Revisiting Training Strategies and Generalization in Deep Metric Learning