Implementation:Snorkel team Snorkel RandomPolicy Init

Knowledge Sources	Snorkel
Domains	Data_Augmentation, Sampling
Last Updated	2026-02-14 20:00 GMT

Overview

Concrete tool for configuring augmentation policies that control TF sampling and sequencing, provided by the Snorkel library.

Description

RandomPolicy and MeanFieldPolicy control how transformation functions are sampled during augmentation. Both extend the Policy base class which provides generate() (sample a TF sequence) and generate_for_example() (sample multiple sequences per data point).

• RandomPolicy: Samples uniformly from available TFs
• MeanFieldPolicy: Samples according to user-specified probability distribution

Usage

Import these classes when configuring how your TFs should be combined. Use RandomPolicy for uniform exploration, MeanFieldPolicy for weighted selection.

Code Reference

Source Location

• Repository: snorkel
• File: snorkel/augmentation/policy/sampling.py (RandomPolicy L70-116, MeanFieldPolicy L8-67), snorkel/augmentation/policy/core.py (Policy L4-76)

Signature

class RandomPolicy(MeanFieldPolicy):
    def __init__(
        self,
        n_tfs: int,
        sequence_length: int = 1,
        n_per_original: int = 1,
        keep_original: bool = True,
    ) -> None:
        """
        Args:
            n_tfs: Number of available TFs.
            sequence_length: TFs to chain per augmentation.
            n_per_original: Augmented copies per original.
            keep_original: Retain originals in output.
        """

class MeanFieldPolicy(Policy):
    def __init__(
        self,
        n_tfs: int,
        sequence_length: int = 1,
        p: Optional[Sequence[float]] = None,
        n_per_original: int = 1,
        keep_original: bool = True,
    ) -> None:
        """
        Args:
            n_tfs: Number of available TFs.
            sequence_length: TFs to chain per augmentation.
            p: Probability distribution over TFs (must sum to 1).
            n_per_original: Augmented copies per original.
            keep_original: Retain originals in output.
        """

Import

from snorkel.augmentation import RandomPolicy, MeanFieldPolicy

I/O Contract

Inputs

Name	Type	Required	Description
n_tfs	int	Yes	Number of transformation functions
sequence_length	int	No	TFs chained per augmentation (default 1)
n_per_original	int	No	Augmented copies per original (default 1)
keep_original	bool	No	Keep originals in output (default True)
p	Optional[Sequence[float]]	No	TF probability distribution (MeanFieldPolicy only)

Outputs

Name	Type	Description
Policy instance	Policy	Configured policy for PandasTFApplier
generate()	List[int]	Sequence of TF indices to apply
generate_for_example()	List[List[int]]	Multiple TF sequences for one data point

Usage Examples

from snorkel.augmentation import RandomPolicy, MeanFieldPolicy

# Random policy: 2 augmented copies, chain 2 TFs each
policy = RandomPolicy(
    n_tfs=5,
    sequence_length=2,
    n_per_original=2,
    keep_original=True,
)

# Mean field policy: weight certain TFs more
policy = MeanFieldPolicy(
    n_tfs=3,
    p=[0.5, 0.3, 0.2],  # First TF applied 50% of the time
    n_per_original=1,
)

Related Pages

Implements Principle

• Principle:Snorkel_team_Snorkel_Augmentation_Policy_Configuration