Reference for ultralytics/utils/export/axelera.py

def torch2axelera(
    model: torch.nn.Module,
    file: str | Path,
    calibration_dataset: torch.utils.data.DataLoader,
    transform_fn: Callable[[Any], np.ndarray],
    metadata: dict | None = None,
    prefix: str = "",
) -> Path:
    """Convert a YOLO model to Axelera format.

    Args:
        model (torch.nn.Module): Source YOLO model for quantization.
        file (str | Path): Source model file path used to derive output names.
        calibration_dataset (torch.utils.data.DataLoader): Calibration dataloader for quantization.
        transform_fn (Callable[[Any], np.ndarray]): Calibration preprocessing transform function.
        metadata (dict | None, optional): Optional metadata to save as YAML. Defaults to None.
        prefix (str, optional): Prefix for log messages. Defaults to "".

    Returns:
        (Path): Path to exported Axelera model directory.
    """
    prev_protobuf = os.environ.get("PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION")
    os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
    try:
        from axelera import compiler
    except ImportError:
        check_requirements(
            "axelera-devkit==1.6.0rc3",
            cmds="--extra-index-url https://software.axelera.ai/artifactory/api/pypi/axelera-pypi/simple",
        )
        from axelera import compiler

    from axelera.compiler import CompilerConfig
    from axelera.compiler.config.model_specific import extract_ultralytics_metadata

    LOGGER.info(f"\n{prefix} starting export with Axelera compiler...")

    file = Path(file)
    model_name = file.stem
    export_path = Path(f"{model_name}_axelera_model")
    export_path.mkdir(exist_ok=True)

    axelera_model_metadata = extract_ultralytics_metadata(model)
    config = CompilerConfig(
        model_metadata=axelera_model_metadata,
        model_name=model_name,
        resources_used=0.25,
        aipu_cores_used=1,
        multicore_mode="batch",
        output_axm_format=True,
    )
    qmodel = compiler.quantize(
        model=model,
        calibration_dataset=calibration_dataset,
        config=config,
        transform_fn=transform_fn,
    )
    compiler.compile(model=qmodel, config=config, output_dir=export_path)

    for artifact in [f"{model_name}.axm", "compiler_config_final.toml"]:
        artifact_path = Path(artifact)
        if artifact_path.exists():
            artifact_path.replace(export_path / artifact_path.name)

    # Remove intermediate compiler artifacts, keeping only the compiled model and config.
    keep_suffixes = {".axm"}
    keep_names = {"compiler_config_final.toml", "metadata.yaml"}
    for f in export_path.iterdir():
        if f.is_file() and f.suffix not in keep_suffixes and f.name not in keep_names:
            f.unlink()

    if metadata is not None:
        YAML.save(export_path / "metadata.yaml", metadata)

    # Restore original PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION value
    if prev_protobuf is None:
        os.environ.pop("PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION", None)
    else:
        os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = prev_protobuf

    return export_path