Fairness meets Cross-Domain Learning: a Benchmark of Models and Metrics

Official code of our work "Fairness meets Cross-Domain Learning: a Benchmark of Models and Metrics"

Introduction

Deep learning-based recognition systems are deployed at scale for real-world applications that inevitably involve our social life. Although of great support when making complex decisions, they might capture spurious data correlations and leverage sensitive attributes (e.g. age, gender, ethnicity). How to factor out this information while maintaining high performance is a problem with several open questions, many of which are shared with those of the domain adaptation and generalization literature which aims at avoiding visual domain biases. In this work, we propose an in-depth study of the relationship between cross-domain learning (CD) and model fairness, by experimentally evaluating 14 CD approaches together with 3 state-of-the-art fairness algorithms on 5 datasets of faces and medical images spanning several demographic groups. We consider attribute classification and landmark detection tasks: the latter is introduced here for the first time in the fairness literature, showing how keypoint localization may be affected by sensitive attribute biases. To assess the analyzed methods, we adopt widely used evaluation metrics while also presenting their limits with a detailed review. Moreover, we propose a new Harmonic Fairness (HF) score that can ease unfairness mitigation model comparisons. Overall, our work shows how CD approaches can outperform state-of-the-art fairness algorithms and defines a framework with dataset and metrics as well as a code suite to pave the way for a more systematic analysis of fairness problems in computer vision.

Setting up the environment

Requirements

Make sure to have a CUDA capable device, running at learst CUDA 11.3. Throughout our experiments we used Python version 3.7.6

General Dependencies

To install all the required dependencies go to the root folder of this project and run:

pip install -r requirements.txt

Classification Datasets

CelebA

1. Download and extract the CelebA "Align&Cropped Images" dataset and the Attributes Annotations from https://mmlab.ie.cuhk.edu.hk/projects/CelebA.html
2. Place image_align_celeba folder, list_attr_celeba.txt and list_eval_partition.txt inside the data/celeba folder.

COVID-19 Chest X-Ray

1. Download the COVID-19 Chest X-Ray dataset from https://github.com/ieee8023/covid-chestxray-dataset
2. Place the decompressed covid-chestxray-dataset folder inside the data folder.

Fitzpatrick17k

1. Download fitzpatrick17k.csv from https://github.com/mattgroh/fitzpatrick17k and follow their instructions to download the images of the Fitzpatrick17k dataset.
2. Place fitzpatrick17k.csv in the data/fitzpatrick17k folder.
3. Place the downloaded images inside the data/fitzpatrick17k/images folder.

Landmark Detection Datasets

UTKFace

1. Download the UTKFace "Aligned&Cropped Faces" dataset and the Landmarks Annotations from https://susanqq.github.io/UTKFace/
2. Place all the images from UTKFace/align folder inside the data/UTKFace folder.

Experiments

Once you've correctly setup the datasets and the environment you can start running the experiments:

python main.py [--arguments...]

To view the general list of arguments, please refer to parse_args.py. Moreover, each specific metric, dataset and experiment, may have its own separate list of additional arguments which can be found in their respective .py file in the codebase.

Full examples on how to run the experiments can be found in the scripts folder.

Contributing

In order to contribute with your own task, dataset, metric or experiment, follow the provided codebase folder structure. Fork this repository and send a pull request.

Adding a Task

To add a new computer vision task, follow the folder structure and make a new folder (with the name of the task) inside the datasets, experiments, metrics and models folders. Specifically, inside the datasets/<new-task-name> folder create two files:

• base_datasets.py : containing the Dataset objects inheriting from torch.utils.data.Dataset
• base_transforms.py : containing the classes from which the build_transform(args) function gets called (refer to the already provided examples)

Adding a Dataset

To add a new dataset, make a new folder with the name of the dataset inside the datasets/<task-name> folder. Then, create a file named build.py inside that folder (file structure example: datasets/<task-name>/<dataset-name>/build.py). The build.py file contains two functions:

• build_splits(args) : is responsible for reading and separating the data into splits. This function shall return a python dict (each value of the dict is a list containing the examples for each split)
• [ Optional ] additional_arguments(parser) : is responsible for parsing additional arguments related to the dataset itself.

If you need a custom Dataset object or a custom Transform (other than those already provided) refer to the base_datasets.py and base_transforms.py files, respectively.

Adding a Metric

To add a new dataset, make a new file or edit the already present file named meters.py inside the metrics/<task-name> folder. Add your code inside meters.py file. Each Metric is an object in the form:

class NewMetricName:

    def __init__(self, args, meters_dict):
        self.args = args
        self.meters_dict = meters_dict
        # Value of the metric is stored and accessed via self.value
        self.value = None

    def compute(self, predicted, target, group):
        # Perform calculation of the metric
        self.value = ...
        return self.value

    @staticmethod
    def additional_arguments(parser):
        # If you need additional metric-specific arguments
        parser.add_argument(...)

    @staticmethod
    def compare(current, best):
        return current > best # Example usage

Adding an Experiment

To add a new dataset, make a new folder with the name of the method (containing a file named experiment.py) inside the experiments/<task-name> folder. (eg. experiments/<task-name>/<method-name>/experiment.py) Add your code inside experiment.py file. Each Experiment is an object in the form:

# NOTE: the name of the class MUST BE 'Experiment'
class Experiment:
    # How datasets/dataloaders shall be configured for this experiment (you can add additional entries inside this dict)
    # NOTE: 'train', 'val', 'test' are reserved keys used by the main training/validation/testing loops, respectively.
    data_config = {
        'train': {'dataset': 'BaseDataset', 'set': 'train_set', 'transform': 'BaseTrainTransform', 'filter': None, 'shuffle': True,  'drop_last': False},
        'val':   {'dataset': 'BaseDataset', 'set': 'val_set',   'transform': 'BaseTestTransform',  'filter': None, 'shuffle': False, 'drop_last': False},
        'test':  {'dataset': 'BaseDataset', 'set': 'test_set',  'transform': 'BaseTestTransform',  'filter': None, 'shuffle': False, 'drop_last': False},
        ...
    }

    @staticmethod
    def additional_arguments(parser):
        # If you need additional experiment-specific arguments
        parser.add_argument('--additional-argument-1', ...)
        parser.add_argument('--additional-argument-2', ...)
        ...
        # If you want to auto-set the output path according also to the values of each additional argument
        # you can return a list with the names of the names of the arguments to be logged. (otherwise return
        # an empty list)
        return ['additional-argument-1', 'additional-argument-2']

    def __init__(self, args, dataloaders):
        self.args = args
        self.dataloaders = dataloaders

        # NOTE: Each Experiment object MUST contain these two attributes
        self.iteration = 0
        self.best_metric = None

        # Model setup
        self.model = ...

        # Optimizers and Schedulers setup
        ...

    def save(self, path):
        # Save a checkpoint for the experiment (called at the end of each validation, will save both the best and last checkpoints)
        torch.save({
            'model': self.model.state_dict(),
            'iteration': self.iteration,
            'best_metric': self.best_metric,
            ...
        }, path)

    def load(self, path):
        # Restore a checkpoint (called before training if the run already exists and before testing to load the best checkpoint found)
        checkpoint = torch.load(path)
        self.model.load_state_dict(checkpoint['model'])
        self.iteration = checkpoint['iteration']
        self.best_metric = checkpoint['best_metric']
        ...

    def train_iteration(self, data):
        # Called at each training iteration

        # Unpack data based on which Dataset object you decided to use
        x, y, g = data
        ...

        # Run your loss calculation + update
        loss = ...

        self.optimizer.zero_grad()
        loss.backward()
        self.optimizer.step()

        # Return a dict containing anything you want to log
        return {'loss': loss.item()}

    @torch.no_grad()
    def evaluate(self, loader):
        # Called at validation-time and test-time
        self.model.eval()

        predicted = []
        target = []
        group = []

        cls_loss = 0

        for x, y, g in loader:
            x, y, g = x.to(DEVICE), y.to(DEVICE), g.to(DEVICE)

            _, pred = self.model(x)

            cls_loss += F.cross_entropy(pred, y).item()

            predicted.append(pred)
            target.append(y)
            group.append(g)

        predicted = torch.cat(predicted)
        target = torch.cat(target)
        group = torch.cat(group)

        self.model.train()

        # Return 3 torch.tensors to compute metrics + all validation losses to be logged
        return predicted, target, group, {'classification_loss': cls_loss / predicted.size(0)}

NOTE: additional models shall go inside the models/<task-name> folder

Citation

@article{iurada2024fairness,
  author={Iurada, Leonardo and Bucci, Silvia and Hospedales, Timothy M. and Tommasi, Tatiana},
  journal={IEEE Access}, 
  title={Fairness meets Cross-Domain Learning: a Benchmark of Models and Metrics}, 
  year={2024},
  volume={},
  number={},
  pages={1-1},
  doi={10.1109/ACCESS.2024.3383841}
}