Metrics Module

The metrics class is the framework’s training-analytics engine. It registers one analytics_t object per model/kfold pair and captures loss, accuracy, and invariant-mass histograms during each epoch, then dumps ROOT/PDF plots to the output directory.

Struct: `analytics_t`

Header: <metrics/metrics.h>

analytics_t holds all ROOT histograms for one model/kfold pair.

Field	Type	Description
`model`	`model_template*`	Pointer to the associated model (not owned).
`report`	`model_report*`	Pointer to the model-report struct (owned; freed by `purge()`).
`this_epoch`	`int`	Current epoch counter for this model/kfold.
`loss_graph`	`std::map<mode_enum, std::map<std::string, TH1F*>>`	Per-mode (train/valid/eval) loss histograms for graph-level features.
`loss_node`	same type	Node-level loss histograms.
`loss_edge`	same type	Edge-level loss histograms.
`accuracy_graph`	same type	Graph-level accuracy histograms.
`accuracy_node`	same type	Node-level accuracy histograms.
`accuracy_edge`	same type	Edge-level accuracy histograms.
`pred_mass_edge`	same type	Predicted invariant-mass histograms for edge features.
`truth_mass_edge`	same type	Truth invariant-mass histograms for edge features.

Method	Description
`void purge()`	Deletes all histogram pointers and the `report` pointer.
`void destroy(std::map<mode_enum, std::map<std::string, TH1F>> data)`	Iterates and deletes all `TH1F` pointers in data.

Class: `metrics`

Header: <metrics/metrics.h>

Inheritance: tools, notification

Public Fields

Field	Type	Description
`output_path`	`std::string`	Directory to which plots are written.
`m_settings`	`settings_t`	Framework settings struct (kfolds, batch-size, etc.).
`colors_h`	`std::vector<Color_t>`	ROOT colour list for multi-kfold overlays (red, green, blue, cyan, violet, orange, coffee, aurora).

Public Methods

Signature	Description
`model_report* register_model(model_template* model, int kfold)`	Creates and stores an `analytics_t` entry for model at fold kfold. Returns the freshly allocated `model_report`.
`void capture(mode_enum mode, int kfold, int epoch, int smpl_len)`	Captures loss, accuracy, and mass histograms from the current model outputs for mode (train/valid/eval) at the given kfold/epoch. smpl_len is the number of events in the batch (used for normalisation).
`void dump_plots(int k)`	Writes all loss + accuracy + mass plots for fold k to `output_path`.
`void dump_loss_plots(int k)`	Writes only loss plots for fold k.
`void dump_accuracy_plots(int k)`	Writes only accuracy plots for fold k.
`void dump_mass_plots(int k)`	Writes only invariant-mass plots for fold k.