LSTMs & LSTMs

model_qn1 = Sequential()
model_qn1.add(Embedding(len(word_indices) + 1,
                     300,
                     weights=[input_layer_matrix],
                     input_length=40,
                     trainable=False))
model_qn1.add(LSTM(300, dropout_W=0.2, dropout_U=0.2))


model_qn2 = Sequential()
model_qn2.add(Embedding(len(word_indices) + 1,
                     300,
                     weights=[input_layer_matrix],
                     input_length=40,
                     trainable=False))
model_qn2.add(LSTM(300, dropout_W=0.2, dropout_U=0.2))


mixed_model = Sequential()
mixed_model.add(Merge([model_qn1, model_qn2], mode='concat'))
print(mixed_model.layers[-1].output_shape)

mixed_model.add(BatchNormalization())
print(mixed_model.layers[-1].output_shape)

mixed_model.add(Reshape((1, 600)))
print(mixed_model.layers[-1].output_shape)

mixed_model.add(LSTM(600, dropout_W=0.2, dropout_U=0.2))
mixed_model.add(Dense(1))
mixed_model.add(Activation('sigmoid'))

mixed_model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])
mixed_model.fit([qn1, qn2], y=responses, batch_size=300, nb_epoch=1, verbose=1, validation_split=0.1, shuffle=True)