Adapt via Bayesian Nonparametric Clustering (ABC)

Overview

Adapt via Bayesian Nonparametric Clustering (ABC) is a novel framework designed for Source Free Domain Adaptation scenarios where unknown target classes are present.

Dependencies & Setup

Please run in Python3.10+ environment. We recommend using a virtual environment to manage dependencies.

Install all dependencies:

pip install -r requirements.txt

Before running any experiments, compile the Cython extension:

python setup.py build_ext --inplace

Training

1. Download Datasets

Download the following datasets from their official websites and unzip them into the ./data folder with the structure below:

./data
├── office
│   ├── amazon
│   │   └── ...
│   ├── amazon.txt
│   └── ...
├── OfficeHome
│   ├── Art
│   │   └── ...
│   ├── Art.txt
│   └── ...
└── visda
    ├── train
    │   └── train.txt
    └── validation
        └── validation.txt

2. Pretrain Source Models

We use the source model pretraining procedure from GLC. Place the pretrained models in pretrained_source/OPDA/ with the naming convention {dataset}_{source}.pkl. For example:

pretrained_source/OPDA/office_amazon.pkl

3. Run Training

Run open-partial domain adaptation on Office, OfficeHome, and VisDA:

bash train.sh

License

This code is released under the MIT License.