def train(self, x_train, y_train, x_valid=None, y_valid=None,
epochs=1, batch_size=32, verbose=1, callbacks=None, shuffle=True):
"""Trains the model for a fixed number of epochs (iterations on a dataset).
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.
"""
train_seq = NERSequence(x_train, y_train, batch_size, self._preprocessor.transform)
if x_valid and y_valid:
valid_seq = NERSequence(x_valid, y_valid, batch_size, self._preprocessor.transform)
f1 = F1score(valid_seq, preprocessor=self._preprocessor)
es = EarlyStopping(monitor='loss', mode='min', verbose=1)
callbacks = [f1,es] + callbacks if callbacks else [f1,es]
self._model.fit_generator(generator=train_seq,
epochs=epochs,
callbacks=callbacks,
verbose=verbose,
shuffle=shuffle)
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 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)
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)