Skip to content

Improve structure and documentation of deployment functionality.

Created by: asogaard

As pointed out here, the current version of the deployment functionality is entirely specific to IceCube, and it is not clear how it should be extended to other experiments.

I suggest:

  • We introduce a minimal, generic InferenceModule class, that loads pretrained GraphNeT models, builds graphs, and applies the former to the latter to get predictions.
  • Refactor I3InferenceModule to have it inherit from InferenceModule as one example of how to apply the same principles in an experiment-specific software stack.
  • Write SQLiteInferenceModule and ParquetInferenceModule classes that inherit from InferenceModule and shows how to to sequential inference on generic data formats.
  • Provide self-contained examples of how to apply the to above inference modules.
  • Add a note in graphnet.deployment that we are happy to assist users wanting to extend this type of inference to other, experiment-specific data formats.

cc @RasmusOrsoe, @JostMigenda