Full Stack Approach For Efficient Deep Learning Inference