Practicalrecommendationsforgradient-basedtrainingofdeeparchitectures