Model Module

The spine.model module contains the deep learning architectures used by SPINE for semantic segmentation, clustering, endpoint finding, graph construction, and end-to-end reconstruction tasks.

Machine learning models for neutrino physics reconstruction.

This module handles the construction, training, and execution of deep learning models for liquid argon time projection chamber (LArTPC) data analysis.

Model management:

ModelManager coordinates model instantiation, training, evaluation, and checkpoints.

Supported model families:

UResNet for semantic segmentation.
PPN for endpoint proposals.
SPICE for point-cloud instance clustering.
GrapPA and related graph models for relational reconstruction.
End-to-end chain models for full reconstruction workflows.

Key features:

Modular configuration-driven model assembly.
Support for sparse and dense convolutions.
Graph neural network components.
Mixed precision and distributed training support.

Example

from spine.model import ModelManager

manager = ModelManager(config)
manager.train(train_loader, val_loader)
results = manager.forward(data_batch)

The module integrates with the broader SPINE ecosystem for data I/O, visualization, and post-processing workflows.

Module Index

The model package is built around configuration-driven instantiation through spine.model.ModelManager. It includes convolutional, graph-based, and hybrid architectures tailored to sparse detector reconstruction rather than generic ML utilities.

`manager`	Centralize all methods associated with a machine-learning model.
`factories`
`uresnet`	UResNet segmentation model and its loss.
`uresnet_ppn`	Module that defines a model and a loss to jointly train the semantic segmentation task and the point proposal task.
`spice`
`full_chain`	Module with the core full reconstruction chain.
`image`	Whole-image classification/regression tasks.
`singlep`	Image classification module.
`vertex`
`grappa`	GrapPA: Graph Neural Network for Particle Aggregation.
`graph_spice`	Supervi dense clustering model and its loss.
`bayes_uresnet`
`layer`	Contains reusable layers for various models.