Learn from the Learnt for SF-ADA (ECCV 2024)

[Paper]

Introduction

This repo provides the official implementation of paper Learn from the Learnt: Source-Free Active Domain Adaptation via Contrastive Sampling and Visual Persistence.

Benchmarks

• VisDA-C
• office-home
• office

Usage

Requirements

• python=3.10
• torch=2.3.0
• torchvision

Dataset

Datasets are placed in data/<dataset>/<domain>. Then generate the filepath2label lists following tim-learn/SHOT.

Training and Evaluation

Take the VISDA-C for example,

• Train the source model

python train_source.py \
    --dset VISDA-C  --net resnet101 --output ckps/source/ \
    --lr 1e-3 --max_epoch 10
# For reference, my source ckpt got 53.67% on the target domain before adaptation. 

• Iteratively query and train on the target domain

python adapt_target.py --dset VISDA-C --init_pool source \
    --source_dir ckps/source/uda/VISDA-C/T --seed {seed} \
    --lambda_topk 0.1 --cd_ratio 0.03 --topk 500 \
    --beta_vpa 10.0 --beta_im 0.9 --sfada_ubl 1 \
    --num_round 10 --max_epoch 10

# for the office dataset
python adapt_target.py --dset office --init_pool source \
    --s 0 --t 1 --seed {seed} \
    --lambda_topk 0.1 --cd_ratio 0.03 --topk 100 \
    --beta_vpa 1.0 --beta_im 0.9 \
    --num_round 10 --ratio_per_round 0.01 \
    --net resnet50 --max_epoch 15

The results reported in the paper were conducted with seeds [2023, 2024, 2025], with smaller datasets naturally exhibiting some variance.

Acknowledgment

The codebase is based on the SF-DA work tim-learn/SHOT.

Citation

@inproceedings{lftl2024,
    title={Learn from the learnt: source-free active domain adaptation via contrastive sampling and visual persistence},
    author={Lyu, Mengyao and Hao, Tianxiang and Xu, Xinhao and Chen, Hui and Lin, Zijia and Han, Jungong and Ding, Guiguang},
    booktitle={European Conference on Computer Vision},
    pages={228--246},
    year={2024},
    organization={Springer}
}