v0.2.0

New features:

• Nuisance parameters to model systematic uncertainties, currently only from PDF / scale reweighting and for SALLY / SALLINO.
• New "score" mode for ensemble Fisher information calculation.
• No need for the Pythia / Delphes patch anymore: DelphesProcessor can extract the event weights from the LHE file.
• Additional (non-morphing) benchmarks can be defined in MadMiner without breaking or resetting the morphing.
• DelphesProcessor allows k factors.
• New function plot_distributions() to plot distributions of observables and systematic error bands.
• Option to save full pyTorch models.
• Option to limit the training sample size.
• Dynamic binning for 2D histograms.

Breaking / API changes:

• DelphesProcessor.add_sample() replaces DelphesProcessor.add_hepmc_sample().
• MadMiner.set_morphing() replaces MadMiner.set_benchmarks_from_morphing().

Documentation and examples:

• New toy example.
• Updated docs.

Bug fixes:

• Fixed critical bug in the evaluation of the score with ensemble methods, which led to wrong results for the detector-level Fisher information.
• Fixed bug in the calculation of the covariance of truth-type Fisher information matrices.
• Fixed gradient clipping.
• Many small bug fixes.

Internal changes:

• MadMiner file format changed to include systematics / nuisance parameter data.
• Significant speed-up of training for CARL, ROLR, and ALICE estimators.
• No unnecessary running of Delphes in DelphesProcessor.
• Restructured logging, now the user is responsible for setting up handlers.
• Renamed some submodules in madminer.utils.interfaces.
• Removed unnecessary static functions in MadMiner.
• Rewrote profiling algorithm.