Dual-discriminator adversarial framework for data-free quantization

doi:10.1016/j.neucom.2022.09.076

CORC > 自动化研究所 > 中国科学院自动化研究所 > 精密感知与控制研究中心 > 精密感知与控制

	Dual-discriminator adversarial framework for data-free quantization
	Li, Zhikai 1,2; Ma, Liping 2; Long, Xianlei 1,2; Xiao, Junrui 1,2; Gu, Qingyi 2
刊名	NEUROCOMPUTING
	2022-10-28
卷号	511 页码:67-77
关键词	Model compression Quantized neural networks Data-free quantization
ISSN号	0925-2312
DOI	10.1016/j.neucom.2022.09.076
通讯作者	Gu, Qingyi(qingyi.gu@ia.ac.cn)
英文摘要	Thanks to the potential to address the privacy and security issues, data-free quantization that generates samples based on the prior information in the model has recently been widely investigated. However, existing methods failed to adequately utilize the prior information and thus cannot fully restore the real-data characteristics and provide effective supervision to the quantized model, resulting in poor performance. In this paper, we propose Dual-Discriminator Adversarial Quantization (DDAQ), a novel data-free quantization framework with an adversarial learning style that enables effective sample generation and learning of the quantized model. Specifically, we employ a generator to produce meaningful and diverse samples directed by two discriminators, aiming to facilitate the matching of the batch normalization (BN) distribution and maximizing the discrepancy between the full-precision model and the quantized model, respectively. Moreover, inspired by mixed-precision quantization, i.e., the importance of each layer is different, we introduce layer importance prior to both discriminators, allowing us to make better use of the information in the model. Subsequently, the quantized model is trained with the generated samples under the supervision of the full-precision model. We evaluate DDAQ on various network structures for different vision tasks, including image classification and object detection, and the experimental results show that DDAQ outperforms all baseline methods with good generality. (C) 2022 Elsevier B.V. All rights reserved.
资助项目	Scientific Instrument Developing Project of the Chinese Academy of Sciences[YJKYYQ20200045]
WOS研究方向	Computer Science
语种	英语
出版者	ELSEVIER
WOS记录号	WOS:000871948700006
资助机构	Scientific Instrument Developing Project of the Chinese Academy of Sciences
内容类型	期刊论文
源URL	[http://ir.ia.ac.cn/handle/173211/50523]
专题	精密感知与控制研究中心_精密感知与控制
通讯作者	Gu, Qingyi
作者单位	1.Univ Chinese Acad Sci, Sch Artificial Intelligence, Jingjia Rd, Beijing, Peoples R China 2.Chinese Acad Sci, Inst Automat, East Zhongguancun Rd, Beijing, Peoples R China
推荐引用方式 GB/T 7714	Li, Zhikai,Ma, Liping,Long, Xianlei,et al. Dual-discriminator adversarial framework for data-free quantization[J]. NEUROCOMPUTING,2022,511:67-77.
APA	Li, Zhikai,Ma, Liping,Long, Xianlei,Xiao, Junrui,&Gu, Qingyi.(2022).Dual-discriminator adversarial framework for data-free quantization.NEUROCOMPUTING,511,67-77.
MLA	Li, Zhikai,et al."Dual-discriminator adversarial framework for data-free quantization".NEUROCOMPUTING 511(2022):67-77.