检索结果-内蒙古大学图书馆

AutoQNN: An End-to-End Framework for Automatically Quantizing Neural Networks

Journal of computer Science & technology 2024年第2期39卷 401-420页

作者：龚成卢冶代素蓉邓倩杜承昆李涛 College of Software Nankai UniversityTianjin 300350China College of Computer Science Nankai UniversityTianjin 300350China State Key Laboratory of Processors Institute of Computing TechnologyChinese Academy of Sciences Beijing 100190China

Exploring the expected quantizing scheme with suitable mixed-precision policy is the key to compress deep neural networks(DNNs)in high efficiency and *** exploration implies heavy workloads for domain experts,and an automatic compression method is ***,the huge search space of the automatic method introduces plenty of computing budgets that make the automatic process challenging to be applied in real *** this paper,we propose an end-to-end framework named AutoQNN,for automatically quantizing different layers utilizing different schemes and bitwidths without any human *** can seek desirable quantizing schemes and mixed-precision policies for mainstream DNN models efficiently by involving three techniques:quantizing scheme search(QSS),quantizing precision learning(QPL),and quantized architecture generation(QAG).QSS introduces five quantizing schemes and defines three new schemes as a candidate set for scheme search,and then uses the Differentiable Neural architecture Search(DNAS)algorithm to seek the layer-or model-desired scheme from the *** is the first method to learn mixed-precision policies by reparameterizing the bitwidths of quantizing schemes,to the best of our *** optimizes both classification loss and precision loss of DNNs efficiently and obtains the relatively optimal mixed-precision model within limited model size and memory *** is designed to convert arbitrary architectures into corresponding quantized ones without manual intervention,to facilitate end-to-end neural network *** have implemented AutoQNN and integrated it into *** experiments demonstrate that AutoQNN can consistently outperform state-of-the-art *** 2-bit weight and activation of AlexNet and ResNet18,AutoQNN can achieve the accuracy results of 59.75%and 68.86%,respectively,and obtain accuracy improvements by up to 1.65%and 1.74%,respectively,compared with state-of-the-art ***,c

关键词： automatic quantization mixed precision quantizing scheme search quantizing precision learning quan-tized architecture generation

来源：评论

学校读者我要写书评

暂无评论

Dynamic inference in probabilistic graphical models 12

Dynamic inference in probabilistic graphical models

引用

12th Innovations in Theoretical computer Science Conference, ITCS 2021

作者： Feng, Weiming He, Kun Sun, Xiaoming Yin, Yitong State Key Laboratory for Novel Software Technology Nanjing University China Shenzhen University China Shenzhen Institute of Computing Sciences China State Key Laboratory of Computer Architecture Institute of Computing Technology Chinese Academy of Sciences Beijing China University of Chinese Academy of Sciences Beijing China

ISBN: (纸本)9783959771771

Probabilistic graphical models, such as Markov random fields (MRFs), are useful for describing high-dimensional distributions in terms of local dependence structures. The probabilistic inference is a fundamental problem related to graphical models, and sampling is a main approach for the problem. In this paper, we study probabilistic inference problems when the graphical model itself is changing dynamically with time. Such dynamic inference problems arise naturally in today’s application, e.g. multivariate time-series data analysis and practical learning procedures. We give a dynamic algorithm for sampling-based probabilistic inferences in MRFs, where each dynamic update can change the underlying graph and all parameters of the MRF simultaneously, as long as the total amount of changes is bounded. More precisely, suppose that the MRF has n variables and polylogarithmic-bounded maximum degree, and N(n) independent samples are sufficient for the inference for a polynomial function N(·). Our algorithm dynamically maintains an answer to the inference problem using Oe(nN(n)) space cost, and Oe(N(n) + n) incremental time cost upon each update to the MRF, as long as the Dobrushin-Shlosman condition is satisfied by the MRFs. This well-known condition has long been used for guaranteeing the efficiency of Markov chain Monte Carlo (MCMC) sampling in the traditional static setting. Compared to the static case, which requires Ω(nN(n)) time cost for redrawing all N(n) samples whenever the MRF changes, our dynamic algorithm gives a Ω(min e {n, N(n)})-factor speedup. Our approach relies on a novel dynamic sampling technique, which transforms local Markov chains (a.k.a. single-site dynamics) to dynamic sampling algorithms, and an "algorithmic Lipschitz" condition that we establish for sampling from graphical models, namely, when the MRF changes by a small difference, samples can be modified to reflect the new distribution, with cost proportional to the difference on MRF. © Weiming F

关键词： Sampling

来源：评论

学校读者我要写书评

暂无评论

Common substructure extraction of proteins by geometric invariants

Common substructure extraction of proteins by geometric inva...

引用

10th International Conference on computer-Aided Design and computer Graphics

作者： Xu, Dong Li, Hua Gu, Tongjun Key Laboratory of Intelligent Information Processing Institute of Computing Technology Chinese Academy of Sciences Key Laboratory of Computer System and Architecture Institute of Computing Technology Chinese Academy of Sciences National Research Center for Intelligent Computing Systems Institute of Computing Technology Chinese Academy of Sciences Center for Systems Biology Institute of Biophysics Chinese Academy of Sciences Graduate University Chinese Academy of Sciences

ISBN: (纸本)9781424415786

Structure alignment could help to find shape similarities between proteins and guide structure classification and fold recognition. Common substructure detection and extraction are especially important, for which could guide the biologist to discover binding site or active site. We represent each segment of alpha-carbon backbone by using dihedral angles and curve moment invariants. Then, local and global structure alignment could be performed by iterative closest point algorithm. Maximum common substructures between a pair of proteins or within a protein could be found. Active sites also could be detected by the proposed algorithm.

关键词： Proteins

来源：评论

学校读者我要写书评

暂无评论

Software and hardware co-designed multi-level TLBs for chip multiprocessors

Software and hardware co-designed multi-level TLBs for chip ...

引用

11th IEEE International Conference on computer and Information technology, CIT 2011 and 11th IEEE International Conference on Scalable computing and Communications, SCALCOM 2011

作者： Zhang, Xiaohui Cong, Ming Chen, Guangqiang Key Laboratory of Computer System and Architecture Institute of Computing Technology Graduate University Beijing China National University of Defense Technology Changsha China

ISBN: (纸本)9780769543888

Translation Lookaside Buffers ( TLBs ) have a significant impact on system performance. Numerous prior studies focus on TLBs design for uniprocessors. With the advent of chip multiprocessors (CMPs), we need a shift to TLBs for chip multiprocessors. This paper presents a software-implemented level-two TLB-SL2-TLB which is a shared level-two TLB for multiprocessors. It not only reduces the cost of TLB refill handler for every processor core, but also reduces the redundant TLB misses' cost for CMPs effectively. Today, CMPs typically employ private per-core TLBs. SL2-TLB together with the hardware TLBs make up a software-hardware co-designed multilevel TLB system which brings great benefit to system performance while avoiding changing the hardware TLB. So it is a convenient and efficient method for CMPs' TLB performance improvement. The benefit brought by SL2-TLB to SPECCPU2000 is less than that to SPECCPU2006, about 5% and 7% separately. Thereinto, the average performance improvement of SPECint 2006 reaches about 12.7%. That is because the overhead for TLB refill is small when the cache is large enough to avoid a miss as walking the page table of applications with small memory footprints. The further optimization for SL2-TLB is kept the SL2-TLB table stay in L2 cache forever by the cache locking scheme. SL2-TLB together with cache locking scheme improves the performances by over 13% for SPECint 2006. And an average performance improvement of over 7% is brought to the new emerging parallel benchmark suite- Princeton Application Repository for Shared-Memory computers (PARSEC). And all the above evaluations are done on Godson-3 processors which is the latest generation of China's most powerful microprocessor family. © 2011 IEEE.

关键词： Benchmarking

来源：评论

学校读者我要写书评

暂无评论

An Eye-Gaze Tracking Method Based on a 3D Ocular Surface Fitting Model 4th

An Eye-Gaze Tracking Method Based on a 3D Ocular Surface Fit...

引用

4th International Conference on Machine Learning for Cyber Security, ML4CS 2022

作者： Ling, Zhang Zongxin, Ma MingYu, Yan Wenchao, Jiang Muhammad Computer Faculty Guangdong University of Technology Guangzhou China State Key Laboratory of Computer Architecture Institute of Computing Technology Chinese Academy of Sciences Beijing China University of Management and Technology Lahore54770 Pakistan

ISBN: (纸本)9783031200984

Nowadays gaze tracking systems have several common drawbacks: low suitability, high complexity and low accuracy. To solve these problems, this paper presents a novel gaze tracking method, which improves the performance of gaze tracking system in two aspects. A 3D ocular quadratic surface model is constructed with eyeball’s point cloud data, which are collected by a pair of binocular cameras in a free space, allowing head movement and without any wearable devices. Geometrical features of ocular surface, including principal direction, Gaussian curvature and radius are exploited to calculate ocular optical axis direction and person-specific parameters, and then to estimate ocular visual axis direction, i.e. gaze direction without any personal calibration procedures. For this purpose, two person-specific parameters, the radius of cornea curvature and the distance between the center of cornea curvature and the pupil on the ocular model, are inputted into a GRNN network to estimate the value of kappa angle, which is an angle between ocular optical axis and visual axis and has important effect on estimation accuracy of the ocular visual axis direction by the ocular optical axis direction. Compared with the pupil center cornea reflection method, the following experimental results show that the method presented in this paper is effective and accurate in kappa angle estimating and gaze tracking. © 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.

关键词： Eye tracking

来源：评论

学校读者我要写书评

暂无评论

Predoop: Preempting reduce task for job execution accelerations

Predoop: Preempting reduce task for job execution accelerati...

引用

4th and 5th Workshops on Big Data Benchmarks, Performance Optimization, and Emerging Hardware, BPOE 2014

作者： Liang, Yi Wang, Yufeng Fan, Minglu Zhang, Chen Zhu, Yuqing Department of Computer Science Beijing University of Technology Beijing China State Key Laboratory of Computer Architecture Institute of Computing Technology Chinese Academy Sciences Beijing China

ISBN: (纸本)9783319130200

Map/Reduce is a popular parallel processing framework for data intensive computing. For overlapping the Map task’s execution phase and the Reduce task’s intermediate data fetching and merging phase, existing Map/ Reduce schedulers always pre-launch the Reduce task at the specific threshold where its map tasks have been launched, and this pattern incurs the occupation of the consuming resources of the reduce task during its idle time on waiting for fetching the intermediate data from map tasks. To address this issue, we propose an extension version of Hadoop map/reduce framework, called Predoop, in this paper. The basic idea of Predoop is to preempt the reduce task during its idle time and allocate the released resource to the map tasks on schedule. To achieve this goal, first, we introduce the preemptive mechanism for reduce tasks and map tasks respectively to enable Map/Reduce tasks to be preempted or resumed with correct status;second, we adopt the preempting-resuming model for the reduce task with the consideration of the progress of Reduce task data fetching & merging and the Map task execution so as to determine the timing of Reduce task preemption and resuming;third, we introduce the preemption-aware task scheduling strategy to allocate the released resources to the on-schedule Map tasks with the consideration of data locality. Experimental result demonstrates that Predoop outperforms Hadoop on various workload and the average job turnaround time can be reduced by maximum of 66.57 %. © Springer International Publishing Switzerland 2014.

关键词： Multitasking

来源：评论

学校读者我要写书评

暂无评论

Near 100%spectral-purity photons from reconfigurable micro-rings

引用

Chinese Physics B 2020年第11期29卷 135-139页

作者： Pingyu Zhu Yingwen Liu Chao Wu Shichuan Xue Xinyao Yu Qilin Zheng Yang Wang Xiaogang Qiang Junjie Wu Ping Xu Institute for Quantum Information&State Key Laboratory of High Performance Computing College of ComputerNational University of Defense TechnologyChangsha 410073China National Innovation Institute of Defense Technology AMSBeijing 100071China National Laboratory of Solid State Microstructures and School of Physics Nanjing UniversityNanjing 210093China

We propose an on-chip reconfigurable micro-ring to engineer the spectral-purity of *** micro-ring resonator is designed to be coupled by one or two asymmetric Mach-Zehnder interferometers and the coupling coefficients hence the quality-factors of the pump and the converted photons can be dynamically changed by the interferometer’s internal *** calculate the joint-spectrum function and obtain the spectral-purity of photons and Schmidt number under different *** show that it is a dynamical method to adjust the spectral-purity and can optimize the spectralpurity of photons up to near 100%.The condition for high-spectral-purity photons is ensured by the micro-ring itself,so it overcomes the trade-off between spectral purity and brightness in the traditional post-filtering *** scheme is robust to fabrication variations and can be successfully applied in different fabrication labs and different *** high-spectral-purity photons will be beneficial for quantum information processing like Boson sampling and other quantum algorithms.

关键词： spectral-purity micro-ring Mach–Zehnder interferometer

来源：评论

学校读者我要写书评

暂无评论

Highly Optimized Code Generation for Stencil Codes with Computation Reuse for GPUs

引用

Journal of computer Science & technology 2016年第6期31卷 1262-1274页

作者： Wen-Jing Ma Kan Gao Guo-Ping Long Laboratory of Parallel Software and Computing Science Institute of Software Chinese Academy of Sciences Beijing 100190 China State Key Laboratory of Computer Science Institute of Software Chinese Academy of Sciences Beijing 100190 China Information Center China Association for Science and Technology Beijing 100863 China

Computation reuse is known as an effective optimization technique. However, due to the complexity of modern GPU architectures, there is yet not enough understanding regarding the intriguing implications of the interplay of compu- ration reuse and hardware specifics on application performance. In this paper, we propose an automatic code generator for a class of stencil codes with inherent computation reuse on CPUs. For such applications, the proper reuse of intermediate results, combined with careful register and on-chip local memory usage, has profound implications on performance. Current state of the art does not address this problem in depth, partially due to the lack of a good program representation that can expose all potential computation reuse. In this paper, we leverage the computation overlap graph （COG）, a simple representation of data dependence and data reuse with ＂element view＂, to expose potential reuse opportunities. Using COG, we propose a portable code generation and tuning framework for GPUs. Compared with current state-of-the-art code generators, our experimental results show up to 56.7% performance improvement on modern GPUs such as NVIDIA C2050.

关键词： GPGPU OpenCL stencil code generation computation reuse

来源：评论

学校读者我要写书评

暂无评论

High Performance computing for CS-InSAR Data Processing

High Performance Computing for CS-InSAR Data Processing

引用

2021 SAR in Big Data Era, BIGSARDATA 2021

作者： Zhang, Weikang You, Haihang Tang, Yixian Wang, Chao Zhang, Hong Chinese Academy of Sciences State Key Laboratory of Computer Architecture Institute of Computing Technology Beijing China Chinese Academy of Sciences Key Laboratory of Digital Earth Science Aerospace Information Research Institute Beijing China

ISBN: (纸本)9781665401234

In this paper, we have developed a highperformance data processing system for big SAR data based on parallel and distributed computation using supercomputer. Aiming at the pipeline of InSAR, we mainly focus on the parallelization and optimization of computing hotspot to accelerate the operation. Furthermore, we take advantage of the technology of distributed computation and multitask automation to get higher performance - process more data in less time. With the high-performance system, we obtained the nationwide measurement of China's surface movement from September 2018 to February 2020 and from January to December in 2020 using 20-30 computing nodes, processing nearly 1PB of data in total. The calculation time of the two images of national scale surface movement is completed in only 7 and 5 days, respectively. © 2021 IEEE.

关键词： Supercomputers

来源：评论

学校读者我要写书评

暂无评论

NDANN: Efficient SSD-Based Approximate Nearest Neighbor Search through Navigation

NDANN: Efficient SSD-Based Approximate Nearest Neighbor Sear...

引用

2022 International Conference on Mechanisms and Robotics, ICMAR 2022

作者： Yang, Kaixiang Wang, Hongya Tan, Zongyuan Zhang, Jie College of Computer Science and Technology Donghua University Shanghai China State Key Laboratory of Computer Architecture Institute of Computing Technology Chinese Academy of Sciences China Institute of Artificial Intelligence Donghua University Shanghai China

ISBN: (纸本)9781510657328

Graph-based Approximate Nearest Neighbor Search (ANNS) algorithms have attracted more attention due to their better performance. Numerous ANNS optimization methods have been proposed by researchers. However, Current graph-based ANNS algorithms still cannot index billion-scale datasets on a single server with 256GB RAM. Although several researchers have investigated on this problem, we believe there is still an improvement in reducing disk accesses. In this paper, we provide a fast navigation layer to decrease the number of disk accesses by assisting query points in swiftly reaching the range of strongly connected components. Compared with the state-of-the-art ANNS algorithms, NDANN reduces the mean latency by about 20% under the same recall. © 2022 SPIE.

关键词： Graphic methods

来源：评论

学校读者我要写书评

暂无评论

建议与咨询留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

时间限定

文献类型

馆藏选择

核心期刊

语言

文献类型

帮助

文字说明：

检索规则说明：

检索范例：

分类表

所选分类

限定检索结果

文献类型

馆藏范围

日期分布

学科分类号

主题

机构

作者

语言

请选择保存的检索档案：

请选择收藏分类：

通借通还

建议与咨询 留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

时间限定

文献类型

馆藏选择

核心期刊

语言

文献类型

帮助

文字说明：

检索规则说明：

检索范例：

分类表

所选分类

限定检索结果

文献类型

馆藏范围

日期分布

学科分类号

主题

机构

作者

语言

请选择保存的检索档案： 新增检索档案 确定 取消

请选择收藏分类： 新增自定义分类 确定 取消

通借通还

建议与咨询留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

请选择保存的检索档案：

请选择收藏分类：