Contribute / Shared definitions

Shared definitions¶

This page lists the steps that training, evaluation and prediction must do the same way. Each has one implementation. To change one, edit that implementation; never repeat it in a caller.

The definitions¶

Definition	Implementation	Callers
Which stars are used, and the split	`load_training_data()`, `split()`	`dorothy train`, `dorothy evaluate`, the Python API
Spectrum preprocessing	`zero_invalid_pixels()`, `normalize_spectra()`, `encode_precision()`	Every catalogue loader, `predict_from_fits`
Label normalization	`transform_labels()`	Training, evaluation
Predicted scatter	`predicted_scatter()`	The loss, validation, evaluation, prediction, saliency
Calibration of the predicted scatter	`calibrate_scatter()`	Evaluation, prediction
Model forward pass per label source	`forward_label_sources()`	Training, validation, evaluation

Data formats¶

Every function passes data in these shapes (Preprocessing defines them).

Array	Shape	Channels
Spectra	`(N, 3, n_wavelengths)`	flux, \(\rho = \ln(1 + 1/\sigma)\), mask
Labels	`(N, 3, n_parameters)`	values, errors, mask
Model output	`(N, 2, n_parameters)`	\(\mu\), \(\ln s\), in normalized units

A masked entry can hold any value, NaN and infinity included. Select with the mask (torch.where, boolean indexing) before arithmetic; never multiply by the mask.
The parameter names, and their number, come from the catalogue’s parameter_names attribute. Never assume 11 parameters.
Random numbers come from the configured seed: numpy.random.RandomState(seed) for the split, and seeded generators for augmentation and initialization.

The test that guards them¶

tests/test_end_to_end.py checks that evaluation reproduces the validation metrics of training for every data layout. A definition that drifts between two callers fails this test.