Metrics

Performance metrics for model evaluation and physics analysis.

Overview

The metrics module provides metrics for assessing model performance:

• accuracy: Classification accuracy metrics

• pagerank: Graph-based ranking metrics

Accuracy Metrics

The accuracy metric provides standard classification performance measures:

• Binary accuracy

• Multi-class accuracy

• Top-k accuracy

• Per-class accuracy

Usage: - Accumulates predictions across batches - Computes final accuracy statistics - Supports weighted and unweighted modes

PageRank Metrics

The pagerank metric uses the PageRank algorithm for graph analysis:

• Node importance ranking

• Graph centrality measures

• Connection strength analysis

Usage: - Evaluates graph structure quality - Ranks nodes by importance - Useful for understanding learned representations

Custom Metrics

Users can define custom metrics by:

1. Inheriting from metric_template

2. Implementing accumulation logic

3. Implementing reduction logic

4. Implementing metric computation

5. Registering with the framework

Metric Usage

Metrics are used during:

• Training: Monitor training progress

• Validation: Select best models

• Testing: Final performance evaluation

• Analysis: Understanding model behavior

Metrics can be:

• Computed per-batch

• Accumulated across batches

• Reduced across epochs

• Logged and visualized