dataeval¶

DataEval provides a simple interface to characterize visual data and its impact on model performance.

It works across classification and object-detection tasks. It also provides capabilities to select and curate datasets to test and train performant, robust, unbiased and reliable AI models and monitor for data shifts that impact performance of deployed models.

Submodules¶

`bias`	Check for skewed or imbalanced datasets and incomplete feature representation.
`config`	Global configuration settings for DataEval.
`core`	Core stateless functions for performing dataset, metadata and model evaluation.
`data`	Dataset organization tools: conform, filter, split, and reshape dataset views.
`exceptions`	Exception and warning classes for DataEval.
`extractors`	Feature extractors that transform input data into arrays.
`flags`	Module for flag enums that control function behavior.
`models`	MAITE Model implementations for opinionated ONNX/LiteRT prediction.
`performance`	Determine whether a problem is feasible and how much data is needed.
`protocols`	Common type protocols used for interoperability with DataEval.
`quality`	Identify potential issues in training and test data.
`scope`	Evaluate data completeness and coverage of a dataset’s label and embedding space.
`shift`	Detect changes in data between different datasets.
`types`	Data types used in DataEval.
`utils`	DataEval utilities organized by domain.

Classes¶

`Embeddings`	Collection of image embeddings from a dataset.
`Metadata`	Collection of binned metadata using Polars DataFrames.
`Ontology`	An immutable, in-memory directed acyclic graph of `OntologyConcept`.

Functions¶

`log`	Add a handler to the logger quickly for debugging.