"""Check that the torch port reproduces the official Paddle Devanagari recognizer. Usage: parity.py ``paddle_preds.tsv`` holds ``relpathtext`` from the Paddle model on the same crops. Both BGR (PaddleX convention) and RGB channel orders are scored. """ from __future__ import annotations import sys from pathlib import Path import torch from common import batch_tensor, ctc_greedy_decode, load_character_list, load_model, read_bgr, resized_width def main() -> int: model_dir, crop_root, labels, paddle_preds = (Path(x) for x in sys.argv[1:5]) chars = load_character_list(model_dir) model = load_model(model_dir, torch.device("cpu")).eval() paddle = dict(line.split("\t", 1) for line in paddle_preds.read_text(encoding="utf-8").splitlines() if "\t" in line) rows = [line.split("\t", 1) for line in labels.read_text(encoding="utf-8").splitlines() if "\t" in line] for order in ("bgr", "rgb"): same = 0 mismatches = [] for rel, _ in rows: image = read_bgr(crop_root / rel) if order == "rgb": image = image[:, :, ::-1].copy() with torch.no_grad(): probs = model(batch_tensor([image], [resized_width(image)])).last_hidden_state text = ctc_greedy_decode(probs, chars)[0][0] if text == paddle.get(rel, "").strip(): same += 1 elif len(mismatches) < 4: mismatches.append((rel, paddle.get(rel), text)) print(f"{order}: torch == paddle on {same}/{len(rows)} crops") for item in mismatches: print(" ", item) return 0 if __name__ == "__main__": sys.exit(main())