config.yaml not working wih anemoi-training train

[albertosanma]

Hi everyone,
first of all, thanks a lot for your work on this project!

I’m currently setting up a training run using a configuration adapted from the example on huggingface. Working through several schema mismatches, I now have a configuration that successfully validates through all the Pydantic layers and proceeds into graph construction and model setup.

However, the process fails during the instantiation of the model, with the following traceback (truncated for clarity):

hydra.errors.InstantiationException: Error in call to target
'anemoi.models.models.encoder_processor_decoder.AnemoiModelEncProcDec':
InstantiationException("Error in call to target
'anemoi.models.layers.mapper.GraphTransformerForwardMapper': InstantiationException()")

It looks like it reaches the model-building stage before this happens, so I suspect it might be related to one of the submodules’ arguments or how the encoder/decoder components are defined. If it helps, I can share my config file or sections of it as an example to get more detailed help.

At this point, I mainly want to make sure I’m using a configuration structure that’s actually compatible with the current version of anemoi-training.
Would you happen to have an example YAML config that’s known to work with the latest release? Having a stable reference would make it much easier to pinpoint whether the problem is due to my edits or to version differences.

Thanks again for your time and for maintaining such a complex framework!

[HCookie]

Hi,
If you are looking at recreating it from the huggingface, I would suggest pinning your versions to that setup.
Otherwise, using anemoi-training config generate will create working configs to adapt from