The Group Robustness is in the Details: Revisiting Finetuning under Spurious Correlations

Official codebase for the NeurIPS 2024 paper: https://arxiv.org/abs/2407.13957

Installation

conda update -n base -c defaults conda
conda create -n milkshake python==3.10
conda activate milkshake
conda install pytorch==2.2.0 torchvision==0.17.0 pytorch-cuda=11.8 -c pytorch -c nvidia
python -m pip install -e .

Instructions

To run an experiment, specify the config with -c. For example,

python exps/finetune.py -c cfgs/waterbirds.yaml

By default, the program will run ERM finetuning with no class-balancing. Here is an example of a run with a different class-balancing method and model size:

python exps/finetune.py -c cfgs/waterbirds.yaml --convnextv2_version nano --balance_erm mixture --mixture_ratio 2

After models are finetuned, run eigenvalue computations with exps/postprocess.py.

Citation and License

This codebase uses Milkshake as a template and inherits its MIT License. Please consider using the following citation:

@inproceedings{labonte24revisiting,
  author={Tyler LaBonte and John C. Hill and Xinchen Zhang and Vidya Muthukumar and Abhishek Kumar},
  title={The Group Robustness is in the Details: Revisiting Finetuning under Spurious Correlations},
  booktitle={Conference on Neural Information Processing Systems (NeurIPS)},
  year={2024},
}