| import os |
| """PaddleOCR-VL-1.6 (default pipeline) on KDoc+Kolm pages -> score_external.py jsonl. Resumable. |
| Usage: run_paddlevl.py OUT_JSONL [LIMIT] [SHARD_INDEX SHARD_COUNT]""" |
| import json, os, sys, time |
| from pathlib import Path |
| from paddleocr import PaddleOCRVL |
| M = Path(os.environ.get("MANIFEST_DIR", "manifests")) |
| out = Path(sys.argv[1]); limit = int(sys.argv[2]) if len(sys.argv) > 2 else 0 |
| si, sc = (int(sys.argv[3]), int(sys.argv[4])) if len(sys.argv) > 4 else (0, 1) |
| pages = [p for b in os.environ.get("BENCHES", "kdoc,kolm").split(",") for p in json.loads((M / f"{b}.json").read_text())["pages"]][si::sc] |
| if limit: pages = pages[:limit] |
| done = {json.loads(l)["id"] for l in out.open()} if out.exists() else set() |
| pipe = PaddleOCRVL(pipeline_version="v1.6", vl_rec_backend="vllm-server", vl_rec_server_url="http://127.0.0.1:8050/v1", vl_rec_api_model_name="PaddleOCR-VL-1.6-0.9B", vl_rec_max_concurrency=64) |
| with out.open("a") as f: |
| for p in pages: |
| if p["id"] in done: continue |
| t = time.time(); text, err, boxes = "", None, [] |
| try: |
| for res in pipe.predict(p["image"]): |
| md = res.markdown |
| text += (md.get("markdown_texts", "") if isinstance(md, dict) else str(md)) |
| r = res.json.get("res", res.json) |
| for b in r.get("parsing_res_list", []): |
| boxes.append({"label": b.get("block_label"), "bbox": b.get("block_bbox")}) |
| boxes.append({"image_size": [r.get("width"), r.get("height")]}) |
| except Exception as e: |
| err = repr(e)[:300] |
| f.write(json.dumps({"id": p["id"], "text": text, "boxes": boxes, "error": err, "seconds": round(time.time() - t, 2)}, ensure_ascii=False) + "\n"); f.flush() |
|
|
