http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
- 中文 을 입력하시려면 zhongwen을 입력하시고 space를누르시면됩니다.
- 北京 을 입력하시려면 beijing을 입력하시고 space를 누르시면 됩니다.
RISS 인기검색어
- 검색키워드
- 저자 : Sunghun Hwangbo (검색결과 1 건)
- 무료
- 기관 내 무료
- 유료
-
Compression of deep learning models through global weight pruning using ADMM
Sunghun Hwangbo,Dongwook Yang,Geonseok Lee,Kichun Lee 한국산업경영시스템학회 2021 한국산업경영시스템학회 학술대회 Vol.2021 No.춘계
Deep learning, which has recently shown excellent performance, has a problem that the amount of computation and required memory are large. Model compression is very useful because it saves memory and reduces storage size while maintaining model performance. Model compression methods reduce the number of edges by pruning weights that are deemed unnecessary in the calculation. Existing weight pruning methods using ADMM construct an optimization problem by a layer-by-layer addition of pre-defined removal-ratio constraints. Decomposing into two subproblems through the ADMM process, one can solve them through gradient descent and projection. However, the layer-by-layer removal ratios must be structurally specified, causing a sharp increase in training time due to a large number of parameters, and hardly feasible to use for large models that actually require weight pruning. Our proposed method performs weight pruning, producing similar performance, by setting a global removal ratio for the entire model without prior knowledge of structural characteristics in order to solve the shortcomings of the existing ADMM weight-pruning methods. To effectively avoid performance degradation, the method removes a relatively small number of previous layers in charge of feature extraction. Experiments show high-quality performance, not necessarily setting layer-by-layer removal ratios. Additionally, experiments increasing layers yield an insight for feature extraction in pruned layers. The experiment of the proposed method to the LeNet-5 model using MNIST data results in a higher compression ratio of 99.3% outperforming those of other existing algorithms. We also demonstrate the effectiveness of the proposed method in YOLOv4, an object detection model requiring substantial computation.
연관 검색어 추천
이 검색어로 많이 본 자료
활용도 높은 자료
