Overview

The WILDS package is an open-source Python package that provides a simple, standardized interface for all datasets in the benchmark. It contains:

1. Data loaders that automatically handle data downloading, processing, and splitting, and
2. Dataset evaluators that standardize model evaluation for each dataset.

In addition, the example scripts contain default models, optimizers, schedulers, and training/evaluation code. New algorithms can be easily added and run on all of the WILDS datasets.

Installation

We recommend using pip to install WILDS:

pip install wilds

If you have already installed it, please check that you have the latest version:

python -c "import wilds; print(wilds.__version__)"
# This should print "2.0.0". If it doesn't, update by running:
pip install -U wilds

If you plan to edit or contribute to WILDS, you should install from source:

git clone git@github.com:p-lambda/wilds.git
cd wilds
pip install -e .

In examples/, we provide a set of scripts that can be used to train models on the WILDS datasets. These scripts were also used to benchmark baselines in our papers [1, 2]. These scripts are not part of the installed WILDS package. To use them, you should install from source, as described above.

Requirements

The WILDS package depends on the following requirements:

• numpy>=1.19.1
• ogb>=1.2.6
• outdated>=0.2.0
• pandas>=1.1.0
• pillow>=7.2.0
• pytz>=2020.4
• torch>=1.7.0
• torch-scatter>=2.0.5
• torch-geometric>=2.0.1
• torchvision>=0.8.2
• tqdm>=4.53.0
• scikit-learn>=0.20.0
• scipy>=1.5.4

Running pip install wilds or pip install -e . will automatically check for and install all of these requirements except for the torch-scatter and torch-geometric packages, which require a quick manual install.

Example script requirements

To run the example scripts, you will also need to install these additional dependencies:

• transformers>=3.5.0
• SwAV requires Apex. To install Apex, please follow the README from the official SwAV repository.
• Our code supports the optional use of Weights & Biases to track and monitor experiments. To install the Weights and Biases Python package, run pip install wandb.

All baseline experiments in the paper were run on Python 3.8.5 and CUDA 10.1.

Datasets

WILDS currently includes 10 datasets, which we’ve briefly listed below. For full dataset descriptions, please see our papers (1, 2).

Using the WILDS package

Data

The WILDS package provides a simple, standardized interface for all datasets in the benchmark. This short Python snippet covers all of the steps of getting started with a WILDS dataset, including dataset download and initialization, accessing various splits, and preparing a user-customizable data loader. We discuss data loading in more detail in #Data loading.

from wilds import get_dataset
from wilds.common.data_loaders import get_train_loader
import torchvision.transforms as transforms

# Load the full dataset, and download it if necessary
dataset = get_dataset(dataset="iwildcam", download=True)

# Get the training set
train_data = dataset.get_subset(
    "train",
    transform=transforms.Compose(
        [transforms.Resize((448, 448)), transforms.ToTensor()]
    ),
)

# Prepare the standard data loader
train_loader = get_train_loader("standard", train_data, batch_size=16)

# (Optional) Load unlabeled data
dataset = get_dataset(dataset="iwildcam", download=True, unlabeled=True)
unlabeled_data = dataset.get_subset(
    "test_unlabeled",
    transform=transforms.Compose(
        [transforms.Resize((448, 448)), transforms.ToTensor()]
    ),
)
unlabeled_loader = get_train_loader("standard", unlabeled_data, batch_size=16)

# Train loop
for labeled_batch, unlabeled_batch in zip(train_loader, unlabeled_loader):
    x, y, metadata = labeled_batch
    unlabeled_x, unlabeled_metadata = unlabeled_batch
    ...

The metadata contains information like the domain identity, e.g., which camera a photo was taken from, or which hospital the patient’s data came from, etc., as well as other metadata.

Domain information

To allow algorithms to leverage domain annotations as well as other groupings over the available metadata, the WILDS package provides Grouper objects. These Grouper objects are helper objects that extract group annotations from metadata, allowing users to specify the grouping scheme in a flexible fashion. They are used to initialize group-aware data loaders (as discussed in #Data loading) and to implement algorithms that rely on domain annotations (e.g., Group DRO). In the following code snippet, we initialize and use a Grouper that extracts the domain annotations on the iWildCam dataset, where the domain is location.

from wilds.common.grouper import CombinatorialGrouper

# Initialize grouper, which extracts domain information
# In this example, we form domains based on location
grouper = CombinatorialGrouper(dataset, ['location'])

# Train loop
for x, y_true, metadata in train_loader:
    z = grouper.metadata_to_group(metadata)
    ...

Data loading

For training, the WILDS package provides two types of data loaders. The standard data loader shuffles examples in the training set, and is used for the standard approach of empirical risk minimization (ERM), where we minimize the average loss.

from wilds.common.data_loaders import get_train_loader

# Prepare the standard data loader
train_loader = get_train_loader('standard', train_data, batch_size=16)

To support other algorithms that rely on specific data loading schemes, we also provide the group data loader. In each minibatch, the group loader first samples a specified number of groups, and then samples a fixed number of examples from each of those groups. (By default, the groups are sampled uniformly at random, which upweights minority groups as a result. This can be toggled by the uniform_over_groups parameter.) We initialize group loaders as follows, using Grouper that specifies the grouping scheme.

# Prepare a group data loader that samples from user-specified groups
train_loader = get_train_loader(
    "group", train_data, grouper=grouper, n_groups_per_batch=2, batch_size=16
)

Lastly, we also provide a data loader for evaluation, which loads examples without shuffling (unlike the training loaders).

from wilds.common.data_loaders import get_eval_loader

# Get the test set
test_data = dataset.get_subset(
    "test",
    transform=transforms.Compose(
        [transforms.Resize((224, 224)), transforms.ToTensor()]
    ),
)

# Prepare the evaluation data loader
test_loader = get_eval_loader("standard", test_data, batch_size=16)

Evaluators

The WILDS package standardizes and automates evaluation for each dataset. Invoking the eval method of each dataset yields all metrics reported in the paper and on the leaderboard.

from wilds.common.data_loaders import get_eval_loader

# Get the test set
test_data = dataset.get_subset(
    "test",
    transform=transforms.Compose(
        [transforms.Resize((224, 224)), transforms.ToTensor()]
    ),
)

# Prepare the data loader
test_loader = get_eval_loader("standard", test_data, batch_size=16)

# Get predictions for the full test set
for x, y_true, metadata in test_loader:
    y_pred = model(x)
    # Accumulate y_true, y_pred, metadata

# Evaluate
dataset.eval(all_y_pred, all_y_true, all_metadata)
# {'recall_macro_all': 0.66, ...}

Most eval methods take in predicted labels for all_y_pred by default, but the default inputs vary across datasets and are documented in the eval docstrings of the corresponding dataset class.

Using the example scripts

In examples/, we provide a set of scripts that can be used to train models on the WILDS datasets.

python examples/run_expt.py --dataset iwildcam --algorithm ERM --root_dir data
python examples/run_expt.py --dataset civilcomments --algorithm groupDRO --root_dir data
python examples/run_expt.py --dataset fmow --algorithm DANN --unlabeled_split test_unlabeled --root_dir data

The scripts are configured to use the default models and reasonable hyperparameters. For exact hyperparameter settings used in our papers, please see our CodaLab executable paper.

Downloading and training on the WILDS datasets

The first time you run these scripts, you might need to download the datasets. You can do so with the --download argument, for example:

# downloads (labeled) dataset
python examples/run_expt.py --dataset globalwheat --algorithm groupDRO --root_dir data --download

# additionally downloads all unlabeled data
python examples/run_expt.py --dataset globalwheat --algorithm groupDRO --root_dir data --download  --unlabeled_split [...]

Note that downloading the large amount of unlabeled data is optional; unlabeled data will only be downloaded if some --unlabeled_split is set. (It does not matter which --unlabeled_split is set; all unlabeled data will be downloaded together.)

Alternatively, you can use the standalone wilds/download_datasets.py script to download the datasets, for example:

# downloads (labeled) data
python wilds/download_datasets.py --root_dir data

# downloads (unlabeled) data
python wilds/download_datasets.py --root_dir data --unlabeled

This will download all datasets to the specified data folder. You can also use the --datasets argument to download particular datasets.

These are the sizes of each of our datasets, as well as their approximate time taken to train and evaluate the default model for a single ERM run using a NVIDIA V100 GPU.

The following are the sizes of the unlabeled data bundles:

^{* These unlabeled datasets are downloaded simultaneously with the labeled data and do not need to be downloaded separately.}

While the camelyon17 dataset is small and fast to train on, we advise against using it as the only dataset to prototype methods on, as the test performance of models trained on this dataset tend to exhibit a large degree of variability over random seeds.

The image datasets (iwildcam, camelyon17, rxrx1, globalwheat, fmow, and poverty) tend to have high disk I/O usage. If training time is much slower for you than the approximate times listed above, consider checking if I/O is a bottleneck (e.g., by moving to a local disk if you are using a network drive, or by increasing the number of data loader workers). To speed up training, you could also disable evaluation at each epoch or for all splits by toggling --evaluate_all_splits and related arguments.

Algorithms

In the examples/algorithms folder, we provide implementations of the adaptation algorithms benchmarked in our papers (1, 2). All algorithms train on labeled data from a WILDS dataset’s train split. Some algorithms are designed to also leverage unlabeled data. To load unlabeled data, specify an --unlabeled_split when running.

In addition to shared hyperparameters such as lr, weight_decay, batch_size, and unlabeled_batch_size, the scripts also take in command line arguments for algorithm-specific hyperparameters.

The repository is set up to facilitate general-purpose algorithm development: new algorithms can be added to examples/algorithms and then run on all of the WILDS datasets using the default models.

Evaluating trained models

We also provide an evaluation script that aggregates prediction CSV files for different replicates and reports on their combined evaluation. To use this, run:

python examples/evaluate.py <predictions_dir> <output_dir> --root-dir <root_dir>

where <predictions_dir> is the path to your predictions directory, <output_dir> is where the results JSON will be writte, and <root_dir> is the dataset root directory. The predictions directory should have a subdirectory for each dataset (e.g. iwildcam) containing prediction CSV files to evaluate; see our submission guidelines for the format. The evaluation script will skip over any datasets that has missing prediction files. Any dataset not in <root_dir> will be downloaded to <root_dir>.

Reproducibility

We have an executable version of our paper on CodaLab that contains the exact commands, code, and data for the experiments reported in our paper, which rely on these scripts. Trained model weights for all datasets can also be found there. All configurations and hyperparameters can also be found in the examples/configs folder of this repo, and dataset-specific parameters are in examples/configs/datasets.py.

Citing WILDS (Bibtex)

  
  @inproceedings{wilds2021,
    title = {{WILDS}: A Benchmark of in-the-Wild Distribution Shifts},
    author = {Pang Wei Koh and Shiori Sagawa and Henrik Marklund and Sang Michael Xie and Marvin Zhang and Akshay Balsubramani and Weihua Hu and Michihiro Yasunaga and Richard Lanas Phillips and Irena Gao and Tony Lee and Etienne David and Ian Stavness and Wei Guo and Berton A. Earnshaw and Imran S. Haque and Sara Beery and Jure Leskovec and Anshul Kundaje and Emma Pierson and Sergey Levine and Chelsea Finn and Percy Liang},
    booktitle = {International Conference on Machine Learning (ICML)},
    year = {2021}
  }
  @inproceedings{sagawa2022extending,
    title = {Extending the WILDS Benchmark for Unsupervised Adaptation},
    author = {Shiori Sagawa and Pang Wei Koh and Tony Lee and Irena Gao and Sang Michael Xie and Kendrick Shen and Ananya Kumar and Weihua Hu and Michihiro Yasunaga and Henrik Marklund and Sara Beery and Etienne David and Ian Stavness and Wei Guo and Jure Leskovec and Kate Saenko and Tatsunori Hashimoto and Sergey Levine and Chelsea Finn and Percy Liang},
    booktitle = {International Conference on Learning Representations (ICLR},
    year = {2022}
  }
  

If you use WILDS datasets in your work, please cite our paper:

If you use unlabeled data from the WILDS datasets, please also cite:

Please also cite the original papers that introduced the datasets, as listed on the datasets page.