def main(args):
print('Loading dataset...')
x_train, y_train = load_data_and_labels(args.train_data)
x_valid, y_valid = load_data_and_labels(args.valid_data)
x_test, y_test = load_data_and_labels(args.test_data)
x_train = np.r_[x_train, x_valid]
y_train = np.r_[y_train, y_valid]
print('Transforming datasets...')
p = ELMoTransformer()
p.fit(x_train, y_train)
print('Loading word embeddings...')
embeddings = load_glove(EMBEDDING_PATH)
embeddings = filter_embeddings(embeddings, p._word_vocab.vocab, 100)
print('Building a model.')
model = ELModel(char_embedding_dim=args.char_emb_size,
word_embedding_dim=args.word_emb_size,
char_lstm_size=args.char_lstm_units,
word_lstm_size=args.word_lstm_units,
char_vocab_size=p.char_vocab_size,
word_vocab_size=p.word_vocab_size,
num_labels=p.label_size,
embeddings=embeddings,
dropout=args.dropout)
model, loss = model.build()
model.compile(loss=loss, optimizer='adam')
print('Training the model...')
trainer = Trainer(model, preprocessor=p)
trainer.train(x_train, y_train, x_test, y_test)
print('Saving the model...')
model.save(args.weights_file, args.params_file)
def main(args):
print('Loading datasets...')
X, y = load_data_and_labels(args.data_path)
x_train, x_valid, y_train, y_valid = train_test_split(X,
y,
test_size=0.1,
random_state=42)
embeddings = KeyedVectors.load(args.embedding_path).wv
print('Transforming datasets...')
p = IndexTransformer()
p.fit(X, y)
embeddings = filter_embeddings(embeddings, p._word_vocab,
embeddings.vector_size)
print('Building a model...')
model = BiLSTMCRF(char_vocab_size=p.char_vocab_size,
word_vocab_size=p.word_vocab_size,
num_labels=p.label_size,
embeddings=embeddings,
char_embedding_dim=50)
model.build()
print('Training the model...')
trainer = Trainer(model, preprocessor=p)
trainer.train(x_train, y_train, x_valid, y_valid)
print('Saving the model...')
model.save(args.weights_file, args.params_file)
p.save(args.preprocessor_file)
def fit(self, x_train, y_train, x_valid=None, y_valid=None,
epochs=1, batch_size=32, verbose=1, callbacks=None, shuffle=True):
"""Fit the model for a fixed number of epochs.
Args:
x_train: list of training data.
y_train: list of training target (label) data.
x_valid: list of validation data.
y_valid: list of validation target (label) data.
batch_size: Integer.
Number of samples per gradient update.
If unspecified, `batch_size` will default to 32.
epochs: Integer. Number of epochs to train the model.
verbose: Integer. 0, 1, or 2. Verbosity mode.
0 = silent, 1 = progress bar, 2 = one line per epoch.
callbacks: List of `keras.callbacks.Callback` instances.
List of callbacks to apply during training.
shuffle: Boolean (whether to shuffle the training data
before each epoch). `shuffle` will default to True.
"""
p = IndexTransformer(initial_vocab=self.initial_vocab, use_char=self.use_char)
p.fit(x_train, y_train)
embeddings = filter_embeddings(self.embeddings, p._word_vocab.vocab, self.word_embedding_dim)
model = BiLSTMCRF(char_vocab_size=p.char_vocab_size,
word_vocab_size=p.word_vocab_size,
num_labels=p.label_size,
word_embedding_dim=self.word_embedding_dim,
char_embedding_dim=self.char_embedding_dim,
word_lstm_size=self.word_lstm_size,
char_lstm_size=self.char_lstm_size,
fc_dim=self.fc_dim,
dropout=self.dropout,
embeddings=embeddings,
use_char=self.use_char,
use_crf=self.use_crf)
model, loss = model.build()
model.compile(loss=loss, optimizer=self.optimizer)
trainer = Trainer(model, preprocessor=p)
trainer.train(x_train, y_train, x_valid, y_valid,
epochs=epochs, batch_size=batch_size,
verbose=verbose, callbacks=callbacks,
shuffle=shuffle)
self.p = p
self.model = model
def training(train, test):
x_train = [x.split() for x in train['sentence'].tolist()]
y_train = train['tag'].tolist()
x_train, x_val, y_train, y_val = train_test_split(x_train,
y_train,
train_size=0.8,
random_state=233)
print('Transforming datasets...')
p = IndexTransformer(use_char=True)
p.fit(x_train, y_train)
embeddings = load_glove(config.glove_file)
embeddings = filter_embeddings(embeddings, p._word_vocab.vocab,
config.glove_size)
model = BiLSTMCRF(char_vocab_size=p.char_vocab_size,
word_vocab_size=p.word_vocab_size,
num_labels=p.label_size,
word_embedding_dim=300,
char_embedding_dim=100,
word_lstm_size=100,
char_lstm_size=50,
fc_dim=100,
dropout=0.5,
embeddings=embeddings,
use_char=True,
use_crf=True)
opt = Adam(lr=0.001)
model, loss = model.build()
model.compile(loss=loss, optimizer=opt, metrics=[crf_viterbi_accuracy])
filepath = '../models/' + 'best_model'
ckp = ModelCheckpoint(filepath + '.h5',
monitor='val_crf_viterbi_accuracy',
verbose=1,
save_best_only=True,
mode='max',
save_weights_only=True)
es = EarlyStopping(monitor='val_crf_viterbi_accuracy',
min_delta=0.00001,
patience=3,
verbose=1,
mode='max')
rlr = ReduceLROnPlateau(monitor='val_crf_viterbi_accuracy',
factor=0.2,
patience=2,
verbose=1,
mode='max',
min_delta=0.0001)
callbacks = [ckp, es, rlr]
train_seq = NERSequence(x_train, y_train, config.batch_size, p.transform)
if x_val and y_val:
valid_seq = NERSequence(x_val, y_val, config.batch_size, p.transform)
f1 = F1score(valid_seq, preprocessor=p)
callbacks.append(f1)
model.fit_generator(generator=train_seq,
validation_data=valid_seq,
epochs=config.nepochs,
callbacks=callbacks,
verbose=True,
shuffle=True,
use_multiprocessing=True,
workers=42)
print("got ", len(x_train), " entries for training and ", len(x_valid),
" entries for testing")
entities = set()
for s in y_train:
for w in s:
entities.add(w)
print("Defined entities are :", entities)
preprocessor = IndexTransformer(use_char=True)
x = x_train + x_valid
y = y_train + y_valid
preprocessor.fit(x, y)
print(len(x_train), 'train sequences')
print(len(x_valid), 'valid sequences')
embeddings = filter_embeddings(wv_model, preprocessor._word_vocab.vocab,
wv_model.vector_size)
# Use pre-trained word embeddings
model = anago.models.BiLSTMCRF(
embeddings=embeddings,
use_crf=False,
use_char=True,
num_labels=preprocessor.label_size,
word_vocab_size=preprocessor.word_vocab_size,
char_vocab_size=preprocessor.char_vocab_size,
dropout=.5,
word_lstm_size=120)
model.build()
model.compile(loss=model.get_loss(), optimizer='adam', metrics=["acc"])
model.summary()
def training(train, test, fold):
x_train = [x.split() for x in train['sentence'].tolist()]
y_train = train['tag'].tolist()
x_test = [x.split() for x in test['sentence'].tolist()]
print('Transforming datasets...')
p = IndexTransformer(use_char=True)
p.fit(x_train + x_test, y_train)
skf = KFold(n_splits=config.nfolds, random_state=config.seed, shuffle=True)
embeddings = load_glove(config.glove_file)
# embeddings_fast = load_glove(config.glove_file)
embeddings_wang = load_glove(config.wang_file)
embeddings = filter_embeddings(embeddings, p._word_vocab.vocab,
config.glove_size)
# embeddings_fast = filter_embeddings(embeddings_fast, p._word_vocab.vocab, config.fasttext_size)
embeddings_wang = filter_embeddings(embeddings_wang, p._word_vocab.vocab,
config.wang_size)
embeddings = np.concatenate((embeddings, embeddings_wang), axis=1)
for n_fold, (train_indices, val_indices) in enumerate(skf.split(x_train)):
if n_fold >= fold:
print("Training fold: ", n_fold)
x_val = list(np.array(x_train)[val_indices])
y_val = list(np.array(y_train)[val_indices])
x_train_spl = list(np.array(x_train)[train_indices])
y_train_spl = list(np.array(y_train)[train_indices])
model = BiLSTMCRF(char_vocab_size=p.char_vocab_size,
word_vocab_size=p.word_vocab_size,
num_labels=p.label_size,
word_embedding_dim=1200,
char_embedding_dim=50,
word_lstm_size=300,
char_lstm_size=300,
fc_dim=50,
dropout=0.5,
embeddings=embeddings,
use_char=True,
use_crf=True)
opt = Adam(lr=0.001)
model, loss = model.build()
model.compile(loss=loss,
optimizer=opt,
metrics=[crf_viterbi_accuracy])
es = EarlyStopping(monitor='val_crf_viterbi_accuracy',
patience=3,
verbose=1,
mode='max',
restore_best_weights=True)
rlr = ReduceLROnPlateau(monitor='val_crf_viterbi_accuracy',
factor=0.2,
patience=2,
verbose=1,
mode='max')
callbacks = [es, rlr]
train_seq = NERSequence(x_train_spl, y_train_spl,
config.batch_size, p.transform)
if x_val and y_val:
valid_seq = NERSequence(x_val, y_val, config.batch_size,
p.transform)
f1 = F1score(valid_seq, preprocessor=p, fold=n_fold)
callbacks.append(f1)
model.fit_generator(generator=train_seq,
validation_data=valid_seq,
epochs=config.nepochs,
callbacks=callbacks,
verbose=True,
shuffle=True,
use_multiprocessing=True,
workers=12)
p.save('../models/best_transform.it')
model.load_weights('../models/best_model_' + str(n_fold) + '.h5')
predict(model, p, x_test, n_fold)
def fit(self, X, y):
""" Trains the NER model. Input is list of AnnotatedDocuments.
Parameters
----------
X : list(list(str))
list of list of tokens
y : list(list(str))
list of list of BIO tags
Returns
-------
self
"""
if self.embeddings is None and self.embeddings_file is None:
raise ValueError(
"Either embeddings or embeddings_file should be provided, exiting."
)
log.info("Preprocessing dataset...")
self.preprocessor_ = ELMoTransformer()
self.preprocessor_.fit(X, y)
if self.embeddings is None:
self.embeddings = load_glove(self.embeddings_file)
embeddings_dim != self.embeddings[list(
self.embeddings.keys())[0]].shape[0]
self.embeddings = filter_embeddings(
self.embeddings, self.preprocessor_._word_vocab.vocab,
embeddings_dim)
log.info("Building model...")
self.model_ = ELModel(
char_embedding_dim=self.char_embedding_dim,
word_embedding_dim=self.word_embedding_dim,
char_lstm_size=self.char_lstm_size,
word_lstm_size=self.word_lstm_size,
char_vocab_size=self.preprocessor_.char_vocab_size,
word_vocab_size=self.preprocessor_.word_vocab_size,
num_labels=self.preprocessor_.label_size,
embeddings=self.embeddings,
dropout=self.dropout)
self.model_, loss = self.model_.build()
optimizer = Adam(lr=self.learning_rate)
self.model_.compile(loss=loss, optimizer=optimizer)
self.model_.summary()
log.info('Training the model...')
self.trainer_ = Trainer(self.model_, preprocessor=self.preprocessor_)
x_train, x_valid, y_train, y_valid = train_test_split(X,
y,
test_size=0.1,
random_state=42)
self.trainer_.train(x_train,
y_train,
x_valid=x_valid,
y_valid=y_valid,
batch_size=self.batch_size,
epochs=self.max_iter)
self.tagger_ = Tagger(self.model_, preprocessor=self.preprocessor_)
return self
def bertFitV2(self,
x_train,
y_train,
x_valid=None,
y_valid=None,
epochs=1,
batch_size=32,
verbose=1,
callbacks=None,
shuffle=True):
sess = tf.Session()
bert_path = "https://tfhub.dev/google/bert_multi_cased_L-12_H-768_A-12/1"
max_seq_length = self._bretMaxLen
p = IndexTransformer(initial_vocab=self.initial_vocab,
use_char=self.use_char)
p.fit(x_train, y_train)
embeddings = filter_embeddings(self.embeddings, p._word_vocab.vocab,
self.word_embedding_dim)
#tokenizer = create_tokenizer_from_hub_module()
#print("tokenizar done")
#train_examples = convert_text_to_examples(x_train, y_train)
#(train_input_ids, train_input_masks, train_segment_ids, train_labels) = convert_examples_to_features(tokenizer,train_examples,max_seq_length=max_seq_length)
model = ABM.BertBiLSTMCRF(num_labels=p.label_size,
char_embedding_dim=self.char_embedding_dim,
word_lstm_size=self.word_lstm_size,
char_lstm_size=self.char_lstm_size,
fc_dim=self.fc_dim,
use_char=self.use_char,
char_vocab_size=None,
use_crf=self.use_crf,
layer2Flag=self._layer2Flag,
layerdropout=self._layerdropout,
bretFlag=self._bretFlag,
bretMaxLen=self._bretMaxLen,
bert_path=self._bert_path)
model, loss = model.build()
# Instantiate variables
ABM.initialize_vars(sess)
model.compile(loss=loss, optimizer=self.optimizer)
trainer = Trainer(model, preprocessor=p)
trainer.train(x_train,
y_train,
x_valid,
y_valid,
epochs=epochs,
batch_size=batch_size,
verbose=verbose,
callbacks=callbacks,
shuffle=shuffle)
self.p = p
self.model = model
x_train, y_train = pickle.load(f)
with open("data_valid.pkl", "rb") as f:
x_valid, y_valid = pickle.load(f)
with open("data_test.pkl", "rb") as f:
x_test, y_test = pickle.load(f)
x_train = np.r_[x_train, x_valid]
y_train = np.r_[y_train, y_valid]
print('Transforming datasets...')
p = ELMoTransformer()
p.fit(x_train, y_train)
print('Loading word embeddings...')
embeddings = load_glove(EMBEDDING_PATH)
embeddings = filter_embeddings(embeddings, p._word_vocab.vocab, EMBEDDING_DIM)
print('Building a model.')
model = ELModel(char_embedding_dim=32,
word_embedding_dim=EMBEDDING_DIM,
char_lstm_size=32,
word_lstm_size=EMBEDDING_DIM,
char_vocab_size=p.char_vocab_size,
word_vocab_size=p.word_vocab_size,
num_labels=p.label_size,
embeddings=embeddings)
model, loss = model.build()
model.compile(loss=loss, optimizer='adam')
print('Training the model...')
trainer = Trainer(model, preprocessor=p)
