Metric Template

metric_template is the base class for all user-defined training metrics. It inherits from both tools and notification, and provides helpers for computing per-event and per-batch statistics during model training, validation, and evaluation. It also exposes ROOT tree output helpers (register_output, write) and the same particle combination utilities as selection_template.

Struct: `metric_t`

Header: <templates/metric_template.h>

metric_t is constructed by the framework and passed to metric_template::define_metric for each graph in the batch.

Field	Type	Description
`kfold`	`int`	Current k-fold index.
`epoch`	`int`	Current training epoch.
`device`	`int`	CUDA device index (0 = first GPU, -1 = CPU).

Public Method

Signature	Description
`template<g> g get(graph_enum grx, std::string name)`	Retrieves the named graph/node/edge variable of type g from the variable store for this event. Prints an error and returns default if the variable is not found.
`std::string mode()`	Returns the training mode string (`"train"`, `"valid"`, `"eval"`).
`std::string* get_filename(long unsigned int idx)`	Returns a pointer to the filename of graph at batch position idx.

Class: `metric_template`

Header: <templates/metric_template.h>

Inheritance: tools, notification

Properties

Property	Type	Description
`name`	`std::string`	Metric class name. Settable/Gettable.
`output_path`	`std::string`	Directory for ROOT output files. Gettable.
`variables`	`std::vector<std::string>`	List of variable names requested from `graph_t` (formatted as `"graph_enum/feature_name"`, e.g. `"data_node/pt"`).
`run_names`	`std::map<std::string, std::string>`	Maps model-name → run-label used for output file organisation.

Virtual Methods (Override in Subclass)

Signature	Description
`virtual metric_template* clone()`	Returns a heap-allocated copy of the metric.
`virtual void define_variables()`	Called once before training. Use `variables` to request which graph/node/edge features to expose in `metric_t::get`.
`virtual void define_metric(metric_t* v)`	Called once per graph during training/validation/evaluation. Use `v->get<T>(graph_enum, "name")` to read features and compute derived quantities.
`virtual void event()`	Called after `define_metric` for each event.
`virtual void batch()`	Called after a full batch is processed.
`virtual void end()`	Called at the end of an epoch.

Output Methods

Signature	Description
`template<T> void register_output(std::string tree, std::string name, T* t)`	Creates an output ROOT branch `tree/name` backed by t. Must be called in `define_variables` (once per run).
`template<T> void write(std::string tree, std::string name, T* t, bool fill = false)`	Fills the branch `tree/name` with current value of t. Pass `fill=true` to also flush the tree entry.

Particle Utilities (same as `selection_template`)

metric_template exposes the same sum, safe_delete, make_unique, vectorize helpers as selection_template; see that page for documentation.

Additionally:

Signature	Description
`std::vector<particle_template*> make_particle(...)`	Constructs a vector of `particle_template` objects from four `std::vector<std::vector<double>>` arrays of (pt, eta, phi, E) in cylindrical coordinates (MeV).