Make RapidOcrOptions ignore `artifacts_path`

[simonschoe]

Requested feature

Currently, when using RapidOCR, the initialization of RapidOcrModel considers artifacts_path. That is, it searches for model artifacts under artifacts_path as defined here:

• docling/docling/models/auto_ocr_model.py

  Lines 65 to 74 in 6a04e27

  	self._engine = RapidOcrModel(
  	enabled=self.enabled,
  	artifacts_path=artifacts_path,
  	options=RapidOcrOptions(
  	backend="onnxruntime",
  	bitmap_area_threshold=self.options.bitmap_area_threshold,
  	force_full_page_ocr=self.options.force_full_page_ocr,
  	),
  	accelerator_options=accelerator_options,
  	)

• docling/docling/models/rapid_ocr_model.py

  Lines 125 to 149 in 6a04e27

  	if artifacts_path is not None:
  	det_model_path = (
  	det_model_path
  	or artifacts_path
  	/ self._model_repo_folder
  	/ self._default_models[backend_enum.value]["det_model_path"]["path"]
  	)
  	cls_model_path = (
  	cls_model_path
  	or artifacts_path
  	/ self._model_repo_folder
  	/ self._default_models[backend_enum.value]["cls_model_path"]["path"]
  	)
  	rec_model_path = (
  	rec_model_path
  	or artifacts_path
  	/ self._model_repo_folder
  	/ self._default_models[backend_enum.value]["rec_model_path"]["path"]
  	)
  	rec_keys_path = (
  	rec_keys_path
  	or artifacts_path
  	/ self._model_repo_folder
  	/ self._default_models[backend_enum.value]["rec_keys_path"]["path"]
  	)

When installing rapidocr it already ships with the base onnx models available under ...\Lib\site-packages\rapidocr\models. Therefore, we usually do not need downloading these models separately from Modelscope. However, when setting artifacts_path in the pipeline, e.g., for loading the layout detection or table structure model, we are currently not able to load the default onnx models shipped with rapidocr.

@geoHeil Would it be possible to either deliberately skip a globally defined artifacts_path in order to preload the shipped models and skip downloading from Modelscope or b) make loading from ...\Lib\site-packages\rapidocr\models the default?

[geoHeil]

Sounds like it should be. Would you want to raise a PR?

[simonschoe]

@geoHeil I was hoping you could fill in. 😁 Until I have set everything up to prepare a proper PR... I am not sure I currently find the time unfortunately.

[geoHeil]

Hm I am a bit tight in the next weeks just as well.

[simonschoe]

@cau-git , @geoHeil : I dug into this a little bit since I found it quite annoying to have the Modelscope dependency if the artifacts are already readily available inside rapidocr. Maybe it is too simple, but by leaving the *.model_path parameters unset in the params mapping for the init of RapidOCR, the artifacts should be automatically loaded from the repo folder:

~\Lib\site-packages\rapidocr\models\ch_PP-OCRv4_det_infer.onnx
~\Lib\site-packages\rapidocr\models\ch_ppocr_mobile_v2.0_cls_infer.onnx
~\Lib\site-packages\rapidocr\models\ch_PP-OCRv4_rec_infer.onnx

If my assumption is correct, you shouldn't need to resolve paths and consider artifacts_path.

class RapidOcrModel(BaseOcrModel):
    # ... 

    def __init__(
        self,
        enabled: bool,
        artifacts_path: Optional[Path],
        options: RapidOcrOptions,
        accelerator_options: AcceleratorOptions,
    ):
        # ... 

            det_model_path = self.options.det_model_path
            cls_model_path = self.options.cls_model_path
            rec_model_path = self.options.rec_model_path
            rec_keys_path = self.options.rec_keys_path
            font_path = self.options.font_path

            # skip auto-resolve
            for model_path in (
                det_model_path,
                cls_model_path,
                rec_model_path,
                rec_keys_path,
                font_path,
            ):
                if model_path is not None and not Path(model_path).exists():
                    _log.warning(f"The provided model path {model_path} is not found.")

            params = {
                "Global.text_score": self.options.text_score,
                # Detection
                "Det.use_cuda": use_cuda,
                "Det.use_dml": use_dml,
                "Det.intra_op_num_threads": intra_op_num_threads,
                # Classification
                "Cls.use_cuda": use_cuda,
                "Cls.use_dml": use_dml,
                "Cls.intra_op_num_threads": intra_op_num_threads,
                # Recognition
                "Rec.use_cuda": use_cuda,
                "Rec.use_dml": use_dml,
                "Rec.intra_op_num_threads": intra_op_num_threads,
                # Engine types and GPU selection
                "Det.engine_type": backend_enum,
                "Cls.engine_type": backend_enum,
                "Rec.engine_type": backend_enum,
                "EngineConfig.paddle.use_cuda": use_cuda,
                "EngineConfig.paddle.gpu_id": gpu_id,
                "EngineConfig.torch.use_cuda": use_cuda,
                "EngineConfig.torch.gpu_id": gpu_id,
            }

            # override rapidocr default if paths are explicitely provided
            if font_path is not None:
                params["Global.font_path"] = font_path
            if det_model_path is not None:
                params["Det.model_path"] = det_model_path
            if cls_model_path is not None:
                params["Cls.model_path"] = cls_model_path
            if rec_model_path is not None:
                params["Rec.model_path"] = rec_model_path
            if rec_keys_path is not None:
                params["Rec.rec_keys_path"] = rec_keys_path
            if self.options.rec_font_path is not None:
                _log.warning(
                    "The 'rec_font_path' option for RapidOCR is deprecated. Please use 'font_path' instead."
                )
                params["Rec.font_path"] = self.options.rec_font_path

            user_params = self.options.rapidocr_params
            self.reader = RapidOCR(params=params)

    # ... 

[geoHeil]

@dolfim-ibm I see you self-assigned it but never published a PR - possibly mine (from just now) is useful to address this ticket.

[dolfim-ibm]

@geoHeil we might have forgotten to link the issue with the changes, and I think my intent was first to verify the actual issue with the latest version.

In general, the Docling logic is that artifacts_path, when specified, shouldn't allow models to download checkpoints, i.e. not having the model weights is considered an error.

If to reconsider this behavior, we should do it for all model weights consistently and in a way it is explicit and transparent for the users.

[geoHeil]

#2896