tf.keras.backend.batch_flatten