Machine learning–based inference toolkit for particle physics
histo_theta_batchsize sets number of parameter points for which histogram data is loaded in parallel.n_observables for AsymptoticLimits.observed_limits() allows overwriting the number of events (= rescaling the log likelihood).DoubleParameterizedRatioEstimator and LikelihoodEstimator instances to crashLHEReader and DelphesReader
LHEReader.add_observable_from_function() now expects functions with signature observable(truth_particles, leptons, photons, jets, met), where truth_particles are the original particles in the LHE file, and the other arguments have smearing applied to them.dof keyword in limit functions to overwrite degrees of freedomLHEReader
sample_only_from_closest_benchmark in the SampleAugmenter functions and plot_distributions() then allows the user to restrict the unweighting / resampling at some parameter point to events from the closest benchmark point. This can significantly reduce the weights of individual events and thus reduce the variance.madminer.sampling.madminer.limits submodule that calculates expected and observed p-values.madminer.analysis submodule with a generic DataAnalyzer class that can be used to access the general setup, weighted events, and cross sections. The classes SampleAugmenter, FisherInformation and AsymptoticLimits subclass it, leading to a more unified interface.n_processes>1 for any sampling call.MLForge is replaced by four different classes ParameterizedRatioEstimator, DoubleParameterizedRatioEstimator, LikelihoodEstimator, and ScoreEstimator.EnsembleForge is now Ensemble, with renamed functions and less clutter (at the cost of some unimportant functionality)DelphesProcessor -> DelphesReader, LHEProcessor -> LHEReader
madminer.morphing now lives in madminer.utils.morphing, and the Morpher class was renamed PhysicsMorpher (since we also have a NuisanceMorpher class)madminer.sampling has changed in a number of ways. The high-level functions have been renamed: extract_samples_train_plain() -> sample_train_plain(), extract_samples_train_local() -> sample_train_local(), extract_samples_train_global() -> train_samples_density(), extract_samples_train_ratio() -> sample_train_ratio(), extract_samples_train_more_ratios() -> sample_train_more_ratios(), extract_samples_test() -> sample_test(), extract_cross_sections() -> cross_sections(). In addition to the physical parameters theta, they all now take descriptions of the nuisance parameters nu as input argument, for which new helper functions exist.madminer.sampling are also different: constant_benchmark_theta() -> benchmark(), multiple_benchmark_thetas() -> benchmarks(), constant_morphing_theta() -> morphing_point(), multiple_morphing_thetas() -> morphing_points, random_morphing_thetas() -> random_morphing_points()
madminer.ml largely rewrittenmadminer.sampling largely rewrittenmadminere.utils.analysis got removed, the old functions are now split between madminer.analysis and madminer.sampling
trainer in MLForge.train() has been renamed to optimizer
MLForge.evaluate() and MLForge.train() are goneMLForge has new methods evaluate_log_likelihood(), evaluate_log_likelihood_ratio(), and evaluate_score(); MLForge.evaluate() just wraps around them nowplot_uncertainty() functionplot_distributions()
MLForge.evaluate() and EnsembleForge.evaluate(), the keyword x_filename is replaced by x, which supports ndarrays as an alternative to filenamesLHEProcessor when visible particle pTs are smearedplot_distributions()