def test_elmo_token_representation(self):
# Load the test words and convert to char ids
with open(os.path.join(self.elmo_fixtures_path, 'vocab_test.txt'), 'r') as fin:
tokens = fin.read().strip().split('\n')
indexer = ELMoTokenCharactersIndexer()
indices = [indexer.token_to_indices(Token(token), Vocabulary()) for token in tokens]
# There are 457 tokens. Reshape into 10 batches of 50 tokens.
sentences = []
for k in range(10):
sentences.append(
indexer.pad_token_sequence(
indices[(k * 50):((k + 1) * 50)], desired_num_tokens=50, padding_lengths={}
)
)
batch = torch.from_numpy(numpy.array(sentences))
elmo_token_embedder = _ElmoCharacterEncoder(self.options_file, self.weight_file)
elmo_token_embedder_output = elmo_token_embedder(batch)
# Reshape back to a list of words and compare with ground truth. Need to also
# remove <S>, </S>
actual_embeddings = remove_sentence_boundaries(
elmo_token_embedder_output['token_embedding'],
elmo_token_embedder_output['mask']
)[0].data.numpy()
actual_embeddings = actual_embeddings.reshape(-1, actual_embeddings.shape[-1])
embedding_file = os.path.join(self.elmo_fixtures_path, 'elmo_token_embeddings.hdf5')
with h5py.File(embedding_file, 'r') as fin:
expected_embeddings = fin['embedding'][...]
assert numpy.allclose(actual_embeddings[:len(tokens)], expected_embeddings, atol=1e-6)
def test_elmo_token_representation(self):
# Load the test words and convert to char ids
with open(os.path.join(self.fixtures_path, 'vocab_test.txt'), 'r') as fin:
tokens = fin.read().strip().split('\n')
indexer = ELMoTokenCharactersIndexer()
indices = [indexer.token_to_indices(Token(token), Vocabulary()) for token in tokens]
# There are 457 tokens. Reshape into 10 batches of 50 tokens.
sentences = []
for k in range(10):
sentences.append(
indexer.pad_token_sequence(
indices[(k * 50):((k + 1) * 50)], desired_num_tokens=50, padding_lengths={}
)
)
batch = Variable(torch.from_numpy(numpy.array(sentences)))
elmo_token_embedder = _ElmoCharacterEncoder(self.options_file, self.weight_file)
elmo_token_embedder_output = elmo_token_embedder(batch)
# Reshape back to a list of words and compare with ground truth. Need to also
# remove <S>, </S>
actual_embeddings = remove_sentence_boundaries(
elmo_token_embedder_output['token_embedding'],
elmo_token_embedder_output['mask']
)[0].data.numpy()
actual_embeddings = actual_embeddings.reshape(-1, actual_embeddings.shape[-1])
embedding_file = os.path.join(self.fixtures_path, 'elmo_token_embeddings.hdf5')
with h5py.File(embedding_file, 'r') as fin:
expected_embeddings = fin['embedding'][...]
assert numpy.allclose(actual_embeddings[:len(tokens)], expected_embeddings, atol=1e-6)
def __init__(self):
from allennlp.modules.elmo import _ElmoCharacterEncoder
if not path.isdir(self.path('elmo')):
makedirs(self.path('elmo'))
self.fweights = self.ensure_file(path.join('elmo', 'weights.hdf5'),
url=self.settings['weights'])
self.foptions = self.ensure_file(path.join('elmo', 'options.json'),
url=self.settings['options'])
self.embeddings = _ElmoCharacterEncoder(self.foptions, self.fweights)
def test_elmo_token_representation_bos_eos(self):
# The additional <S> and </S> embeddings added by the embedder should be as expected.
indexer = ELMoTokenCharactersIndexer()
elmo_token_embedder = _ElmoCharacterEncoder(self.options_file, self.weight_file)
for correct_index, token in [[0, '<S>'], [2, '</S>']]:
indices = indexer.token_to_indices(Token(token), Vocabulary())
indices = torch.from_numpy(numpy.array(indices)).view(1, 1, -1)
embeddings = elmo_token_embedder(indices)['token_embedding']
assert numpy.allclose(embeddings[0, correct_index, :].data.numpy(), embeddings[0, 1, :].data.numpy())
def test_elmo_token_representation_bos_eos(self):
# The additional <S> and </S> embeddings added by the embedder should be as expected.
indexer = ELMoTokenCharactersIndexer()
elmo_token_embedder = _ElmoCharacterEncoder(self.options_file, self.weight_file)
for correct_index, token in [[0, '<S>'], [2, '</S>']]:
indices = indexer.tokens_to_indices([Token(token)], Vocabulary(), "correct")
indices = torch.from_numpy(numpy.array(indices["correct"])).view(1, 1, -1)
embeddings = elmo_token_embedder(indices)['token_embedding']
assert numpy.allclose(embeddings[0, correct_index, :].data.numpy(), embeddings[0, 1, :].data.numpy())
def __init__(
self,
options_file: str,
weight_file: str,
requires_grad: bool = False,
vocab_to_cache: List[str] = None,
) -> None:
super(_ElmoBiLm, self).__init__()
self._token_embedder = _ElmoCharacterEncoder(
options_file, weight_file, requires_grad=requires_grad
)
self._requires_grad = requires_grad
if requires_grad and vocab_to_cache:
logging.warning(
"You are fine tuning ELMo and caching char CNN word vectors. "
"This behaviour is not guaranteed to be well defined, particularly. "
"if not all of your inputs will occur in the vocabulary cache."
)
# This is an embedding, used to look up cached
# word vectors built from character level cnn embeddings.
self._word_embedding = None
self._bos_embedding: torch.Tensor = None
self._eos_embedding: torch.Tensor = None
if vocab_to_cache:
logging.info("Caching character cnn layers for words in vocabulary.")
# This sets 3 attributes, _word_embedding, _bos_embedding and _eos_embedding.
# They are set in the method so they can be accessed from outside the
# constructor.
self.create_cached_cnn_embeddings(vocab_to_cache)
with open(cached_path(options_file), "r") as fin:
options = json.load(fin)
if not options["lstm"].get("use_skip_connections"):
raise ConfigurationError("We only support pretrained biLMs with residual connections")
self._elmo_lstm = ElmoLstm(
input_size=options["lstm"]["projection_dim"],
hidden_size=options["lstm"]["projection_dim"],
cell_size=options["lstm"]["dim"],
num_layers=options["lstm"]["n_layers"],
memory_cell_clip_value=options["lstm"]["cell_clip"],
state_projection_clip_value=options["lstm"]["proj_clip"],
requires_grad=requires_grad,
)
self._elmo_lstm.load_weights(weight_file)
# Number of representation layers including context independent layer
self.num_layers = options["lstm"]["n_layers"] + 1
def test_elmo_token_representation(self):
# Load the test words and convert to char ids
with open(os.path.join(self.elmo_fixtures_path, "vocab_test.txt"),
"r") as fin:
words = fin.read().strip().split("\n")
vocab = Vocabulary()
indexer = ELMoTokenCharactersIndexer()
tokens = [Token(word) for word in words]
indices = indexer.tokens_to_indices(tokens, vocab)
# There are 457 tokens. Reshape into 10 batches of 50 tokens.
sentences = []
for k in range(10):
char_indices = indices["elmo_tokens"][(k * 50):((k + 1) * 50)]
sentences.append(
indexer.as_padded_tensor_dict(
{"elmo_tokens": char_indices},
padding_lengths={"elmo_tokens": 50})["elmo_tokens"])
batch = torch.stack(sentences)
elmo_token_embedder = _ElmoCharacterEncoder(self.options_file,
self.weight_file)
elmo_token_embedder_output = elmo_token_embedder(batch)
# Reshape back to a list of words and compare with ground truth. Need to also
# remove <S>, </S>
actual_embeddings = remove_sentence_boundaries(
elmo_token_embedder_output["token_embedding"],
elmo_token_embedder_output["mask"])[0].data.numpy()
actual_embeddings = actual_embeddings.reshape(
-1, actual_embeddings.shape[-1])
embedding_file = os.path.join(self.elmo_fixtures_path,
"elmo_token_embeddings.hdf5")
with h5py.File(embedding_file, "r") as fin:
expected_embeddings = fin["embedding"][...]
assert numpy.allclose(actual_embeddings[:len(tokens)],
expected_embeddings,
atol=1e-6)
def __init__(self,
options_file: str,
weight_file: str,
requires_grad: bool = False,
vocab_to_cache: List[str] = None,
combine_method="weighted-sum",
random_init=False) -> None:
super(ElmoBiLm, self).__init__()
self._token_embedder = _ElmoCharacterEncoder(
options_file, weight_file, requires_grad=requires_grad)
self._requires_grad = requires_grad
self._word_embedding = None
self._bos_embedding: torch.Tensor = None
self._eos_embedding: torch.Tensor = None
with open(cached_path(options_file), "r") as fin:
options = json.load(fin)
if not options["lstm"].get("use_skip_connections"):
raise ConfigurationError(
"We only support pretrained biLMs with residual connections")
self._elmo_lstm = LatticeElmoLstm(
input_size=options["lstm"]["projection_dim"],
hidden_size=options["lstm"]["projection_dim"],
cell_size=options["lstm"]["dim"],
num_layers=options["lstm"]["n_layers"],
memory_cell_clip_value=options["lstm"]["cell_clip"],
state_projection_clip_value=options["lstm"]["proj_clip"],
requires_grad=requires_grad,
combine_method=combine_method)
if not random_init:
self._elmo_lstm.load_weights(weight_file)
else:
print("WARNING!!! ELMo weights will not be loaded!!!")
# Number of representation layers including context independent layer
self.num_layers = options["lstm"]["n_layers"] + 1
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 __init__(self):
super(Model, self).__init__()
options_file = "https://allennlp.s3.amazonaws.com/models/elmo/2x4096_512_2048cnn_2xhighway/elmo_2x4096_512_2048cnn_2xhighway_options.json"
weight_file = "https://allennlp.s3.amazonaws.com/models/elmo/2x4096_512_2048cnn_2xhighway/elmo_2x4096_512_2048cnn_2xhighway_weights.hdf5"
self._token_embedder = _ElmoCharacterEncoder(options_file, weight_file)
def test_elmo_character_encoder_with_allennlp():
allennlp_embedder = _ElmoCharacterEncoder(
ELMO_OPTIONS_FILE,
ELMO_WEIGHT_FILE,
)
embedder = ElmoCharacterEncoderFactory(
ELMO_OPTIONS_FILE,
ELMO_WEIGHT_FILE,
).create()
allennlp_parameters = [
'_char_embedding_weights',
'char_conv_0.bias',
'char_conv_0.weight',
'char_conv_1.bias',
'char_conv_1.weight',
'char_conv_2.bias',
'char_conv_2.weight',
'char_conv_3.bias',
'char_conv_3.weight',
'char_conv_4.bias',
'char_conv_4.weight',
'_projection.bias',
'_projection.weight',
]
embedder_parameters = [
'char_embedding.weight',
'char_conv_0.bias',
'char_conv_0.weight',
'char_conv_1.bias',
'char_conv_1.weight',
'char_conv_2.bias',
'char_conv_2.weight',
'char_conv_3.bias',
'char_conv_3.weight',
'char_conv_4.bias',
'char_conv_4.weight',
'output_proj.bias',
'output_proj.weight',
]
allennlp_parameters_diff = [
'_highways._layers.0.bias',
'_highways._layers.0.weight',
'_highways._layers.1.bias',
'_highways._layers.1.weight',
]
embedder_parameters_diff = [
'highway.layers_0.bias',
'highway.layers_0.weight',
'highway.layers_1.bias',
'highway.layers_1.weight',
]
assert len(allennlp_parameters) == len(embedder_parameters)
assert len(allennlp_parameters_diff) == len(embedder_parameters_diff)
allennlp_embedder_named_parameters = dict(
allennlp_embedder.named_parameters())
# Same.
for allennlp_param, embedder_param in zip(allennlp_parameters,
embedder_parameters):
allennlp_w = allennlp_embedder_named_parameters[allennlp_param].data
embedder_w = embedder.named_parameters()[embedder_param].data
np.testing.assert_array_equal(embedder_w.numpy(), allennlp_w.numpy())
assert embedder_w.dtype == allennlp_w.dtype
# Diff on highway.
for allennlp_param, embedder_param in zip(allennlp_parameters_diff,
embedder_parameters_diff):
allennlp_w = allennlp_embedder_named_parameters[allennlp_param].data
embedder_w = embedder.named_parameters()[embedder_param].data
assert embedder_w.dtype == allennlp_w.dtype
np.testing.assert_raises(
AssertionError,
np.testing.assert_array_equal,
embedder_w.numpy(),
allennlp_w.numpy(),
)
sentences = [
['ELMo', 'helps', 'disambiguate', 'ELMo', 'from', 'Elmo', '.'],
['The', 'sentence', '.'],
]
# `(2, 7, 50)`
character_ids = _sentences_to_ids(sentences)
# AllenNLP.
out = allennlp_embedder(character_ids)
allennlp_token_embedding, _ = remove_sentence_boundaries(
out['token_embedding'], out['mask'])
assert list(allennlp_token_embedding.shape) == [2, 7, 16]
# Ours.
inputs = pack_padded_sequence(character_ids, [7, 3], batch_first=True)
out = embedder(inputs.data)
ours_token_embedding = _unpack(out, inputs.batch_sizes)
assert list(ours_token_embedding.shape) == [2, 7, 16]
np.testing.assert_array_almost_equal(
ours_token_embedding.data.numpy(),
allennlp_token_embedding.data.numpy(),
)
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(), "indices")["indices"]
sentences = []
for k in range((len(indices) // 50) + 1):
sentences.append(indexer.pad_token_sequence(indices[(k * 50):((k + 1) * 50)],
desired_num_tokens=50,
padding_lengths={}))
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):
array = numpy.array(sentences[i * batch_size: (i + 1) * batch_size])
if device != -1:
batch = torch.from_numpy(array).cuda(device)
else:
batch = torch.from_numpy(array)
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 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")
