If you are using keras exclusively with the tensorflow backend, I would recommend to use the keras implementation found in tf.keras rather than the keras module. That way, you won't scratch your head about possible incompatibilities or bugs (see also that question).