Implementation of CNN for IceCube

[sevmag]

Implement the CNN from this paper

Is your feature request related to a problem? Please describe. Needed for direct comparison between the CNN and other architectures all within GraphNet trained on the same data.

Describe the solution you'd like To do this one need to create a new NodeDefinition similar to the one described in #770. Additionally, it needs a model that will be used as the backbone argument of the StandardModel. The first thing its forward() function should call is a function that maps the graph resulting from the new NodeDefinition into the necessary format of the new CNN architecture. After that the forward should run through the CNN Architecture like expected.

Can someone assign me to this?

[pweigel]

This architecture is public, you can find it here: https://github.com/icecube/dnn_reco

They use tensorflow and use special hexagonal convolutions which leverage the (approximate) symmetries in IceCube. You can find the code for those convolutions and other utilities here: https://github.com/icecube/TFScripts

Hope this is useful!