virajprabhu / inclusive_da

PyTorch Code for "Can domain adaptation make object recognition work for everyone?" (L2ID workshop, CVPR 2022)

PyTorch Code for "Can domain adaptation make object recognition work for everyone?"

Learning with Limited and Imperfect Data workshop, CVPR 2022

Table of Contents

• Setup and Dependencies
• Usage
  • Download data
  • Train and adapt model
• Citation

Setup and Dependencies

1. Create an anaconda environment with atleast Python 3.6 and activate:

conda create -n da python=3.6.8
conda activate da

2. Navigate into the code directory: cd inclusive_da/
3. Install dependencies: (Takes ~2-3 minutes)

pip install -r requirements.txt

And you're all set!

Usage

Download data

To download both Dollarstreet-DA and GeoYFCC-DA datasets, run cd data; sh download.sh (note: requires ~12GB of free space). This will download both datasets as a single zip file – unzip and you are set.

Train unsupervised DA model

Run python train.py to train a source model from scratch followed by unsupervised DA, by passing it appropriate arguments.

Hyperparameter configurations for each benchmark are included as yml files inside the config folder:

python train.py --id <experiment_identifier> \
                --load_from_cfg True \ 
                --cfg_file config/dollarstreet/<da_method>.yml

You can pass in in DANN/MMD/SENTRY as <da_method>, or implement your own in adapt/adapt.py. To run a custom train job, you can either i] Create a new config file (ideally don't edit the existing ones), or ii) manually override the value for certain hyperparameters.

Citation

If you found our curated dataset or work useful, please consider citing our paper as well as the papers introducing the original data sources:

@inproceedings{prabhu_2022_CVPR,
    author    = {Prabhu, Viraj and Selvaraju, Ramprasaath R. and Hoffman, Judy and Naik, Nikhil},
    title     = {Can Domain Adaptation Make Object Recognition Work for Everyone?},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2022},
    pages     = {3981-3988}
}

@inproceedings{rojas2022dollar,
  title={The dollar street dataset: Images representing the geographic and socioeconomic diversity of the world},
  author={Rojas, William A Gaviria and Diamos, Sudnya and Kini, Keertan Ranjan and Kanter, David and Reddi, Vijay Janapa and Coleman, Cody},
  booktitle={Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
  year={2022}
}

@inproceedings{dubey2021adaptive,
  title={Adaptive methods for real-world domain generalization},
  author={Dubey, Abhimanyu and Ramanathan, Vignesh and Pentland, Alex and Mahajan, Dhruv},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={14340--14349},
  year={2021}
}