Modeling the Q-Diversity in a Min-max Play Game for Robust Optimization

This code implements the following paper:

Modeling the Q-Diversity in a Min-max Play Game for Robust Optimization

Environment

Create an environment with the following commands:

conda create --name gdro python=3.8
conda activate gdro
pip install -r requirements.txt

Downloading Datasets

We processed all the involved training datasets BiasedSST2, SST2, MultiNLI, CivilComments and they can be downloaded here.

Adding other datasets

Add the following:

• put the dataset file in the folder dataset/
• inherit from the class ConfounderDataset in util_data/confounder_dataset.py (similar to util_data/multinli_dataset.py)
• edit util_data/utils.py to load the new dataset and modify dataset_attributes in util_data/data.py.

Sample Commands for running Q-Diversity

CUDA_VISIBLE_DEVICES=4 python main.py \
    -d BiasedSST2 \
    -t gold_label_random \
    -c has_bias_string \
    --lr 2e-05 --batch_size 32 \
    --meta_epoch 1 \
    --weight_decay 0 \
    --model bert \
    --n_epochs 20 \
    --reweight_groups \
    --robust \
    --generalization_adjustment 0 \
    --mix_alpha 7 \
    --weight_mix 0.5 \
    --log_dir logs/biasedsst_gdro_alpha7_epochs20_lr2e-5_weight1_me1 --save_best