KL Guided Domain Adaptation

This repository is the official implementation for the ICLR 2022 paper KL Guided Domain Adaptation.

Please consider citing our paper as

@inproceedings{
nguyen2022kl,
title={{KL} Guided Domain Adaptation},
author={A. Tuan Nguyen and Toan Tran and Yarin Gal and Philip Torr and Atilim Gunes Baydin},
booktitle={International Conference on Learning Representations},
year={2022},
url={https://openreview.net/forum?id=0JzqUlIVVDd}
}

Credits:

A large part of this repo is modified from the DomainBed codebase https://github.com/facebookresearch/DomainBed

Requirements:

python3, pytorch 1.7.0 or higher, torchvision 0.8.0 or higher

How to run:

To run the experiments with the default hyper-parameters (might get slightly sub-optimal performance)

python -m scripts.train --data_dir [data_dir] --algorithm [algorithm] --seed [seed] --dataset [dataset] --train_envs [source_env] --test_envs [target_env]

Where:

[data_dir] is the /pat/to/your/data/directory
[algorithm] is any of the algorithm reported in our paper, namely KL, ERM, PERM, DANN, MMD, CORAL, WD.
[seed] is the random seed (0,1,2).
[dataset] is any of the dataset reported in the paper: RotatedMNIST, SVHNMNIST, MNISTUSPS, VisDA17, etc.
- If RotatedMNIST: [source_env] is 0 and [target_env] is either 1, 2, 3, 4 or 5.
- If SVHNMNIST: [source_env] is 0 (SVHN) and [target_env] is 1 (MNIST).
- If MNISTUSPS: [source_env] is 0 (MNIST) and [target_env] is 1 (USPS) or vice versa.
- If VisDA17: [source_env] is 0 and [target_env] is 1.

Alternatively, you can also do a full sweep to find the best hyper-parameters as we did

python -m scripts.sweep launch \
    --datasets RotatedMNIST \
    --algorithms KL \
    --output_dir ./results_sweep \
    --data_dir /path/to/your/data/ \
    --gpus 0 1 2 3 4 5 6 7  \
    --command_launcher multi_gpu

You can set the --datasets flag to other datasets such as SVHNMNIST, MNISTUSPS, VisDA17, etc.
You can set the --algorithms flag to other algorithms in the paper such as ERM, PERM, DANN, MMD, CORAL, WD.
The --gpus flag include all possible gpus of your system.

To check the results: