Add 323 files (ocr-tech-mps-integration/tools/prepare_native_spotcheck_bundle.py ...)
0cfd653 verified Download ocr-deva-eval-20260924/predict_both.py from prosperious/ocr-projects-archive: direct link, hf CLI and curl.
- Browser
- Download file 3.76 kB
-
https://huggingface.co/prosperious/ocr-projects-archive/resolve/main/ocr-deva-eval-20260924/predict_both.py
- Command line
-
hf download hf://prosperious/ocr-projects-archive/ocr-deva-eval-20260924/predict_both.py
-
curl -L -o predict_both.py https://huggingface.co/prosperious/ocr-projects-archive/resolve/main/ocr-deva-eval-20260924/predict_both.py
3.76 kB
| #!/usr/bin/env python3 | |
| """Run the deployed deva-v2 recognizer and the official devanagari_PP-OCRv5_mobile_rec | |
| over labelled line sets, with the same preprocessing as the k2 OCR worker. | |
| Writes <out>/<model>/<set>.tsv rows: image_path, reference, prediction, score. | |
| Scoring is done separately with tools/eval_metrics.py. | |
| """ | |
| from __future__ import annotations | |
| import json | |
| import sys | |
| import time | |
| from pathlib import Path | |
| import numpy as np | |
| from paddleocr._models.text_recognition import TextRecognition | |
| from PIL import Image, ImageOps | |
| HOME = Path.home() | |
| MODELS = { | |
| "deva-v2": ("PP-OCRv5_mobile_rec", HOME / "ampixa-demo/models/ocr/space/models/deva-v2/artifacts"), | |
| "official": ("devanagari_PP-OCRv5_mobile_rec", HOME / ".paddlex/official_models/devanagari_PP-OCRv5_mobile_rec"), | |
| } | |
| BATCH_SIZE = 16 | |
| def load_set(root: Path, labels: Path) -> list[tuple[str, str]]: | |
| rows: list[tuple[str, str]] = [] | |
| for line_no, line in enumerate(labels.read_text(encoding="utf-8").splitlines(), start=1): | |
| if not line.strip(): | |
| continue | |
| parts = line.split("\t", 1) | |
| if len(parts) != 2: | |
| raise ValueError(f"{labels}:{line_no}: expected '<path>\\t<text>'") | |
| image_path = root / parts[0] | |
| if not image_path.is_file(): | |
| raise FileNotFoundError(f"{labels}:{line_no}: missing image {image_path}") | |
| rows.append((parts[0], parts[1])) | |
| return rows | |
| def to_bgr(path: Path) -> np.ndarray: | |
| # Mirrors the worker: exif_transpose -> RGB -> BGR contiguous ndarray. | |
| with Image.open(path) as probe: | |
| image = ImageOps.exif_transpose(probe).convert("RGB") | |
| return np.asarray(image)[:, :, ::-1].copy() | |
| def main() -> int: | |
| out = Path(sys.argv[1]) | |
| sets = { | |
| name: (Path(root), Path(labels)) | |
| for name, root, labels in (arg.split(":", 2) for arg in sys.argv[2:]) | |
| } | |
| timings: dict[str, dict[str, float]] = {} | |
| for model, (model_name, model_dir) in MODELS.items(): | |
| for required in ("inference.json", "inference.pdiparams", "inference.yml"): | |
| if not (model_dir / required).is_file(): | |
| raise FileNotFoundError(f"{model}: missing {model_dir / required}") | |
| predictor = TextRecognition(model_name=model_name, model_dir=str(model_dir), device="cpu") | |
| timings[model] = {} | |
| for set_name, (root, labels) in sets.items(): | |
| rows = load_set(root, labels) | |
| crops = [to_bgr(root / rel) for rel, _ in rows] | |
| started = time.perf_counter() | |
| results = list(predictor.predict(crops, batch_size=BATCH_SIZE)) | |
| elapsed = time.perf_counter() - started | |
| if len(results) != len(rows): | |
| raise RuntimeError(f"{model}/{set_name}: {len(results)} results for {len(rows)} crops") | |
| (out / model).mkdir(parents=True, exist_ok=True) | |
| with (out / model / f"{set_name}.tsv").open("w", encoding="utf-8") as handle: | |
| for (rel, reference), result in zip(rows, results, strict=True): | |
| payload = result.json.get("res", result.json) if hasattr(result, "json") else result | |
| text = str(payload.get("rec_text") or "").strip() | |
| score = payload.get("rec_score") | |
| clean_ref = reference.replace("\t", " ") | |
| handle.write(f"{rel}\t{clean_ref}\t{text.replace(chr(9), ' ')}\t{score}\n") | |
| timings[model][set_name] = round(elapsed / len(rows) * 1000, 2) | |
| print(f"{model} {set_name}: {len(rows)} lines, {timings[model][set_name]} ms/line", flush=True) | |
| (out / "timings_ms_per_line.json").write_text(json.dumps(timings, indent=2), encoding="utf-8") | |
| return 0 | |
| if __name__ == "__main__": | |
| raise SystemExit(main()) | |