def test_inverse_transform_one_cat(self):
x_train, y_train = [['a']], [['O']]
it = IndexTransformer()
it.fit(self.x, self.y)
_, y = it.transform(x_train, y_train)
inv_y = it.inverse_transform(y)
self.assertNotEqual(inv_y, self.y)
def test_inverse_transform_unknown_token(self):
x_train, y_train = [['a', 'b']], [['X', 'O']]
it = IndexTransformer()
it.fit(self.x, self.y)
_, y = it.transform(x_train, y_train)
inv_y = it.inverse_transform(y)
self.assertNotEqual(inv_y, self.y)
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 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)
print('Transforming datasets...')
p = IndexTransformer(use_char=args.no_char_feature)
p.fit(x_train, y_train)
print('Building a model.')
model = BiLSTMCRF(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,
dropout=args.dropout,
use_char=args.no_char_feature,
use_crf=args.no_use_crf)
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_valid, y_valid)
print('Saving the model...')
model.save(args.weights_file, args.params_file)
p.save(args.preprocessor_file)
def test_vocab_size_lower_off(self):
word_vocab_size = 5
char_vocab_size = 4
label_size = 3
# lower is not effective.
it = IndexTransformer(lower=False)
it.fit(self.x, self.y)
self.assertEqual(it.word_vocab_size, word_vocab_size + 2) # pad, unk
self.assertEqual(it.char_vocab_size, char_vocab_size + 2) # pad, unk
self.assertEqual(it.label_size, label_size + 1) # pad
def test_vocab_size_with_initial_vocab(self):
vocab = {'aaa', 'aab', 'aac'}
word_vocab_size = 4 + len(vocab)
char_vocab_size = 4
label_size = 3
# Add initial vocab.
it = IndexTransformer(lower=True, initial_vocab=vocab)
it.fit(self.x, self.y)
self.assertEqual(it.word_vocab_size, word_vocab_size + 2) # pad, unk
self.assertEqual(it.char_vocab_size, char_vocab_size + 2) # pad, unk
self.assertEqual(it.label_size, label_size + 1) # pad
def test_batch_iter(self):
X, y = load_data_and_labels(self.filename)
batch_size = 32
p = IndexTransformer()
p.fit(X, y)
gen = NERSequence(X, y, batch_size, preprocess=p.transform)
y_gen = []
for i in range(len(gen)):
x1, y1 = gen[i]
y_gen.extend(y1)
self.assertEqual(len(y_gen), len(y))
def test_train_no_character(self):
p = IndexTransformer(use_char=False)
p.fit(self.x_train, self.y_train)
model = BiLSTMCRF(word_vocab_size=p.word_vocab_size,
num_labels=p.label_size,
use_crf=False,
use_char=False)
model, loss = model.build()
model.compile(loss=loss, optimizer='adam')
trainer = Trainer(model, preprocessor=p)
trainer.train(self.x_train,
self.y_train,
x_valid=self.x_valid,
y_valid=self.y_valid)
def test_batch_iter(self):
X, y = load_data_and_labels(self.filename)
batch_size = 32
p = IndexTransformer()
p.fit(X, y)
steps, generator = batch_iter(X,
y,
batch_size,
shuffle=False,
preprocessor=p)
y_gen = []
for _ in range(steps):
x1, y1 = next(generator)
y_gen.extend(y1)
self.assertEqual(len(y_gen), len(y))
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 test_transform_unknown_token(self):
it = IndexTransformer()
it.fit(self.x, self.y)
x_train, y_train = [['aaa']], [['X']]
X, y = it.transform(x_train, y_train)
words, chars, length = X
# Check sequence length.
self.assertEqual(len(words), len(x_train))
self.assertEqual(len(chars), len(x_train))
self.assertEqual(len(y), len(y_train))
# Check sequence type.
self.assertIsInstance(words, np.ndarray)
self.assertIsInstance(chars, np.ndarray)
self.assertIsInstance(y, np.ndarray)
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
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)
class BiLstmCrfNER(NERModel):
def __init__(self,
word_embedding_dim=100,
char_embedding_dim=25,
word_lstm_size=100,
char_lstm_size=25,
fc_dim=100,
dropout=0.5,
embeddings=None,
use_char=True,
use_crf=True,
batch_size=16,
learning_rate=0.001,
max_iter=10):
""" Construct a BiLSTM-CRF NER model. Model is augmented with character
level embeddings as well as word embeddings by default. Implementation
is provided by the Anago project.
Parameters
----------
word_embedding_dim : int, optional, default 100
word embedding dimensions.
char_embedding_dim : int, optional, default 25
character embedding dimensions.
word_lstm_size : int, optional, default 100
character LSTM feature extractor output dimensions.
char_lstm_size : int, optional, default 25
word tagger LSTM output dimensions.
fc_dim : int, optional, default 100
output fully-connected layer size.
dropout : float, optional, default 0.5
dropout rate.
embeddings : numpy array
word embedding matrix.
use_char : bool, optional, default True
add char feature.
use_crf : bool, optional, default True
use crf as last layer.
batch_size : int, optional, default 16
training batch size.
learning_rate : float, optional, default 0.001
learning rate for Adam optimizer
max_iter : int
number of epochs of training
Attributes
----------
preprocessor_ : reference to preprocessor
model_ : reference to generated model
trainer_ : internal reference to Anago Trainer (model)
tagger_ : internal reference to Anago Tagger (predictor)
"""
super().__init__()
self.word_embedding_dim = word_embedding_dim
self.char_embedding_dim = char_embedding_dim
self.word_lstm_size = word_lstm_size
self.char_lstm_size = char_lstm_size
self.fc_dim = fc_dim
self.dropout = dropout
self.embedding = None
self.use_char = True
self.use_crf = True
self.batch_size = batch_size
self.learning_rate = learning_rate
self.max_iter = max_iter
# populated by fit() and load(), expected by save() and transform()
self.preprocessor_ = None
self.model_ = None
self.trainer_ = None
self.tagger_ = None
def fit(self, X, y):
""" Trains the NER model. Input is list of list of tokens and tags.
Parameters
----------
X : list(list(str))
list of list of tokens
y : list(list(str))
list of list of BIO tags
Returns
-------
self
"""
log.info("Preprocessing dataset...")
self.preprocessor_ = IndexTransformer(use_char=self.use_char)
self.preprocessor_.fit(X, y)
log.info("Building model...")
self.model_ = BiLSTMCRF(
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,
dropout=self.dropout,
use_char=self.use_char,
use_crf=self.use_crf)
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 predict(self, X):
""" Predicts using the NER model.
Parameters
----------
X : list(list(str))
list of list of tokens.
Returns
-------
y : list(list(str))
list of list of predicted BIO tags.
"""
if self.tagger_ is None:
raise ValueError("No tagger found, either run fit() to train or load() a trained model")
log.info("Predicting from model...")
ypreds = [self.tagger_.predict(" ".join(x)) for x in X]
return ypreds
def save(self, dirpath):
""" Saves model to local disk, given a dirpath
Parameters
----------
dirpath : str
a directory where model artifacts will be saved.
Model saves a weights.h5 weights file, a params.json parameter
file, and a preprocessor.pkl preprocessor file.
Returns
-------
None
"""
if self.model_ is None or self.preprocessor_ is None:
raise ValueError("No model artifacts to save, either run fit() to train or load() a trained model")
if not os.path.exists(dirpath):
os.makedirs(dirpath)
weights_file = os.path.join(dirpath, "weights.h5")
params_file = os.path.join(dirpath, "params.json")
preprocessor_file = os.path.join(dirpath, "preprocessor.pkl")
save_model(self.model_, weights_file, params_file)
self.preprocessor_.save(preprocessor_file)
write_param_file(self.get_params(), os.path.join(dirpath, "params.yaml"))
def load(self, dirpath):
""" Loads a trained model from local disk, given the dirpath
Parameters
----------
dirpath : str
a directory where model artifacts are saved.
Returns
-------
self
"""
if not os.path.exists(dirpath):
raise ValueError("Model directory not found: {:s}".format(dirpath))
weights_file = os.path.join(dirpath, "weights.h5")
params_file = os.path.join(dirpath, "params.json")
preprocessor_file = os.path.join(dirpath, "preprocessor.pkl")
if not (os.path.exists(weights_file) or
os.path.exists(params_file) or
os.path.exists(preprocessor_file)):
raise ValueError("Model files may be corrupted, exiting")
self.model_ = load_model(weights_file, params_file)
self.preprocessor_ = IndexTransformer.load(preprocessor_file)
self.tagger_ = Tagger(self.model_, preprocessor=self.preprocessor_)
return self
print('Loading data...')
x_train, y_train = load_data_and_labels(train_path)
x_valid, y_valid = load_data_and_labels(valid_path)
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,
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)
class TestTrainer(unittest.TestCase):
@classmethod
def setUpClass(cls):
if not os.path.exists(LOG_ROOT):
os.mkdir(LOG_ROOT)
if not os.path.exists(SAVE_ROOT):
os.mkdir(SAVE_ROOT)
cls.weights_file = os.path.join(SAVE_ROOT, 'weights.h5')
cls.params_file = os.path.join(SAVE_ROOT, 'params.json')
cls.preprocessor_file = os.path.join(SAVE_ROOT, 'preprocessor.pickle')
def setUp(self):
# Load datasets.
train_path = os.path.join(DATA_ROOT, 'train.txt')
valid_path = os.path.join(DATA_ROOT, 'valid.txt')
self.x_train, self.y_train = load_data_and_labels(train_path)
self.x_valid, self.y_valid = load_data_and_labels(valid_path)
# Fit transformer.
self.p = IndexTransformer()
self.p.fit(self.x_train, self.y_train)
# Build a model.
self.model = BiLSTMCRF(char_vocab_size=self.p.char_vocab_size,
word_vocab_size=self.p.word_vocab_size,
num_labels=self.p.label_size)
self.model, loss = self.model.build()
self.model.compile(loss=loss, optimizer='adam')
def test_train(self):
trainer = Trainer(self.model, preprocessor=self.p)
trainer.train(self.x_train,
self.y_train,
x_valid=self.x_valid,
y_valid=self.y_valid)
def test_train_no_valid(self):
trainer = Trainer(self.model, preprocessor=self.p)
trainer.train(self.x_train, self.y_train)
def test_train_no_crf(self):
model = BiLSTMCRF(char_vocab_size=self.p.char_vocab_size,
word_vocab_size=self.p.word_vocab_size,
num_labels=self.p.label_size,
use_crf=False)
model, loss = model.build()
model.compile(loss=loss, optimizer='adam')
trainer = Trainer(model, preprocessor=self.p)
trainer.train(self.x_train,
self.y_train,
x_valid=self.x_valid,
y_valid=self.y_valid)
def test_train_no_character(self):
p = IndexTransformer(use_char=False)
p.fit(self.x_train, self.y_train)
model = BiLSTMCRF(word_vocab_size=p.word_vocab_size,
num_labels=p.label_size,
use_crf=False,
use_char=False)
model, loss = model.build()
model.compile(loss=loss, optimizer='adam')
trainer = Trainer(model, preprocessor=p)
trainer.train(self.x_train,
self.y_train,
x_valid=self.x_valid,
y_valid=self.y_valid)
def test_save(self):
# Train the model.
trainer = Trainer(self.model, preprocessor=self.p)
trainer.train(self.x_train, self.y_train)
# Save the model.
save_model(self.model, self.weights_file, self.params_file)
self.p.save(self.preprocessor_file)