#!/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 //.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 '\\t'") 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())