PyTorch dataset wrappers for the several popular datasets from fair machine learning research.
The following datasets are wrapped:
- Adult (Census Income).
- Default
- Law School (data from here)
- SouthGerman
pip install fairness-datasetsfrom fairnessdatasets import Adult
# load (if necessary, download) the Adult training dataset
train_set = Adult(root="datasets", download=True)
# load the test set
test_set = Adult(root="datasets", train=False, download=True)
inputs, target = train_set[0] # retrieve the first sample of the training set
# iterate over the training set
for inputs, target in iter(train_set):
... # Do something with a single sample
# use a PyTorch data loader
from torch.utils.data import DataLoader
loader = DataLoader(test_set, batch_size=32, shuffle=True)
for epoch in range(100):
for inputs, targets in iter(loader):
... # Do something with a batch of samplesYou can use Adult(..., raw=True) to turn off the one-hot encoding
and z-score normalization applied by the Adult class by default.
The remaining dataset classes can be used in the same way as Adult.
However, these datasets don't come with a fixed train/test split,
so that the dataset instances always contain all data.
To create a train/test split, use
from fairnessdatasets import Default
from torch.utils.data import random_split
dataset = Default(root="datasets", download=True)
rng = torch.Generator().manual_seed(42) # for reproducible results
train_set, test_set = random_split(dataset, [0.7, 0.3], generator=rng)Turn off status messages while downloading the dataset:
Adult(root=..., output_fn=None)Use the logging module for logging status messages while downloading the
dataset instead of placing the status messages on sys.stdout.
import logging
Adult(root=..., output_fn=logging.info)
