Reproduce “DLIP: Distilling Language-Image Pre-training” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Huafeng Kuang, Jie Wu, Xiawu Zheng, Ming Li, Xuefeng Xiao, Rui Wang, Min Zheng, Rongrong Ji
A new framework, DLIP, systematically distills Vision-Language Pre-training (VLP) models, achieving up to 77.6% parameter reduction and 2.7x inference acceleration while retaining over 95% of the teacher model's performance on various multimodal tasks. The approach provides practical guidelines for VLP distillation and outperforms existing methods for compressing these models.
Key insights
- Both image and text encoders are equally important for VLP performance under compression, while multimodal fusion modules achieve optimal performance with a moderate number of layers.
- Representation-based knowledge transfer is more efficient than attention-based, and transferring multimodal fusion information proves more effective than unimodal information for VLP distillation.
- Dedicated distillation of the decoder module is essential for improving performance on generative VLP tasks, and a pre-trained visual encoder is critical for overall VLP model performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.