This repository contains the implementation for the BMVC'21 paper "Surprisingly Simple Semi-Supervised Domain Adaptation with Pretraining and Consistency".
pip install -r requirements.txt
Our implementation follows a similar data format to MME. To download and set up DomainNet data for the real and sketch domains, run
sh download_data.sh
For setting up data for clipart and painting, uncomment corresponding lines in download_data.sh.
Images go into the following directories:
./data/multi/real/<category_name>,
./data/multi/sketch/<category_name>
And file lists in the txt format are available in the following directories:
./data/txt/multi/labeled_source_images_real.txt,
./data/txt/multi/unlabeled_target_images_sketch_3.txt,
./data/txt/multi/validation_target_images_sketch_3.txt.
Office, Office-Home and VisDA-17 datasets, can be set up in a similar manner,
with filelists provided in their respective directories in data/.
For pretraining with rotation prediction
python addnl_scripts/pretrain/rot_pred.py --batch_size=16 --steps=5001 --dataset=multi --source=real --target=sketch --save_dir=expts/rot_pred
For semi-supervised domain adaptation (SSDA) training
python main.py --steps=50001 --dataset=multi --source=real --target=sketch --backbone=expts/rot_pred/checkpoint.pth.tar
The following are commands to run the experiments for the corresponding results in the paper.
- Virtual Adversarial Training (VAT) : Adversarial perturbation using VAT instead of image augmentation can be used as follows:
python main.py --vat_tw=0.01 --steps=10001 --net=alexnet --aug_level=0 --dataset=office_home --source=Real --target=Clipart
- Pretraining with Momentum Contrast (MoCo)
python addnl_scripts/pretrain/moco.py --steps=5001 --dataset=office_home --source=Real --target=Clipart --save_dir=expts/moco_pretraining
- Evaluation scripts : To compute the
-distance, use (while providing the correct
dataset,source,targetandnet)
python addnl_scripts/eval/compute_proxy_distance.py --backbone_path=path/to/backbone
For computing nearest neighbors classifier accuracy, use
python addnl_scripts/eval/nn_classifier.py --backbone_path=path/to/backbone
If you find this repository useful for your work, please consider citing:
@article{mishra2021surprisingly,
title={Surprisingly Simple Semi-Supervised Domain Adaptation with Pretraining and Consistency},
author={Mishra, Samarth and Saenko, Kate and Saligrama, Venkatesh},
journal={arXiv preprint arXiv:2101.12727},
year={2021}
}