SfMamba

<div align="center"> <h1>SfMmaba </h1> <h3>SfMamba: Efficient Source-Free Domain Adaptation via Selective Scan Modeling</h3> <h3>https://arxiv.org/abs/2601.08608</h3> </div>

Overview

• SfMamba

<p align="center"> <img src="assets/backbone.png" alt="arch" width="90%"> </p>

Main Results

Results

| Name | Backbone | Office-Home | VisDA-C | Office | DomainNet-126| Params | FLOPs | Th. | | :---: | :---:| :---:| :---:| :---: |:---: | :---: | :---: | :---: | | SfMamba-S |VMamba-S[s2l15] | 81.7| 89.3| 93.3 | 77.9 | 58.9M | 9.2G | 503 | | SfMamba-T |VMamba-T[s1l8] | 79.2 | 88.5 | 91.2 | 76.2 | 39.0M | 5.4G | 701|

Getting Started

Environment and Dataset Setup

1. Create Environment
   Follow the VMamba repository instructions to set up your Python environment.
2. Download Datasets
   Obtain the following datasets from their official sources:

• Office

• Office-Home

• VisDA-C

• DomainNet-126

  Set the data_root path in our training scripts (run_sfda_*.sh) to your dataset directory.

3. Download Pretrained Backbones

• Get the checkpoint files from the links provided in the above table
• Place them in the utils/ folder
• Configure model_path and config in the corresponding run_sfda_*.sh file

Training Pipeline

We provide complete training scripts for both source model pre-training and target domain adaptation. The repository contains dedicated bash files for each benchmark dataset:

bash run_sfda_office.sh       # For Office dataset
bash run_sfda_officeHome.sh   # For Office-Home
bash run_sfda_visda.sh        # For VisDA-C
bash run_sfda_domainnet126.sh # For DomainNet-126

1. Source Model Training

Example: Train source model on DomainNet126 Clipart domain (domain 0)

python main_source.py 
    --cfg $config_file \
    --data_path ${data_root}/domainnet126/ \
    --lr $lr \
    --output $source_dir \
    --dataset domainnet126 \
    --test_envs 1 2 3 \
    --pretrained $model_path \
    --dg_aug \
    --batch_size 64 \
    --test_batch_size 32 \
    TRAIN.WARMUP_EPOCHS 3 \
    TRAIN.EPOCHS 30

--test_envs: Domains held out for testing

2. Target Model Training

Example: Adapt from Clipart (domain 0) to Sketch (domain 3)

st_cas="03"   # Source-Target pair (0→3)
for pair in $st_cas; do  
    s=${pair:0:1} # Source domain index
    t=${pair:1:1} # Target domain index
    echo "s=$s, t=$t"
    python main_target.py --cfg $config_file \
        --data_path ${data_root}/domainnet126/ \
        --lr $lr \
        --output $out_dir \
        --dataset domainnet126 \
        --target_env $t \
        --issave \
        --batch_size 32 \
        --source_env $s \
        --test_batch_size 16 \
        MODEL.SOURCE_DIR $source_dir \
        TRAIN.EPOCHS $epoch \
        TRAIN.WARMUP_EPOCHS 1 
done

Inference

Examples for inference the model transfered from source domain 0 to target domain 1 in the Office-Home dataset with seed 1 is offered in below bash file.

bash inference.sh

Citation

@misc{chen2026sfmambaefficientsourcefreedomain,
      title={SfMamba: Efficient Source-Free Domain Adaptation via Selective Scan Modeling}, 
      author={Xi Chen and Hongxun Yao and Sicheng Zhao and Jiankun Zhu and Jing Jiang and Kui Jiang},
      year={2026},
      eprint={2601.08608},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2601.08608}, 
}