Source-Free-Domain-Generalization

An open-world scenario domain generalization code base

You can download the .zip file for all code directly from here. Anonymous links are for the convenience of double-blind review and will not be LTS.

Installation

# create virtual environment and install packages
conda env create -f environment.yaml

# activate virtual environment
conda activate cae

# install tllib 
python setup.py install

We offered anonymous links PACS, Terra for review.

you should put *.zip files in ./data and unzip.

└─data
    ├─domainnet
    ├─office-home
    ├─PACS
    ├─Terra
    └─VLCS

Usage

Scripts for experiment.

# exp on all the datasets 
sh [DG_method].sh

# for example
sh cae.sh

SFDG experiment.

# SFDG experiments 
python cae.py [data_path] -d [dataset] -t [target domain] -a [backbone_of_CLIP] --seed [seed] --log [log_path]

# for example 
python cae.py data/PACS -d PACS  -t S -a vitb16 --seed 0 --log logs/cae/PACS_S

Open-world experiment

We collected two extra domains ('X' for pixel_style and 'G' for geometric) for PACS dataset to test open-world performance of our method.

# SFDG
python cae.py data/PACS -d PACS  -t X -a vitb16 --seed 0 --log logs/cae/PACS_X

# DG
python erm.py data/PACS -d PACS  -s P A C -t G -a resnet50 --seed 0 --log logs/erm/PACS_G

DG experiment.

# DG experiments  
python [DG_method].py [data_path] -d [dataset] -s [source domains] -t [target domain] -a [backbone_of_CLIP] --seed [seed] --log [log_path]

# for example 
python erm.py data/PACS -d PACS -s P A C  -t S -a resnet50 --seed 0 --log logs/cae/PACS_S --freeze-bn

TLlib experiment.

TLlib is a public toolbox for transfer learning, we modified these files for experiments on Terra and VLCS datasets.

./tllib/vision/datasets/terra.py

./tllib/vision/datasets/vlcs.py

./tllib/vision/datasets/__init__.py