SCER

Code Implementation for Spurious Correlation Aware Embedding Regularization for Worst Group Robustness

Code Implementation

Installation

First, clone our repository and set up the environment:

# Clone our repository first
# Create and activate the conda environment
conda env create -f environment

To download the required data, run the following command:

python -m scer.scripts.download --data_path <data_path> --download

Training

To train SCER for one time, use the following command:

python -m scer.train --algorithm SCER --dataset "Your Data" --train_attr yes \
    --data_dir "Your_path" --output_dir "Your_path"

To run a sweep for SCER, use the following command:

python -m scer.sweep launch --algorithms SCER --dataset "Your Data" \
    --train_attr no --n_hparams "Your Num" --n_trials 1

Collecting Results

After training, you can collect and summarize the results using the following command:

python -m scer.collect_results --input_dir "Your output path" 

Acknowledgements

This implementation is primarily based on SubpopBench (GitHub Repository).
We would like to acknowledge and thank the authors of SubpopBench for their work and contributions.