Skip to content

Manifold Regularized Dynamic Network Pruning #29

Description

@ryoherisson

一言でいうと

インスタンスの複雑度と類似度を考慮したDynamic Pruning手法を提案.

論文リンク

著者/所属機関

Yehui Tang, Yunhe Wang, Yixing Xu, Yiping Deng, Chao Xu, Dacheng Tao, Chang Xu
(Peking University, Huawei Technologies, University of Sydney)

投稿日付(yyyy/MM/dd)

2021/03/10

概要

フィルタレベルのプルニーニングにおいて,入力データの複雑度と類似度を考慮してフィルタの重要性を判断し,インスタンスごとに適切なサブネットワークを割り当てる手法.
ImageNet,ResNet-34で,精度劣化を0.57%に抑えながらFLOPSを55.3%削減.

新規性・差分

  • 入力データの複雑度と類似度を考慮してフィルタの重要性を判断し、インスタンスごとに適切なサブネットワークを割り当てる手法を提案.

手法

スクリーンショット 2021-06-03 23 09 36
スクリーンショット 2021-06-04 0 40 22

学習時の目的関数
スクリーンショット 2021-06-04 0 45 02

結果

スクリーンショット 2021-06-04 0 41 25

コメント

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions