Open tkuri opened 3 years ago
DNNの新しいモデル圧縮手法。フィルタの冗長性を排除するために、個々のフィルタ重みの重要度を見るのではなく、重みの相関が大きいペアを反復的に同定して各ペアから1つずつ削除していく。削除の前に各ペアの相関を高めるような最適化をし情報の損失を抑えているのがキモ。
https://openaccess.thecvf.com/content_WACV_2020/html/Singh_Leveraging_Filter_Correlations_for_Deep_Model_Compression_WACV_2020_paper.html
未確認。
論文概要
DNNの新しいモデル圧縮手法。フィルタの冗長性を排除するために、個々のフィルタ重みの重要度を見るのではなく、重みの相関が大きいペアを反復的に同定して各ペアから1つずつ削除していく。削除の前に各ペアの相関を高めるような最適化をし情報の損失を抑えているのがキモ。
https://openaccess.thecvf.com/content_WACV_2020/html/Singh_Leveraging_Filter_Correlations_for_Deep_Model_Compression_WACV_2020_paper.html
Code
未確認。