Initial Public Release

Release v0.1.0 — Initial Public Release

Highlights

• First open release of Dual-Region Foreground–Background Augmentation (DRA)
• Two reference implementations:
  • SFDA/: Source-Free Domain Adaptation on PACS (Hydra-driven)
  • Person_ReID/: Person re-identification baselines with DRA
• Augmentation utilities and visualization helpers
• Configurable training via CLI overrides

Repository Layout

• SFDA/ — core SFDA pipeline (main_win.py, configs/, datasets/PACS/* lists)
• Person_ReID/ — training, testing, and augmentation for ReID
• assets/ — figures for README
• SFDA/scripts_win/ — example commands (Windows)

Install

• Python 3.8+
• SFDA: pip install -r SFDA/requirements.txt (install PyTorch/torchvision per CUDA from pytorch.org)
• ReID: pip install -r Person_ReID/requirements.txt

Quickstart

• SFDA smoke test (no W&B, small batch, 1 epoch):
  • python SFDA/main_win.py learn.epochs=1 data.batch_size=16 use_wandb=false memo=release-smoke
  • Datasets: place PACS under datasets/PACS/ as described in SFDA/README.md
• ReID baseline (Market-1501 example):
  • Train: python Person_ReID/train.py --gpu_ids 0 --use_rn18 --batchsize 32 --data_dir Person_ReID/data/Market/pytorch
  • Test: python Person_ReID/test.py --gpu_ids 0 --use_rn18 --test_dir Person_ReID/data/Market/pytorch

Notes & Known Issues

• Windows scripts in SFDA/scripts_win/, run them by pasting commands into your terminal.
• Multi-GPU/distributed runs require CUDA devices; for single-GPU/CPU users, keep defaults conservative and reduce workers/batch size if needed.
• Mask generation (U^2-Net for SFDA; MediaPipe + SAM2 for ReID) is optional preprocessing; see subproject READMEs for instructions.

Licensing & Citation

• License: MIT (see LICENSE)
• Please cite the paper linked in the top-level README.md if this work helps your research.

Support

• Issues and questions: open a GitHub issue with environment, command, and logs. Include dataset layout if relevant.