Semi-supervised Domain Adaptation via Sample-to-Sample Self-Distillation (WACV 2022)

Acknowledgements

The code is based on the official pytorch implementation of Semi-supervised Domain Adaptation via Minimax Entropy (ICCV 2019) (GitHub link: SSDA_MME).

Also, the official code of Attract, Perturb, and Explore: Learning a Feature Alignment Network for Semi-supervised Domain Adaptation (ECCV 2020) is used for several experiments in our paper (GitHub link: APE).

Environment

• CUDA 11.1
• Python 3.7.6
• Pytorch 1.8.0
• numpy / torchvision / tqdm / Pillow

You can create a conda environment through the commands below:

conda create -n s3d python=3.7.6
conda activate s3d
conda install pytorch==1.8.0 torchvision==0.9.0 cudatoolkit=11.1 -c pytorch -c conda-forge
pip install -r requirements.txt

Dataset preparation

You can follow the instruction of SSDA_MME to prepare the datasets (DomainNet and Office-Home).

All datasets should be in the data folder.

For example, images of real domain from DomainNet are located in,

./data/multi/real/<category_name>

Also, the dataset split files of DomainNet are located in,

./data/multi

Training

There are two stages for training the model.

Stage 1: Supervised learning using labeled datasets

python main_stage1.py --dataset multi --source real --target clipart --net resnet34 --save_check --num 1 --save_interval 1000 --exp_variation s1_s+t

Stage 2: Sample-to-sample self-distillation

python main_stage2.py --dataset multi --source real --target clipart --net resnet34 --save_check --num 1 --save_interval 1000 --s1_exp_variation s1_s+t --exp_variation s2_s3d --pseudo_interval 100 --kd_lambda 8 --sty_layer layer4