Describir: On Compressing Deep Neural Networks