Adding Customized Datasets

Step 1. Implement the dataset class

A custom dataset class must implement the load_data function. Take a look at this sample implementation; the split is stored in a directory self.data_dir, which is determined by the args.data_dir and split. args.data_dir is either loaded from the arguments --data_dir or from the default value, which will be introduced into details later. While split has three possible string values, {"train", "dev", "test"}, indicating the split that will be loaded.

Then the load_data function must assign the value of self.X as inputs, self.y as target labels, and self.protected_label as the information for debiasing, such as gender, age, and race.

from fairlib.dataloaders.utils import BaseDataset
from pathlib import Path

class SampleDataset(BaseDataset):

    def load_data(self):
        # Path to the data split.

        self.filename = "{}_df.pkl".format(self.split)
        data = pd.read_pickle(Path(self.args.data_dir) / self.filename)

        # Assuming the size of the split is N
        self.X = data["X"] # N * input_size
        self.y = data["y"] # N, numerical labels, for example, {0,1}
        self.protected_label = data["protected_label"] # N, numerical labels

For more example, please take a look at fairlib/src/dataloader/loaders.py.

Debiasing methods such as instance reweighting will be automatically applied to the loaded dataset, which is implemented in fairlib.src.dataloaders.utils.BaseDataset.

Step 2. Register the dataset

• default_dataset_roots

To avoid identifying the data root every time, we could predefine the default root the our dataset. we just need to add the path to the dictionary default_dataset_roots in fairlib/src/dataloaders/__init__.py. Take a look at the following example.

    default_dataset_roots = dict(
        Moji='/data/deepmoji/split2/',
        # New dataset
        New_dataset='./data/path_to_dataset',
    )

• get_dataloaders New, we need to register the dataset class for later use. Specially, all we need to do is adding an extra else if statement for the new dataset.

def get_dataloaders(args):
    # Add the new dataset name here to avoid assert
    assert args.dataset in ["Moji", "Bios_gender", "Bios_economy", "Bios_both", "New_dataset"], "Not implemented"

    if args.dataset == "Moji":
        task_dataloader = DeepMojiDataset
    elif args.dataset in ["Bios_gender", "Bios_economy", "Bios_both"]:
        task_dataloader = BiosDataset
        args.protected_task = args.dataset.split("_")[1]
    # Add an additional else if statement for the new dataset
    elif args.dataset == "New_dataset":
        task_dataloader = SampleDataset # The SampleDataset class is the introduced in the previous section.
    else:
        pass