Eventdisplay-ML - a toolkit to interface and run machine learning methods together with the Eventdisplay software package for gamma-ray astronomy data analysis.
Stereo (direction and energy) reconstruction tested and validated on both VERITAS and CTAO simulations plus VERITAS data.
Compatible with:
Changelog
New Features
-
Add
--max_tel_per_type 10argument to restrict the number of telescope parameters per telescope type.
Fix bug in indexing arrays with non-continuous telescope identifiers. (#49) -
Improve stereo reconstruction by adding the geometrical feature img2_ang.
Change clipping min for size to '1' (applicable for small images in SSTs).
Add preview_rows as command line parameter to allow flexible printout for debugging. (#51) -
Algorithm improvements
- Switch to residual learning (predict corrections to baseline reconstructions)
- Add target standardization for balanced multi-target training
- Introduce energy-bin weighting with low-statistics suppression
- Refine XGBoost training (regularization, early stopping, updated hyperparameters)
-
New features
- Training diagnostics with cached metrics (generalization gap, residual normality)
- SHAP feature importance caching per target
- Diagnostic scripts and CLI tools for evaluation and interpretability
- Reproducible diagnostics via model metadata reconstruction
- Expanded test suite and improved error handling
(#53)
Maintenance
- Update g/h separation to new sorting scheme of telescope-dependent variables. (#45)
- Add early stopping to classification. Increase number of estimators. (#48)
- Add detailed copilot instructions. (#50)
Bugfixes
- Correct log10 handling for energy residuals
- Fix scaler loading/inversion in apply pipeline
- Fix energy-bin weighting logic
- Ensure safe energy validation (ErecS) without dropping rows
- Align evaluation metrics with residual formulation
- Resolve pandas/sklearn warnings and compatibility issues