Config.cfg Execution Logic

[koayst]

Hi,

I am trying to understand the logical flow on how Spacy uses config.cfg to create a model by studying rel_componnent. I also did the tutorial found in thinc.ai.

Based on this config, how does Spacy knows it needs to build the model that is labelled as "rel_model.v1" ?

From thinc.ai tutorial, there is a common pattern (as shown below) after the config file is read and resolved.

  model = loaded_config ["model"]
  optimizer = loaded_config ["optimizer"]
  n_iter = loaded_config ["training"]["n_iter"]
  batch_size = loaded_config ["training"]["batch_size"]
  train_model(model, optimizer, n_iter, batch_size)

Can't seem to find similar pattern in the "rel_components" files.

Could you please help me to understand the flow? I think I have looked every documentation I can read. Something still does not connect.

Thanks.

[koayst]

In the project.yml file, under the "commands" section for "train_cpu", there is an option "-c"

script:
- "python -m spacy train ${vars.tok2vec_config} --output training --paths.train ${vars.train_file} --paths.dev ${vars.dev_file} -c ./scripts/custom_functions.py"

Wonder what is the significance of "custom_functions.py" when it comes to training the model?

[thomashacker]

Hey koayst ✨,
I think, the missing piece to better understand how the config system works is the Registry, where you can map strings like rel_model.v1 to a function which, in this case, would return a thinc model. You can also assign additional parameters when creating new registries. Here's the example from the REL code:

@spacy.registry.architectures("rel_model.v1")
def create_relation_model(
    create_instance_tensor: Model[List[Doc], Floats2d],
    classification_layer: Model[Floats2d, Floats2d],
) -> Model[List[Doc], Floats2d]:
    with Model.define_operators({">>": chain}):
        model = create_instance_tensor >> classification_layer
        model.attrs["get_instances"] = create_instance_tensor.attrs["get_instances"]
    return model

The -c parameter, when using spacy train, is important to resolve custom registry strings in your config, such as rel_model.v1.
You can read more about the different parameters in this train section