def __init__(
self,
word_embeddings: TextFieldEmbedder,
encoders: torch.nn.ModuleDict,
vocab: Vocabulary,
tasks: List[Task],
) -> None:
super().__init__(vocab)
self._tasks = tasks
self.word_embeddings = word_embeddings
self.encoders = encoders
# self.hidden2tag = dict()
self.crftagger = torch.nn.ModuleDict()
self.metrics = dict()
self._inference_mode = False
for task in tasks:
tag_namespace = task.tag_namespace
self.crftagger[tag_namespace] = CrfTagger(
vocab=vocab,
text_field_embedder=self.word_embeddings,
encoder=self.encoders[task.task_type],
label_namespace=tag_namespace,
label_encoding=task.label_encoding,
calculate_span_f1=task.calculate_span_f1,
)
def __init__(self,
vocab: Vocabulary,
params: Params,
regularizer: RegularizerApplicator = None):
super(LayerNer, self).__init__(vocab = vocab, regularizer = regularizer)
# Base Text Field Embedder
text_field_embedder_params = params.pop("text_field_embedder")
text_field_embedder = BasicTextFieldEmbedder.from_params(vocab=vocab,
params=text_field_embedder_params)
self._text_field_embedder = text_field_embedder
############
# NER Stuffs
############
ner_params = params.pop("ner")
# Encoder
encoder_ner_params = ner_params.pop("encoder")
encoder_ner = Seq2SeqEncoder.from_params(encoder_ner_params)
self._encoder_ner = encoder_ner
# Tagger NER - CRF Tagger
tagger_ner_params = ner_params.pop("tagger")
tagger_ner = CrfTagger(vocab = vocab,
text_field_embedder = self._text_field_embedder,
encoder = self._encoder_ner,
label_namespace = tagger_ner_params.pop("label_namespace", "labels"),
constraint_type = tagger_ner_params.pop("constraint_type", None),
dropout = tagger_ner_params.pop("dropout", None),
regularizer = regularizer)
self._tagger_ner = tagger_ner
logger.info("Multi-Task Learning Model has been instantiated.")
def __init__(
self,
word_embeddings: TextFieldEmbedder,
encoders: torch.nn.ModuleDict,
vocab: Vocabulary,
tasks: List[Task],
) -> None:
super().__init__(vocab)
self._tasks = tasks
self.word_embeddings = word_embeddings
self.encoders = encoders
# self.hidden2tag = dict()
self.tagger_or_classifier = torch.nn.ModuleDict()
self.metrics = dict()
self._inference_mode = False
for task in tasks:
tag_namespace = task.tag_namespace
if task.task_type in TAGGING_TASKS:
self.tagger_or_classifier[tag_namespace] = CrfTagger(
vocab=vocab,
text_field_embedder=self.word_embeddings,
encoder=self.encoders[task.task_type],
label_namespace=tag_namespace,
label_encoding=task.label_encoding,
calculate_span_f1=task.calculate_span_f1,
)
elif task.task_type in CLASSIFICATION_TASKS:
self.tagger_or_classifier[tag_namespace] = BasicClassifierWithMetrics(
vocab=vocab,
text_field_embedder=self.word_embeddings,
seq2vec_encoder=self.encoders[task.task_type],
label_namespace=tag_namespace,
)
else:
raise NotImplementedError(f"model for task.task_type={task.task_type} not implemented.")
def __init__(self,
vocab: Vocabulary,
params: Params,
regularizer: RegularizerApplicator = None):
super(LayerNerEmdCoref, self).__init__(vocab=vocab,
regularizer=regularizer)
# Base text Field Embedder
text_field_embedder_params = params.pop("text_field_embedder")
text_field_embedder = BasicTextFieldEmbedder.from_params(
vocab=vocab, params=text_field_embedder_params)
self._text_field_embedder = text_field_embedder
############
# NER Stuffs
############
ner_params = params.pop("ner")
# Encoder
encoder_ner_params = ner_params.pop("encoder")
encoder_ner = Seq2SeqEncoder.from_params(encoder_ner_params)
self._encoder_ner = encoder_ner
# Tagger NER - CRF Tagger
tagger_ner_params = ner_params.pop("tagger")
tagger_ner = CrfTagger(
vocab=vocab,
text_field_embedder=self._text_field_embedder,
encoder=self._encoder_ner,
label_namespace=tagger_ner_params.pop("label_namespace", "labels"),
constraint_type=tagger_ner_params.pop("constraint_type", None),
dropout=tagger_ner_params.pop("dropout", None),
regularizer=regularizer)
self._tagger_ner = tagger_ner
############
# EMD Stuffs
############
emd_params = params.pop("emd")
# Encoder
encoder_emd_params = emd_params.pop("encoder")
encoder_emd = Seq2SeqEncoder.from_params(encoder_emd_params)
self._encoder_emd = encoder_emd
shortcut_text_field_embedder = ShortcutConnectTextFieldEmbedder(
base_text_field_embedder=self._text_field_embedder,
previous_encoders=[self._encoder_ner])
self._shortcut_text_field_embedder = shortcut_text_field_embedder
# Tagger: EMD - CRF Tagger
tagger_emd_params = emd_params.pop("tagger")
tagger_emd = CrfTagger(
vocab=vocab,
text_field_embedder=self._shortcut_text_field_embedder,
encoder=self._encoder_emd,
label_namespace=tagger_emd_params.pop("label_namespace", "labels"),
constraint_type=tagger_emd_params.pop("constraint_type", None),
dropout=tagger_ner_params.pop("dropout", None),
regularizer=regularizer)
self._tagger_emd = tagger_emd
##############
# Coref Stuffs
##############
coref_params = params.pop("coref")
# Encoder
encoder_coref_params = coref_params.pop("encoder")
encoder_coref = Seq2SeqEncoder.from_params(encoder_coref_params)
self._encoder_coref = encoder_coref
shortcut_text_field_embedder_coref = ShortcutConnectTextFieldEmbedder(
base_text_field_embedder=self._text_field_embedder,
previous_encoders=[self._encoder_ner, self._encoder_emd])
self._shortcut_text_field_embedder_coref = shortcut_text_field_embedder_coref
# Tagger: Coreference
tagger_coref_params = coref_params.pop("tagger")
eval_on_gold_mentions = tagger_coref_params.pop_bool(
"eval_on_gold_mentions", False)
init_params = tagger_coref_params.pop("initializer", None)
initializer = (InitializerApplicator.from_params(init_params)
if init_params is not None else InitializerApplicator())
tagger_coref = CoreferenceCustom(
vocab=vocab,
text_field_embedder=self._shortcut_text_field_embedder_coref,
context_layer=self._encoder_coref,
mention_feedforward=FeedForward.from_params(
tagger_coref_params.pop("mention_feedforward")),
antecedent_feedforward=FeedForward.from_params(
tagger_coref_params.pop("antecedent_feedforward")),
feature_size=tagger_coref_params.pop_int("feature_size"),
max_span_width=tagger_coref_params.pop_int("max_span_width"),
spans_per_word=tagger_coref_params.pop_float("spans_per_word"),
max_antecedents=tagger_coref_params.pop_int("max_antecedents"),
lexical_dropout=tagger_coref_params.pop_float(
"lexical_dropout", 0.2),
initializer=initializer,
regularizer=regularizer,
eval_on_gold_mentions=eval_on_gold_mentions)
self._tagger_coref = tagger_coref
if eval_on_gold_mentions:
self._tagger_coref._eval_on_gold_mentions = True
logger.info("Multi-Task Learning Model has been instantiated.")
def __init__(self,
vocab: Vocabulary,
params: Params,
regularizer: RegularizerApplicator = None):
super(LayerEmdRelation, self).__init__(vocab=vocab,
regularizer=regularizer)
# Base text Field Embedder
text_field_embedder_params = params.pop("text_field_embedder")
text_field_embedder = BasicTextFieldEmbedder.from_params(
vocab=vocab, params=text_field_embedder_params)
self._text_field_embedder = text_field_embedder
############
# EMD Stuffs
############
emd_params = params.pop("emd")
# Encoder
encoder_emd_params = emd_params.pop("encoder")
encoder_emd = Seq2SeqEncoder.from_params(encoder_emd_params)
self._encoder_emd = encoder_emd
# Tagger EMD - CRF Tagger
tagger_emd_params = emd_params.pop("tagger")
tagger_emd = CrfTagger(
vocab=vocab,
text_field_embedder=self._text_field_embedder,
encoder=self._encoder_emd,
label_namespace=tagger_emd_params.pop("label_namespace", "labels"),
label_encoding=tagger_emd_params.pop("label_encoding", None),
dropout=tagger_emd_params.pop("dropout", None),
regularizer=regularizer,
)
self._tagger_emd = tagger_emd
############################
# Relation Extraction Stuffs
############################
relation_params = params.pop("relation")
# Encoder
encoder_relation_params = relation_params.pop("encoder")
encoder_relation = Seq2SeqEncoder.from_params(encoder_relation_params)
self._encoder_relation = encoder_relation
shortcut_text_field_embedder_relation = ShortcutConnectTextFieldEmbedder(
base_text_field_embedder=self._text_field_embedder,
previous_encoders=[self._encoder_emd])
self._shortcut_text_field_embedder_relation = shortcut_text_field_embedder_relation
# Tagger: Relation
tagger_relation_params = relation_params.pop("tagger")
tagger_relation = RelationExtractor(
vocab=vocab,
text_field_embedder=self._shortcut_text_field_embedder_relation,
context_layer=self._encoder_relation,
d=tagger_relation_params.pop_int("d"),
l=tagger_relation_params.pop_int("l"),
n_classes=tagger_relation_params.pop("n_classes"),
activation=tagger_relation_params.pop("activation"),
)
self._tagger_relation = tagger_relation
logger.info("Multi-Task Learning Model has been instantiated.")
token_embedding = PretrainedBertEmbedder(
pretrained_model="bert-base-uncased",
top_layer_only=True, # conserve memory
)
else:
token_embedding = Embedding(num_embeddings=vocab.get_vocab_size('tokens'),
embedding_dim=EMBEDDING_DIM)
word_embeddings = BasicTextFieldEmbedder({"tokens": token_embedding})
lstm = PytorchSeq2SeqWrapper(
torch.nn.LSTM(word_embeddings.get_output_dim(),
hidden_dim // 2,
bidirectional=True,
batch_first=True))
model = CrfTagger(vocab, word_embeddings, lstm)
if torch.cuda.is_available():
cuda_device = 0
model = model.cuda(cuda_device)
else:
cuda_device = -1
optimizer = optim.Adam(model.parameters(), lr=learning_rate)
iterator = BucketIterator(batch_size=batch_size,
sorting_keys=[("tokens", "num_tokens")])
iterator.index_with(vocab)
# print(next(iter(iterator(train_dataset))))
trainer = Trainer(model=model,
optimizer=optimizer,
iterator=iterator,
def __init__(self,
vocab: Vocabulary,
params: Params,
regularizer: RegularizerApplicator = None):
super(LayerNerEmd, self).__init__(vocab=vocab, regularizer=regularizer)
# Base text Field Embedder
text_field_embedder_params = params.pop("text_field_embedder")
text_field_embedder = BasicTextFieldEmbedder.from_params(
vocab=vocab, params=text_field_embedder_params)
self._text_field_embedder = text_field_embedder
############
# NER Stuffs
############
ner_params = params.pop("ner")
# Encoder
encoder_ner_params = ner_params.pop("encoder")
encoder_ner = Seq2SeqEncoder.from_params(encoder_ner_params)
self._encoder_ner = encoder_ner
# Tagger NER - CRF Tagger
tagger_ner_params = ner_params.pop("tagger")
tagger_ner = CrfTagger(
vocab=vocab,
text_field_embedder=self._text_field_embedder,
encoder=self._encoder_ner,
label_namespace=tagger_ner_params.pop("label_namespace", "labels"),
label_encoding=tagger_ner_params.pop("label_encoding", None),
dropout=tagger_ner_params.pop("dropout", None),
regularizer=regularizer,
)
self._tagger_ner = tagger_ner
############
# EMD Stuffs
############
emd_params = params.pop("emd")
# Encoder
encoder_emd_params = emd_params.pop("encoder")
encoder_emd = Seq2SeqEncoder.from_params(encoder_emd_params)
self._encoder_emd = encoder_emd
shortcut_text_field_embedder = ShortcutConnectTextFieldEmbedder(
base_text_field_embedder=self._text_field_embedder,
previous_encoders=[self._encoder_ner])
self._shortcut_text_field_embedder = shortcut_text_field_embedder
# Tagger: EMD - CRF Tagger
tagger_emd_params = emd_params.pop("tagger")
tagger_emd = CrfTagger(
vocab=vocab,
text_field_embedder=self._shortcut_text_field_embedder,
encoder=self._encoder_emd,
label_namespace=tagger_emd_params.pop("label_namespace", "labels"),
label_encoding=tagger_emd_params.pop("label_encoding", None),
dropout=tagger_ner_params.pop("dropout", None),
regularizer=regularizer,
)
self._tagger_emd = tagger_emd
logger.info("Multi-Task Learning Model has been instantiated.")
weight_file = 'https://s3-us-west-2.amazonaws.com/allennlp/models/elmo/2x1024_128_2048cnn_1xhighway/elmo_2x1024_128_2048cnn_1xhighway_weights.hdf5'
token_embedding = ElmoTokenEmbedder(options_file, weight_file)
else:
token_embedding = Embedding(
num_embeddings=vocab.get_vocab_size('tokens'),
embedding_dim=config.embedding_dim)
word_embeddings = BasicTextFieldEmbedder({"tokens": token_embedding})
lstm = PytorchSeq2SeqWrapper(
torch.nn.LSTM(word_embeddings.get_output_dim(),
400 // 2,
bidirectional=True,
batch_first=True))
model = CrfTagger(vocab, word_embeddings, lstm)
with open(model_path, 'rb') as f:
model.load_state_dict(torch.load(f))
predictor = SentenceTaggerPredictor(model, reader)
predictions = predictor.predict_batch_instance(dataset)
golds = []
logits = []
for d in dataset:
# golds.append(d)
# print(list(d.fields['tags']))
golds.extend(list(d.fields['tags']))
for p in predictions:
# print(p['tags'])
logits.extend(p['tags'])
