Skip to main content
Video s3
    Details
    Presenter(s)
    Hana Kim Headshot
    Display Name
    Hana Kim
    Affiliation
    Affiliation
    Ewha Womans University
    Country
    Author(s)
    Display Name
    Hana Kim
    Affiliation
    Affiliation
    Ewha Womans University
    Display Name
    Hyun Eun
    Affiliation
    Affiliation
    OPENEDGES Technology, Inc.
    Display Name
    Jung Hwan Choi
    Affiliation
    Affiliation
    OPENEDGES Technology, Inc.
    Display Name
    Ji-Hoon Kim
    Affiliation
    Affiliation
    Ewha Womans University
    Abstract

    In this paper, we present two simple methods for fast analysis in mixed-precision determination and for computational complexity reduction in inference, respectively. With the proposed SQNR-based analysis, we can significantly reduce the time required in mixed-precision scheme determination compared to conventional mAP-based scheme with negligible loss of accuracy. Also, by combining the hyperparameters of the target neural networks in the process of mixed-precision determination, we can reduce the computational complexity in inference. We applied these two proposed methods to the SSDlite network with MobileNet-v2 and YOLOv2 network and evaluate the quantized networks with the proposed mixed-precision schemes on the Pascal-VOC dataset.

    Slides
    • SQNR-Based Layer-Wise Mixed-Precision Schemes with Computational Complexity Consideration (application/pdf)