Automatic Feature Reweighting (AFR)

This repository contains the code for the paper Simple and Fast Group Robustness by Automatic Feature Reweighting by Shikai Qiu, Andres Potapczynski, Pavel Izmailov and Andrew Gordon Wilson (ICML 2023).

Introduction

Robustness to spurious correlations is of foundational importance for the reliable deployment of machine learning systems in the real-world, where it is common that deep learning models fail to generalize on unusual examples and minority groups in the data. Standard methods for reducing the reliance on spurious features typically assume that we know what the spurious feature is, which is rarely true in the real world. Methods that attempt to alleviate this limitation are complex, hard to tune, and lead to a significant computational overhead compared to standard training. To address these limitations, we propose Automatic Feature Reweighting (AFR), a simple method that is easy to implement, does not require group annotations on the training data or manual intervention from experts, and is computationally cheap to run and tune.

(Left) Test worst-group accuracy and (right) training time comparison on Waterbirds for state-of-the-art methods that do not use group information during training. AFR outperforms the baselines, while only requiring a small fraction of compute time.

AFR achieves state-of-the-art results across a wide variety of benchmarks without requiring group labels through training, and with fast training times that are only marginally longer than standard ERM. We summarize the worst group accuracy achieved by AFR and competing methods on benchmark datasets in the table below.

Method	Waterbirds	CelebA	MultiNLI	CivilComments	Chest X-Ray
ERM	72.6	47.2	67.9	57.4	4.4
JTT	86.7	81.1	72.6	69.3	52.3
CNC	88.5±0.3	88.8±0.9	-	68.9±2.1	-
AFR (Ours)	90.4±1.1	82.0±0.5	73.4±0.6	68.7±0.6	56.0±3.4

Citation

If you find AFR useful, please cite our work:

@article{qiu2023afr,
    title={{Simple and Fast Group Robustness by Automatic Feature Reweighting}},
    author={Shikai Qiu and Andres Potapczynski and Pavel Izmailov and Andrew Gordon Wilson},
    journal={International Conference on Machine Learning (ICML)},
    year={2023}
}

Installation and Requirements

Simply clone the repo and install the requirements by sourcing scripts/requirements.sh. We opted for having the requirements in a bash script rather than the usual formats to make it easier to see and install the version of PyTorch that we used in the project. The most important requirements are:

• PyTorch. ML framework used in the project.
• transformers. To load language models used for text benchmarks.
• timm. To load vision models used for image benchmarks.
• wandb. To perform logging and hyperparameter sweeps.

It is also required that you have access to the datasets that you want to run as benchmarks.

File Structure

.
├── data --> groups several utils for loading dataset and augmentations
│   ├── augmix_transforms.py
│   ├── dataloaders.py
│   ├── datasets.py --> contains the base clases for managing the datasets and groups
│   ├── data_transforms.py
│   ├── __init__.py
├── logs --> where experimental results get saved
├── losses --> comprises the loss functions used
│   └── __init__.py
├── models --> incorporates all the model definitions used in the experiments
│   ├── __init__.py  --> most model definitions
│   ├── model_utils.py
│   ├── preresnet.py
│   └── text_models.py
├── optimizers --> has the optimizers used in the experiments
│   └── __init__.py
├── notebooks --> provides easy-to-use notebooks for running AFR on Waterbirds, CelebA, MultiNLI, and CivilComments
│   ├── wb_celeba.ipynb
│   ├── multnli.ipynb
│   └── civil.ipynb
├── README.md
├── scripts
│   └── requirements.sh
├── setup.cfg --> yapf linter settings used in the project
├── train_embeddings.py --> main script to train from embeddings (second stage)
├── train_supervised.py --> main script used to run checkpoints (first stage)
└── utils --> utils to run experiments, log results and track metrics
    ├── common_utils.py
    ├── general.py
    ├── __init__.py
    ├── logging.py
    ├── logging_utils.py
    └── supervised_utils.py

Running AFR

We now show the command line arguments for running AFR's first and second stage. We explain the arguments in detail after each command. We also provide notebook implementations of AFR's second stage that take care of hyperparameter tuning and provide special support for MultiNLI and CivilComments due to different models and data formats. See "Running AFR in a Notebook".

First Stage

Using waterbirds as an example, for the first stage run

ARCH='imagenet_resnet50_pretrained'
DATA_DIR='/datasets/waterbirds_official'
SEED=21
PROP=80
OMP_NUM_THREADS=4 python3 train_supervised.py \
    --output_dir=logs/waterbirds/${PROP}_${SEED} \
    --project=waterbirds \
    --seed=${SEED} \
    --eval_freq=10 \
    --save_freq=10 \
    --data_dir=${DATA_DIR} \
    --data_transform=AugWaterbirdsCelebATransform \
    --model=${ARCH} \
    --train_prop=${PROP} \
    --num_epochs=50 \
    --batch_size=32 \
    --optimizer=sgd_optimizer \
    --scheduler=constant_lr_scheduler \
    --init_lr=0.003 \
    --weight_decay=1e-4

Name	Description
output_dir	Specifies where the results are saved.
project	Name of the wandb project.
seed	Seed to use.
eval_freq	How often (in epochs) to evaluate the current model on the validation set.
save_freq	How often (in epochs) to save a model checkpoint.
data_dir	File path where the dataset is located.
data_transform	Type of augmentation being used.
model	Model architecture.
train_prop	% of train dataset to use for the 1st stage.
num_epochs	Number of epochs to run.
batch_size	Size of the mini-batches.
optimizer	Type of optimizer to use.
scheduler	Type of scheduler to use.
init_lr	Initial learning rate (starting point for the scheduler).
weight_decay	Weight decay value.

Second Stage

To run the second stage in our waterbirds example use the following command

python3 train_embeddings.py \
--use_wandb \
--output_dir=logs/waterbirds/emb \
--project=waterbirds \
--seed=0 \
--base_model_dir=<first_stage_checkpoint_dir> \
--model=imagenet_resnet50_pretrained \
--data_dir=/datasets/waterbirds_official \
--data_transform=NoAugWaterbirdsCelebATransform \
--num_augs=1 \
--num_epochs=500 \
--batch_size=128 \
--emb_batch_size=-1 \
--optimizer=sgd_optimizer \
--scheduler=constant_lr_scheduler \
--init_lr=0.01 \
--momentum=0. \
--weight_decay=0. \
--grad_norm=1.0 \
--loss=afr \
--tune_on=train \
--train_prop=-20 \
--gamma=<afr_gamma> \
--reg_coeff=<afr_lambda>

Name	Description
use_wandb	use wandb for logging (remove this flag if not using).
output_dir	Specifies where the results are saved.
project	Name of the wandb project.
seed	Seed to use.
base_model_dir	Location of the first stage checkpoint.
model	Model architecture.
data_dir	File path where the dataset is located.
data_transform	Type of augmentation being used to compute embeddings.
num_augs	Number of augmentations to use for embeddings.
num_epochs	Number of epochs to run.
batch_size	Size of the mini-batches for computing embeddings.
optimizer	Type of optimizer to use.
scheduler	Type of scheduler to use.
init_lr	Initial learning rate (starting point for the scheduler).
weight_decay	Weight decay value (should be 0).
loss	Name of loss function to be used (should be afr).
train_prop	% of train dataset use for 2nd stage (-x% means the last x%).
gamma	$\gamma$ in AFR.
reg_coeff	$\lambda$ in AFR for ${\ell }_2$ regularization to first stage checkpoint.
grad_norm	Gradient norm cap (shouldb be 1).

Examples

Waterbirds

# 1st stage
python3 train_supervised.py --use_wandb --output_dir=logs/waterbirds/80_0 --project=waterbirds --seed=0 --eval_freq=10 --save_freq=10 --data_dir='/datasets/waterbirds_official' --dataset=SpuriousDataset --data_transform=AugWaterbirdsCelebATransform --model='imagenet_resnet50_pretrained' --max_prop=1.0 --train_prop=80 --num_epochs=50 --batch_size=32 --optimizer=sgd_optimizer --scheduler=constant_lr_scheduler --init_lr=0.003 --weight_decay=1e-4
# 2nd stage
python3 train_embeddings.py --use_wandb --output_dir=logs/waterbirds/emb --project=waterbirds --seed=0 --base_model_dir=logs/waterbirds/80_0 --model=imagenet_resnet50_pretrained --data_dir=/datasets/waterbirds_official --data_transform=NoAugWaterbirdsCelebATransform --num_epochs=500 --batch_size=128 --emb_batch_size=-1 --optimizer=sgd_optimizer --scheduler=constant_lr_scheduler --init_lr=0.01 --momentum=0. --weight_decay=0. --loss=afr --tune_on=train --train_prop=-20 --gamma=10 --num_augs=1 --grad_norm=1.0 --reg_coeff=0.0

CelebA

# 1st stage
python3 train_supervised.py --use_wandb --output_dir=logs/celeba/80_1 --project=celeba --seed=1 --eval_freq=10 --save_freq=10 --data_dir='/datasets/CelebA' --data_transform=AugWaterbirdsCelebATransform --model='imagenet_resnet50_pretrained' --max_prop=1.0 --train_prop=80 --num_epochs=20 --batch_size=100 --optimizer=sgd_optimizer --scheduler=cosine_lr_scheduler --init_lr=3e-3 --weight_decay=1e-4
# 2nd stage
python3 train_embeddings.py --use_wandb --output_dir=logs/celeba/emb --project=celeba --seed=0 --base_model_dir=logs/celeba/80_1 --model='imagenet_resnet50_pretrained' --data_dir='/datasets/CelebA' --data_transform=NoAugWaterbirdsCelebATransform --num_epochs=1000 --batch_size=128 --emb_batch_size=-1 --optimizer=sgd_optimizer --scheduler=constant_lr_scheduler --init_lr=0.02 --momentum=0.0 --weight_decay=0. --loss=afr --tune_on=train --train_prop=-0.2 --gamma=1.44 --num_augs=1 --grad_norm=1.0 --reg_coeff=0.001 --checkpoint=final_checkpoint.pt

CivilComments

# 1st stage
OMP_NUM_THREADS=4 python3 wilds_exps/run_expt.py --dataset civilcomments --algorithm ERM --root_dir /data/users/pavel_i/datasets/ --log_dir 'logs/bert_civilcomments_dfrdrop_2/' --seed 2 --dfr_reweighting_seed 2 --dfr_reweighting_drop --dfr_reweighting_frac 0.2 --model bert-base-uncased
# 2nd stage to be run in the notebook

MultiNLI

# 1st stage
OMP_NUM_THREADS=4 python3 gdro_fork/run_expt.py -s confounder -d MultiNLI -t gold_label_random -c sentence2_has_negation --lr 2e-05 --batch_size 32 --weight_decay 0 --model bert --n_epochs 3 --seed 1 --log_dir=logs/multinli/erm_dfrdrop1 --root_dir=/data/users/pavel_i/datasets/multinli/ --dfr_reweighting_drop --dfr_reweighting_seed=1 --dfr_reweighting_frac=0.2 --save_last --save_best
# 2nd stage to be run in the notebook

Running AFR in a Notebook

To facilitate reproducing results in the paper, the notebooks directory provide easy-to-use notebooks with self-contained code to run and tune AFR on Waterbirds, CelebA, MultiNLI, and CivilComments. You only need to modify a few path variables to use your own model checkpoints. Those variables can be located by searching for TODO. Separate notebooks exist for MultiNLI and CivilComments because they use models and data formats different from Waterbirds and CelebA that are not natively supported in our code base.

Data Access

Waterbirds and CelebA

Please follow the instructions in the DFR repo to prepare the Waterbirds and CelebA datasets.

CivilComments

For CivilComments, we use the version of the dataset natively available in the WILDS package.

We use the standard WILDS ERM training scrtipts to train models on this dataset.

MultiNLI

To run experiments on the MultiNLI dataset, manually download and unzip the dataset from this link. Then copy the dataset_files/utils_glue.py to the root directory of the dataset.

We train models on CivilComments using the code available in the spurious_feature_learning.

Chest X-Ray

The chest X-ray data that we used combines two sources: CheXpert and NIH. Please place both datasets on the same folder. Our dataloader is expecting a chexpert folder containing the CheXpert data and a nihcxr folder containing the NIH data.