Deep learning model compression using network sensitivity and gradients