Compression of Deep Convolutional Neural Networks under Joint Sparsity Constraints

TL;DR

This paper introduces a universal CNN compression framework that produces a single sparse model adaptable to different hardware platforms and convolution methods, significantly reducing model size and computational cost.

Contribution

The proposed method creates a universal compressed CNN model with joint sparsity in spatial and Winograd domains, enabling deployment without re-training and further complexity reduction.

Findings

Achieved up to 47.7x compression on AlexNet

Reduced computational cost by up to 23.5x

Enabled universal deployment across platforms

Abstract

We consider the optimization of deep convolutional neural networks (CNNs) such that they provide good performance while having reduced complexity if deployed on either conventional systems utilizing spatial-domain convolution or lower complexity systems designed for Winograd convolution. Furthermore, we explore the universal quantization and compression of these networks. In particular, the proposed framework produces one compressed model whose convolutional filters can be made sparse either in the spatial domain or in the Winograd domain. Hence, one compressed model can be deployed universally on any platform, without need for re-training on the deployed platform, and the sparsity of its convolutional filters can be exploited for further complexity reduction in either domain. To get a better compression ratio, the sparse model is compressed in the spatial domain which has a less number of parameters. From our experiments, we obtain $24.2\times$, $47.7\times$ and $35.4\times$ compressed models for ResNet-18, AlexNet and CT-SRCNN, while their computational cost is also reduced by $4.5\times$, $5.1\times$ and $23.5\times$, respectively.