from keras.layers import Dense, Embedding, LSTM, TimeDistributed, Input, Bidirectional, Dropout
from keras.models import Model
from keras_contrib.layers import CRF
def create_model(maxlen, chars, word_size, infer=False):
"""
:param infer:
:param maxlen:
:param chars:
:param word_size:
:return:
"""
sequence = Input(shape=(maxlen,), dtype='int32')
embedded = Embedding(len(chars) + 1, word_size, input_length=maxlen, mask_zero=True)(sequence)
blstm = Bidirectional(LSTM(64, return_sequences=True), merge_mode='sum')(embedded)
output = TimeDistributed(Dense(5, activation='softmax'))(blstm)
model = Model(input=sequence, output=output)
if not infer:
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
return model
