Reference for ultralytics/engine/exporter.py

def __call__(self, model=None) -> str:
    """Returns list of exported files/dirs after running callbacks."""
    self.run_callbacks("on_export_start")
    t = time.time()
    fmt = self.args.format.lower()  # to lowercase
    if fmt in {"tensorrt", "trt"}:  # 'engine' aliases
        fmt = "engine"
    if fmt in {"mlmodel", "mlpackage", "mlprogram", "apple", "ios", "coreml"}:  # 'coreml' aliases
        fmt = "coreml"
    fmts = tuple(export_formats()["Argument"][1:])  # available export formats
    if fmt not in fmts:
        import difflib

        # Get the closest match if format is invalid
        matches = difflib.get_close_matches(fmt, fmts, n=1, cutoff=0.6)  # 60% similarity required to match
        if not matches:
            raise ValueError(f"Invalid export format='{fmt}'. Valid formats are {fmts}")
        LOGGER.warning(f"WARNING ⚠️ Invalid export format='{fmt}', updating to format='{matches[0]}'")
        fmt = matches[0]
    flags = [x == fmt for x in fmts]
    if sum(flags) != 1:
        raise ValueError(f"Invalid export format='{fmt}'. Valid formats are {fmts}")
    (
        jit,
        onnx,
        xml,
        engine,
        coreml,
        saved_model,
        pb,
        tflite,
        edgetpu,
        tfjs,
        paddle,
        mnn,
        ncnn,
        imx,
    ) = flags  # export booleans
    is_tf_format = any((saved_model, pb, tflite, edgetpu, tfjs))

    # Device
    dla = None
    if fmt == "engine" and self.args.device is None:
        LOGGER.warning("WARNING ⚠️ TensorRT requires GPU export, automatically assigning device=0")
        self.args.device = "0"
    if fmt == "engine" and "dla" in str(self.args.device):  # convert int/list to str first
        dla = self.args.device.split(":")[-1]
        self.args.device = "0"  # update device to "0"
        assert dla in {"0", "1"}, f"Expected self.args.device='dla:0' or 'dla:1, but got {self.args.device}."
    self.device = select_device("cpu" if self.args.device is None else self.args.device)

    # Checks
    if imx and not self.args.int8:
        LOGGER.warning("WARNING ⚠️ IMX only supports int8 export, setting int8=True.")
        self.args.int8 = True
    if not hasattr(model, "names"):
        model.names = default_class_names()
    model.names = check_class_names(model.names)
    if self.args.half and self.args.int8:
        LOGGER.warning("WARNING ⚠️ half=True and int8=True are mutually exclusive, setting half=False.")
        self.args.half = False
    if self.args.half and onnx and self.device.type == "cpu":
        LOGGER.warning("WARNING ⚠️ half=True only compatible with GPU export, i.e. use device=0")
        self.args.half = False
        assert not self.args.dynamic, "half=True not compatible with dynamic=True, i.e. use only one."
    self.imgsz = check_imgsz(self.args.imgsz, stride=model.stride, min_dim=2)  # check image size
    if self.args.int8 and engine:
        self.args.dynamic = True  # enforce dynamic to export TensorRT INT8
    if self.args.optimize:
        assert not ncnn, "optimize=True not compatible with format='ncnn', i.e. use optimize=False"
        assert self.device.type == "cpu", "optimize=True not compatible with cuda devices, i.e. use device='cpu'"
    if self.args.int8 and tflite:
        assert not getattr(model, "end2end", False), "TFLite INT8 export not supported for end2end models."
    if edgetpu:
        if not LINUX:
            raise SystemError("Edge TPU export only supported on Linux. See https://coral.ai/docs/edgetpu/compiler")
        elif self.args.batch != 1:  # see github.com/ultralytics/ultralytics/pull/13420
            LOGGER.warning("WARNING ⚠️ Edge TPU export requires batch size 1, setting batch=1.")
            self.args.batch = 1
    if isinstance(model, WorldModel):
        LOGGER.warning(
            "WARNING ⚠️ YOLOWorld (original version) export is not supported to any format.\n"
            "WARNING ⚠️ YOLOWorldv2 models (i.e. 'yolov8s-worldv2.pt') only support export to "
            "(torchscript, onnx, openvino, engine, coreml) formats. "
            "See https://docs.ultralytics.com/models/yolo-world for details."
        )
    if self.args.int8 and not self.args.data:
        self.args.data = DEFAULT_CFG.data or TASK2DATA[getattr(model, "task", "detect")]  # assign default data
        LOGGER.warning(
            "WARNING ⚠️ INT8 export requires a missing 'data' arg for calibration. "
            f"Using default 'data={self.args.data}'."
        )

    # Input
    im = torch.zeros(self.args.batch, 3, *self.imgsz).to(self.device)
    file = Path(
        getattr(model, "pt_path", None) or getattr(model, "yaml_file", None) or model.yaml.get("yaml_file", "")
    )
    if file.suffix in {".yaml", ".yml"}:
        file = Path(file.name)

    # Update model
    model = deepcopy(model).to(self.device)
    for p in model.parameters():
        p.requires_grad = False
    model.eval()
    model.float()
    model = model.fuse()

    if imx:
        from ultralytics.utils.torch_utils import FXModel

        model = FXModel(model)
    for m in model.modules():
        if isinstance(m, Classify):
            m.export = True
        if isinstance(m, (Detect, RTDETRDecoder)):  # includes all Detect subclasses like Segment, Pose, OBB
            m.dynamic = self.args.dynamic
            m.export = True
            m.format = self.args.format
            m.max_det = self.args.max_det
        elif isinstance(m, C2f) and not is_tf_format:
            # EdgeTPU does not support FlexSplitV while split provides cleaner ONNX graph
            m.forward = m.forward_split
        if isinstance(m, Detect) and imx:
            from ultralytics.utils.tal import make_anchors

            m.anchors, m.strides = (
                x.transpose(0, 1)
                for x in make_anchors(
                    torch.cat([s / m.stride.unsqueeze(-1) for s in self.imgsz], dim=1), m.stride, 0.5
                )
            )

    y = None
    for _ in range(2):
        y = model(im)  # dry runs
    if self.args.half and onnx and self.device.type != "cpu":
        im, model = im.half(), model.half()  # to FP16

    # Filter warnings
    warnings.filterwarnings("ignore", category=torch.jit.TracerWarning)  # suppress TracerWarning
    warnings.filterwarnings("ignore", category=UserWarning)  # suppress shape prim::Constant missing ONNX warning
    warnings.filterwarnings("ignore", category=DeprecationWarning)  # suppress CoreML np.bool deprecation warning

    # Assign
    self.im = im
    self.model = model
    self.file = file
    self.output_shape = (
        tuple(y.shape)
        if isinstance(y, torch.Tensor)
        else tuple(tuple(x.shape if isinstance(x, torch.Tensor) else []) for x in y)
    )
    self.pretty_name = Path(self.model.yaml.get("yaml_file", self.file)).stem.replace("yolo", "YOLO")
    data = model.args["data"] if hasattr(model, "args") and isinstance(model.args, dict) else ""
    description = f'Ultralytics {self.pretty_name} model {f"trained on {data}" if data else ""}'
    self.metadata = {
        "description": description,
        "author": "Ultralytics",
        "date": datetime.now().isoformat(),
        "version": __version__,
        "license": "AGPL-3.0 License (https://ultralytics.com/license)",
        "docs": "https://docs.ultralytics.com",
        "stride": int(max(model.stride)),
        "task": model.task,
        "batch": self.args.batch,
        "imgsz": self.imgsz,
        "names": model.names,
    }  # model metadata
    if model.task == "pose":
        self.metadata["kpt_shape"] = model.model[-1].kpt_shape

    LOGGER.info(
        f"\n{colorstr('PyTorch:')} starting from '{file}' with input shape {tuple(im.shape)} BCHW and "
        f'output shape(s) {self.output_shape} ({file_size(file):.1f} MB)'
    )

    # Exports
    f = [""] * len(fmts)  # exported filenames
    if jit or ncnn:  # TorchScript
        f[0], _ = self.export_torchscript()
    if engine:  # TensorRT required before ONNX
        f[1], _ = self.export_engine(dla=dla)
    if onnx:  # ONNX
        f[2], _ = self.export_onnx()
    if xml:  # OpenVINO
        f[3], _ = self.export_openvino()
    if coreml:  # CoreML
        f[4], _ = self.export_coreml()
    if is_tf_format:  # TensorFlow formats
        self.args.int8 |= edgetpu
        f[5], keras_model = self.export_saved_model()
        if pb or tfjs:  # pb prerequisite to tfjs
            f[6], _ = self.export_pb(keras_model=keras_model)
        if tflite:
            f[7], _ = self.export_tflite(keras_model=keras_model, nms=False, agnostic_nms=self.args.agnostic_nms)
        if edgetpu:
            f[8], _ = self.export_edgetpu(tflite_model=Path(f[5]) / f"{self.file.stem}_full_integer_quant.tflite")
        if tfjs:
            f[9], _ = self.export_tfjs()
    if paddle:  # PaddlePaddle
        f[10], _ = self.export_paddle()
    if mnn:  # MNN
        f[11], _ = self.export_mnn()
    if ncnn:  # NCNN
        f[12], _ = self.export_ncnn()
    if imx:
        f[13], _ = self.export_imx()

    # Finish
    f = [str(x) for x in f if x]  # filter out '' and None
    if any(f):
        f = str(Path(f[-1]))
        square = self.imgsz[0] == self.imgsz[1]
        s = (
            ""
            if square
            else f"WARNING ⚠️ non-PyTorch val requires square images, 'imgsz={self.imgsz}' will not "
            f"work. Use export 'imgsz={max(self.imgsz)}' if val is required."
        )
        imgsz = self.imgsz[0] if square else str(self.imgsz)[1:-1].replace(" ", "")
        predict_data = f"data={data}" if model.task == "segment" and fmt == "pb" else ""
        q = "int8" if self.args.int8 else "half" if self.args.half else ""  # quantization
        LOGGER.info(
            f'\nExport complete ({time.time() - t:.1f}s)'
            f"\nResults saved to {colorstr('bold', file.parent.resolve())}"
            f'\nPredict:         yolo predict task={model.task} model={f} imgsz={imgsz} {q} {predict_data}'
            f'\nValidate:        yolo val task={model.task} model={f} imgsz={imgsz} data={data} {q} {s}'
            f'\nVisualize:       https://netron.app'
        )

    self.run_callbacks("on_export_end")
    return f  # return list of exported files/dirs

def add_callback(self, event: str, callback):
    """Appends the given callback."""
    self.callbacks[event].append(callback)

@try_export
def export_coreml(self, prefix=colorstr("CoreML:")):
    """YOLO CoreML export."""
    mlmodel = self.args.format.lower() == "mlmodel"  # legacy *.mlmodel export format requested
    check_requirements("coremltools>=6.0,<=6.2" if mlmodel else "coremltools>=7.0")
    import coremltools as ct  # noqa

    LOGGER.info(f"\n{prefix} starting export with coremltools {ct.__version__}...")
    assert not WINDOWS, "CoreML export is not supported on Windows, please run on macOS or Linux."
    assert self.args.batch == 1, "CoreML batch sizes > 1 are not supported. Please retry at 'batch=1'."
    f = self.file.with_suffix(".mlmodel" if mlmodel else ".mlpackage")
    if f.is_dir():
        shutil.rmtree(f)
    if self.args.nms and getattr(self.model, "end2end", False):
        LOGGER.warning(f"{prefix} WARNING ⚠️ 'nms=True' is not available for end2end models. Forcing 'nms=False'.")
        self.args.nms = False

    bias = [0.0, 0.0, 0.0]
    scale = 1 / 255
    classifier_config = None
    if self.model.task == "classify":
        classifier_config = ct.ClassifierConfig(list(self.model.names.values())) if self.args.nms else None
        model = self.model
    elif self.model.task == "detect":
        model = IOSDetectModel(self.model, self.im) if self.args.nms else self.model
    else:
        if self.args.nms:
            LOGGER.warning(f"{prefix} WARNING ⚠️ 'nms=True' is only available for Detect models like 'yolov8n.pt'.")
            # TODO CoreML Segment and Pose model pipelining
        model = self.model

    ts = torch.jit.trace(model.eval(), self.im, strict=False)  # TorchScript model
    ct_model = ct.convert(
        ts,
        inputs=[ct.ImageType("image", shape=self.im.shape, scale=scale, bias=bias)],
        classifier_config=classifier_config,
        convert_to="neuralnetwork" if mlmodel else "mlprogram",
    )
    bits, mode = (8, "kmeans") if self.args.int8 else (16, "linear") if self.args.half else (32, None)
    if bits < 32:
        if "kmeans" in mode:
            check_requirements("scikit-learn")  # scikit-learn package required for k-means quantization
        if mlmodel:
            ct_model = ct.models.neural_network.quantization_utils.quantize_weights(ct_model, bits, mode)
        elif bits == 8:  # mlprogram already quantized to FP16
            import coremltools.optimize.coreml as cto

            op_config = cto.OpPalettizerConfig(mode="kmeans", nbits=bits, weight_threshold=512)
            config = cto.OptimizationConfig(global_config=op_config)
            ct_model = cto.palettize_weights(ct_model, config=config)
    if self.args.nms and self.model.task == "detect":
        if mlmodel:
            # coremltools<=6.2 NMS export requires Python<3.11
            check_version(PYTHON_VERSION, "<3.11", name="Python ", hard=True)
            weights_dir = None
        else:
            ct_model.save(str(f))  # save otherwise weights_dir does not exist
            weights_dir = str(f / "Data/com.apple.CoreML/weights")
        ct_model = self._pipeline_coreml(ct_model, weights_dir=weights_dir)

    m = self.metadata  # metadata dict
    ct_model.short_description = m.pop("description")
    ct_model.author = m.pop("author")
    ct_model.license = m.pop("license")
    ct_model.version = m.pop("version")
    ct_model.user_defined_metadata.update({k: str(v) for k, v in m.items()})
    try:
        ct_model.save(str(f))  # save *.mlpackage
    except Exception as e:
        LOGGER.warning(
            f"{prefix} WARNING ⚠️ CoreML export to *.mlpackage failed ({e}), reverting to *.mlmodel export. "
            f"Known coremltools Python 3.11 and Windows bugs https://github.com/apple/coremltools/issues/1928."
        )
        f = f.with_suffix(".mlmodel")
        ct_model.save(str(f))
    return f, ct_model

@try_export
def export_edgetpu(self, tflite_model="", prefix=colorstr("Edge TPU:")):
    """YOLO Edge TPU export https://coral.ai/docs/edgetpu/models-intro/."""
    LOGGER.warning(f"{prefix} WARNING ⚠️ Edge TPU known bug https://github.com/ultralytics/ultralytics/issues/1185")

    cmd = "edgetpu_compiler --version"
    help_url = "https://coral.ai/docs/edgetpu/compiler/"
    assert LINUX, f"export only supported on Linux. See {help_url}"
    if subprocess.run(cmd, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL, shell=True).returncode != 0:
        LOGGER.info(f"\n{prefix} export requires Edge TPU compiler. Attempting install from {help_url}")
        sudo = subprocess.run("sudo --version >/dev/null", shell=True).returncode == 0  # sudo installed on system
        for c in (
            "curl https://packages.cloud.google.com/apt/doc/apt-key.gpg | sudo apt-key add -",
            'echo "deb https://packages.cloud.google.com/apt coral-edgetpu-stable main" | '
            "sudo tee /etc/apt/sources.list.d/coral-edgetpu.list",
            "sudo apt-get update",
            "sudo apt-get install edgetpu-compiler",
        ):
            subprocess.run(c if sudo else c.replace("sudo ", ""), shell=True, check=True)
    ver = subprocess.run(cmd, shell=True, capture_output=True, check=True).stdout.decode().split()[-1]

    LOGGER.info(f"\n{prefix} starting export with Edge TPU compiler {ver}...")
    f = str(tflite_model).replace(".tflite", "_edgetpu.tflite")  # Edge TPU model

    cmd = (
        "edgetpu_compiler "
        f'--out_dir "{Path(f).parent}" '
        "--show_operations "
        "--search_delegate "
        "--delegate_search_step 30 "
        "--timeout_sec 180 "
        f'"{tflite_model}"'
    )
    LOGGER.info(f"{prefix} running '{cmd}'")
    subprocess.run(cmd, shell=True)
    self._add_tflite_metadata(f)
    return f, None

@try_export
def export_engine(self, dla=None, prefix=colorstr("TensorRT:")):
    """YOLO TensorRT export https://developer.nvidia.com/tensorrt."""
    assert self.im.device.type != "cpu", "export running on CPU but must be on GPU, i.e. use 'device=0'"
    f_onnx, _ = self.export_onnx()  # run before TRT import https://github.com/ultralytics/ultralytics/issues/7016

    try:
        import tensorrt as trt  # noqa
    except ImportError:
        if LINUX:
            check_requirements("tensorrt>7.0.0,!=10.1.0")
        import tensorrt as trt  # noqa
    check_version(trt.__version__, ">=7.0.0", hard=True)
    check_version(trt.__version__, "!=10.1.0", msg="https://github.com/ultralytics/ultralytics/pull/14239")

    # Setup and checks
    LOGGER.info(f"\n{prefix} starting export with TensorRT {trt.__version__}...")
    is_trt10 = int(trt.__version__.split(".")[0]) >= 10  # is TensorRT >= 10
    assert Path(f_onnx).exists(), f"failed to export ONNX file: {f_onnx}"
    f = self.file.with_suffix(".engine")  # TensorRT engine file
    logger = trt.Logger(trt.Logger.INFO)
    if self.args.verbose:
        logger.min_severity = trt.Logger.Severity.VERBOSE

    # Engine builder
    builder = trt.Builder(logger)
    config = builder.create_builder_config()
    workspace = int(self.args.workspace * (1 << 30)) if self.args.workspace is not None else 0
    if is_trt10 and workspace > 0:
        config.set_memory_pool_limit(trt.MemoryPoolType.WORKSPACE, workspace)
    elif workspace > 0 and not is_trt10:  # TensorRT versions 7, 8
        config.max_workspace_size = workspace
    flag = 1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)
    network = builder.create_network(flag)
    half = builder.platform_has_fast_fp16 and self.args.half
    int8 = builder.platform_has_fast_int8 and self.args.int8

    # Optionally switch to DLA if enabled
    if dla is not None:
        if not IS_JETSON:
            raise ValueError("DLA is only available on NVIDIA Jetson devices")
        LOGGER.info(f"{prefix} enabling DLA on core {dla}...")
        if not self.args.half and not self.args.int8:
            raise ValueError(
                "DLA requires either 'half=True' (FP16) or 'int8=True' (INT8) to be enabled. Please enable one of them and try again."
            )
        config.default_device_type = trt.DeviceType.DLA
        config.DLA_core = int(dla)
        config.set_flag(trt.BuilderFlag.GPU_FALLBACK)

    # Read ONNX file
    parser = trt.OnnxParser(network, logger)
    if not parser.parse_from_file(f_onnx):
        raise RuntimeError(f"failed to load ONNX file: {f_onnx}")

    # Network inputs
    inputs = [network.get_input(i) for i in range(network.num_inputs)]
    outputs = [network.get_output(i) for i in range(network.num_outputs)]
    for inp in inputs:
        LOGGER.info(f'{prefix} input "{inp.name}" with shape{inp.shape} {inp.dtype}')
    for out in outputs:
        LOGGER.info(f'{prefix} output "{out.name}" with shape{out.shape} {out.dtype}')

    if self.args.dynamic:
        shape = self.im.shape
        if shape[0] <= 1:
            LOGGER.warning(f"{prefix} WARNING ⚠️ 'dynamic=True' model requires max batch size, i.e. 'batch=16'")
        profile = builder.create_optimization_profile()
        min_shape = (1, shape[1], 32, 32)  # minimum input shape
        max_shape = (*shape[:2], *(int(max(1, workspace) * d) for d in shape[2:]))  # max input shape
        for inp in inputs:
            profile.set_shape(inp.name, min=min_shape, opt=shape, max=max_shape)
        config.add_optimization_profile(profile)

    LOGGER.info(f"{prefix} building {'INT8' if int8 else 'FP' + ('16' if half else '32')} engine as {f}")
    if int8:
        config.set_flag(trt.BuilderFlag.INT8)
        config.set_calibration_profile(profile)
        config.profiling_verbosity = trt.ProfilingVerbosity.DETAILED

        class EngineCalibrator(trt.IInt8Calibrator):
            def __init__(
                self,
                dataset,  # ultralytics.data.build.InfiniteDataLoader
                batch: int,
                cache: str = "",
            ) -> None:
                trt.IInt8Calibrator.__init__(self)
                self.dataset = dataset
                self.data_iter = iter(dataset)
                self.algo = trt.CalibrationAlgoType.ENTROPY_CALIBRATION_2
                self.batch = batch
                self.cache = Path(cache)

            def get_algorithm(self) -> trt.CalibrationAlgoType:
                """Get the calibration algorithm to use."""
                return self.algo

            def get_batch_size(self) -> int:
                """Get the batch size to use for calibration."""
                return self.batch or 1

            def get_batch(self, names) -> list:
                """Get the next batch to use for calibration, as a list of device memory pointers."""
                try:
                    im0s = next(self.data_iter)["img"] / 255.0
                    im0s = im0s.to("cuda") if im0s.device.type == "cpu" else im0s
                    return [int(im0s.data_ptr())]
                except StopIteration:
                    # Return [] or None, signal to TensorRT there is no calibration data remaining
                    return None

            def read_calibration_cache(self) -> bytes:
                """Use existing cache instead of calibrating again, otherwise, implicitly return None."""
                if self.cache.exists() and self.cache.suffix == ".cache":
                    return self.cache.read_bytes()

            def write_calibration_cache(self, cache) -> None:
                """Write calibration cache to disk."""
                _ = self.cache.write_bytes(cache)

        # Load dataset w/ builder (for batching) and calibrate
        config.int8_calibrator = EngineCalibrator(
            dataset=self.get_int8_calibration_dataloader(prefix),
            batch=2 * self.args.batch,  # TensorRT INT8 calibration should use 2x batch size
            cache=str(self.file.with_suffix(".cache")),
        )

    elif half:
        config.set_flag(trt.BuilderFlag.FP16)

    # Free CUDA memory
    del self.model
    gc.collect()
    torch.cuda.empty_cache()

    # Write file
    build = builder.build_serialized_network if is_trt10 else builder.build_engine
    with build(network, config) as engine, open(f, "wb") as t:
        # Metadata
        meta = json.dumps(self.metadata)
        t.write(len(meta).to_bytes(4, byteorder="little", signed=True))
        t.write(meta.encode())
        # Model
        t.write(engine if is_trt10 else engine.serialize())

    return f, None

@try_export
def export_imx(self, prefix=colorstr("IMX:")):
    """YOLO IMX export."""
    gptq = False
    assert LINUX, "export only supported on Linux. See https://developer.aitrios.sony-semicon.com/en/raspberrypi-ai-camera/documentation/imx500-converter"
    if getattr(self.model, "end2end", False):
        raise ValueError("IMX export is not supported for end2end models.")
    if "C2f" not in self.model.__str__():
        raise ValueError("IMX export is only supported for YOLOv8 detection models")
    check_requirements(("model-compression-toolkit==2.1.1", "sony-custom-layers==0.2.0", "tensorflow==2.12.0"))
    check_requirements("imx500-converter[pt]==3.14.3")  # Separate requirements for imx500-converter

    import model_compression_toolkit as mct
    import onnx
    from sony_custom_layers.pytorch.object_detection.nms import multiclass_nms

    try:
        out = subprocess.run(
            ["java", "--version"], check=True, capture_output=True
        )  # Java 17 is required for imx500-converter
        if "openjdk 17" not in str(out.stdout):
            raise FileNotFoundError
    except FileNotFoundError:
        subprocess.run(["sudo", "apt", "install", "-y", "openjdk-17-jdk", "openjdk-17-jre"], check=True)

    def representative_dataset_gen(dataloader=self.get_int8_calibration_dataloader(prefix)):
        for batch in dataloader:
            img = batch["img"]
            img = img / 255.0
            yield [img]

    tpc = mct.get_target_platform_capabilities(
        fw_name="pytorch", target_platform_name="imx500", target_platform_version="v1"
    )

    config = mct.core.CoreConfig(
        mixed_precision_config=mct.core.MixedPrecisionQuantizationConfig(num_of_images=10),
        quantization_config=mct.core.QuantizationConfig(concat_threshold_update=True),
    )

    resource_utilization = mct.core.ResourceUtilization(weights_memory=3146176 * 0.76)

    quant_model = (
        mct.gptq.pytorch_gradient_post_training_quantization(  # Perform Gradient-Based Post Training Quantization
            model=self.model,
            representative_data_gen=representative_dataset_gen,
            target_resource_utilization=resource_utilization,
            gptq_config=mct.gptq.get_pytorch_gptq_config(n_epochs=1000, use_hessian_based_weights=False),
            core_config=config,
            target_platform_capabilities=tpc,
        )[0]
        if gptq
        else mct.ptq.pytorch_post_training_quantization(  # Perform post training quantization
            in_module=self.model,
            representative_data_gen=representative_dataset_gen,
            target_resource_utilization=resource_utilization,
            core_config=config,
            target_platform_capabilities=tpc,
        )[0]
    )

    class NMSWrapper(torch.nn.Module):
        def __init__(
            self,
            model: torch.nn.Module,
            score_threshold: float = 0.001,
            iou_threshold: float = 0.7,
            max_detections: int = 300,
        ):
            """
            Wrapping PyTorch Module with multiclass_nms layer from sony_custom_layers.

            Args:
                model (nn.Module): Model instance.
                score_threshold (float): Score threshold for non-maximum suppression.
                iou_threshold (float): Intersection over union threshold for non-maximum suppression.
                max_detections (float): The number of detections to return.
            """
            super().__init__()
            self.model = model
            self.score_threshold = score_threshold
            self.iou_threshold = iou_threshold
            self.max_detections = max_detections

        def forward(self, images):
            # model inference
            outputs = self.model(images)

            boxes = outputs[0]
            scores = outputs[1]
            nms = multiclass_nms(
                boxes=boxes,
                scores=scores,
                score_threshold=self.score_threshold,
                iou_threshold=self.iou_threshold,
                max_detections=self.max_detections,
            )
            return nms

    quant_model = NMSWrapper(
        model=quant_model,
        score_threshold=self.args.conf or 0.001,
        iou_threshold=self.args.iou,
        max_detections=self.args.max_det,
    ).to(self.device)

    f = Path(str(self.file).replace(self.file.suffix, "_imx_model"))
    f.mkdir(exist_ok=True)
    onnx_model = f / Path(str(self.file).replace(self.file.suffix, "_imx.onnx"))  # js dir
    mct.exporter.pytorch_export_model(
        model=quant_model, save_model_path=onnx_model, repr_dataset=representative_dataset_gen
    )

    model_onnx = onnx.load(onnx_model)  # load onnx model
    for k, v in self.metadata.items():
        meta = model_onnx.metadata_props.add()
        meta.key, meta.value = k, str(v)

    onnx.save(model_onnx, onnx_model)

    subprocess.run(
        ["imxconv-pt", "-i", str(onnx_model), "-o", str(f), "--no-input-persistency", "--overwrite-output"],
        check=True,
    )

    # Needed for imx models.
    with open(f / "labels.txt", "w") as file:
        file.writelines([f"{name}\n" for _, name in self.model.names.items()])

    return f, None

@try_export
def export_mnn(self, prefix=colorstr("MNN:")):
    """YOLOv8 MNN export using MNN https://github.com/alibaba/MNN."""
    f_onnx, _ = self.export_onnx()  # get onnx model first

    check_requirements("MNN>=2.9.6")
    import MNN  # noqa
    from MNN.tools import mnnconvert

    # Setup and checks
    LOGGER.info(f"\n{prefix} starting export with MNN {MNN.version()}...")
    assert Path(f_onnx).exists(), f"failed to export ONNX file: {f_onnx}"
    f = str(self.file.with_suffix(".mnn"))  # MNN model file
    args = ["", "-f", "ONNX", "--modelFile", f_onnx, "--MNNModel", f, "--bizCode", json.dumps(self.metadata)]
    if self.args.int8:
        args.extend(("--weightQuantBits", "8"))
    if self.args.half:
        args.append("--fp16")
    mnnconvert.convert(args)
    # remove scratch file for model convert optimize
    convert_scratch = Path(self.file.parent / ".__convert_external_data.bin")
    if convert_scratch.exists():
        convert_scratch.unlink()
    return f, None

@try_export
def export_ncnn(self, prefix=colorstr("NCNN:")):
    """YOLO NCNN export using PNNX https://github.com/pnnx/pnnx."""
    check_requirements("ncnn")
    import ncnn  # noqa

    LOGGER.info(f"\n{prefix} starting export with NCNN {ncnn.__version__}...")
    f = Path(str(self.file).replace(self.file.suffix, f"_ncnn_model{os.sep}"))
    f_ts = self.file.with_suffix(".torchscript")

    name = Path("pnnx.exe" if WINDOWS else "pnnx")  # PNNX filename
    pnnx = name if name.is_file() else (ROOT / name)
    if not pnnx.is_file():
        LOGGER.warning(
            f"{prefix} WARNING ⚠️ PNNX not found. Attempting to download binary file from "
            "https://github.com/pnnx/pnnx/.\nNote PNNX Binary file must be placed in current working directory "
            f"or in {ROOT}. See PNNX repo for full installation instructions."
        )
        system = "macos" if MACOS else "windows" if WINDOWS else "linux-aarch64" if ARM64 else "linux"
        try:
            release, assets = get_github_assets(repo="pnnx/pnnx")
            asset = [x for x in assets if f"{system}.zip" in x][0]
            assert isinstance(asset, str), "Unable to retrieve PNNX repo assets"  # i.e. pnnx-20240410-macos.zip
            LOGGER.info(f"{prefix} successfully found latest PNNX asset file {asset}")
        except Exception as e:
            release = "20240410"
            asset = f"pnnx-{release}-{system}.zip"
            LOGGER.warning(f"{prefix} WARNING ⚠️ PNNX GitHub assets not found: {e}, using default {asset}")
        unzip_dir = safe_download(f"https://github.com/pnnx/pnnx/releases/download/{release}/{asset}", delete=True)
        if check_is_path_safe(Path.cwd(), unzip_dir):  # avoid path traversal security vulnerability
            shutil.move(src=unzip_dir / name, dst=pnnx)  # move binary to ROOT
            pnnx.chmod(0o777)  # set read, write, and execute permissions for everyone
            shutil.rmtree(unzip_dir)  # delete unzip dir

    ncnn_args = [
        f'ncnnparam={f / "model.ncnn.param"}',
        f'ncnnbin={f / "model.ncnn.bin"}',
        f'ncnnpy={f / "model_ncnn.py"}',
    ]

    pnnx_args = [
        f'pnnxparam={f / "model.pnnx.param"}',
        f'pnnxbin={f / "model.pnnx.bin"}',
        f'pnnxpy={f / "model_pnnx.py"}',
        f'pnnxonnx={f / "model.pnnx.onnx"}',
    ]

    cmd = [
        str(pnnx),
        str(f_ts),
        *ncnn_args,
        *pnnx_args,
        f"fp16={int(self.args.half)}",
        f"device={self.device.type}",
        f'inputshape="{[self.args.batch, 3, *self.imgsz]}"',
    ]
    f.mkdir(exist_ok=True)  # make ncnn_model directory
    LOGGER.info(f"{prefix} running '{' '.join(cmd)}'")
    subprocess.run(cmd, check=True)

    # Remove debug files
    pnnx_files = [x.split("=")[-1] for x in pnnx_args]
    for f_debug in ("debug.bin", "debug.param", "debug2.bin", "debug2.param", *pnnx_files):
        Path(f_debug).unlink(missing_ok=True)

    yaml_save(f / "metadata.yaml", self.metadata)  # add metadata.yaml
    return str(f), None