Evaluating Differentially Private Machine Learning in Practice