DMC: Differentiable Model Compression for Hardware-Efficient Convolutional Neural Network

Abstract

Hardware-efficient CNN model design can be divided into two stages: training of a large baseline network to achieve high accuracy and applying model compression to create a smaller network, at the possible expense of a slight reduction in accuracy. This paper proposes a new differential model compression (DMC) method based on bilevel optimization to find the importance of channels in a pretrained CNN. Experimental results show that, for model compression for an image classification task, DMC requires only 12 GPU minutes to achieve a similar compression ratio, but with increased image classification accuracy, when cmpared to the previous best method.