocab_file:
tokens = vocab_file.read().strip().split("\n")
# Insert the sentence boundary tokens which elmo uses at positions 1 and 2.
if tokens[0] != DEFAULT_OOV_TOKEN and not use_custom_oov_token:
raise ConfigurationError(
"ELMo embeddings require the use of a OOV token.")
tokens = [tokens[0]] + ["<S>", "</S>"] + tokens[1:]
indexer = ELMoTokenCharactersIndexer()
indices = indexer.tokens_to_indices([Token(token) for token in tokens],
Vocabulary())["tokens"]
sentences = []
for k in range((len(indices) // 50) + 1):
sentences.append(
indexer.as_padded_tensor_dict(indices[(k * 50):((k + 1) * 50)],
padding_lengths={"tokens": 50}))
last_batch_remainder = 50 - (len(indices) % 50)
if device != -1:
elmo_token_embedder = _ElmoCharacterEncoder(
elmo_config_path, elmo_weights_path).cuda(device)
else:
elmo_token_embedder = _ElmoCharacterEncoder(elmo_config_path,
elmo_weights_path)
all_embeddings = []
for i in range((len(sentences) // batch_size) + 1):
batch = torch.stack(sentences[i * batch_size:(i + 1) * batch_size])
if device != -1:
batch = batch.cuda(device)
token_embedding = elmo_token_embedder(batch)["token_embedding"].data
# Reshape back to a list of words of shape (batch_size * 50, encoding_dim)
# We also need to remove the <S>, </S> tokens appended by the encoder.
per_word_embeddings = (token_embedding[:, 1:-1, :].contiguous().view(
-1, token_embedding.size(-1)))
all_embeddings.append(per_word_embeddings)
# Remove the embeddings associated with padding in the last batch.
all_embeddings[-1] = all_embeddings[-1][:-last_batch_remainder, :]
embedding_weight = torch.cat(all_embeddings, 0).cpu().numpy()
# Write out the embedding in a glove format.
os.makedirs(output_dir, exist_ok=True)
with gzip.open(os.path.join(output_dir, "elmo_embeddings.txt.gz"),
"wb") as embeddings_file:
for i, word in enumerate(tokens):
string_array = " ".join(
str(x) for x in list(embedding_weight[i, :]))
embeddings_file.write(f"{word} {string_array}\n".encode("utf-8"))
# Write out the new vocab with the <S> and </S> tokens.
_, vocab_file_name = os.path.split(vocab_path)
with open(os.path.join(output_dir, vocab_file_name),
"w") as new_vocab_file:
for word in tokens:
new_vocab_file.write(f"{word}\n")
def test_elmo_as_array_produces_token_sequence(self):
indexer = ELMoTokenCharactersIndexer()
tokens = [Token("Second"), Token(".")]
indices = indexer.tokens_to_indices(tokens, Vocabulary())
padded_tokens = indexer.as_padded_tensor_dict(
indices, padding_lengths={"elmo_tokens": 3})
expected_padded_tokens = [
[
259,
84,
102,
100,
112,
111,
101,
260,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
],
[
259,
47,
260,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
261,
],
[
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
],
]
assert padded_tokens["elmo_tokens"].tolist() == expected_padded_tokens
def main(
vocab_path: str,
elmo_config_path: str,
elmo_weights_path: str,
output_dir: str,
batch_size: int,
device: int,
use_custom_oov_token: bool = False,
):
with open(vocab_path, "r") as vocab_file:
tokens = vocab_file.read().strip().split("\n")
if tokens[0] != DEFAULT_OOV_TOKEN and not use_custom_oov_token:
raise ConfigurationError("ELMo embeddings require the use of a OOV token.")
tokens = [tokens[0]] + ["<S>", "</S>"] + tokens[1:]
indexer = ELMoTokenCharactersIndexer()
indices = indexer.tokens_to_indices([Token(token) for token in tokens], Vocabulary())["tokens"]
sentences = []
for k in range((len(indices) // 50) + 1):
sentences.append(
indexer.as_padded_tensor_dict(
indices[(k * 50) : ((k + 1) * 50)], padding_lengths={"tokens": 50}
)
)
last_batch_remainder = 50 - (len(indices) % 50)
if device != -1:
elmo_token_embedder = _ElmoCharacterEncoder(elmo_config_path, elmo_weights_path).cuda(
device
)
else:
elmo_token_embedder = _ElmoCharacterEncoder(elmo_config_path, elmo_weights_path)
all_embeddings = []
for i in range((len(sentences) // batch_size) + 1):
batch = torch.stack(sentences[i * batch_size : (i + 1) * batch_size])
if device != -1:
batch = batch.cuda(device)
token_embedding = elmo_token_embedder(batch)["token_embedding"].data
per_word_embeddings = (
token_embedding[:, 1:-1, :].contiguous().view(-1, token_embedding.size(-1))
)
all_embeddings.append(per_word_embeddings)
all_embeddings[-1] = all_embeddings[-1][:-last_batch_remainder, :]
embedding_weight = torch.cat(all_embeddings, 0).cpu().numpy()
os.makedirs(output_dir, exist_ok=True)
with gzip.open(os.path.join(output_dir, "elmo_embeddings.txt.gz"), "wb") as embeddings_file:
for i, word in enumerate(tokens):
string_array = " ".join(str(x) for x in list(embedding_weight[i, :]))
embeddings_file.write(f"{word} {string_array}\n".encode("utf-8"))
_, vocab_file_name = os.path.split(vocab_path)
with open(os.path.join(output_dir, vocab_file_name), "w") as new_vocab_file:
for word in tokens:
new_vocab_file.write(f"{word}\n")
