中国科学院机构知识库网格
Chinese Academy of Sciences Institutional Repositories Grid
首页 机构 成果 学者
  
  1. CAS IR Grid
  2. 自动化研究所
  3. 中国科学院自动化研究所
  4. 中国科学院分子影像重点实验室
Differentiable RandAugment: Learning Selecting Weights and Magnitude Distributions of Image Transformations

文献类型：期刊论文

作者Xiao, Anqi1; Shen, Biluo1; Tian, Jie1,2,3; Hu, Zhenhua1
刊名IEEE TRANSACTIONS ON IMAGE PROCESSING
出版日期2023
卷号32页码:2413-2427
ISSN号1057-7149
关键词Task analysis Training Data models Costs Optimization Search problems Upper bound Data augmentation automated machine learning differentiable optimization random augmentation
DOI10.1109/TIP.2023.3265266
通讯作者Tian, Jie(tian@ieee.org)
英文摘要Automatic data augmentation is a technique to automatically search for strategies for image transformations, which can improve the performance of different vision tasks. RandAugment (RA), one of the most widely used automatic data augmentations, achieves great success in different scales of models and datasets. However, RA randomly selects transformations with equivalent probabilities and applies a single magnitude for all transformations, which is suboptimal for different models and datasets. In this paper, we develop Differentiable RandAugment (DRA) to learn selecting weights and magnitudes of transformations for RA. The magnitude of each transformation is modeled following a normal distribution with both learnable mean and standard deviation. We also introduce the gradient of transformations to reduce the bias in gradient estimation and KL divergence as part of the loss to reduce the optimization gap. Experiments on CIFAR-10/100 and ImageNet demonstrate the efficiency and effectiveness of DRA. Searching for only 0.95 GPU hours on ImageNet, DRA can reach a Top-1 accuracy of 78.19% with ResNet-50, which outperforms RA by 0.28% under the same settings. Transfer learning on object detection also demonstrates the power of DRA. The proposed DRA is one of the few that surpasses RA on ImageNet and has great potential to be integrated into modern training pipelines to achieve state-of-the-art performance. Our code will be made publicly available for out-of-the-box use.
资助项目National Natural Science Foundation of China (NSFC)[92059207] ; National Natural Science Foundation of China (NSFC)[62027901] ; National Natural Science Foundation of China (NSFC)[81930053] ; National Natural Science Foundation of China (NSFC)[81227901] ; Chinese Academy of Sciences (CAS) Youth Interdisciplinary[JCTD-2021-08] ; Zhuhai High-Level Health Personnel Team Project[HLHPTP201703] ; Cloud TensorProcessing Unit (TPUs) from Google's TPU Research Cloud (TRC)
WOS研究方向Computer Science ; Engineering
语种英语
出版者IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC
WOS记录号WOS:000981890900002
资助机构National Natural Science Foundation of China (NSFC) ; Chinese Academy of Sciences (CAS) Youth Interdisciplinary ; Zhuhai High-Level Health Personnel Team Project ; Cloud TensorProcessing Unit (TPUs) from Google's TPU Research Cloud (TRC)
源URL[http://ir.ia.ac.cn/handle/173211/53263]  
专题自动化研究所_中国科学院分子影像重点实验室
通讯作者Tian, Jie
作者单位1.Chinese Acad Sci, Inst Automat, CAS Key Lab Mol Imaging, Beijing Key Lab Mol Imaging, Beijing 100190, Peoples R China
2.Beihang Univ, Beijing Adv Innovat Ctr Big Data Based Precis Med, Sch Engn Med, Beijing 100191, Peoples R China
3.Xidian Univ, Engn Res Ctr Mol & Neuro Imaging, Sch Life Sci & Technol, Minist Educ, Xian 710071, Peoples R China
推荐引用方式
GB/T 7714		 Xiao, Anqi,Shen, Biluo,Tian, Jie,et al. Differentiable RandAugment: Learning Selecting Weights and Magnitude Distributions of Image Transformations[J]. IEEE TRANSACTIONS ON IMAGE PROCESSING,2023,32:2413-2427.
APA Xiao, Anqi,Shen, Biluo,Tian, Jie,&Hu, Zhenhua.(2023).Differentiable RandAugment: Learning Selecting Weights and Magnitude Distributions of Image Transformations.IEEE TRANSACTIONS ON IMAGE PROCESSING,32,2413-2427.
MLA Xiao, Anqi,et al."Differentiable RandAugment: Learning Selecting Weights and Magnitude Distributions of Image Transformations".IEEE TRANSACTIONS ON IMAGE PROCESSING 32(2023):2413-2427.

入库方式： OAI收割

来源：自动化研究所

浏览0
下载0
收藏0
其他版本

除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。