A new framework for training state-of-the-art visual classification models