Versatile, full‐spectrum, and swift network sampling for model generation. (September 2022)

• Record Type: Journal Article
• Authors: Wang, Huanyu; Zhang, Yongshun; Wu, Jianxin
• Is Part Of: Pattern recognition. Volume 129(2022)
• Publication Date: 2022-09
• Subjects: Model generation -- Convolutional neural networks -- Structured pruning -- Model compression; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2022.108729
• Languages: English
• ISSNs: 0031-3203

Abstract

Abstract: Given one task, it is difficult to generate CNN models for many different hardware platforms with extremely diverse computing power for this task. Repeating network pruning or architecture search for each platform is very time-consuming. In this paper, we propose properties that are required for this model generation problem: versatile (fits diverse applications and network structures), full-spectrum (generates models for devices with tiny to gigantic computing power), and swift (total training time for all platforms is short, and generated models have low latency). We show that existing methods do not satisfy these requirements and propose a VFS method (the V/F/S represents Versatile/Full-spectrum/Swift, respectively). VFS uses importance sampling to sample many submodels with versatile structures and with different input image resolutions. We propose new fine-tuning strategies that only need to fine-tune a best candidate submodel for few epochs for each platform. VFS satisfies all three requirements. It generates versatile models with low latency for diverse applications, is suitable for devices with a wide range of computing power differences, and the models which are generated by VFS achieve state-of-the-art accuracy.