model.py

Embedding(self.vocabulary_size + 1, self.word_dim,
                            trainable=True, mask_zero=True,
                            dropout=self.dropout,
                            input_length=self.max_len)(sequence)
forward = LSTM(self.hidden_size, return_sequences=True)(emb)
backward = LSTM(self.hidden_size, go_backwards=True, return_sequences=True)(emb)
merged = merge([forward, backward], mode='concat')#, concat_axis=-1) # for theano uncomment
output = TimeDistributed(Dense(len(self.tag2ind), activation="softmax"))(merged)