def __init__(self, config, ntags=None):
# build input, directly feed with word embedding by the data generator
word_input = Input(shape=(None, config.word_embedding_size), name='word_input')
# build character based embedding
char_input = Input(shape=(None, config.max_char_length), dtype='int32', name='char_input')
char_embeddings = TimeDistributed(Embedding(input_dim=config.char_vocab_size,
output_dim=config.char_embedding_size,
mask_zero=True,
#embeddings_initializer=RandomUniform(minval=-0.5, maxval=0.5),
name='char_embeddings'
))(char_input)
chars = TimeDistributed(Bidirectional(LSTM(config.num_char_lstm_units, return_sequences=False)))(char_embeddings)
# length of sequence not used for the moment (but used for f1 communication)
length_input = Input(batch_shape=(None, 1), dtype='int32', name='length_input')
# combine characters and word embeddings
x = Concatenate()([word_input, chars])
x = Dropout(config.dropout)(x)
x = Bidirectional(LSTM(units=config.num_word_lstm_units,
return_sequences=True,
recurrent_dropout=config.recurrent_dropout))(x)
x = Dropout(config.dropout)(x)
x = Dense(config.num_word_lstm_units, activation='tanh')(x)
x = Dense(ntags)(x)
self.crf = ChainCRF()
pred = self.crf(x)
self.model = Model(inputs=[word_input, char_input, length_input], outputs=[pred])
self.config = config
def __init__(self, config, ntags=None):
# build input, directly feed with word embedding by the data generator
word_input = Input(shape=(None, config.word_embedding_size), name='word_input')
# build character based embedding
char_input = Input(shape=(None, config.max_char_length), dtype='int32', name='char_input')
char_embeddings = TimeDistributed(
Embedding(input_dim=config.char_vocab_size,
output_dim=config.char_embedding_size,
mask_zero=True,
name='char_embeddings'
))(char_input)
dropout = Dropout(config.dropout)(char_embeddings)
conv1d_out = TimeDistributed(Conv1D(kernel_size=3, filters=30, padding='same',activation='tanh', strides=1))(dropout)
maxpool_out = TimeDistributed(GlobalMaxPooling1D())(conv1d_out)
chars = Dropout(config.dropout)(maxpool_out)
# custom features input and embeddings
casing_input = Input(batch_shape=(None, None,), dtype='int32', name='casing_input')
"""
casing_embedding = Embedding(input_dim=config.case_vocab_size,
output_dim=config.case_embedding_size,
mask_zero=True,
trainable=False,
name='casing_embedding')(casing_input)
casing_embedding = Dropout(config.dropout)(casing_embedding)
"""
# length of sequence not used for the moment (but used for f1 communication)
length_input = Input(batch_shape=(None, 1), dtype='int32')
# combine words, custom features and characters
x = Concatenate(axis=-1)([word_input, chars])
x = Dropout(config.dropout)(x)
x = Bidirectional(LSTM(units=config.num_word_lstm_units,
return_sequences=True,
recurrent_dropout=config.recurrent_dropout))(x)
x = Dropout(config.dropout)(x)
x = Dense(config.num_word_lstm_units, activation='tanh')(x)
x = Dense(ntags)(x)
self.crf = ChainCRF()
pred = self.crf(x)
self.model = Model(inputs=[word_input, char_input, casing_input, length_input], outputs=[pred])
self.config = config