• 제목/요약/키워드: tensor computation

검색결과 25건 처리시간 0.026초

NVIDIA Jetson TX1 기반의 사람 표정 판별을 위한 YOLO 모델 FPS 향상 방법 (YOLO Model FPS Enhancement Method for Determining Human Facial Expression based on NVIDIA Jetson TX1)

  • 배승주;최현준;정구민
    • 한국정보전자통신기술학회논문지
    • /
    • 제12권5호
    • /
    • pp.467-474
    • /
    • 2019
  • 본 이 논문에서는 NVIDIA Jetson TX1에서 YOLO v2 모델의 정확도를 유지하면서 FPS를 개선하는 방법을 제안한다. 일반적으로, 딥러닝 모델에서는 연산량을 줄여 처리 속도를 높이기 위해 파라미터들을 실수형에서 정수형으로 변환하여 정수 연산을 통해 속도를 높이거나 네트워크의 깊이를 감소시키는 방법을 사용한다. 그러나 이 방법들은 인식 정확도가 떨어질 수 있다. 이 논문에서는 YOLO v2 모델을 이용해 표정인식기를 개발하고 정확도 유지 시키기 위해 정수 연산이나 네트워크 깊이 감소를 사용하는 대신, 다음 세 가지 방법을 통해 연산량 및 메모리 소모를 줄인다. 첫 번째, $3{\times}3$ 필터를 $1{\times}1$ 필터로 교체하여 각 Layer 당 매개 변수 수를 9 분의 1로 줄인다. 두 번째, TensorRT의 추론 가속 기능 중 CBR (Convolution-Add Bias-Relu)을 통해 연산량을 줄이고, 마지막으로 TensorRT를 사용하여 반복되는 동일한 연산구조를 가진 레이어를 통합하여 메모리 소비를 줄인다. 시뮬레이션 결과, 기존 YOLO v2 모델에 비해 정확도는 1 % 감소했지만 FPS는 기존 3.9 FPS에서 11 FPS로 282%의 속도 향상을 보였다.

적분방정식을 사용한 3차원 MT 모델링에서의 텐서 그린 적분의 계산 (Computation of Green's Tensor Integrals in Three-Dimensional Magnetotelluric Modeling Using Integral Equations)

  • 김희준;이동성
    • 자원환경지질
    • /
    • 제27권1호
    • /
    • pp.41-47
    • /
    • 1994
  • 적분방정식을 사용한 3차원 전자기 모델링에 나오는 많은 텐서 그린 적분의 수치계산에 신속 한겔변환 (FHT) 아르고리즘 (Anderson, 1982)을 적용하였다. 한겔변환은 FHT에서 사용가능한 연관 및 지연 중합으로 효율적으로 계산할 수 있다. 먼저 수평 층서모형에 대한 텐서 그린 적분을 보여주고 난 다음 이들을 FHT로 신속하게 계산할 수 있도록 서로 연관된 형태의 함수로 고쳐쓴다. FHT로 연관된 한겔변환의 전행열이 단일 직접 중합과 거의 비슷한 계산시간으로 신속 정확하게 구해진다. 5층 수평 층서모형에 대한 컴퓨터실험의 결과, FHT는 직접 및 지연 중합법에 비하여 각각 117 및 4배 빠르다.

  • PDF

XEM: Tensor accelerator for AB21 supercomputing artificial intelligence processor

  • Won Jeon;Mi Young Lee;Joo Hyun Lee;Chun-Gi Lyuh
    • ETRI Journal
    • /
    • 제46권5호
    • /
    • pp.839-850
    • /
    • 2024
  • As computing systems become increasingly larger, high-performance computing (HPC) is gaining importance. In particular, as hyperscale artificial intelligence (AI) applications, such as large language models emerge, HPC has become important even in the field of AI. Important operations in hyperscale AI and HPC are mainly linear algebraic operations based on tensors. An AB21 supercomputing AI processor has been proposed to accelerate such applications. This study proposes a XEM accelerator to accelerate linear algebraic operations in an AB21 processor effectively. The XEM accelerator has outer product-based parallel floating-point units that can efficiently process tensor operations. We provide hardware details of the XEM architecture and introduce new instructions for controlling the XEM accelerator. Additionally, hardware characteristic analyses based on chip fabrication and simulator-based functional verification are conducted. In the future, the performance and functionalities of the XEM accelerator will be verified using an AB21 processor.

브레이드 프리폼의 투과율 계수 예측 (Prediction of Permeability for Braided Preform)

  • Youngseok Song;Youn, Jae-Roun
    • 한국복합재료학회:학술대회논문집
    • /
    • 한국복합재료학회 2003년도 춘계학술발표대회 논문집
    • /
    • pp.184-187
    • /
    • 2003
  • Complete prediction of second order permeability tensor for three dimensional circular braided preform is critical to understand the resin transfer molding process of composites. The permeability can be predicted by considering resin flow through the multi-axial fiber structure. In this study, permeability tensor for a 3-D circular braided preform is calculated by solving a boundary problem of a periodic unit cell. Flow field through the unit cell is obtained by using a 3-D finite volume method (FVM) and Darcy's law is utilized to obtain permeability tensor. Flow analysis for two cases that a fiber tow is regarded as impermeable solid and permeable porous medium is carried out respectively. It is found that the flow within the intra-tow region of the braided preform is negligible if inter-tow porosity is relatively high but the flow through the tow must be considered when the porosity is low. To avoid checkerboard pressure field and improve the efficiency of numerical computation, a new interpolation function for velocity variation is proposed on the basis of analytic solutions. Permeability of the braided preform is measured through a radial flow experiment and compared with the permeability predicted numerically.

  • PDF

Analysis of Contact Force in Eddy-current System Using the Virtual Air-Gap Concept

  • Park, Byung Su;Kim, Hwi Dae;Choi, Hong Soon;Park, Il Han
    • Journal of Electrical Engineering and Technology
    • /
    • 제10권3호
    • /
    • pp.1349-1355
    • /
    • 2015
  • It is difficult to calculate the magnetic force of an object of magnetic material in contact with other objects using the existing methods, such as Maxwell stress tensor method, magnetic charge method, or magnetizing current method. These methods are applicable for force computation only when the object is surrounded by air. The virtual air-gap concept has been proposed for calculating the contact force. However, its application is limited to magneto-static system. In this paper, we present the virtual air-gap concept for contact surface force in the eddy-current system. Its validity and usefulness are shown by comparison between numerical and experimental examples.

정자력 계산을 위한 적응 유한 요소법 (An Adaptive Finite Element Method for Magnetostatic Force Computations)

  • 박용규;박일한;정형석;정현교;이기식;한송엽
    • 대한전기학회논문지
    • /
    • 제38권2호
    • /
    • pp.100-105
    • /
    • 1989
  • This paper presents an adaptive finite element method for magnetostatic force computation using Maxwell's stress tensor. Mesh refinements are performed automatically by interelement magnetic field intensity discontinuity errors and element force errors. In initial mesh, the computed forces for different integration paths give great differences, but converge to a certain value as mesh division is performed by the adaptive scheme, We obtained good agreement between analytic solutions and numerical values in typical examples.

  • PDF

적응격자계를 이용한 경계층의 확산제어천이 예측 (Prediction of the Diffusion Controlled Boundary Layer Transition with an Adaptive Grid)

  • 조지룡
    • 한국전산유체공학회지
    • /
    • 제6권4호
    • /
    • pp.15-25
    • /
    • 2001
  • Numerical prediction of the diffusion controlled transition in a turbine gas pass is important because it can change the local heat transfer rate over a turbine blade as much as three times. In this study, the gas flow over turbine blade is simplified to the flat plate boundary layer, and an adaptive grid scheme redistributing grid points within the computation domain is proposed with a great emphasis on the construction of the grid control function. The function is sensitized to the second invariant of the mean strain tensor, its spatial gradient, and the interaction of pressure gradient and flow deformation. The transition process is assumed to be described with a κ-ε turbulence model. An elliptic solver is employed to integrate governing equations. Numerical results show that the proposed adaptive grid scheme is very effective in obtaining grid independent numerical solution with a very low grid number. It is expected that present scheme is helpful in predicting actual flow within a turbine to improve computation efficiency.

  • PDF

Mathematical Modeling and Control for A Single Winding Bearingless Flywheel Motor in Electric/Suspension Mode

  • Yuan, Ye;Huang, Yonghong;Xiang, Qianwen;Sun, Yukun
    • Journal of Electrical Engineering and Technology
    • /
    • 제13권5호
    • /
    • pp.1935-1944
    • /
    • 2018
  • With the increase of the production of energy from renewable, it becomes important to look at techniques to store this energy. Therefore, a single winding bearingless flywheel motor (SWBFM) specially for flywheel energy storage system is introduced. For the control system of SWBFM, coupling between the torque and the suspension subsystems exists inevitably. It is necessary to build a reasonable radial force mathematical model to precisely control SWBFM. However, SWBFM has twelve independently controlled windings which leads to high-order matrix transformation and complex differential calculation in the process of mathematical modeling based on virtual displacement method. In this frame, a Maxwell tensor modeling method which is no need the detailed derivation and complex theoretical computation is present. Moreover, it possesses advantages of universality, accuracy, and directness. The fringing magnetic path is improved from straight and circular lines to elliptical line and the rationality of elliptical line is verified by virtual displacement theory according to electromagnetic torque characteristics. A correction function is taken to increase the model accuracy based on finite element analysis. Simulation and experimental results show that the control system of SWBFM with radial force mathematical model based on Maxwell tensor method is feasible and has high precision.

Fundamental theory of curved structures from a non-tensorial point of view

  • Paavola, Juha;Salonen, Eero-Matti
    • Structural Engineering and Mechanics
    • /
    • 제7권2호
    • /
    • pp.159-180
    • /
    • 1999
  • The present paper shows a new non-tensorial approach to derive basic equations for various structural analyses. It can be used directly in numerical computation procedures. The aim of the paper is, however, to show that the approach serves as an excellent tool for analytical purposes also, working as a link between analytical and numerical techniques. The paper gives a method to derive, at first, expressions for strains in general beam and shell analyses, and secondly, the governing equilibrium equations. The approach is based on the utilization of local fixed Cartesian coordinate systems. Applying these, all the definitions required are the simple basic ones, well-known from the analyses in common global coordinates. In addition, the familiar principle of virtual work has been adopted. The method will be, apparently, most powerful in teaching the theories of curved beam and shell structures for students not familiar with tensor analysis. The final results obtained have no novelty value in themselves, but the procedure developed opens through its systematic and graphic progress a new standpoint to theoretical considerations.

비압축성 유동계산을 위한 계층 요소 사용에 대한 연구 (A Study on the Use of Hierarchical Elements for Incompressible Flow Computations)

  • 김진환
    • 대한기계학회:학술대회논문집
    • /
    • 대한기계학회 2001년도 춘계학술대회논문집E
    • /
    • pp.422-429
    • /
    • 2001
  • A two dimensional hierarchical elements are investigated for a use on the incompressible flow computation. The construction of hierarchical elements are explained through the tensor product of 1-D hierarchical functions, and a systematic treatment of essential boundary values has been developed for the degrees of freedom corresponding to higher order terms. The numerical study for the poisson problem showed that the present scheme can increase the convergence and accuracy of finite element solutions, and can be more efficient than the standard first order with many elements. Also, for Stokes and cavity flow cases, solutions from hierarchical elements showed better resolutions and future promises for higher order solutions.

  • PDF