On using pretext tasks to learn representations from network logs