Add 1000 files (ocr-tech/tests/test_audit_limbu_book_missed_line_review_coverage.py ...)
d82db76 verified Download ocr-tech/tools/deva_rec_torch/torch_predict.py from prosperious/ocr-projects-archive: direct link, hf CLI and curl.
- Browser
- Download file 2.28 kB
-
https://huggingface.co/prosperious/ocr-projects-archive/resolve/main/ocr-tech/tools/deva_rec_torch/torch_predict.py
- Command line
-
hf download hf://prosperious/ocr-projects-archive/ocr-tech/tools/deva_rec_torch/torch_predict.py
-
curl -L -o torch_predict.py https://huggingface.co/prosperious/ocr-projects-archive/resolve/main/ocr-tech/tools/deva_rec_torch/torch_predict.py
2.28 kB
| """Predict eval sets with a torch recognizer checkpoint (HF format) on MPS/CPU. | |
| Usage: torch_predict.py <model_dir> <out_dir> name:root:labels [name:root:labels ...] | |
| Writes ``<out_dir>/<name>.tsv`` rows ``relpath<TAB>prediction<TAB>score``. 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()) | |