Reference / API reference / dorothy

dorothy

DOROTHY: neural networks that infer stellar parameters and abundances from spectra.

The package trains networks on APOGEE and GALAH labels to predict stellar parameters, abundances and their uncertainties from DESI, BOSS and LAMOST (LRS and MRS) spectra.

Main components:

• config: Configuration schema and loading utilities

• data: Data loading, normalization, and preprocessing

• models: Neural network architectures

• losses: Loss functions including heteroscedastic loss

• training: Training loop

• inference: Prediction and evaluation

• analysis: Saliency and k-NN anomaly detection for trained models, and label-source comparison

• visualization: Plots of training and evaluation results

• cli: Command-line interface

Modules

analysis	Analysis tools: saliency, anomaly detection and label-source comparison.
cli	Command-line interface for DOROTHY.
config	Configuration management for DOROTHY experiments.
data	Data loading, normalization, and preprocessing utilities.
inference	Prediction and evaluation utilities.
losses	Loss functions for training stellar parameter models.
models	Neural network architectures for stellar parameter inference.
training	Training infrastructure for DOROTHY models.
utils	Small shared utilities.
visualization	Visualization utilities for DOROTHY training and evaluation.