Version 0.2.0
Changes:
- Reduced memory footprint
- DOM exclusions are now handled directly in DNN I3Module
- Additional configuration checks when applying an exported DNN model
Added Functionality:
- Define event weights in loss function
- Label smoothing for fuzzy labels
- GNN comparison architecture and config
- Option to load FilterMask in general misc data loader
- File weighting for files from different datasets
- Gradient clipping and NaN replacement in gradients
- DNN LLh classes to obtain directional error contours
- Biased sampling of training events based on misc data or NN prediction
Added Models and Configs:
- NuGen L2 event selection
- Various starting event selection models
- Muon classification
- Track reconstruction at L2 (track length, direction, ..)
- HESE Spice3.2 reconstruction (used for cascade real-time alert stream)
- Muon multiplicity
- Muon scattering event selection
- Starting event selection model based on Cascade-based input data
Bug Fixes:
- Fix naming of I3Modules: allows inclusion of multiple DNN reco segments
- Throw error when pulses are not in frame
- Stabilize loss and architectures
- Fix global time offset shift
- Fix python3 compatibility
- Automatically find packages in setup.py