"""Predict eval sets with a torch recognizer checkpoint (HF format) on MPS/CPU. Usage: torch_predict.py name:root:labels [name:root:labels ...] Writes ``/.tsv`` rows ``relpathpredictionscore``. Only crops with the identical PaddleX pad width share a batch, so every crop sees exactly the padding it gets from single-image PaddleX inference (SVTR attention is padding-sensitive). """ from __future__ import annotations import sys from collections import defaultdict from pathlib import Path import torch from common import batch_tensor, ctc_greedy_decode, load_character_list, load_model, padded_width, pick_device, read_bgr, resized_width def main() -> int: model_dir, out_dir = Path(sys.argv[1]), Path(sys.argv[2]) out_dir.mkdir(parents=True, exist_ok=True) device = pick_device("auto") chars = load_character_list(model_dir) model = load_model(model_dir, device).eval() for spec in sys.argv[3:]: name, root, labels = spec.split(":") rows = [line.split("\t", 1)[0] for line in Path(labels).read_text(encoding="utf-8").splitlines() if "\t" in line] images = [read_bgr(Path(root) / rel) for rel in rows] widths = [resized_width(img) for img in images] by_pad: dict[int, list[int]] = defaultdict(list) for i, width in enumerate(widths): by_pad[padded_width(width)].append(i) preds: dict[int, tuple[str, float]] = {} with torch.no_grad(): for group in by_pad.values(): for start in range(0, len(group), 32): idx = group[start:start + 32] probs = model(batch_tensor([images[i] for i in idx], [widths[i] for i in idx]).to(device)).last_hidden_state for i, result in zip(idx, ctc_greedy_decode(probs.float().cpu(), chars), strict=True): preds[i] = result with (out_dir / f"{name}.tsv").open("w", encoding="utf-8") as handle: for i, rel in enumerate(rows): text, score = preds[i] handle.write(f"{rel}\t{text.replace(chr(9), ' ')}\t{score:.4f}\n") print(f"{name}: {len(rows)} predictions", flush=True) return 0 if __name__ == "__main__": sys.exit(main())