radnets

PyTorch-based spectral gamma-ray anomaly detection and identification, used in the paper "Neural Network Approaches for Mobile Spectroscopic Gamma-ray Source Detection," Bilton et al., 2021.

Overview

A gamma-ray spectrum can be treated as a d-dimensional vector x, which consists of benign background and potentially sources of interest. The goal of spectral anomaly detection is to flag spectra which deviate from ordinary background. Additionally, spectral identification classifies which type of radioactive sources are found in x, if any. This project uses pytorch for both spectral anomaly detection and identification.

Features

• Spectral anomaly detection using autoencoders (see radnets.models.autoencoders)
• Spectral identification (see radnets.models.identification)
• Both detection and identification methods can be either feedforward or recurrent. Using recurrent models allows temporal information (i.e., features across series of spectra) to be leveraged to enhance detection capabilities.
• Models are based on specifying YAML configuration files: the architecture is specified, and the the model class will automatically generate the model.

Structure

• radnets.data
  • Tools for preprocessing data and custom pytorch Datasets and Dataloaders used in training models.
• radnets.models
  • pytorch model classes for detection and identification. A distinction is made betweeen feedforward and recurrent models since they require a different treatment of data.
• radnets.training
  • Tools used in training models.
• radnets.utils
  • General utilities.

Setup

Note: this has been developed and tested using python 3.7. Install dependencies:

pip install -r requirements.txt

Install the radnets package:

pip install -e .

If you're developing, install the development dependencies:

pip install -r requirements-dev.txt