Bridging the gap between deep learning and sparse matrix format selection