Accuracy Metric (metrics.accuracy)

Import with:

from AnalysisG.metrics.accuracy import AccuracyMetric

AccuracyMetric

AccuracyMetric is a MetricTemplate subclass that computes top-multiplicity classification accuracy over training, validation, and evaluation modes.

For each epoch and k-fold it reads three ROOT trees: event_accuracy_training, event_accuracy_validation, and event_accuracy_evaluation, each with leaves ntop_truth, ntop_scores, and edge.

Results

After Postprocessing():

• ROC curves for each top-multiplicity class are saved to ./figures/epoch-{N}/{model_name}/ntops.{ext}.

• Summary AUC-vs-epoch line plots are saved to ./figures/summary/ntop-{class}.{ext}.

• ./figures/summary/roc.txt contains the AUC table.

Attributes available after Postprocessing():

Attribute	Description
auc	dict[model_name][epoch] — AUC results from the ROC curve objects for each model at each epoch.

``root_leaves`` keys (set in __cinit__):

{
    "event_accuracy_training":   ["ntop_truth", "ntop_scores", "edge"],
    "event_accuracy_validation": ["ntop_truth", "ntop_scores", "edge"],
    "event_accuracy_evaluation": ["ntop_truth", "ntop_scores", "edge"],
}

The corresponding root_fx callbacks populate the internal collector C++ object which aggregates truth labels and score vectors.