AnyDA: Anytime Domain Adaptation [Paper] [Website]

This repository contains the implementation details of our ICLR 2023 paper AnyDA: Anytime Domain Adaptation.

Omprakash Chakraborty*, Aadarsh Sahoo, Rameswar Panda and Abir Das, "AnyDA: Anytime Domain Adaptation"\

If you use the codes and models from this repo, please cite our work. Thanks!

@inproceedings{chakraborty2022anyda,
  title={{AnyDA: Anytime Domain Adaptation}},
  author={Chakraborty, Omprakash and Sahoo, Aadarsh and Panda, Rameswar and Das, Abir},
  booktitle={The Eleventh International Conference on Learning Representations},
  year={2023}
}

Preparing the Environment

Conda

Please use the anyda.yml file to create the conda environment anyda as:

conda env create -f anyda.yml

Data Directory Structure

All the datasets should be stored in the folder ./data following the convention ./data/<dataset_name>/<domain_names>. E.g. for Office31 the structure would be as follows:

.
├── ...
├── data
│   ├── Office31
│   │    ├── amazon
│   │    ├── webcam
│   │    ├── dslr
│   └── ...
└── ...

For using datasets stored in some other directories, please update the path to the data accordingly in the txt files inside the folder ./data_labels.

The official download links for the datasets used for this paper are:

Office31: https://people.eecs.berkeley.edu/~jhoffman/domainadapt/#datasets_code

OfficeHome: http://hemanthdv.org/OfficeHome-Dataset/

DomainNet: http://ai.bu.edu/M3SDA/

Hyperparameters

All the training hyperparameters for a given dataset should be saved as .yml file and stored in the folder ./apps. A sample file has been provided in the ./apps folder name resnet50_office31.yml

Training AnyDA

Here is a sample and recomended command to train AnyDA for the transfer task of Amazon -> Webcam from Office31 dataset: