Evidential Uncertainty Quantification
for Active Domain Adaptation

Overview

An active domain adaptation framework based on evidential deep learning (EDL) implemented with

• two sampling strategies: uncertainty sampling and certainty sampling
• two uncertainty quantification methods: entropy-based and variance-based
• three EDL loss functions: negative log-likelihood, cross-entropy, and sum-of-squares

Paper

Official implementation of

Evidential Uncertainty Quantification: A Variance-Based Perspective [WACV 2024]
Ruxiao Duan¹, Brian Caffo¹, Harrison X. Bai², Haris I. Sair², Craig Jones^1
1Johns Hopkins University, ²Johns Hopkins University School of Medicine

paper | code | slides | poster | abstract

Datasets

• Office-Home
• Visda-2017

Main Files

• src/config.py: parameter configurations.
• src/data.py: data preprocessing and loading.
• src/loss.py: loss functions.
• src/model.py: model architecture.
• src/sampling.py: uncertainty and certainty sampling strategies.
• src/train.py: model training.
• src/transforms.py: image transformations.
• src/uncertainty.py: uncertainty quantification.
• src/utils.py: utility functions.

Getting Started

• Install environment:

git clone https://github.com/KerryDRX/EvidentialADA.git
conda create -y --name active python=3.7.5
conda activate active
pip install -r requirements.txt

• Download dataset from Office-Home or Visda-2017 to local environment. Image files should be stored in the hierarchy of
  • <dataset-folder>/<domain>/<class>/<image-filename>.
• In src/config.py:
  • Set DATASET.NAME to "Office-Home" or "Visda-2017".
  • Set PATHS.DATA_DIR to <dataset-folder>.
  • Set PATHS.OUTPUT_DIR to the output folder.
  • Set other parameters.
• Train the model:

python src/train.py

• Results are saved to PATHS.OUTPUT_DIR.

Acknowledgement

The active learning framework is partially adapted from Dirichlet-based Uncertainty Calibration.

Citation

@inproceedings{duan2024evidential,
  title={Evidential Uncertainty Quantification: A Variance-Based Perspective},
  author={Duan, Ruxiao and Caffo, Brian and Bai, Harrison X and Sair, Haris I and Jones, Craig},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  year={2024}
}