Implementation:Snorkel team Snorkel SliceAwareClassifier Make Slice Dataloader

Knowledge Sources	Snorkel
Domains	Data_Slicing, Data_Preparation
Last Updated	2026-02-14 20:00 GMT

Overview

Concrete tool for creating slice-aware dataloaders with indicator and prediction labels from a dataset and slice matrix, provided by the Snorkel library.

Description

The SliceAwareClassifier.make_slice_dataloader() method combines a DictDataset with a slice matrix (from PandasSFApplier) to produce a DictDataLoader augmented with slice-specific labels. Internally, it calls add_slice_labels() which modifies the datasets Y_dict in-place, adding two label tensors per slice:

• task_slice:{name}_ind - indicator labels
• task_slice:{name}_pred - masked prediction labels

Usage

Use this method after initializing a SliceAwareClassifier and obtaining slice matrices. Create separate dataloaders for train/valid/test splits.

Code Reference

Source Location

• Repository: snorkel
• File: snorkel/slicing/sliceaware_classifier.py (make_slice_dataloader L93-125), snorkel/slicing/utils.py (add_slice_labels L15-55)

Signature

class SliceAwareClassifier(MultitaskClassifier):
    def make_slice_dataloader(
        self,
        dataset: DictDataset,
        S: np.recarray,
        **dataloader_kwargs: Any,
    ) -> DictDataLoader:
        """
        Create DictDataLoader with slice labels.

        Args:
            dataset: Base DictDataset with features and labels.
            S: [n_examples, n_slices] slice recarray from PandasSFApplier.
            **dataloader_kwargs: Passed to DictDataLoader (batch_size, shuffle, etc.).
        Returns:
            DictDataLoader with slice-specific labels added to Y_dict.
        """

def add_slice_labels(
    dataloader: DictDataLoader,
    base_task: Task,
    S: np.recarray,
) -> None:
    """
    Modify dataloader in-place, adding indicator and prediction labels per slice.
    """

Import

from snorkel.slicing import SliceAwareClassifier
from snorkel.slicing.utils import add_slice_labels

I/O Contract

Inputs

Name	Type	Required	Description
dataset	DictDataset	Yes	Base dataset with X_dict and Y_dict
S	np.recarray	Yes	Slice matrix from PandasSFApplier
**dataloader_kwargs	Any	No	batch_size, shuffle, num_workers, etc.

Outputs

Name	Type	Description
dataloader	DictDataLoader	Augmented dataloader with slice indicator/prediction labels in Y_dict

Usage Examples

import torch
from snorkel.classification import DictDataset

# Create base dataset
X_dict = {"input_data": torch.randn(100, 50)}
Y_dict = {"task": torch.randint(0, 2, (100,))}
dataset = DictDataset(name="train_data", split="train", X_dict=X_dict, Y_dict=Y_dict)

# S from PandasSFApplier
# model is a SliceAwareClassifier instance

train_dl = model.make_slice_dataloader(
    dataset=dataset,
    S=S_train,
    batch_size=32,
    shuffle=True,
)

# The dataloader now contains labels like:
# "task_slice:sf_short_ind", "task_slice:sf_short_pred", etc.

Related Pages

Implements Principle

• Principle:Snorkel_team_Snorkel_Slice_Aware_Data_Preparation