On the Choice of Data for Efficient Training and Validation of End-to-End Driving Models