Reproduce “Adapter-X: A Novel General Parameter-Efficient Fine-Tuning Framework for Vision” 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

Minglei Li, Peng Ye, Yongqi Huang, Lin Zhang, Tao Chen, Tong He, Jiayuan Fan, Wanli Ouyang

Adapter-X adapts pretrained vision models to classify image and 3D point-cloud datasets by freezing the backbone and routing tokens through shared adapters with block-specific prompts. On VTAB-1K, AdaptFormer-X reaches 76.2 accuracy with 0.17M parameters versus 68.9 with 85.84M for full fine-tuning. On ScanObjectNN PB_T50_RS, DAPT-X reaches 88.45% with 0.42M versus 88.31% with 22.1M for reproduced Point-MAE fine-tuning, limited to tests.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.