Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

3 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Category-Specific Channel Pruning for Effective Modality Bridging in Text-only Tuning

Introduction

This repo officially implements Category-Specific Channel Pruning for Effective Modality Bridging in Text-only Tuning.

Modality Bridging Pruning (MBP) selectively activates task-relevant channels, leading to performance improvements across single-label and multi-label tasks.

Contact us with pubowei@mail.ustc.edu.cn

Fig.1 Our two-step strategy: Mean-base initialization avoids overfitting, and pruning channels aims to bridge the modality gap, which leads to fit vision embeddings better in TOT. .

Performance Comparison

DTD Flowers IN-R Aircraft UCF101 Food101 EuroSat SUN397 Mean
ZSCLIP 40.3 64.0 65.8 18.1 61.0 77.1 35.9 60.8 52.8
cls-only 37.4 60.4 63.1 15.4 56.4 74.3 39.1 56.3 50.3
CLIP (DST) 42.4 66.6 68.6 19.5 62.4 79.3 45.8 61.7 55.8
DST 41.6 64.5 67.5 17.2 62.0 77.2 43.6 61.0 54.3
LaFTer 44.3 64.2 63.6 18.3 63.7 78.1 44.2 60.7 54.6
TAP 51.1 66.9 68.9 21.6 66.3 78.5 54.2 63.4 58.9
Our-TF 51.7 69.7 69.8 22.3 64.6 80.5 53.3 62.2 59.3
Our 52.3 72.6 70.0 22.7 65.5 80.6 55.0 63.0 60.2

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

1 watching

Forks

Releases

Packages

Contributors

Languages