Codebase overview

anemoi-inference is built on top of a registry pattern. The structure of the code follows the structure of the configuration files.

Each registry-backed component has a set of classes defined under src/anemoi/inference/<component_name>/. Each class inherits from the base class src/anemoi/inference/<component_name>.py. For example, one can define plot as an output. The code which runs the plot output is found in src/anemoi/inference/outputs/plot.py.

The baseclass Output in src/anemoi/inference/output.py defines all shared or default behavior, which is then overridden as needed by the subclasses.

How is the correct subclass found and described? Each subclass uses the @<component_name>_registry.register("<name>") decorator to register itself into the component registry. Each component has its own registry, which is defined under src/anemoi/inference/<component_name>/__init__.py.

So, to indicate the the plot output should be registered to the output_registry, it is tagged with @output_registry.register("plot"). When an end user configures the yaml to use a plot output, the runner retrieves the matching class from the registry and returns it.

output:
  plot:
    dir: test_dir/

If the above configuration snippet is passed in, the class registered under "plot" (in this case, PlotOutput) is instantiated, with any additonal configuration passed into the matching argument on the class (in this case, it would be instantiated as PlotOutput(..., dir="test_dir/")).

These are all the registry-backed components that follow this pattern:

  • Inputs: input.py, src/anemoi/inference/inputs/

  • Outputs: output.py, src/anemoi/inference/outputs

  • Runners: runner.py, src/anemoi/inference/runners

  • Transports: transport.py, src/anemoi/inference/transports

  • Processors: processor.py. Note Processor is a special case, with mid-, pre-, and post- processors having their own registries but a shared base class.

For more information on how these classes work, adding a new option to an existing registry, or updating an existing option, check out Adding to Configuration options.

RunConfiguration

RunConfiguration is a pydantic class which describes at a high level the expected configuration formatting.

This organizes the top-level configuration inputs, and can be a useful reference when creating configuration files or implementing features which change configuration. For most features, this file should not need updating.