prosperious's picture
Add 1000 files (ocr-tech/tests/test_audit_limbu_book_missed_line_review_coverage.py ...)
d82db76 verified
Raw History Blame Contribute Delete
3.58 kB
#!/usr/bin/env python3
"""Run GLM-OCR on one image or page-bundle directory."""
from __future__ import annotations
import argparse
import json
from pathlib import Path
from ocrtech.external_outputs import bundle_page_paths, materialize_external_markdown
def build_parser() -> argparse.ArgumentParser:
parser = argparse.ArgumentParser(description="Run GLM-OCR and write ocrtech-compatible outputs.")
parser.add_argument("--input", required=True, help="image path or directory of page images")
parser.add_argument("--out", required=True, help="output directory")
parser.add_argument("--model-id", default="zai-org/GLM-OCR")
parser.add_argument("--prompt", default="Text Recognition:")
parser.add_argument("--max-new-tokens", type=int, default=8192)
return parser
def main() -> int:
args = build_parser().parse_args()
input_path = Path(args.input)
output_dir = Path(args.out)
output_dir.mkdir(parents=True, exist_ok=True)
try:
from transformers import AutoModelForImageTextToText, AutoProcessor
except ImportError as exc:
raise RuntimeError("GLM-OCR wrapper requires transformers with multimodal support installed.") from exc
try:
processor = AutoProcessor.from_pretrained(args.model_id, trust_remote_code=True, use_fast=False)
model = AutoModelForImageTextToText.from_pretrained(
pretrained_model_name_or_path=args.model_id,
torch_dtype="auto",
device_map="auto",
trust_remote_code=True,
)
except Exception as exc:
raise RuntimeError(
"GLM-OCR processor/model initialization failed. The upstream model card currently requires a git "
"transformers build, for example `pip install git+https://github.com/huggingface/transformers.git`, "
f"plus any remote-code dependencies such as torchvision. Root cause: {type(exc).__name__}: {exc}"
) from exc
page_payloads: list[dict[str, str]] = []
markdown_pages: list[str] = []
for page_path in bundle_page_paths(input_path):
messages = [
{
"role": "user",
"content": [
{"type": "image", "url": str(page_path)},
{"type": "text", "text": args.prompt},
],
}
]
inputs = processor.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=True,
return_dict=True,
return_tensors="pt",
).to(model.device)
inputs.pop("token_type_ids", None)
generated_ids = model.generate(**inputs, max_new_tokens=args.max_new_tokens)
generated = generated_ids[0][inputs["input_ids"].shape[1] :]
markdown = _clean_generation(processor.decode(generated, skip_special_tokens=False))
page_payloads.append({"page_path": str(page_path), "markdown": markdown})
markdown_pages.append(markdown)
joined_markdown = "\n\n<!-- pagebreak -->\n\n".join(markdown_pages)
materialize_external_markdown(
input_path,
output_dir,
joined_markdown,
metadata={"engine": "glm-ocr", "model_id": args.model_id, "prompt": args.prompt},
raw_payload={"pages": page_payloads},
)
return 0
def _clean_generation(text: str) -> str:
value = text.strip()
for token in ("<|assistant|>", "<|user|>", "<|endoftext|>", "<|im_end|>"):
value = value.replace(token, "")
return value.strip()
if __name__ == "__main__":
raise SystemExit(main())