API Reference

The domino API Reference is organized into the following sections:

• Slice Discovery Methods

  Most slice discovery methods adhere to a three-step procedure: (1) embed, (2) slice, and (3) describe [eyuboglu_2022]. For each of these steps, the domino package provides implementations of various algorithms under a common API. This makes it easy to compose a custom slice discovery method from different choices for each step.

  • Embed validation data in a representation space. This reference page describes a number of popular encoders implemented in domino.

  • Slice the representation space into underperforming regions. This reference describes the slicing algorithms implemented in domino.

  • Describe slices with natural language. This reference page describes the domino method for describing discovered slices.

• Evaluation

  domino also provides scripts for performing evaluations of slice discovery methods. These scripts are provided as part of the domino.eval sub-package.

  • Model Training

• Utilities

  • GUI - domino includes implementations of simple Jupyter Notebook interfaces to help you explore discovered slices.

eyuboglu_2022

Eyuboglu, S. et al. Domino: Discovering Systematic Errors with Cross-Modal Embeddings. in International Conference on Learning Representations (2022).