Principle:Snorkel team Snorkel Augmentation Policy Configuration

Knowledge Sources	AutoAugment: Learning Augmentation Strategies from Data Population Based Augmentation
Domains	Data_Augmentation, Sampling
Last Updated	2026-02-14 20:00 GMT

Overview

A strategy for configuring how transformation functions are sampled and sequenced to generate augmented data points.

Description

Augmentation Policy Configuration determines how transformation functions are combined and applied during data augmentation. A policy controls:

• Selection: Which TFs to apply to each data point
• Sequencing: How many TFs to chain in sequence
• Repetition: How many augmented copies to generate per original
• Retention: Whether to keep original data points in the output

Two main policy types are supported:

• RandomPolicy: Uniformly samples TF sequences at random
• MeanFieldPolicy: Samples TFs with user-specified probabilities (allowing weighted selection)

Usage

Use this principle after defining transformation functions and before applying them. Choose RandomPolicy for uniform exploration of augmentations, or MeanFieldPolicy when some TFs should be applied more frequently than others.

Theoretical Basis

A policy $\pi$ defines a distribution over TF sequences:

$\pi :\to \text{Seq}(\{1,\dots ,m_{\text{tf}}\},L)$

where $L$ is the sequence length. For MeanFieldPolicy with parameters $p=(p_1,\dots ,p_{m_{\text{tf}}}):$

$P({\text{select TF}}_j)=p_j,\sum _j{p}_j=1$

For RandomPolicy: $p_j=1/m_{\text{tf}}$ for all $j$.

Related Pages

Implemented By

• Implementation:Snorkel_team_Snorkel_RandomPolicy_Init