export_onnx.py

from __future__ import annotations

import argparse
import json
import os

import torch

from configuration_htr import HTRConfig
from modeling_htr import HTRConvTextModel


class InferenceWrapper(torch.nn.Module):
    def __init__(self, model: HTRConvTextModel):
        super().__init__()
        self.model = model

    def forward(self, image: torch.Tensor) -> torch.Tensor:
        return self.model(pixel_values=image, return_dict=True).logits


def main() -> None:
    parser = argparse.ArgumentParser(description="Export HF model to ONNX.")
    parser.add_argument(
        "--hf-model-dir", required=True, help="Directory with HF artifacts."
    )
    parser.add_argument("--output-dir", default="onnx", help="ONNX output directory.")
    parser.add_argument(
        "--onnx-name", default="model.onnx", help="Output ONNX model filename."
    )
    parser.add_argument(
        "--dummy-width", type=int, default=3072, help="Dummy input width for export."
    )
    args = parser.parse_args()

    os.makedirs(args.output_dir, exist_ok=True)
    model = HTRConvTextModel.from_pretrained(
        args.hf_model_dir,
        trust_remote_code=True,
        low_cpu_mem_usage=False,
        torch_dtype=torch.float32,
    )
    model.eval()

    cfg = HTRConfig.from_pretrained(args.hf_model_dir, trust_remote_code=True)
    dummy = torch.randn(1, 1, cfg.image_height, args.dummy_width)

    wrapped = InferenceWrapper(model)
    onnx_path = os.path.join(args.output_dir, args.onnx_name)
    torch.onnx.export(
        wrapped,
        dummy,
        onnx_path,
        input_names=["image"],
        output_names=["logits"],
        dynamic_axes={
            "image": {0: "batch", 3: "width"},
            "logits": {0: "batch", 1: "timesteps"},
        },
        opset_version=18,
        do_constant_folding=True,
        export_params=True,
    )

    alphabet_path = os.path.join(args.hf_model_dir, "alphabet.json")
    if os.path.isfile(alphabet_path):
        with open(alphabet_path, "r", encoding="utf-8") as f:
            alph = json.load(f)
        with open(
            os.path.join(args.output_dir, "alphabet.json"), "w", encoding="utf-8"
        ) as f:
            json.dump(alph, f, ensure_ascii=False, indent=2)

    print(f"ONNX exported to: {onnx_path}")


if __name__ == "__main__":
    main()