FedSCAl: Leveraging Server and Client Alignment for Unsupervised Federated Source-Free Domain Adaptation

A guide to setting up, training, and running SCAl

---

🎯 Overview

an FL framework leveraging our proposed Server-Client Alignment (SCAl) mechanism to regularize client updates by aligning the clients’ and server model’s predictions.

This repository supports:

• Source training
• SCAl:Federated source-Free Domain Adaptation (FFreeDA)
• SCAL with adaptive or fixed thresholds
• Baseline (no SCAl)
• BMD(as base sfda)
• ViT-Small and ViT-Base backbones
• OfficeHome and DomainNet-small dataset experiments

---

📂 Directory Structure

./data/
    └── <dataset_folder>
        ├── OfficeHome
        ├── DomainNetS
        └── ...

Ensure all datasets are downloaded inside ./data/<dataset_folder>.

---

Step 1: Train Server (Source) Models

python train_centralDA_source.py \
    --domain_s 'art' --domain_u 'clipart' \
    --model_name VITs --d_mode new --device cuda:0 \
    --var_lr 0.03 --scheduler_name 'CosineAnnealingLR' --cycles 50 \
    --control_name "1_sup-ft-fix_15_0.3_iid_5-5_0.07_1" \
    --resume_mode 1 --init_seed 2020 --backbone_arch vit-small

python train_centralDA_sourceOC.py \
    --data_name 'DomainNetS' --domain_s 'clipart' \
    --data_name_unsup 'DomainNetS' --domain_u 'infograph' \
    --model_name VITs --d_mode new --device cuda:0 \
    --var_lr 0.1 --scheduler_name 'CosineAnnealingLR' --cycles 50 \
    --control_name "1_sup-ft-fix_8_0.5_iid_5-5_0.07_1" \
    --resume_mode 1 --init_seed 2020 --backbone_arch vit-b

---

Step 2: Run SCAl

🔵 SCAl on OfficeHome

python train_classifier_ssDA_target.py \
    --domain_s 'art' --unsup_doms 'product-clipart-realworld' \
    --model_name VITs --d_mode new --device cuda:1 \
    --var_lr 0.03 --scheduler_name 'CosineAnnealingLR' --cycles 60 \
    --control_name "1110022_sup-ft-fix_15_0.3_iid_5-5_0.07_1" \
    --resume_mode 1 --init_seed 2025 --run_shot 1 --par 1 \
    --tag_ "2020_art_0.03_VITs_1_sup-ft-fix" \
    --client_test 1 --pick 'checkpoint' --add_fix 1 \
    --backbone_arch vit-small --lambda 3 --g_lambda 0.3 \
    --global_reg 1 --adpt_thr 1

---

SCAL Variants

🟣 Fixed Threshold Instead of Adaptive

--adpt_thr 0 --threshold <set_threshold>

🟢 Client Alignment Only (No Global Regularizer)

--global_reg 0 

⚪ Baseline (No SCAl)

--add_fix 0

🔶 BMD(as base sfda)

--run_shot 0 --add_fix 1 --global_reg 1 --adpt_thr 1

🔷 ViT-Base Backbone

--backbone_arch vit-b

---

Run SCAL on DomainNet-Small

python train_classifier_ssDA_target_DN.py \
    --data_name 'DomainNetS' --data_name_unsup 'DomainNetS' \
    --domain_s 'sketch' \
    --unsup_doms 'clipart-quickdraw-real-painting-infograph' \
    --model_name VITs --d_mode new --device cuda:0 \
    --var_lr 0.03 --scheduler_name 'CosineAnnealingLR' --cycles 50 \
    --control_name "99001_sup-ft-fix_10_0.3_iid_5-5_0.07_1" \
    --resume_mode 1 --init_seed 2020 --run_shot 1 --par 1 \
    --tag_ "2020_sketch_0.1_VITs_991_sup-ft-fix" \
    --client_test 1 --pick 'checkpoint' --backbone_arch vit-small

---

🛠️ Tips for Reproducibility

• Keep init_seed identical across runs for consistent results.
• Use the same control_name when resuming.
• Verify dataset folder structure before training.
• Ensure GPU selection (cuda:0, cuda:1, ...) matches your hardware.