Add 1000 files (ocr-tech/tests/test_audit_limbu_book_missed_line_review_coverage.py ...)
d82db76 verified Download ocr-tech/tools/deva_rec_torch/export_onnx.py from prosperious/ocr-projects-archive: direct link, hf CLI and curl.
- Browser
- Download file 2.63 kB
-
https://huggingface.co/prosperious/ocr-projects-archive/resolve/main/ocr-tech/tools/deva_rec_torch/export_onnx.py
- Command line
-
hf download hf://prosperious/ocr-projects-archive/ocr-tech/tools/deva_rec_torch/export_onnx.py
-
curl -L -o export_onnx.py https://huggingface.co/prosperious/ocr-projects-archive/resolve/main/ocr-tech/tools/deva_rec_torch/export_onnx.py
2.63 kB
| """Export a fine-tuned torch recognizer to a PaddleX-servable ONNX model directory. | |
| Usage: export_onnx.py <hf_model_dir> <reference_onnx_dir> <out_dir> | |
| ``reference_onnx_dir`` is PaddlePaddle/devanagari_PP-OCRv5_mobile_rec_onnx: its | |
| ``inference.yml`` (pre/post-processing and character dictionary) is copied unchanged, and | |
| the exported graph keeps its I/O contract: input ``x`` [N, 3, 48, W] float32, output | |
| ``fetch_name_0`` [N, T, 570] softmax probabilities. The export is verified against the torch | |
| model on random inputs of several widths before the directory is considered valid. | |
| """ | |
| from __future__ import annotations | |
| import shutil | |
| import sys | |
| from pathlib import Path | |
| import numpy as np | |
| import onnxruntime as ort | |
| import torch | |
| from common import load_model | |
| class Probs(torch.nn.Module): | |
| def __init__(self, model: torch.nn.Module) -> None: | |
| super().__init__() | |
| self.model = model | |
| def forward(self, x: torch.Tensor) -> torch.Tensor: | |
| return self.model(x).last_hidden_state | |
| def main() -> int: | |
| model_dir, reference_dir, out_dir = Path(sys.argv[1]), Path(sys.argv[2]), Path(sys.argv[3]) | |
| out_dir.mkdir(parents=True, exist_ok=True) | |
| model = Probs(load_model(model_dir, torch.device("cpu")).eval()).eval() | |
| onnx_path = out_dir / "inference.onnx" | |
| torch.onnx.export( | |
| model, (torch.randn(1, 3, 48, 320),), str(onnx_path), | |
| input_names=["x"], output_names=["fetch_name_0"], | |
| dynamic_axes={"x": {0: "batch", 3: "width"}, "fetch_name_0": {0: "batch", 1: "steps"}}, | |
| opset_version=17, dynamo=False, | |
| ) | |
| shutil.copyfile(reference_dir / "inference.yml", out_dir / "inference.yml") | |
| session = ort.InferenceSession(str(onnx_path), providers=["CPUExecutionProvider"]) | |
| worst = 0.0 | |
| for batch, width in ((1, 320), (4, 480), (2, 1200)): | |
| x = torch.randn(batch, 3, 48, width) | |
| with torch.no_grad(): | |
| ref = model(x).numpy() | |
| got = np.asarray(session.run(["fetch_name_0"], {"x": x.numpy()})[0]) | |
| if got.shape != ref.shape: | |
| raise RuntimeError(f"shape mismatch at width {width}: onnx {got.shape} vs torch {ref.shape}") | |
| diff = float(np.abs(got - ref).max()) | |
| argmax_agree = float((got.argmax(-1) == ref.argmax(-1)).mean()) | |
| print(f"batch {batch} width {width}: T={got.shape[1]} max|diff|={diff:.2e} argmax agree={argmax_agree:.4f}") | |
| worst = max(worst, diff) | |
| if worst > 1e-3: | |
| raise RuntimeError(f"ONNX export deviates from torch (max |diff| {worst:.2e})") | |
| print(f"exported and verified: {out_dir}") | |
| return 0 | |
| if __name__ == "__main__": | |
| sys.exit(main()) | |