請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/59117
標題: | 基於通道修剪的一種高效能深度神經網路模型壓縮 An Efficient and Effective Channel Pruning Based Model Compression Methodology For DNN |
作者: | Min-Jhih Huang 黃敏智 |
指導教授: | 吳家麟(Ja-Ling Wu) |
關鍵字: | 深度學習,卷積神經網路,模型壓縮,物件偵測,分類任務,通道剪枝,模型利用率, Deep Learning,Convolution Neural Network,Model Compression,Object Detection,Classification,Channel Pruning,Model Utilization, |
出版年 : | 2020 |
學位: | 碩士 |
摘要: | 深度學習在電腦視覺領域上獲得了重大突破,許多電腦視覺的任務交由深度學習的方法都獲得了前所未有的成果,但深度學習模型往往令人詬病的就是需要的運算資源龐大,而這往往代表著很高的成本,雖然在學術上的任務有重大突破,但也導致難以普 及到各種商業情境。 基於這樣的動機,而開始鑽研模型壓縮這塊領域,為了能夠讓 深度學習模型在商用情境中能夠被廣泛使用。 本文基於通道剪枝法去對深度卷積神 經網路(CNN)去進行壓縮,讓模型能夠在算力資源低的設備上運行又能符合情境上的要求,在本文中也有一個章節去做個案探討並付諸實現。 而本文在方法論也提出了一個全新的壓縮模型概念跟方法,其方法都優於同樣基於通道剪枝的其他方法並做 實驗示之。本論文充分的展現模型壓縮對於深度學習網路在未來普及上扮演了重大的角色,讓它不僅僅只是為了論文上那些漂亮的實驗數據,而是為了讓它能夠真正的用在生活之中。 Deep learning has made a breakthrough in the field of computer vision. Many computer vision tasks handed over to deep learning methods have achieved unprecedented results, but the deep learning model is often criticized for the huge computing resources required, and This often represents a high cost. Although it's impressive in academics, it also makes it difficult to spread to various business scenarios. Based on this motivation, I started to study the field of deep learning model compression to facilitate deep learning models to be widely used in commercial scenarios. This article is based on the channel pruning method to compress the deep convolutional neural network (CNN) so that the model can be affordable on devices with low computation resources and satisfy the requirements of the situation. There is also a chapter in this article to do case study. In this paper, the methodology also proposes a new concept and novel method of model compression, which are superior to other methods based on channel pruning. This paper fully demonstrates that model compression plays a major role in the future popularization of deep learning networks, making it not only for the artistic experimental data in the paper but also for it to be truly used in life. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/59117 |
DOI: | 10.6342/NTU202003445 |
全文授權: | 有償授權 |
顯示於系所單位: | 資訊工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
U0001-1408202016123500.pdf 目前未授權公開取用 | 6.63 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。