arrow
Return

Simultaneously Training and Compressing Vision-and-Language Pre-Training Model

delete2023-01-01
delete0
PRE
AI
Q
Qiaosong Qi
A
Aixi Zhang
Y
Yue Liao *
W
Wenyu Sun
王永亮 cover
王永亮 (Yongliang Wang)
X
Xiaobo Li
S
Si Liu
DOI:10.1109/TMM.2022.3233258delete
deleteOriginal
deleteOriginal request for help
deleteShare
deleteSave
Abstract

Abstract

En 中文
Model compression is an essential step for large-scale pre-training models toward practical application and deployment on the edge device. However, when conventional compression methods following 'pre-training then compressing' two-phase pipeline are applied to Vision-and-Language Pre-training (VLP) models, it will lead to a high calculation and memory overhead. In this work, we break the two-phase pipeline and propose an efficient and effective one-phase VLP model compression mechanism, named REDUCER, which stands for 'simultaneously training and compREssing' VLP model via progressive moDUle replaCing and nEtwork Rewiring. Specifically, REDUCER consists of three insightful designs. Firstly, we design a one-phase compression framework to train and compress the VLP model simultaneously to avoid the extra calculation and memory cost caused by an isolated model compression phase in the conventional two-phase pipeline. Secondly, we propose an adaptive progressive module replacing mechanism to compress the model depth free from explicit knowledge distillation losses, relieving the multi-task optimization problems. Thirdly, we integrate pruning techniques into VLP model compression to simultaneously compress the model in width and depth. Overall, we obtain a lightweight VLP model with only one pre-training phase, and it is the first one-phase compression method for VLP models. Extensive experiments have been conducted on representative VLP models, i.e., ClipBERT and VICTOR, and the experimental results show a superior trade-off between performance and efficiency.
Keywords:
Transmission line measurements
Adaptation models
Computational modeling
Training
Task analysis
Pipelines
Costs
Vision-and-Language
model compression
pre-training Model

Journal

IEEE Transactions on Multimedia cover
IEEE Transactions on Multimedia
IF:
9.7
Papers:
4.6K
Citations:
2.4W

Organization

A
alibaba group
Scholars:
1.1K
Papers: 789
Citations: 0
B
Beihang University
Scholars:
5.2W
Papers: 4.1W
Citations: 37
Cited Papers

Cited Papers

errShare
errSave
errShare
errSave
Frame Augmented Alternating Attention Network for Video Question Answering
err2020-04-01
err39
PREAI
errZhang, Wenqiao; Tang, Siliang; Cao, Yanpeng; Pu, Shiliang; Wu, Fei; Zhuang, Yueting
errShare
errSave
errShare
errSave
Visual Genome: Connecting Language and Vision Using Crowdsourced Dense Image Annotations
err2017-02-06
err3.1K
errOAAI
errKrishna, Ranjay; Zhu, Yuke; Groth, Oliver; Johnson, Justin; Hata, Kenji; Kravitz, Joshua; Chen, Stephanie; Kalantidis, Yannis; Li, Li-Jia; Shamma, David A.; Bernstein, Michael S.; Li Fei-Fei
errShare
errSave
Priming Effects of Vegetable Juice on Endogenous Production of Tumor Necrosis Factor
err2014-06-12
err0
PREAI
errMasatoshi Yamazaki; Hiroshi Ueda; Koutaro Fukuda; Miki Okamoto; Satoru Yui
errShare
errSave
Mechanisms of corticosteroid action in immune thrombocytopenic purpura (ITP): experimental studies using ITP-prone mice, (NZW x BXSB) F1
err1992-02-15
err0
errOAAI
errH Mizutani; T Furubayashi; Y Imai; H Kashiwagi; S Honda; H Take; Y Kurata; T Yonezawa; S Tarui; S Ikehara
errShare
errSave
researcher View more