EUDA: An Efficient Unsupervised Domain Adaptation via Self-Supervised Vision Transformer

Introduction

EUDA (Efficient Unsupervised Domain Adaptation) is a novel approach leveraging self-supervised Vision Transformers for domain adaptation tasks. This method significantly reduces the number of learnable parameters, making the adaptation process more efficient while maintaining high performance.

The implementation includes a submodule from DINOv2, which provides the self-supervised pre-training backbone for our Vision Transformer model.

Key Features

• Efficient Parameter Usage: Utilizes 42% to 99.7% fewer learnable parameters.
• Self-Supervised Learning: Leverages Vision Transformers pre-trained in a self-supervised manner.
• Scalable: Designed to be scalable and adaptable to various domain adaptation scenarios.
• Flexible Architecture: Built with modularity and flexibility, allowing easy modifications and extensions.

Architecture

The architecture consists of a Vision Transformer backbone pre-trained using self-supervised learning techniques provided by DINOv2. This backbone is then adapted using our efficient unsupervised domain adaptation methodology.

Installation

Prerequisites

• Python >= 3.8
• Additional dependencies as listed in requirements.txt

Steps

1. Clone the Repository

git clone --recurse-submodules https://github.com/A-Abedi/EUDA
cd EUDA

2. Install Dependencies

pip install -r requirements.txt

Usage

Data Preparation

Download the datasets from Office-31, Office-Home, VisDA-2017 and place them in a directory of your choice.

Specify the path to this directory using the --dataset_path argument in the running command.

Training

Run training process using this command:

python3 Main.py --train_batch_size BATCHSIZE --dataset DATASET_NAME --name EXPERIEMENT_NAME --train_list TRAIN_LIST --test_list TEST_LIST --num_classes NUM_CLASSES --backbone_size BACKBONE_SIZE --bottleneck_size BOTTLENECK_SIZE --num_steps NUM_STEPS --img_size IMAGE_SZIE --gpu_id GPU_ID --dataset_path /Datasets/

For example:

python3 Main.py --train_batch_size 32 --dataset office-home --name office_home_da --train_list Data/office-home/Art.txt --test_list Data/office-home/Clipart.txt --num_classes 65 --backbone_size base --bottleneck_size base --num_steps 2000 --img_size 224 --gpu_id 0 --dataset_path /Datasets/

Test

Test the trained model using command like this:

python3 Main.py --eval_batch_size 64 --test_list Data/office/amazon_list.txt --num_classes 65 --backbone_type DINOv2 --backbone_size base --bottleneck_size base --img_size 224 --gpu_id 0 --is_test --pretrained_dir /Output/checkpoint.bin

Acknowledgements

• DINOv2
• Our code is based on the TVT code base. We extend our gratitude to the authors for their contributions.

References

Include references to the paper:

@article{abedi2024euda,
  title={EUDA: An Efficient Unsupervised Domain Adaptation via Self-Supervised Vision Transformer},
  author={Abedi, Ali and Wu, Q M Jonathan and Zhang, Ning and
                   Pourpanah, Farhad},
  journal={arXiv preprint arXiv:2407.21311},
  year={2024}
}