Skip to main content

Pipeline Overview

This page presents the complete workflow of the AstroSpectro project, from raw data acquisition to the generation of classification results. The pipeline is designed as a sequence of modular steps.

Workflow diagram​

The diagram below illustrates the main phases of the pipeline. Each phase is implemented by one or more specific modules in the source code.

Phase description​

1. Acquisition (optional)​

The starting point. SmartDownloader retrieves raw spectra from the LAMOST servers and stores them locally. It logs completed observation plans in downloaded_plans.csv to avoid redundant downloads.

2. Session preparation​

At each run, DatasetBuilder selects a new batch of unused spectra by consulting the trained_spectra.csv log. A local catalogue is then generated from the FITS headers of this batch, so the analysis relies on precise, up-to-date metadata.

3. Processing and modelling​

ProcessingPipeline takes the batch and its local catalogue and runs the processing chain: preprocessing, line detection and extraction of the feature vector. The resulting feature table goes to SpectralClassifier, which handles hyperparameter tuning and training (the reference model is XGBoost; see Model Card).

4. Finalisation​

Once the model is trained, DatasetBuilder updates trained_spectra.csv with the spectra just used. The session artefacts (the .pkl model and a detailed JSON report) are saved in data/models/ and data/reports/ for traceability and reproducibility.