def __init__(self,
config,
w_embeddings=None,
c_embeddings=None,
pos_embeddings=None,
ntags=None):
# build word embedding
word_ids = Input(batch_shape=(None, None), dtype='int32')
if w_embeddings is None:
word_embeddings = Embedding(input_dim=config.vocab_size,
output_dim=config.word_embedding_size,
mask_zero=True)(word_ids)
else:
word_embeddings = Embedding(input_dim=w_embeddings.shape[0],
output_dim=w_embeddings.shape[1],
mask_zero=True,
weights=[w_embeddings])(word_ids)
# combine characters and word
x = Bidirectional(
LSTM(units=config.num_word_lstm_units,
return_sequences=True))(word_embeddings)
x = Dropout(config.dropout)(x)
x = Dense(ntags)(x)
self.crf = ChainCRF()
pred = self.crf(x)
sequence_lengths = Input(batch_shape=(None, 1), dtype='int32')
self.model = Model(inputs=[word_ids, sequence_lengths], outputs=[pred])
self.config = config
def __init__(self,
config,
w_embeddings=None,
c_embeddings=None,
pos_embeddings=None,
ntags=None):
# build word embedding
word_ids = Input(batch_shape=(None, None), dtype='int32')
if w_embeddings is None:
word_embeddings = Embedding(input_dim=config.vocab_size,
output_dim=config.word_embedding_size,
mask_zero=True)(word_ids)
else:
word_embeddings = Embedding(input_dim=w_embeddings.shape[0],
output_dim=w_embeddings.shape[1],
mask_zero=True,
weights=[w_embeddings])(word_ids)
# build character based word embedding
char_ids = Input(batch_shape=(None, None, None), dtype='int32')
char_embeddings = Embedding(input_dim=config.char_vocab_size,
output_dim=config.char_embedding_size,
mask_zero=True)(char_ids)
s = K.shape(char_embeddings)
char_embeddings = Lambda(lambda x: K.reshape(
x, shape=(-1, s[-2], config.char_embedding_size)))(char_embeddings)
fwd_state = LSTM(config.num_char_lstm_units,
return_state=True)(char_embeddings)[-2]
bwd_state = LSTM(config.num_char_lstm_units,
return_state=True,
go_backwards=True)(char_embeddings)[-2]
char_embeddings = Concatenate(axis=-1)([fwd_state, bwd_state])
# shape = (batch size, max sentence length, char hidden size)
char_embeddings = Lambda(lambda x: K.reshape(
x, shape=[-1, s[1], 2 * config.num_char_lstm_units]))(
char_embeddings)
hand_feature = Input(batch_shape=(None, None,
config.hand_feature_size),
dtype='float32')
# combine characters and word
x = Concatenate(axis=-1)([word_embeddings, char_embeddings])
x = Dropout(config.dropout)(x)
x = Bidirectional(
LSTM(units=config.num_word_lstm_units, return_sequences=True))(x)
x = Dropout(config.dropout)(x)
x = Concatenate(axis=-1)([x, hand_feature])
x = Dense(ntags)(x)
self.crf = ChainCRF()
pred = self.crf(x)
sequence_lengths = Input(batch_shape=(None, 1), dtype='int32')
self.model = Model(
inputs=[word_ids, char_ids, hand_feature, sequence_lengths],
outputs=[pred])
self.config = config
def test_chain_crf(self):
vocab_size = 20
n_classes = 11
model = Sequential()
model.add(Embedding(vocab_size, n_classes))
layer = ChainCRF()
model.add(layer)
model.compile(loss=layer.loss, optimizer='sgd')
# Train first mini batch
batch_size, maxlen = 2, 2
x = np.random.randint(1, vocab_size, size=(batch_size, maxlen))
y = np.random.randint(n_classes, size=(batch_size, maxlen))
y = np.eye(n_classes)[y]
model.train_on_batch(x, y)
print(x)
print(y)