group-conditional-DRO

This repository contains code for experiments in the ICML2021 paper Examining and Combating Spurious Features under Distribution Shift.

Data

Data processing scripts can be found under fairseq_gdro/process.

MNLI

(1) Generate meta data of the MNLI dataset

python process/gen_mnli_meta.py

(2) Generate imperfect partitions, under the output folder "*.fg.labels" are labels for clean partitions, and "train.resplit.labels" are labels for imperfect partions

python process/gen_mnli.py

(3) make binarized data as inputs to fairseq

bash process/make_bin_mnli.sh
Toxicity Detection: FDCL18 ((Fortuna & Nunes, 2018)), besides the clean partition, we also explore imperfect partitions created by a supervised classifier (fairseq_gdro/process/gen_resplite_labels_toxic.py) and with unsupervised clustering using BERT-sentence embeddings (fairseq_gdro/process/clustering_with_pretrained_models.py).
CelebA (Liu et al., 2015): Data loader for clean and imperfect partitions can be found in image_classification/data/celeba.py

Training

For text experiments, please install fairseq under fairseq_gdro:

cd fairseq_gdro; pip install --editable ./

Important arguments can be found in Line 488 of fairseq_gdro/fairseq/options.py and fairseq_gdro/fairseq/criterion/group_dro_loss.py.

Baseline Models: the scripts to run baseline approaches can be found under fairseq_gdro/baseline_jobs.
GC-DRO: the scripts to run GC-DRO can found under fariseq_gdro/jobs/.

For image classification, the scripts to run baseline approaches and GC-DRO can be found under image_classification/scripts/. For clean partition, use --group_split 'confounder'; for imperfect partition, use --group_split 'domain'.

For details of the methods and results, please refer to our paper.

@inproceedings{zhou21icml,
    title = {Examining and Combating Spurious Features under Distribution Shift},
    author = {Chunting Zhou and Xuezhe Ma and Paul Michel and Graham Neubig},
    booktitle = {International Conference on Machine Learning (ICML)},
    address = {Virtual},
    month = {July},
    url = {http://arxiv.org/abs/2106.07171},
    year = {2021}
}
}

violet-zct/group-conditional-DRO

group-conditional-DRO

Data

Training