Feature-space domain adaptation

FPS-UDA trains decision planes over frozen feature banks.

Build or download dataset-level H5 feature banks, select explicit source/target views, and train Feature-Space Planes Searcher (FPS) from NumPy, Torch, or H5 features.

Run the example Use feature banks Paper

Documentation

Choose the path you need

Quick Start

Install FPS-UDA, run the packaged Office31 smoke test, and move from editable install to the fps-uda CLI.

Feature Banks

Download released H5 banks, inspect view roles, analyze views, and extract new banks from images.

Training

Configure io, views, optimization, schedules, losses, train/sweep commands, and Python API usage.

Reference

Look up YAML fields, CLI commands, dataset helpers, benchmark utilities, and repository layout.

Story

Read the origin narrative and inspect the historical prototype H5 and original notebooks.

Advanced

Add custom losses, use custom backbones, use HuggingFace AutoModel vision encoders, and understand public release choices.

Paper

Feature-Space Planes Searcher

This repository accompanies Feature-Space Planes Searcher: A Universal Domain Adaptation Framework for Interpretability and Computational Efficiency.

Zhitong Cheng+, Yiran Jiang+, Yulong Ge, Yufeng Li, Zhongheng Qin, Rongzhi Lin, and Jianwei Ma*.

+ Equal contribution. * Corresponding author.

Citation

Cite the IEEE Xplore paper

@article{cheng_fps_uda,
  title  = {Feature-Space Planes Searcher: A Universal Domain Adaptation Framework for Interpretability and Computational Efficiency},
  author = {Cheng, Zhitong and Jiang, Yiran and Ge, Yulong and Li, Yufeng and Qin, Zhongheng and Lin, Rongzhi and Ma, Jianwei},
  journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
  url    = {https://ieeexplore.ieee.org/abstract/document/11568428},
  note   = {IEEE Xplore document 11568428}
}