Models

Machine learning model implementations for physics analysis.

Overview

The models module contains neural network architectures:

  • RecursiveGraphNeuralNetwork: Recursive message-passing GNN

  • grift: Graph Inference with Feature Transformation

Model Architecture

Recursive Graph Neural Network

A recursive message-passing architecture designed for physics graphs:

  • Multiple message-passing rounds

  • Node and edge updates

  • Graph-level readout

  • Customizable aggregation

GRIFT Model

Graph Inference with Feature Transformation model for learning on physics graphs.

Model Integration

Models in AnalysisG:

  1. Inherit from model_template

  2. Define network architecture in C++/Python

  3. Implement forward pass

  4. Specify loss functions and metrics

  5. Configure optimizers

Training Workflow

The model template handles:

  • Data Loading: Batching and shuffling

  • Training Loop: Forward/backward passes

  • Validation: Periodic evaluation

  • Checkpointing: Saving best models

  • Logging: Metrics and losses

  • Early Stopping: Based on validation performance

Models can be trained using:

  • Standard supervised learning

  • Semi-supervised learning

  • Self-supervised learning

  • Transfer learning

Inference

Trained models can be used for:

  • Batch inference on new data

  • Real-time event classification

  • Feature extraction

  • Uncertainty quantification