The CosmicTagger project in high-energy particle physics domain deals with detecting neutrino interactions in a detector overwhelmed by cosmic particles. The goal is to classify each pixel to separate cosmic pixels, background pixels, and neutrino pixels in a neutrinos dataset. The technique uses multiple 2D projections of the same image and the raw data is 3 images per event. The training model is a UResNet architecture for multi-plane semantic segmentation and is available in both Torch and Tensorflow with single node and distributed-memory multi-node implementations.

The model is available in a development version with sparse convolutions in the torch framework. It is flexible enough to generate synthetic data on the fly or use real data training datasets, and dense or sparse data. The data for this network is in larcv3 format and is in the public domain. Currently, data is available in full resolution (HxW = 1280 x 2048) of three images per training sample. The image size is large, and the network is large, so to accommodate older hardware or smaller GPUs, the model can be run with a reduced image size. The datasets are kept at full resolution but a downsampling operation is applied prior to feeding images and labels to the network. The code is scaled on several leadership supercomputers, including Theta at ALCF.