Using ds4sd/docling-models from local path

[FBR65]

Hi,

for me it's not possible to use the normal snapshot_download modus wehre the models are downloaded during execution.

The models have to be downloaded in front of the execution to a specific directory.

Is there a parameter which can be used to use local models instead of huggingface download?

[dolfim-ibm]

Yes, it is possible. We will shortly add a concrete example in the documentation (#224) and an explicit option in the CLI (#203).

[AdityaMannu1709]

Hi @dolfim-ibm, any updates on this example for using models from local instead of downloading always before a run?

[vagenas]

@FBR65 you can check how to pass local (i.e. pre-fetched) models in the docs here:
https://ds4sd.github.io/docling/usage/#provide-specific-artifacts-path

[MJonker]

I think I need to supply also the easyOCR model location:

https://github.com/DS4SD/docling/blob/main/docling/datamodel/pipeline_options.py#L144-L146

[MJonker]

Ok after setting the environment variable for local EasyOCR models, I am getting this:

Maybe I am setting the artifacts_path incorrectly...

[MJonker]

somehow it still gives an error when initializing the easyOCR model:

     66     lang_list=self.options.lang,
     67     gpu=use_gpu,
     68     model_storage_directory=self.options.model_storage_directory,
     69     recog_network=self.options.recog_network,
     70     download_enabled=self.options.download_enabled,
     71     verbose=False,

Could really use a hint

[MJonker]

this the code

from docling.document_converter import DocumentConverter, InputFormat, PdfFormatOption
from docling.document_converter import ConversionError
import docling.pipeline.standard_pdf_pipeline
from docling.datamodel.pipeline_options import PdfPipelineOptions
from pathlib import Path
import os

print(os.path.abspath("models/docling"))
pipeline_options = PdfPipelineOptions(artifacts_path=str(os.path.abspath("models/docling")))
os.environ['EASYOCR_MODULE_PATH'] = str(os.path.abspath("models"))

dir_name="pdfonly"
number_of_files_processed=0
number_of_files_failed=0;
for filename in os.listdir(dir_name):
    print("number_of_files_processed: "+str(number_of_files_processed))
    
    
    print('begin:' + filename)

    try: 
        #converter = DocumentConverter()
        converter = DocumentConverter(format_options={
            InputFormat.PDF: PdfFormatOption(pipeline_options=pipeline_options)})
        result = converter.convert(os.path.abspath(dir_name+"/"+filename))
        print(result.document.export_to_markdown())

[MJonker]

This helped:
#381

pipeline_options.ocr_options.model_storage_directory = str(os.path.abspath("models"))