Principle:Snorkel team Snorkel Transformation Application

Knowledge Sources	A Survey on Data Augmentation for Text Classification
Domains	Data_Augmentation, Data_Pipeline
Last Updated	2026-02-14 20:00 GMT

Overview

A process for applying transformation functions to a dataset according to a configured policy to produce an augmented dataset.

Description

Transformation Application executes transformation functions on each data point following the augmentation policy. For each original data point, the policy generates one or more TF sequences. Each sequence is applied to a copy of the data point, and successful transformations produce new augmented examples.

The process handles:

• Copying data points before transformation (to avoid mutation)
• Chaining multiple TFs in sequence
• Skipping transformations that return None
• Optionally retaining original data points
• Memory-efficient generator mode for large datasets

Usage

Use this principle after defining TFs and configuring an augmentation policy. Apply transformations to generate augmented training data.

Theoretical Basis

Given dataset ${𝒟}$, TFs $\{{\tau }_1,\dots ,{\tau }_m\}$, and policy $\pi$:

For each $x_i\in {𝒟}$:

1. Sample $k$ sequences from $\pi$
2. For each sequence $(j_1,\dots ,j_L)$:
   1. Apply $x^{\prime }={\tau }_{j_L}(\dots {\tau }_{j_1}(x_i))$
   2. If $x^{\prime }\ne \text{None}$, add $(x^{\prime },y_i)$ to augmented set

Related Pages

Implemented By

• Implementation:Snorkel_team_Snorkel_PandasTFApplier_Apply