Buckets:

download
raw
3.43 kB
#!/usr/bin/env python3
import argparse
import os
from pathlib import Path
from docling.models.utils.hf_model_download import download_hf_model
from docling.utils.model_downloader import download_models
_parser = argparse.ArgumentParser()
_parser.add_argument("--models-dir", default="./models",
help="Target directory for model files (default: ./models, matching path_base_model='.' in config)")
_args, _ = _parser.parse_known_args()
os.environ["PADDLE_PDX_CACHE_HOME"] = os.path.abspath(_args.models_dir)
from paddlex import create_model
MODELS_PP_OCR = [
"latin_PP-OCRv5_mobile_rec",
"PP-OCRv6_tiny_det",
]
LAYOUT_HERON_ONNX_REPO = "docling-project/docling-layout-heron-onnx"
FIGURE_CLASSIFIER_REPO = "docling-project/DocumentFigureClassifier-v2.5"
def is_model_present(name: str, target_dir: Path) -> bool:
"""Return True and print a skip message if the model directory already exists and is non-empty."""
if target_dir.exists() and any(target_dir.iterdir()):
print(f" -> {name}: already present at {target_dir}, skipping")
return True
return False
def write_rec_dict(model) -> None:
character_dict = model.config["PostProcess"]["character_dict"]
dict_path = os.path.join(model.model_dir, "dict.txt")
with open(dict_path, "w", encoding="utf-8") as f:
for char in character_dict:
f.write((char if char is not None else "") + "\n")
print(f" -> dictionnaire ({len(character_dict)} caractères) écrit dans {dict_path}")
def paddle_ocr_model(models_dir: Path) -> None:
for name in MODELS_PP_OCR:
# PaddleX stores models under {PADDLE_PDX_CACHE_HOME}/official_models/{name}_onnx
target_dir = models_dir / "official_models" / (name + "_onnx")
if is_model_present(name, target_dir):
continue
print(f"\nDownloading: {name}")
model = create_model(name, engine="onnxruntime")
if name.endswith("_rec"):
write_rec_dict(model)
def docling_model(models_dir: Path) -> None:
tableformer_dir = models_dir / "docling-project--docling-models" / "model_artifacts" / "tableformer"
classifier_dir = models_dir / FIGURE_CLASSIFIER_REPO.replace("/", "--")
if not is_model_present("TableFormer", tableformer_dir) or not is_model_present("DocumentFigureClassifier", classifier_dir):
print("\nDownloading docling models (TableFormer + DocumentFigureClassifier)")
download_models(
output_dir=models_dir,
with_layout=False,
with_tableformer=True,
with_tableformer_v2=False,
with_picture_classifier=True,
with_code_formula=False,
with_rapidocr=False,
with_easyocr=False,
progress=True,
)
heron_dir = models_dir / LAYOUT_HERON_ONNX_REPO.replace("/", "--")
if not is_model_present("layout-heron-onnx", heron_dir):
print("\nDownloading layout model (ONNX variant used in pipeline)")
download_hf_model(
repo_id=LAYOUT_HERON_ONNX_REPO,
local_dir=heron_dir,
progress=True,
)
def main(models_dir: str) -> None:
models_dir = Path(models_dir)
print(f"Target directory : {models_dir}")
paddle_ocr_model(models_dir)
docling_model(models_dir)
print("\nDone, the models have been downloaded!")
if __name__ == "__main__":
main(_args.models_dir)

Xet Storage Details

Size:
3.43 kB
·
Xet hash:
0453f4b24dc031af6c0ddb40f8bdb3a257acd1fa1586535e5f40faced0ba88c9

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.