What is your training data? Does it have both vision and descriptive components?