CFA: Compositional Feature Alignment

Code for the paper Enhancing Compositinal Generalization via Compositional Feature Alignment

CREDITS: Our code is heavily based on https://github.com/mlfoundations/wise-ft and https://github.com/locuslab/FLYP. We thank the authors for open sourcing their code.

Setting up conda env

conda create -n cfa python=3.11 anaconda
conda activate cfa
pip install -r requirements.txt

Datasets

We conducted our experiments on DomainBed and Wilds. Put the datasets under ./cfa/data/.

Run CFA

Our CFA is a two stage finetuning method. To run CFA, fill out all the parameters in the lauching files.

Stage-1

To run Stage-1 of CFA and Linear Probing, fill out the configs in linear_probe_exps.py.

python linear_probes_exps.py  # Run Linear Probing Codes

Stage-2

To run Stage-2 of CFA, Finetuning and LP-FT, fill out the configs in launch_exp.py.

python launch_exp.py  # Run Finetuning Codes

WiSE

To perform model interpolation, fill out the configs in interpolate.py.

python interpolate.py  # Run WiSE