lir.datasets.glass module

class lir.datasets.glass.GlassData(cache_dir: PathLike)

Bases: DataProvider

LA-ICP-MS measurements of elemental concentration from floatglass.

The measurements are from reference glass from casework, collected in the past 10 years or so. For more info on the DataProvider, see: https://github.com/NetherlandsForensicInstitute/elemental_composition_glass

This data provider has a pre-defined train/test split, with a training set of three instances per source, and a test set of five instances per source.

Data are retrieved from the web as needed and stored locally for later use.

Parameters:

cache_dir (PathLike) – Cache directory used for storing downloaded dataset files.

get_instances() → FeatureData

Return data with pre-defined assignments of training data and test data.

The training data is read from training.csv and has three instances (replicates) per source. The test data is read from duplo.csv and triplo.csv and has a total of five instances per source.

The features are elemental concentrations on a log_10 basis, and normalized to Si. The elements are: K39, Ti49, Mn55, Rb85, Sr88, Zr90, Ba137, La139, Ce140, Pb208

The source_ids are unique identifiers of a glass particle. Each particle is from a different reference window. An instance is a replicate measurement on a glass particle. Source ids are prefixed with the role assignment, e.g. ‘test-123’ and ‘train-123’. The ids ‘test-123’ and ‘train-123’ refer to different glass particles (and therefore different reference windows).

The instance_ids values of an instance are a concatenation of the filename and a row number, e.g. “training.csv:22”.

The data are returned as a FeatureData object with the following properties: - features: an (n, 10) array of feature values - source_ids: a 1d array of source ids (str) - instance_ids: a 1d array of unique instance ids (str) - role_assignments: a 1d array of role assignments (values “train” or “test”)

Returns:

FeatureData object parsed from the source.

Return type:

FeatureData