FedReDA: Federated Reliability-aware Dual Adapters for Noisy-Label Learning on Vision Foundation Models

This repository contains the reference implementation of FedReDA,
a federated noisy-label learning method built on a frozen DINOv2 backbone with
dual Reins adapters (student/teacher) and noise-aware distillation.

Features

• Frozen DINOv2 ViT-S/14 backbone (_small_variant)
• Two Reins adapters:
  • reins : per-client student adapter (Adapter1)
  • reins2 : global / LOO teacher adapter (Adapter2)
• LOO teacher or shared FedAvg teacher
• GMM + agreement mask 기반 clean/noisy 샘플 구분
• Noisy KD + clean CE + FedProx-style regularization
• ComputeTracker 로 GPU+CPU 시간 및 샘플 수 자동 로깅

---

Repository Structure (예시)

.
├── FedReDA.py
├── dino_variant.py
├── other_repos/
│   └── FedNoRo/
├── rein/
│   └── models/backbones/
│       ├── reins_dinov2.py
│       └── reins.py
├── dataset/
│   └── dataset.py
├── utils/
│   └── utils.py
└── checkpoints/
    └── dinov2_vits14_pretrain.pth