Please wait a minute...
浙江大学学报(工学版)  2021, Vol. 55 Issue (4): 626-638    DOI: 10.3785/j.issn.1008-973X.2021.04.004
计算机技术、电信技术     
边端融合的终端情境自适应深度感知模型
王虹力(),郭斌*(),刘思聪,刘佳琪,仵允港,於志文
西北工业大学 计算机学院,陕西 西安 710072
End context-adaptative deep sensing model with edge-end collaboration
Hong-li WANG(),Bin GUO*(),Si-cong LIU,Jia-qi LIU,Yun-gang WU,Zhi-wen YU
College of Computer Science, Northwestern Polytechnical University, Xi’an 710072, China
 全文: PDF(1756 KB)   HTML
摘要:

研究边端融合的深度模型终端情境自适应问题. 提出边端融合增强的模型压缩方法(X-ADMM),利用模型压缩技术简化模型结构,以层为粒度寻找模型最佳分割点,协同边端设备提高运行效率. 为了实现模型分割的动态自适应,提出基于图的自适应深度模型手术刀算法(GADS). 当模型运行情境(如存储、电量、带宽等)发生变化时,优先在邻近分割状态中快速搜索最能满足资源约束的分割点,实现快速自适应调整. 实验结果表明,该模型平均在0.1 ms内实现了模型分割点的自适应调优,在保证模型精度下降不超过2.5%的情况下,运行总时延最高下降了56.65%.

关键词: 深度学习边缘智能模型压缩模型分割自适应感知    
Abstract:

The end context adaptative of deep models with edge-end collaboration was analyzed. The partition and alternating direction method of multiplier method (X-ADMM) was proposed. The model compression was employed to simplify the model structure, and the model was partitioned at layer granularity to find the best partition point. The model can collaborate with edge-end devices to improve model operation efficiency. The graph based adaptive DNN surgery algorithm (GADS) was proposed in order to realize the dynamic adaptation of model partition. The model will preferentially search for the partition point that best meets resource constraints among surrounding partition states to achieve rapid adaptation when the running context (e.g., storage, power, bandwidth) of the model changes. The experimental results showed that the model realized the adaptive tuning of model partition point in an average of 0.1 ms. The total running latency was reduced by 56.65% at the highest with no more than 2.5% accuracy loss.

Key words: deep learning    edge intelligence    model compression    model partition    adaptive perception
收稿日期: 2021-01-26 出版日期: 2021-05-07
CLC:  TP 391  
基金资助: 国家重点研发计划资助项目(2019YFB1703901);国家自然科学基金资助项目(61772428,61725205)
通讯作者: 郭斌     E-mail: wanghongli@mail.nwpu.edu.cn;guob@nwpu.edu.cn
作者简介: 王虹力(1999—),女,硕士生,从事深度学习模型环境自适应研究. orcid.org/0000-0001-8923-4695. E-mail: wanghongli@mail.nwpu.edu.cn
服务  
把本文推荐给朋友
加入引用管理器
E-mail Alert
作者相关文章  
王虹力
郭斌
刘思聪
刘佳琪
仵允港
於志文

引用本文:

王虹力,郭斌,刘思聪,刘佳琪,仵允港,於志文. 边端融合的终端情境自适应深度感知模型[J]. 浙江大学学报(工学版), 2021, 55(4): 626-638.

Hong-li WANG,Bin GUO,Si-cong LIU,Jia-qi LIU,Yun-gang WU,Zhi-wen YU. End context-adaptative deep sensing model with edge-end collaboration. Journal of ZheJiang University (Engineering Science), 2021, 55(4): 626-638.

链接本文:

http://www.zjujournals.com/eng/CN/10.3785/j.issn.1008-973X.2021.04.004        http://www.zjujournals.com/eng/CN/Y2021/V55/I4/626

图 1  模型压缩的一般流程图
图 2  模型压缩框架图
图 3  最近邻边端模型的分割效应
图 4  基于图的自适应深度模型手术刀算法
图 5  不同网络带宽下对AlexNet进行划分的总时延
图 6  AlexNet各层输出数据量
图 7  经典CNN不同分割方案下总时延及输出数据变化图
图 8  构建2台设备下的模型分割状态图
图 9  2台设备下部分分割状态的图结构
图 10  KD(K维)树构建示例
图 11  模型压缩后推断延迟和精度下降比例
网络 A0 /% T0 /ms γ TRAP-ADMM[19] /ms ARAP-ADMM /% TX-ADMM /ms AX-ADMM /%
Alexnet 85.82 46.8 14.7 38.78 84.21 36.3 84.17
GoogleNet 87.48 943.6 15.6 883.95 84.91 532.6 84.60
Resnet-18 91.60 285.5 15.4 267.70 90.01 213.5 89.80
VGG-16 91.66 203.7 16.0 186.90 89.59 88.3 89.20
MobileNet 89.60 219.2 2.0 207.89 87.96 179.2 87.90
MobileNet 89.60 219.2 4.0 196.69 80.60 160.3 80.83
ShuffleNet 88.14 202.8 4.0 183.79 84.25 153.4 84.19
表 1  不同方法下典型深度学习模型的推断时延和精度
图 12  KNN(K近邻算法)与KD树搜索的效果对比
算法 ${L_{ p{\rm{a} } } }$ ${S_{\rm{t}}}$ /ms
Neurosurgeon 0.0822 0.4412
GADS-KNN 0.0452 0.9927
GADS-KD树(无s) 0.0452 0.1331
GADS 0.0581 0.0997
表 2  KD树搜索的有效性
情境变化 M1 /MB M2 /MB E1 /J E2 /J T ′ /s B /(MB·s?1 变化前分割状态集 变化后分割状态集
移动端存储减少 230 9 0.02 0.02 0.01 6 8,10,9 7,8,10
移动端存储减少 230 4.5 0.02 0.02 0.01 6
移动端电量降低 230 9 0.02 0.02 0.01 6 8,10,9 6,7,8
移动端电量降低 230 9 0.02 0.005 0.01 6
网络带宽降低 230 9 0.02 0.02 0.01 6 8,10,9 10,8,9
网络带宽降低 230 9 0.02 0.02 0.01 1
表 3  GADS自适应能力实验结果
图 13  AlexNet不同分割点下所需的存储量、能耗
图 14  AlexNet不同分割点下总时延与各层中间输出数据量
1 KRIZHEVSKY A, SUTSKEVER I, HINTON G E Imagenet classification with deep convolutional neural networks[J]. Advances in Neural Information Processing Systems, 2012, 25: 1097- 1105
2 HE K, ZHANG X, REN S, et al. Deep residual learning for image recognition [C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas: IEEE, 2016: 770-778.
3 ALOM M Z, TAHA T M, YAKOPCIC C, et al A state-of-the-art survey on deep learning theory and architectures[J]. Electronics, 2019, 8 (3): 292
doi: 10.3390/electronics8030292
4 ZHOU Z, CHEN X, LI E, et al Edge intelligence: paving the last mile of artificial intelligence with edge computing[J]. Proceedings of the IEEE, 2019, 107 (8): 1738- 1762
doi: 10.1109/JPROC.2019.2918951
5 LUO P, ZHU Z, LIU Z, et al. Face model compression by distilling knowledge from neurons [C]// Proceedings of the 30th AAAI Conference on Artificial Intelligence. Arizona: AAAI, 2016: 3560-3566.
6 HAN S, POOL J, TRAN J, et al. Learning both weights and connections for efficient neural networks [C]// Proceedings of the 28th International Conference on Neural Information Processing Systems. Montreal: MIT Press, 2015: 1135-1143.
7 LUO J H, WU J, LIN W. Thinet: a filter level pruning method for deep neural network compression [C]// Proceedings of the IEEE International Conference on Computer Vision. Venice: IEEE, 2017: 5058-5066.
8 LIU S, LIN Y, ZHOU Z, et al. On-demand deep model compression for mobile devices: a usage-driven model selection framework [C]// Proceedings of the 16th Annual International Conference on Mobile Systems, Applications, and Services. Munich: ACM, 2018: 389-400.
9 ZHAO Z, BARIJOUGH K M, GERSTLAUER A DeepThings: distributed adaptive deep learning inference on resource-constrained IoT edge clusters[J]. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2018, 37 (11): 2348- 2359
doi: 10.1109/TCAD.2018.2858384
10 KANG Y, HAUSWALD J, GAO C, et al Neurosurgeon: collaborative intelligence between the cloud and mobile edge[J]. ACM SIGARCH Computer Architecture News, 2017, 45 (1): 615- 629
doi: 10.1145/3093337.3037698
11 LI H, HU C, JIANG J, et al. JALAD: joint accuracy-and latency-aware deep structure decoupling for edge-cloud execution [C]// 2018 IEEE 24th International Conference on Parallel and Distributed Systems. Singapore: IEEE, 2018: 671-678.
12 MAO J, CHEN X, NIXON K W, et al. Modnn: local distributed mobile computing system for deep neural network [C]// Design, Automation and Test in Europe Conference and Exhibition. Lausanne: IEEE, 2017: 1396-1401.
13 KO J H, NA T, AMIR M F, et al. Edge-host partitioning of deep neural networks with feature space encoding for resource-constrained internet-of-things platforms [C]// 2018 15th IEEE International Conference on Advanced Video and Signal Based Surveillance. New Zealand: IEEE, 2018: 1-6.
14 LIU N, MA X, XU Z, et al. AutoCompress: an automatic DNN structured pruning framework for ultra-high compression rates [C]// Proceedings of the AAAI Conference on Artificial Intelligence. New York: AAAI, 2020, 34(04): 4876-4883.
15 HE Y, ZHANG X, SUN J. Channel pruning for accelerating very deep neural networks [C]// Proceedings of the IEEE International Conference on Computer Vision. Venice: IEEE, 2017: 1389-1397.
16 HU C, BAO W, WANG D, et al. Dynamic adaptive DNN surgery for inference acceleration on the edge [C]// IEEE Conference on Computer Communications. Paris: IEEE, 2019: 1423-1431.
17 CHEN Y H, EMER J, SZE V Eyeriss: a spatial architecture for energy-efficient dataflow for convolutional neural networks[J]. ACM SIGARCH Computer Architecture News, 2016, 44 (3): 367- 379
doi: 10.1145/3007787.3001177
18 YANG T J, CHEN Y H, SZE V. Designing energy-efficient convolutional neural networks using energy-aware pruning [C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Hawaii: IEEE, 2017: 5687-5695.
[1] 许佳辉,王敬昌,陈岭,吴勇. 基于图神经网络的地表水水质预测模型[J]. 浙江大学学报(工学版), 2021, 55(4): 601-607.
[2] 张腾,蒋鑫龙,陈益强,陈前,米涛免,陈彪. 基于腕部姿态的帕金森病用药后开-关期检测[J]. 浙江大学学报(工学版), 2021, 55(4): 639-647.
[3] 徐利锋,黄海帆,丁维龙,范玉雷. 基于改进DenseNet的水果小目标检测[J]. 浙江大学学报(工学版), 2021, 55(2): 377-385.
[4] 许豪灿,李基拓,陆国栋. 由LeNet-5从单张着装图像重建三维人体[J]. 浙江大学学报(工学版), 2021, 55(1): 153-161.
[5] 黄毅鹏,胡冀苏,钱旭升,周志勇,赵文露,马麒,沈钧康,戴亚康. SE-Mask-RCNN:多参数MRI前列腺癌分割方法[J]. 浙江大学学报(工学版), 2021, 55(1): 203-212.
[6] 陈巧红,陈翊,李文书,贾宇波. 多尺度SE-Xception服装图像分类[J]. 浙江大学学报(工学版), 2020, 54(9): 1727-1735.
[7] 郑浦,白宏阳,李伟,郭宏伟. 复杂背景下的小目标检测算法[J]. 浙江大学学报(工学版), 2020, 54(9): 1777-1784.
[8] 周登文,田金月,马路遥,孙秀秀. 基于多级特征并联的轻量级图像语义分割[J]. 浙江大学学报(工学版), 2020, 54(8): 1516-1524.
[9] 明涛,王丹,郭继昌,李锵. 基于多尺度通道重校准的乳腺癌病理图像分类[J]. 浙江大学学报(工学版), 2020, 54(7): 1289-1297.
[10] 闫旭,范晓亮,郑传潘,臧彧,王程,程明,陈龙彪. 基于图卷积神经网络的城市交通态势预测算法[J]. 浙江大学学报(工学版), 2020, 54(6): 1147-1155.
[11] 汪周飞,袁伟娜. 基于深度学习的多载波系统信道估计与检测[J]. 浙江大学学报(工学版), 2020, 54(4): 732-738.
[12] 杨冰,莫文博,姚金良. 融合局部特征与深度学习的三维掌纹识别[J]. 浙江大学学报(工学版), 2020, 54(3): 540-545.
[13] 洪炎佳,孟铁豹,黎浩江,刘立志,李立,徐硕瑀,郭圣文. 多模态多维信息融合的鼻咽癌MR图像肿瘤深度分割方法[J]. 浙江大学学报(工学版), 2020, 54(3): 566-573.
[14] 贾子钰,林友芳,张宏钧,王晶. 基于深度卷积神经网络的睡眠分期模型[J]. 浙江大学学报(工学版), 2020, 54(10): 1899-1905.
[15] 王万良,杨小涵,赵燕伟,高楠,吕闯,张兆娟. 采用卷积自编码器网络的图像增强算法[J]. 浙江大学学报(工学版), 2019, 53(9): 1728-1740.