Bug: check for stretched grid breaks running with minimal graph config

[dietervdb-meteo]

What happened?

The merging of PR199 introduced the line
if "lam_resolution" in getattr(config.graph.nodes.hidden, "node_builder", []): in forecaster.py
This will result in an error in case one sets up a config where the graph part is the minimal

graph:
  overwrite: False 
  data: data
  hidden: hidden

which is sufficient to load a graph previously stored on disk. Rather than trying to determine if the model is stretched grid or not from the graph part of the training config this should preferably be checked on the graph directly.

What are the steps to reproduce the bug?

Run training with a config for which the graph part is as above.

Version

current develop

Platform (OS and architecture)

Lumi [irrelevant]

Relevant log output

No response

Accompanying data

No response

Organisation

rmi

[dietervdb-meteo]

@jswijnands , any simple fixes/ways around for this?

[dietervdb-meteo]

As a temporary fix, the following minimal graph config still works [when running from a graph on disk]:

graph:
  overwrite: False #This to take graph from file
  data: data #This and next line needed by AnemoiModelEncProcDec in anemoi-models
  hidden: hidden 
  # below needed to circumvent the bug reported here [it doesn't do anything else in the code]
  nodes:
    hidden:
      node_builder:
        _target_: anemoi.graphs.nodes.LimitedAreaTriNodes # options: ZarrDatasetNodes, NPZFileNodes, TriNodes

[dietervdb-meteo]

As mentioned in ecmwf/anemoi-core#30 , it would be elegant if the graph part of the training config when one wants to run from a graph on disk could be as minimal as:

graph:
   overwrite: False #or even something like create: False

[jswijnands]

Hi Dieter, I see. We have indeed discussed various ways on how to identify whether a model is a stretched grid or not. Your minimal graph config use case was not envisioned then. The temporary fix you suggested indeed works.

[jswijnands]

The if statement below may solve this problem, so you can still use a minimal script:

if "nodes" in getattr(config, "graph", []) and
    "hidden" in getattr(config.graph, "nodes", []) and
    "node_builder" in getattr(config.graph.nodes, "hidden", []) and
    "lam_resolution" in getattr(config.graph.nodes.hidden, "node_builder", []):

[jswijnands]

Perhaps your preferred setup can also be used as one of the test scripts that is run when pushing new code changes. In that way, it gets flagged so it can be solved during development (also for future PRs).

[dietervdb-meteo]

Concerning your proposed if statement above. That will indeed allow users that are not stretched grid and run graph from file to use a minimal config, but what about stretched grid users that would like to run with a minimal config and graph from file?

Maybe we can add some 'type' attribute to the graph, that can take values like 'stretched_grid', 'lam', 'icon' etc. That will allow other parts of the code to make decisions based on that. This type could be set through the config/or automatically detected through some if statements, but upon graph creation, rather than in another part of the code.

As this crosses between graph and training implementing something like that could be done once the repos are merged.

@JPXKQX , any comments/ideas?

[JPXKQX]

This already exists. When you create a graph with anemoi-graphs each set of nodes has a node_type attribute whose value is the name of the class that created it:

>>> g["hidden"].node_type
'StretchedTriNodes'

[dietervdb-meteo]

Great, thanks for pointing that out! @jswijnands , how about using the above to check if the model is stretched grid or not?

[jswijnands]

Good idea, I'll look into this

[jswijnands]

Hi Dieter, I've made this change and successfully tested it for stretched grid. See PR #204