Set artifacts_path for picture classifier

[simonschoe]

Bug

When downloading and pointing to models locally, we commonly use pipeline_options.artifacts_path. For example:
pipeline_options.artifacts_path = "models"

The layout and tableformer model are then imported from the following paths:

• Layout model: models/model_artifacts/layout/...
• Tableformer: models/model_artifacts/tableformer/...

However, when importing the new picture classifier from a local folder, it requires that all model files are in the root folder, i.e., models:

• Picture classifier: models/config.json, models/model.safetensors

I find this very counterintuitive and not very well structured.

EDIT: This is even more annoying when using both the picture classifier and the CodeFormula models and importing from local since both accept the same artifacts_path.

Proposal

Two options that come to mind:
a. When initalizing the DocumentFigureClassifierPredictor, simply point to models as artifacts_path and allow same folder structure as on huggingface. For example:

DocumentFigureClassifierPredictor(
    artifacts_path="models", # model files are in models/ds4sd/DocumentFigureClassifier
    device="cpu",
    num_threads=4,
) 

b. On huggingface, for the DocumentFigureClassifier repo use the model_artifacts/<model_type> repo structure that is similar to the main docling repository and allow import by simply stating the root folder. For example:

DocumentFigureClassifierPredictor(
    artifacts_path="models", # model files are in models/model_artifacts/DocumentFigureClassifier/...
    device="cpu",
    num_threads=4,
) 

Docling version

2.16.0