dgenerate-ultralytics-headless 8.3.196__py3-none-any.whl → 8.3.248__py3-none-any.whl

ultralytics/engine/exporter.py

@@ -20,6 +20,8 @@ MNN                     | `mnn`                     | yolo11n.mnn
 NCNN                    | `ncnn`                    | yolo11n_ncnn_model/
 IMX                     | `imx`                     | yolo11n_imx_model/
 RKNN                    | `rknn`                    | yolo11n_rknn_model/
+ExecuTorch              | `executorch`              | yolo11n_executorch_model/
+Axelera                 | `axelera`                 | yolo11n_axelera_model/
 
 Requirements:
     $ pip install "ultralytics[export]"
@@ -48,6 +50,8 @@ Inference:
                          yolo11n_ncnn_model         # NCNN
                          yolo11n_imx_model          # IMX
                          yolo11n_rknn_model         # RKNN
+                         yolo11n_executorch_model   # ExecuTorch
+                         yolo11n_axelera_model      # Axelera
 
 TensorFlow.js:
     $ cd .. && git clone https://github.com/zldrobit/tfjs-yolov5-example.git && cd tfjs-yolov5-example
@@ -62,7 +66,6 @@ import re
 import shutil
 import subprocess
 import time
-import warnings
 from copy import deepcopy
 from datetime import datetime
 from pathlib import Path
@@ -82,13 +85,17 @@ from ultralytics.utils import (
     ARM64,
     DEFAULT_CFG,
     IS_COLAB,
+    IS_DEBIAN_BOOKWORM,
+    IS_DEBIAN_TRIXIE,
+    IS_DOCKER,
     IS_JETSON,
+    IS_RASPBERRYPI,
+    IS_UBUNTU,
     LINUX,
     LOGGER,
     MACOS,
     MACOS_VERSION,
     RKNN_CHIPS,
-    ROOT,
     SETTINGS,
     TORCH_VERSION,
     WINDOWS,
@@ -98,21 +105,38 @@ from ultralytics.utils import (
     get_default_args,
 )
 from ultralytics.utils.checks import (
+    IS_PYTHON_3_10,
+    IS_PYTHON_MINIMUM_3_9,
+    check_apt_requirements,
     check_imgsz,
-    check_is_path_safe,
     check_requirements,
     check_version,
     is_intel,
     is_sudo_available,
 )
-from ultralytics.utils.downloads import attempt_download_asset, get_github_assets, safe_download
-from ultralytics.utils.export import export_engine, export_onnx
-from ultralytics.utils.files import file_size, spaces_in_path
+from ultralytics.utils.export import (
+    keras2pb,
+    onnx2engine,
+    onnx2saved_model,
+    pb2tfjs,
+    tflite2edgetpu,
+    torch2imx,
+    torch2onnx,
+)
+from ultralytics.utils.files import file_size
 from ultralytics.utils.metrics import batch_probiou
 from ultralytics.utils.nms import TorchNMS
 from ultralytics.utils.ops import Profile
 from ultralytics.utils.patches import arange_patch
-from ultralytics.utils.torch_utils import TORCH_1_13, get_latest_opset, select_device
+from ultralytics.utils.torch_utils import (
+    TORCH_1_10,
+    TORCH_1_11,
+    TORCH_1_13,
+    TORCH_2_1,
+    TORCH_2_4,
+    TORCH_2_9,
+    select_device,
+)
 
 
 def export_formats():
@@ -137,7 +161,7 @@ def export_formats():
             True,
             ["batch", "dynamic", "half", "int8", "simplify", "nms", "fraction"],
         ],
-        ["CoreML", "coreml", ".mlpackage", True, False, ["batch", "half", "int8", "nms"]],
+        ["CoreML", "coreml", ".mlpackage", True, False, ["batch", "dynamic", "half", "int8", "nms"]],
         ["TensorFlow SavedModel", "saved_model", "_saved_model", True, True, ["batch", "int8", "keras", "nms"]],
         ["TensorFlow GraphDef", "pb", ".pb", True, True, ["batch"]],
         ["TensorFlow Lite", "tflite", ".tflite", True, False, ["batch", "half", "int8", "nms", "fraction"]],
@@ -148,13 +172,42 @@ def export_formats():
         ["NCNN", "ncnn", "_ncnn_model", True, True, ["batch", "half"]],
         ["IMX", "imx", "_imx_model", True, True, ["int8", "fraction", "nms"]],
         ["RKNN", "rknn", "_rknn_model", False, False, ["batch", "name"]],
+        ["ExecuTorch", "executorch", "_executorch_model", True, False, ["batch"]],
+        ["Axelera", "axelera", "_axelera_model", False, False, ["batch", "int8", "fraction"]],
     ]
     return dict(zip(["Format", "Argument", "Suffix", "CPU", "GPU", "Arguments"], zip(*x)))
 
 
+def best_onnx_opset(onnx, cuda=False) -> int:
+    """Return max ONNX opset for this torch version with ONNX fallback."""
+    if TORCH_2_4:  # _constants.ONNX_MAX_OPSET first defined in torch 1.13
+        opset = torch.onnx.utils._constants.ONNX_MAX_OPSET - 1  # use second-latest version for safety
+        if cuda:
+            opset -= 2  # fix CUDA ONNXRuntime NMS squeeze op errors
+    else:
+        version = ".".join(TORCH_VERSION.split(".")[:2])
+        opset = {
+            "1.8": 12,
+            "1.9": 12,
+            "1.10": 13,
+            "1.11": 14,
+            "1.12": 15,
+            "1.13": 17,
+            "2.0": 17,  # reduced from 18 to fix ONNX errors
+            "2.1": 17,  # reduced from 19
+            "2.2": 17,  # reduced from 19
+            "2.3": 17,  # reduced from 19
+            "2.4": 20,
+            "2.5": 20,
+            "2.6": 20,
+            "2.7": 20,
+            "2.8": 23,
+        }.get(version, 12)
+    return min(opset, onnx.defs.onnx_opset_version())
+
+
 def validate_args(format, passed_args, valid_args):
-    """
-    Validate arguments based on the export format.
+    """Validate arguments based on the export format.
 
     Args:
         format (str): The export format.
@@ -175,15 +228,6 @@ def validate_args(format, passed_args, valid_args):
             assert arg in valid_args, f"ERROR ❌️ argument '{arg}' is not supported for format='{format}'"
 
 
-def gd_outputs(gd):
-    """Return TensorFlow GraphDef model output node names."""
-    name_list, input_list = [], []
-    for node in gd.node:  # tensorflow.core.framework.node_def_pb2.NodeDef
-        name_list.append(node.name)
-        input_list.extend(node.input)
-    return sorted(f"{x}:0" for x in list(set(name_list) - set(input_list)) if not x.startswith("NoOp"))
-
-
 def try_export(inner_func):
     """YOLO export decorator, i.e. @try_export."""
     inner_args = get_default_args(inner_func)
@@ -194,9 +238,12 @@ def try_export(inner_func):
         dt = 0.0
         try:
             with Profile() as dt:
-                f, model = inner_func(*args, **kwargs)
-            LOGGER.info(f"{prefix} export success ✅ {dt.t:.1f}s, saved as '{f}' ({file_size(f):.1f} MB)")
-            return f, model
+                f = inner_func(*args, **kwargs)  # exported file/dir or tuple of (file/dir, *)
+            path = f if isinstance(f, (str, Path)) else f[0]
+            mb = file_size(path)
+            assert mb > 0.0, "0.0 MB output model size"
+            LOGGER.info(f"{prefix} export success ✅ {dt.t:.1f}s, saved as '{path}' ({mb:.1f} MB)")
+            return f
         except Exception as e:
             LOGGER.error(f"{prefix} export failure {dt.t:.1f}s: {e}")
             raise e
@@ -205,8 +252,7 @@ def try_export(inner_func):
 
 
 class Exporter:
-    """
-    A class for exporting YOLO models to various formats.
+    """A class for exporting YOLO models to various formats.
 
     This class provides functionality to export YOLO models to different formats including ONNX, TensorRT, CoreML,
     TensorFlow, and others. It handles format validation, device selection, model preparation, and the actual export
@@ -256,8 +302,7 @@ class Exporter:
     """
 
     def __init__(self, cfg=DEFAULT_CFG, overrides=None, _callbacks=None):
-        """
-        Initialize the Exporter class.
+        """Initialize the Exporter class.
 
         Args:
             cfg (str, optional): Path to a configuration file.
@@ -269,7 +314,11 @@ class Exporter:
         callbacks.add_integration_callbacks(self)
 
     def __call__(self, model=None) -> str:
-        """Return list of exported files/dirs after running callbacks."""
+        """Export a model and return the final exported path as a string.
+
+        Returns:
+            (str): Path to the exported file or directory (the last export artifact).
+        """
         t = time.time()
         fmt = self.args.format.lower()  # to lowercase
         if fmt in {"tensorrt", "trt"}:  # 'engine' aliases
@@ -284,15 +333,32 @@ class Exporter:
             # 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}")
+                msg = "Model is already in PyTorch format." if fmt == "pt" else f"Invalid export format='{fmt}'."
+                raise ValueError(f"{msg} Valid formats are {fmts}")
             LOGGER.warning(f"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, rknn) = (
-            flags  # export booleans
-        )
+        (
+            jit,
+            onnx,
+            xml,
+            engine,
+            coreml,
+            saved_model,
+            pb,
+            tflite,
+            edgetpu,
+            tfjs,
+            paddle,
+            mnn,
+            ncnn,
+            imx,
+            rknn,
+            executorch,
+            axelera,
+        ) = flags  # export booleans
 
         is_tf_format = any((saved_model, pb, tflite, edgetpu, tfjs))
 
@@ -302,9 +368,10 @@ class Exporter:
             LOGGER.warning("TensorRT requires GPU export, automatically assigning device=0")
             self.args.device = "0"
         if engine and "dla" in str(self.args.device):  # convert int/list to str first
-            dla = self.args.device.rsplit(":", 1)[-1]
+            device_str = str(self.args.device)
+            dla = device_str.rsplit(":", 1)[-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}."
+            assert dla in {"0", "1"}, f"Expected device 'dla:0' or 'dla:1', but got {device_str}."
         if imx and self.args.device is None and torch.cuda.is_available():
             LOGGER.warning("Exporting on CPU while CUDA is available, setting device=0 for faster export on GPU.")
             self.args.device = "0"  # update device to "0"
@@ -313,23 +380,37 @@ class Exporter:
         # Argument compatibility checks
         fmt_keys = fmts_dict["Arguments"][flags.index(True) + 1]
         validate_args(fmt, self.args, fmt_keys)
+        if axelera:
+            if not IS_PYTHON_3_10:
+                raise SystemError("Axelera export only supported on Python 3.10.")
+            if not self.args.int8:
+                LOGGER.warning("Setting int8=True for Axelera mixed-precision export.")
+                self.args.int8 = True
+            if model.task not in {"detect"}:
+                raise ValueError("Axelera export only supported for detection models.")
+            if not self.args.data:
+                self.args.data = "coco128.yaml"  # Axelera default to coco128.yaml
         if imx:
             if not self.args.int8:
                 LOGGER.warning("IMX export requires int8=True, setting int8=True.")
                 self.args.int8 = True
-            if not self.args.nms:
+            if not self.args.nms and model.task in {"detect", "pose", "segment"}:
                 LOGGER.warning("IMX export requires nms=True, setting nms=True.")
                 self.args.nms = True
-            if model.task not in {"detect", "pose"}:
-                raise ValueError("IMX export only supported for detection and pose estimation models.")
+            if model.task not in {"detect", "pose", "classify", "segment"}:
+                raise ValueError(
+                    "IMX export only supported for detection, pose estimation, classification, and segmentation models."
+                )
         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("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("half=True only compatible with GPU export, i.e. use device=0")
+        if self.args.half and jit and self.device.type == "cpu":
+            LOGGER.warning(
+                "half=True only compatible with GPU export for TorchScript, i.e. use device=0, setting half=False."
+            )
             self.args.half = False
         self.imgsz = check_imgsz(self.args.imgsz, stride=model.stride, min_dim=2)  # check image size
         if self.args.optimize:
@@ -346,18 +427,18 @@ class Exporter:
             assert self.args.name in RKNN_CHIPS, (
                 f"Invalid processor name '{self.args.name}' for Rockchip RKNN export. Valid names are {RKNN_CHIPS}."
             )
-        if self.args.int8 and tflite:
-            assert not getattr(model, "end2end", False), "TFLite INT8 export not supported for end2end models."
         if self.args.nms:
             assert not isinstance(model, ClassificationModel), "'nms=True' is not valid for classification models."
             assert not tflite or not ARM64 or not LINUX, "TFLite export with NMS unsupported on ARM64 Linux"
-            if getattr(model, "end2end", False):
+            assert not is_tf_format or TORCH_1_13, "TensorFlow exports with NMS require torch>=1.13"
+            assert not onnx or TORCH_1_13, "ONNX export with NMS requires torch>=1.13"
+            if getattr(model, "end2end", False) or isinstance(model.model[-1], RTDETRDecoder):
                 LOGGER.warning("'nms=True' is not available for end2end models. Forcing 'nms=False'.")
                 self.args.nms = False
             self.args.conf = self.args.conf or 0.25  # set conf default value for nms export
-        if (engine or self.args.nms) and self.args.dynamic and self.args.batch == 1:
+        if (engine or coreml or self.args.nms) and self.args.dynamic and self.args.batch == 1:
             LOGGER.warning(
-                f"'dynamic=True' model with '{'nms=True' if self.args.nms else 'format=engine'}' requires max batch size, i.e. 'batch=16'"
+                f"'dynamic=True' model with '{'nms=True' if self.args.nms else f'format={self.args.format}'}' requires max batch size, i.e. 'batch=16'"
             )
         if edgetpu:
             if not LINUX or ARM64:
@@ -408,9 +489,13 @@ class Exporter:
         model = model.fuse()
 
         if imx:
-            from ultralytics.utils.torch_utils import FXModel
+            from ultralytics.utils.export.imx import FXModel
 
-            model = FXModel(model)
+            model = FXModel(model, self.imgsz)
+        if tflite or edgetpu:
+            from ultralytics.utils.export.tensorflow import tf_wrapper
+
+            model = tf_wrapper(model)
         for m in model.modules():
             if isinstance(m, Classify):
                 m.export = True
@@ -425,27 +510,13 @@ class Exporter:
             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):  # dry runs
             y = NMSModel(model, self.args)(im) if self.args.nms and not coreml and not imx else model(im)
-        if self.args.half and onnx and self.device.type != "cpu":
+        if self.args.half and (onnx or jit) 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
@@ -477,6 +548,8 @@ class Exporter:
             self.metadata["dla"] = dla  # make sure `AutoBackend` uses correct dla device if it has one
         if model.task == "pose":
             self.metadata["kpt_shape"] = model.model[-1].kpt_shape
+            if hasattr(model, "kpt_names"):
+                self.metadata["kpt_names"] = model.kpt_names
 
         LOGGER.info(
             f"\n{colorstr('PyTorch:')} starting from '{file}' with input shape {tuple(im.shape)} BCHW and "
@@ -485,37 +558,41 @@ class Exporter:
         self.run_callbacks("on_export_start")
         # Exports
         f = [""] * len(fmts)  # exported filenames
-        if jit or ncnn:  # TorchScript
-            f[0], _ = self.export_torchscript()
+        if jit:  # TorchScript
+            f[0] = self.export_torchscript()
         if engine:  # TensorRT required before ONNX
-            f[1], _ = self.export_engine(dla=dla)
+            f[1] = self.export_engine(dla=dla)
         if onnx:  # ONNX
-            f[2], _ = self.export_onnx()
+            f[2] = self.export_onnx()
         if xml:  # OpenVINO
-            f[3], _ = self.export_openvino()
+            f[3] = self.export_openvino()
         if coreml:  # CoreML
-            f[4], _ = self.export_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)
+                f[6] = self.export_pb(keras_model=keras_model)
             if tflite:
-                f[7], _ = self.export_tflite()
+                f[7] = self.export_tflite()
             if edgetpu:
-                f[8], _ = self.export_edgetpu(tflite_model=Path(f[5]) / f"{self.file.stem}_full_integer_quant.tflite")
+                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()
+                f[9] = self.export_tfjs()
         if paddle:  # PaddlePaddle
-            f[10], _ = self.export_paddle()
+            f[10] = self.export_paddle()
         if mnn:  # MNN
-            f[11], _ = self.export_mnn()
+            f[11] = self.export_mnn()
         if ncnn:  # NCNN
-            f[12], _ = self.export_ncnn()
+            f[12] = self.export_ncnn()
         if imx:
-            f[13], _ = self.export_imx()
+            f[13] = self.export_imx()
         if rknn:
-            f[14], _ = self.export_rknn()
+            f[14] = self.export_rknn()
+        if executorch:
+            f[15] = self.export_executorch()
+        if axelera:
+            f[16] = self.export_axelera()
 
         # Finish
         f = [str(x) for x in f if x]  # filter out '' and None
@@ -540,7 +617,7 @@ class Exporter:
             )
 
         self.run_callbacks("on_export_end")
-        return f  # return list of exported files/dirs
+        return f  # path to final export artifact
 
     def get_int8_calibration_dataloader(self, prefix=""):
         """Build and return a dataloader for calibration of INT8 models."""
@@ -561,7 +638,9 @@ class Exporter:
                 f"The calibration dataset ({n} images) must have at least as many images as the batch size "
                 f"('batch={self.args.batch}')."
             )
-        elif n < 300:
+        elif self.args.format == "axelera" and n < 100:
+            LOGGER.warning(f"{prefix} >100 images required for Axelera calibration, found {n} images.")
+        elif self.args.format != "axelera" and n < 300:
             LOGGER.warning(f"{prefix} >300 images recommended for INT8 calibration, found {n} images.")
         return build_dataloader(dataset, batch=self.args.batch, workers=0, drop_last=True)  # required for batch loading
 
@@ -580,21 +659,24 @@ class Exporter:
             optimize_for_mobile(ts)._save_for_lite_interpreter(str(f), _extra_files=extra_files)
         else:
             ts.save(str(f), _extra_files=extra_files)
-        return f, None
+        return f
 
     @try_export
     def export_onnx(self, prefix=colorstr("ONNX:")):
         """Export YOLO model to ONNX format."""
-        requirements = ["onnx>=1.12.0"]
+        requirements = ["onnx>=1.12.0,<2.0.0"]
         if self.args.simplify:
-            requirements += ["onnxslim>=0.1.67", "onnxruntime" + ("-gpu" if torch.cuda.is_available() else "")]
+            requirements += ["onnxslim>=0.1.71", "onnxruntime" + ("-gpu" if torch.cuda.is_available() else "")]
         check_requirements(requirements)
-        import onnx  # noqa
+        import onnx
+
+        opset = self.args.opset or best_onnx_opset(onnx, cuda="cuda" in self.device.type)
+        LOGGER.info(f"\n{prefix} starting export with onnx {onnx.__version__} opset {opset}...")
+        if self.args.nms:
+            assert TORCH_1_13, f"'nms=True' ONNX export requires torch>=1.13 (found torch=={TORCH_VERSION})"
 
-        opset_version = self.args.opset or get_latest_opset()
-        LOGGER.info(f"\n{prefix} starting export with onnx {onnx.__version__} opset {opset_version}...")
         f = str(self.file.with_suffix(".onnx"))
-        output_names = ["output0", "output1"] if isinstance(self.model, SegmentationModel) else ["output0"]
+        output_names = ["output0", "output1"] if self.model.task == "segment" else ["output0"]
         dynamic = self.args.dynamic
         if dynamic:
             dynamic = {"images": {0: "batch", 2: "height", 3: "width"}}  # shape(1,3,640,640)
@@ -606,14 +688,14 @@ class Exporter:
             if self.args.nms:  # only batch size is dynamic with NMS
                 dynamic["output0"].pop(2)
         if self.args.nms and self.model.task == "obb":
-            self.args.opset = opset_version  # for NMSModel
+            self.args.opset = opset  # for NMSModel
 
         with arange_patch(self.args):
-            export_onnx(
+            torch2onnx(
                 NMSModel(self.model, self.args) if self.args.nms else self.model,
                 self.im,
                 f,
-                opset=opset_version,
+                opset=opset,
                 input_names=["images"],
                 output_names=output_names,
                 dynamic=dynamic or None,
@@ -638,8 +720,23 @@ class Exporter:
             meta = model_onnx.metadata_props.add()
             meta.key, meta.value = k, str(v)
 
+        # IR version
+        if getattr(model_onnx, "ir_version", 0) > 10:
+            LOGGER.info(f"{prefix} limiting IR version {model_onnx.ir_version} to 10 for ONNXRuntime compatibility...")
+            model_onnx.ir_version = 10
+
+        # FP16 conversion for CPU export (GPU exports are already FP16 from model.half() during tracing)
+        if self.args.half and self.args.format == "onnx" and self.device.type == "cpu":
+            try:
+                from onnxruntime.transformers import float16
+
+                LOGGER.info(f"{prefix} converting to FP16...")
+                model_onnx = float16.convert_float_to_float16(model_onnx, keep_io_types=True)
+            except Exception as e:
+                LOGGER.warning(f"{prefix} FP16 conversion failure: {e}")
+
         onnx.save(model_onnx, f)
-        return f, model_onnx
+        return f
 
     @try_export
     def export_openvino(self, prefix=colorstr("OpenVINO:")):
@@ -649,7 +746,7 @@ class Exporter:
         import openvino as ov
 
         LOGGER.info(f"\n{prefix} starting export with openvino {ov.__version__}...")
-        assert TORCH_1_13, f"OpenVINO export requires torch>=1.13.0 but torch=={TORCH_VERSION} is installed"
+        assert TORCH_2_1, f"OpenVINO export requires torch>=2.1 but torch=={TORCH_VERSION} is installed"
         ov_model = ov.convert_model(
             NMSModel(self.model, self.args) if self.args.nms else self.model,
             input=None if self.args.dynamic else [self.im.shape],
@@ -678,13 +775,6 @@ class Exporter:
             check_requirements("nncf>=2.14.0")
             import nncf
 
-            def transform_fn(data_item) -> np.ndarray:
-                """Quantization transform function."""
-                data_item: torch.Tensor = data_item["img"] if isinstance(data_item, dict) else data_item
-                assert data_item.dtype == torch.uint8, "Input image must be uint8 for the quantization preprocessing"
-                im = data_item.numpy().astype(np.float32) / 255.0  # uint8 to fp16/32 and 0-255 to 0.0-1.0
-                return np.expand_dims(im, 0) if im.ndim == 3 else im
-
             # Generate calibration data for integer quantization
             ignored_scope = None
             if isinstance(self.model.model[-1], Detect):
@@ -703,18 +793,18 @@ class Exporter:
 
             quantized_ov_model = nncf.quantize(
                 model=ov_model,
-                calibration_dataset=nncf.Dataset(self.get_int8_calibration_dataloader(prefix), transform_fn),
+                calibration_dataset=nncf.Dataset(self.get_int8_calibration_dataloader(prefix), self._transform_fn),
                 preset=nncf.QuantizationPreset.MIXED,
                 ignored_scope=ignored_scope,
             )
             serialize(quantized_ov_model, fq_ov)
-            return fq, None
+            return fq
 
         f = str(self.file).replace(self.file.suffix, f"_openvino_model{os.sep}")
         f_ov = str(Path(f) / self.file.with_suffix(".xml").name)
 
         serialize(ov_model, f_ov)
-        return f, None
+        return f
 
     @try_export
     def export_paddle(self, prefix=colorstr("PaddlePaddle:")):
@@ -730,23 +820,24 @@ class Exporter:
                 "x2paddle",
             )
         )
-        import x2paddle  # noqa
-        from x2paddle.convert import pytorch2paddle  # noqa
+        import x2paddle
+        from x2paddle.convert import pytorch2paddle
 
         LOGGER.info(f"\n{prefix} starting export with X2Paddle {x2paddle.__version__}...")
         f = str(self.file).replace(self.file.suffix, f"_paddle_model{os.sep}")
 
         pytorch2paddle(module=self.model, save_dir=f, jit_type="trace", input_examples=[self.im])  # export
         YAML.save(Path(f) / "metadata.yaml", self.metadata)  # add metadata.yaml
-        return f, None
+        return f
 
     @try_export
     def export_mnn(self, prefix=colorstr("MNN:")):
         """Export YOLO model to MNN format using MNN https://github.com/alibaba/MNN."""
-        f_onnx, _ = self.export_onnx()  # get onnx model first
+        assert TORCH_1_10, "MNN export requires torch>=1.10.0 to avoid segmentation faults"
+        f_onnx = self.export_onnx()  # get onnx model first
 
         check_requirements("MNN>=2.9.6")
-        import MNN  # noqa
+        import MNN
         from MNN.tools import mnnconvert
 
         # Setup and checks
@@ -763,98 +854,73 @@ class Exporter:
         convert_scratch = Path(self.file.parent / ".__convert_external_data.bin")
         if convert_scratch.exists():
             convert_scratch.unlink()
-        return f, None
+        return f
 
     @try_export
     def export_ncnn(self, prefix=colorstr("NCNN:")):
         """Export YOLO model to NCNN format using PNNX https://github.com/pnnx/pnnx."""
         check_requirements("ncnn", cmds="--no-deps")  # no deps to avoid installing opencv-python
-        import ncnn  # noqa
+        check_requirements("pnnx")
+        import ncnn
+        import pnnx
 
-        LOGGER.info(f"\n{prefix} starting export with NCNN {ncnn.__version__}...")
+        LOGGER.info(f"\n{prefix} starting export with NCNN {ncnn.__version__} and PNNX {pnnx.__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} 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} 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]}"',
-        ]
+        ncnn_args = dict(
+            ncnnparam=(f / "model.ncnn.param").as_posix(),
+            ncnnbin=(f / "model.ncnn.bin").as_posix(),
+            ncnnpy=(f / "model_ncnn.py").as_posix(),
+        )
+
+        pnnx_args = dict(
+            ptpath=(f / "model.pt").as_posix(),
+            pnnxparam=(f / "model.pnnx.param").as_posix(),
+            pnnxbin=(f / "model.pnnx.bin").as_posix(),
+            pnnxpy=(f / "model_pnnx.py").as_posix(),
+            pnnxonnx=(f / "model.pnnx.onnx").as_posix(),
+        )
+
         f.mkdir(exist_ok=True)  # make ncnn_model directory
-        LOGGER.info(f"{prefix} running '{' '.join(cmd)}'")
-        subprocess.run(cmd, check=True)
+        pnnx.export(self.model, inputs=self.im, **ncnn_args, **pnnx_args, fp16=self.args.half, device=self.device.type)
 
-        # Remove debug files
-        pnnx_files = [x.rsplit("=", 1)[-1] for x in pnnx_args]
-        for f_debug in ("debug.bin", "debug.param", "debug2.bin", "debug2.param", *pnnx_files):
+        for f_debug in ("debug.bin", "debug.param", "debug2.bin", "debug2.param", *pnnx_args.values()):
             Path(f_debug).unlink(missing_ok=True)
 
         YAML.save(f / "metadata.yaml", self.metadata)  # add metadata.yaml
-        return str(f), None
+        return str(f)
 
     @try_export
     def export_coreml(self, prefix=colorstr("CoreML:")):
         """Export YOLO model to CoreML format."""
         mlmodel = self.args.format.lower() == "mlmodel"  # legacy *.mlmodel export format requested
-        check_requirements("coremltools>=8.0")
-        import coremltools as ct  # noqa
+        check_requirements(
+            ["coremltools>=9.0", "numpy>=1.14.5,<=2.3.5"]
+        )  # latest numpy 2.4.0rc1 breaks coremltools exports
+        import coremltools as ct
 
         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'."
+        assert TORCH_1_11, "CoreML export requires torch>=1.11"
+        if self.args.batch > 1:
+            assert self.args.dynamic, (
+                "batch sizes > 1 are not supported without 'dynamic=True' for CoreML export. Please retry at 'dynamic=True'."
+            )
+        if self.args.dynamic:
+            assert not self.args.nms, (
+                "'nms=True' cannot be used together with 'dynamic=True' for CoreML export. Please disable one of them."
+            )
+            assert self.model.task != "classify", "'dynamic=True' is not supported for CoreML classification models."
         f = self.file.with_suffix(".mlmodel" if mlmodel else ".mlpackage")
         if f.is_dir():
             shutil.rmtree(f)
 
-        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()))
             model = self.model
         elif self.model.task == "detect":
-            model = IOSDetectModel(self.model, self.im) if self.args.nms else self.model
+            model = IOSDetectModel(self.model, self.im, mlprogram=not mlmodel) if self.args.nms else self.model
         else:
             if self.args.nms:
                 LOGGER.warning(f"{prefix} 'nms=True' is only available for Detect models like 'yolo11n.pt'.")
@@ -862,13 +928,26 @@ class Exporter:
             model = self.model
         ts = torch.jit.trace(model.eval(), self.im, strict=False)  # TorchScript model
 
+        if self.args.dynamic:
+            input_shape = ct.Shape(
+                shape=(
+                    ct.RangeDim(lower_bound=1, upper_bound=self.args.batch, default=1),
+                    self.im.shape[1],
+                    ct.RangeDim(lower_bound=32, upper_bound=self.imgsz[0] * 2, default=self.imgsz[0]),
+                    ct.RangeDim(lower_bound=32, upper_bound=self.imgsz[1] * 2, default=self.imgsz[1]),
+                )
+            )
+            inputs = [ct.TensorType("image", shape=input_shape)]
+        else:
+            inputs = [ct.ImageType("image", shape=self.im.shape, scale=1 / 255, bias=[0.0, 0.0, 0.0])]
+
         # Based on apple's documentation it is better to leave out the minimum_deployment target and let that get set
         # Internally based on the model conversion and output type.
-        # Setting minimum_depoloyment_target >= iOS16 will require setting compute_precision=ct.precision.FLOAT32.
+        # Setting minimum_deployment_target >= iOS16 will require setting compute_precision=ct.precision.FLOAT32.
         # iOS16 adds in better support for FP16, but none of the CoreML NMS specifications handle FP16 as input.
         ct_model = ct.convert(
             ts,
-            inputs=[ct.ImageType("image", shape=self.im.shape, scale=scale, bias=bias)],  # expects ct.TensorType
+            inputs=inputs,
             classifier_config=classifier_config,
             convert_to="neuralnetwork" if mlmodel else "mlprogram",
         )
@@ -885,12 +964,7 @@ class Exporter:
                 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:
-                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)
+            ct_model = self._pipeline_coreml(ct_model, weights_dir=None if mlmodel else ct_model.weights_dir)
 
         m = self.metadata  # metadata dict
         ct_model.short_description = m.pop("description")
@@ -910,20 +984,21 @@ class Exporter:
             )
             f = f.with_suffix(".mlmodel")
             ct_model.save(str(f))
-        return f, ct_model
+        return f
 
     @try_export
     def export_engine(self, dla=None, prefix=colorstr("TensorRT:")):
         """Export YOLO model to TensorRT format 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
+        f_onnx = self.export_onnx()  # run before TRT import https://github.com/ultralytics/ultralytics/issues/7016
 
         try:
-            import tensorrt as trt  # noqa
+            import tensorrt as trt
         except ImportError:
             if LINUX:
-                check_requirements("tensorrt>7.0.0,!=10.1.0")
-            import tensorrt as trt  # noqa
+                cuda_version = torch.version.cuda.split(".")[0]
+                check_requirements(f"tensorrt-cu{cuda_version}>7.0.0,!=10.1.0")
+            import tensorrt as trt
         check_version(trt.__version__, ">=7.0.0", hard=True)
         check_version(trt.__version__, "!=10.1.0", msg="https://github.com/ultralytics/ultralytics/pull/14239")
 
@@ -931,7 +1006,7 @@ class Exporter:
         LOGGER.info(f"\n{prefix} starting export with TensorRT {trt.__version__}...")
         assert Path(f_onnx).exists(), f"failed to export ONNX file: {f_onnx}"
         f = self.file.with_suffix(".engine")  # TensorRT engine file
-        export_engine(
+        onnx2engine(
             f_onnx,
             f,
             self.args.workspace,
@@ -946,26 +1021,26 @@ class Exporter:
             prefix=prefix,
         )
 
-        return f, None
+        return f
 
     @try_export
     def export_saved_model(self, prefix=colorstr("TensorFlow SavedModel:")):
         """Export YOLO model to TensorFlow SavedModel format."""
         cuda = torch.cuda.is_available()
         try:
-            import tensorflow as tf  # noqa
+            import tensorflow as tf
         except ImportError:
             check_requirements("tensorflow>=2.0.0,<=2.19.0")
-            import tensorflow as tf  # noqa
+            import tensorflow as tf
         check_requirements(
             (
                 "tf_keras<=2.19.0",  # required by 'onnx2tf' package
                 "sng4onnx>=1.0.1",  # required by 'onnx2tf' package
                 "onnx_graphsurgeon>=0.3.26",  # required by 'onnx2tf' package
-                "ai-edge-litert>=1.2.0,<1.4.0",  # required by 'onnx2tf' package
-                "onnx>=1.12.0",
+                "ai-edge-litert>=1.2.0" + (",<1.4.0" if MACOS else ""),  # required by 'onnx2tf' package
+                "onnx>=1.12.0,<2.0.0",
                 "onnx2tf>=1.26.3",
-                "onnxslim>=0.1.67",
+                "onnxslim>=0.1.71",
                 "onnxruntime-gpu" if cuda else "onnxruntime",
                 "protobuf>=5",
             ),
@@ -984,80 +1059,50 @@ class Exporter:
         if f.is_dir():
             shutil.rmtree(f)  # delete output folder
 
-        # Pre-download calibration file to fix https://github.com/PINTO0309/onnx2tf/issues/545
-        onnx2tf_file = Path("calibration_image_sample_data_20x128x128x3_float32.npy")
-        if not onnx2tf_file.exists():
-            attempt_download_asset(f"{onnx2tf_file}.zip", unzip=True, delete=True)
+        # Export to TF
+        images = None
+        if self.args.int8 and self.args.data:
+            images = [batch["img"] for batch in self.get_int8_calibration_dataloader(prefix)]
+            images = (
+                torch.nn.functional.interpolate(torch.cat(images, 0).float(), size=self.imgsz)
+                .permute(0, 2, 3, 1)
+                .numpy()
+                .astype(np.float32)
+            )
 
         # Export to ONNX
+        if isinstance(self.model.model[-1], RTDETRDecoder):
+            self.args.opset = self.args.opset or 19
+            assert 16 <= self.args.opset <= 19, "RTDETR export requires opset>=16;<=19"
         self.args.simplify = True
-        f_onnx, _ = self.export_onnx()
-
-        # Export to TF
-        np_data = None
-        if self.args.int8:
-            tmp_file = f / "tmp_tflite_int8_calibration_images.npy"  # int8 calibration images file
-            if self.args.data:
-                f.mkdir()
-                images = [batch["img"] for batch in self.get_int8_calibration_dataloader(prefix)]
-                images = torch.nn.functional.interpolate(torch.cat(images, 0).float(), size=self.imgsz).permute(
-                    0, 2, 3, 1
-                )
-                np.save(str(tmp_file), images.numpy().astype(np.float32))  # BHWC
-                np_data = [["images", tmp_file, [[[[0, 0, 0]]]], [[[[255, 255, 255]]]]]]
-
-        import onnx2tf  # scoped for after ONNX export for reduced conflict during import
-
-        LOGGER.info(f"{prefix} starting TFLite export with onnx2tf {onnx2tf.__version__}...")
-        keras_model = onnx2tf.convert(
-            input_onnx_file_path=f_onnx,
-            output_folder_path=str(f),
-            not_use_onnxsim=True,
-            verbosity="error",  # note INT8-FP16 activation bug https://github.com/ultralytics/ultralytics/issues/15873
-            output_integer_quantized_tflite=self.args.int8,
-            quant_type="per-tensor",  # "per-tensor" (faster) or "per-channel" (slower but more accurate)
-            custom_input_op_name_np_data_path=np_data,
-            enable_batchmatmul_unfold=True,  # fix lower no. of detected objects on GPU delegate
-            output_signaturedefs=True,  # fix error with Attention block group convolution
-            disable_group_convolution=self.args.format in {"tfjs", "edgetpu"},  # fix error with group convolution
+        f_onnx = self.export_onnx()  # ensure ONNX is available
+        keras_model = onnx2saved_model(
+            f_onnx,
+            f,
+            int8=self.args.int8,
+            images=images,
+            disable_group_convolution=self.args.format in {"tfjs", "edgetpu"},
+            prefix=prefix,
         )
         YAML.save(f / "metadata.yaml", self.metadata)  # add metadata.yaml
-
-        # Remove/rename TFLite models
-        if self.args.int8:
-            tmp_file.unlink(missing_ok=True)
-            for file in f.rglob("*_dynamic_range_quant.tflite"):
-                file.rename(file.with_name(file.stem.replace("_dynamic_range_quant", "_int8") + file.suffix))
-            for file in f.rglob("*_integer_quant_with_int16_act.tflite"):
-                file.unlink()  # delete extra fp16 activation TFLite files
-
         # Add TFLite metadata
         for file in f.rglob("*.tflite"):
-            f.unlink() if "quant_with_int16_act.tflite" in str(f) else self._add_tflite_metadata(file)
+            file.unlink() if "quant_with_int16_act.tflite" in str(file) else self._add_tflite_metadata(file)
 
         return str(f), keras_model  # or keras_model = tf.saved_model.load(f, tags=None, options=None)
 
     @try_export
     def export_pb(self, keras_model, prefix=colorstr("TensorFlow GraphDef:")):
         """Export YOLO model to TensorFlow GraphDef *.pb format https://github.com/leimao/Frozen-Graph-TensorFlow."""
-        import tensorflow as tf  # noqa
-        from tensorflow.python.framework.convert_to_constants import convert_variables_to_constants_v2  # noqa
-
-        LOGGER.info(f"\n{prefix} starting export with tensorflow {tf.__version__}...")
         f = self.file.with_suffix(".pb")
-
-        m = tf.function(lambda x: keras_model(x))  # full model
-        m = m.get_concrete_function(tf.TensorSpec(keras_model.inputs[0].shape, keras_model.inputs[0].dtype))
-        frozen_func = convert_variables_to_constants_v2(m)
-        frozen_func.graph.as_graph_def()
-        tf.io.write_graph(graph_or_graph_def=frozen_func.graph, logdir=str(f.parent), name=f.name, as_text=False)
-        return f, None
+        keras2pb(keras_model, f, prefix)
+        return f
 
     @try_export
     def export_tflite(self, prefix=colorstr("TensorFlow Lite:")):
         """Export YOLO model to TensorFlow Lite format."""
         # BUG https://github.com/ultralytics/ultralytics/issues/13436
-        import tensorflow as tf  # noqa
+        import tensorflow as tf
 
         LOGGER.info(f"\n{prefix} starting export with tensorflow {tf.__version__}...")
         saved_model = Path(str(self.file).replace(self.file.suffix, "_saved_model"))
@@ -1067,7 +1112,111 @@ class Exporter:
             f = saved_model / f"{self.file.stem}_float16.tflite"  # fp32 in/out
         else:
             f = saved_model / f"{self.file.stem}_float32.tflite"
-        return str(f), None
+        return str(f)
+
+    @try_export
+    def export_axelera(self, prefix=colorstr("Axelera:")):
+        """YOLO Axelera export."""
+        os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
+        try:
+            from axelera import compiler
+        except ImportError:
+            check_apt_requirements(
+                ["libllvm14", "libgirepository1.0-dev", "pkg-config", "libcairo2-dev", "build-essential", "cmake"]
+            )
+
+            check_requirements(
+                "axelera-voyager-sdk==1.5.2",
+                cmds="--extra-index-url https://software.axelera.ai/artifactory/axelera-runtime-pypi "
+                "--extra-index-url https://software.axelera.ai/artifactory/axelera-dev-pypi",
+            )
+
+        from axelera import compiler
+        from axelera.compiler import CompilerConfig
+
+        self.args.opset = 17  # hardcode opset for Axelera
+        onnx_path = self.export_onnx()
+        model_name = Path(onnx_path).stem
+        export_path = Path(f"{model_name}_axelera_model")
+        export_path.mkdir(exist_ok=True)
+
+        if "C2PSA" in self.model.__str__():  # YOLO11
+            config = CompilerConfig(
+                quantization_scheme="per_tensor_min_max",
+                ignore_weight_buffers=False,
+                resources_used=0.25,
+                aipu_cores_used=1,
+                multicore_mode="batch",
+                output_axm_format=True,
+                model_name=model_name,
+            )
+        else:  # YOLOv8
+            config = CompilerConfig(
+                tiling_depth=6,
+                split_buffer_promotion=True,
+                resources_used=0.25,
+                aipu_cores_used=1,
+                multicore_mode="batch",
+                output_axm_format=True,
+                model_name=model_name,
+            )
+
+        qmodel = compiler.quantize(
+            model=onnx_path,
+            calibration_dataset=self.get_int8_calibration_dataloader(prefix),
+            config=config,
+            transform_fn=self._transform_fn,
+        )
+
+        compiler.compile(model=qmodel, config=config, output_dir=export_path)
+
+        axm_name = f"{model_name}.axm"
+        axm_src = Path(axm_name)
+        axm_dst = export_path / axm_name
+
+        if axm_src.exists():
+            axm_src.replace(axm_dst)
+
+        YAML.save(export_path / "metadata.yaml", self.metadata)
+
+        return export_path
+
+    @try_export
+    def export_executorch(self, prefix=colorstr("ExecuTorch:")):
+        """Exports a model to ExecuTorch (.pte) format into a dedicated directory and saves the required metadata,
+        following Ultralytics conventions.
+        """
+        LOGGER.info(f"\n{prefix} starting export with ExecuTorch...")
+        assert TORCH_2_9, f"ExecuTorch export requires torch>=2.9.0 but torch=={TORCH_VERSION} is installed"
+
+        # BUG executorch build on arm64 Docker requires packaging>=22.0 https://github.com/pypa/setuptools/issues/4483
+        if LINUX and ARM64 and IS_DOCKER:
+            check_requirements("packaging>=22.0")
+
+        check_requirements("ruamel.yaml<0.19.0")
+        check_requirements("executorch==1.0.1", "flatbuffers")
+        # Pin numpy to avoid coremltools errors with numpy>=2.4.0, must be separate
+        check_requirements("numpy<=2.3.5")
+
+        from executorch.backends.xnnpack.partition.xnnpack_partitioner import XnnpackPartitioner
+        from executorch.exir import to_edge_transform_and_lower
+
+        file_directory = Path(str(self.file).replace(self.file.suffix, "_executorch_model"))
+        file_directory.mkdir(parents=True, exist_ok=True)
+
+        file_pte = file_directory / self.file.with_suffix(".pte").name
+        sample_inputs = (self.im,)
+
+        et_program = to_edge_transform_and_lower(
+            torch.export.export(self.model, sample_inputs), partitioner=[XnnpackPartitioner()]
+        ).to_executorch()
+
+        with open(file_pte, "wb") as file:
+            file.write(et_program.buffer)
+
+        YAML.save(file_directory / "metadata.yaml", self.metadata)
+
+        return str(file_directory)
 
     @try_export
     def export_edgetpu(self, tflite_model="", prefix=colorstr("Edge TPU:")):
@@ -1077,65 +1226,33 @@ class Exporter:
         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 = "sudo " if is_sudo_available() else ""
             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",
+                f"{sudo}mkdir -p /etc/apt/keyrings",
+                f"curl -fsSL https://packages.cloud.google.com/apt/doc/apt-key.gpg | {sudo}gpg --dearmor -o /etc/apt/keyrings/google.gpg",
+                f'echo "deb [signed-by=/etc/apt/keyrings/google.gpg] https://packages.cloud.google.com/apt coral-edgetpu-stable main" | {sudo}tee /etc/apt/sources.list.d/coral-edgetpu.list',
             ):
-                subprocess.run(c if is_sudo_available() else c.replace("sudo ", ""), shell=True, check=True)
-        ver = subprocess.run(cmd, shell=True, capture_output=True, check=True).stdout.decode().rsplit(maxsplit=1)[-1]
+                subprocess.run(c, shell=True, check=True)
+            check_apt_requirements(["edgetpu-compiler"])
 
+        ver = subprocess.run(cmd, shell=True, capture_output=True, check=True).stdout.decode().rsplit(maxsplit=1)[-1]
         LOGGER.info(f"\n{prefix} starting export with Edge TPU compiler {ver}...")
+        tflite2edgetpu(tflite_file=tflite_model, output_dir=tflite_model.parent, prefix=prefix)
         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
+        return f
 
     @try_export
     def export_tfjs(self, prefix=colorstr("TensorFlow.js:")):
         """Export YOLO model to TensorFlow.js format."""
         check_requirements("tensorflowjs")
-        import tensorflow as tf
-        import tensorflowjs as tfjs  # noqa
 
-        LOGGER.info(f"\n{prefix} starting export with tensorflowjs {tfjs.__version__}...")
         f = str(self.file).replace(self.file.suffix, "_web_model")  # js dir
         f_pb = str(self.file.with_suffix(".pb"))  # *.pb path
-
-        gd = tf.Graph().as_graph_def()  # TF GraphDef
-        with open(f_pb, "rb") as file:
-            gd.ParseFromString(file.read())
-        outputs = ",".join(gd_outputs(gd))
-        LOGGER.info(f"\n{prefix} output node names: {outputs}")
-
-        quantization = "--quantize_float16" if self.args.half else "--quantize_uint8" if self.args.int8 else ""
-        with spaces_in_path(f_pb) as fpb_, spaces_in_path(f) as f_:  # exporter can not handle spaces in path
-            cmd = (
-                "tensorflowjs_converter "
-                f'--input_format=tf_frozen_model {quantization} --output_node_names={outputs} "{fpb_}" "{f_}"'
-            )
-            LOGGER.info(f"{prefix} running '{cmd}'")
-            subprocess.run(cmd, shell=True)
-
-        if " " in f:
-            LOGGER.warning(f"{prefix} your model may not work correctly with spaces in path '{f}'.")
-
+        pb2tfjs(pb_file=f_pb, output_dir=f, half=self.args.half, int8=self.args.int8, prefix=prefix)
         # Add metadata
         YAML.save(Path(f) / "metadata.yaml", self.metadata)  # add metadata.yaml
-        return f, None
+        return f
 
     @try_export
     def export_rknn(self, prefix=colorstr("RKNN:")):
@@ -1151,7 +1268,7 @@ class Exporter:
 
         from rknn.api import RKNN
 
-        f, _ = self.export_onnx()
+        f = self.export_onnx()
         export_path = Path(f"{Path(f).stem}_rknn_model")
         export_path.mkdir(exist_ok=True)
 
@@ -1162,30 +1279,30 @@ class Exporter:
         f = f.replace(".onnx", f"-{self.args.name}.rknn")
         rknn.export_rknn(f"{export_path / f}")
         YAML.save(export_path / "metadata.yaml", self.metadata)
-        return export_path, None
+        return export_path
 
     @try_export
     def export_imx(self, prefix=colorstr("IMX:")):
         """Export YOLO model to IMX format."""
-        gptq = False
         assert LINUX, (
-            "export only supported on Linux. "
-            "See https://developer.aitrios.sony-semicon.com/en/raspberrypi-ai-camera/documentation/imx500-converter"
+            "Export only supported on Linux."
+            "See https://developer.aitrios.sony-semicon.com/en/docs/raspberry-pi-ai-camera/imx500-converter?version=3.17.3&progLang="
         )
+        assert not ARM64, "IMX export is not supported on ARM64 architectures."
+        assert IS_PYTHON_MINIMUM_3_9, "IMX export is only supported on Python 3.9 or above."
+
         if getattr(self.model, "end2end", False):
             raise ValueError("IMX export is not supported for end2end models.")
         check_requirements(
-            ("model-compression-toolkit>=2.4.1", "sony-custom-layers>=0.3.0", "edge-mdt-tpc>=1.1.0", "pydantic<=2.11.7")
+            (
+                "model-compression-toolkit>=2.4.1",
+                "edge-mdt-cl<1.1.0",
+                "edge-mdt-tpc>=1.2.0",
+                "pydantic<=2.11.7",
+            )
         )
-        check_requirements("imx500-converter[pt]>=3.16.1")  # Separate requirements for imx500-converter
-        check_requirements("mct-quantizers>=1.6.0")  # Separate for compatibility with model-compression-toolkit
-
-        import model_compression_toolkit as mct
-        import onnx
-        from edgemdt_tpc import get_target_platform_capabilities
-        from sony_custom_layers.pytorch import multiclass_nms_with_indices
 
-        LOGGER.info(f"\n{prefix} starting export with model_compression_toolkit {mct.__version__}...")
+        check_requirements("imx500-converter[pt]>=3.17.3")
 
         # Install Java>=17
         try:
@@ -1194,153 +1311,24 @@ class Exporter:
             java_version = int(version_match.group(1)) if version_match else 0
             assert java_version >= 17, "Java version too old"
         except (FileNotFoundError, subprocess.CalledProcessError, AssertionError):
-            cmd = (["sudo"] if is_sudo_available() else []) + ["apt", "install", "-y", "openjdk-21-jre"]
-            subprocess.run(cmd, 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 = get_target_platform_capabilities(tpc_version="4.0", device_type="imx500")
-
-        bit_cfg = mct.core.BitWidthConfig()
-        if "C2PSA" in self.model.__str__():  # YOLO11
-            if self.model.task == "detect":
-                layer_names = ["sub", "mul_2", "add_14", "cat_21"]
-                weights_memory = 2585350.2439
-                n_layers = 238  # 238 layers for fused YOLO11n
-            elif self.model.task == "pose":
-                layer_names = ["sub", "mul_2", "add_14", "cat_22", "cat_23", "mul_4", "add_15"]
-                weights_memory = 2437771.67
-                n_layers = 257  # 257 layers for fused YOLO11n-pose
-        else:  # YOLOv8
-            if self.model.task == "detect":
-                layer_names = ["sub", "mul", "add_6", "cat_17"]
-                weights_memory = 2550540.8
-                n_layers = 168  # 168 layers for fused YOLOv8n
-            elif self.model.task == "pose":
-                layer_names = ["add_7", "mul_2", "cat_19", "mul", "sub", "add_6", "cat_18"]
-                weights_memory = 2482451.85
-                n_layers = 187  # 187 layers for fused YOLO11n-pose
-
-        # Check if the model has the expected number of layers
-        if len(list(self.model.modules())) != n_layers:
-            raise ValueError("IMX export only supported for YOLOv8n and YOLO11n models.")
-
-        for layer_name in layer_names:
-            bit_cfg.set_manual_activation_bit_width([mct.core.common.network_editors.NodeNameFilter(layer_name)], 16)
-
-        config = mct.core.CoreConfig(
-            mixed_precision_config=mct.core.MixedPrecisionQuantizationConfig(num_of_images=10),
-            quantization_config=mct.core.QuantizationConfig(concat_threshold_update=True),
-            bit_width_config=bit_cfg,
-        )
-
-        resource_utilization = mct.core.ResourceUtilization(weights_memory=weights_memory)
-
-        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, use_hessian_sample_attention=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):
-            """Wrap PyTorch Module with multiclass_nms layer from sony_custom_layers."""
-
-            def __init__(
-                self,
-                model: torch.nn.Module,
-                score_threshold: float = 0.001,
-                iou_threshold: float = 0.7,
-                max_detections: int = 300,
-                task: str = "detect",
-            ):
-                """
-                Initialize NMSWrapper with PyTorch Module and NMS parameters.
-
-                Args:
-                    model (torch.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 (int): The number of detections to return.
-                    task (str): Task type, either 'detect' or 'pose'.
-                """
-                super().__init__()
-                self.model = model
-                self.score_threshold = score_threshold
-                self.iou_threshold = iou_threshold
-                self.max_detections = max_detections
-                self.task = task
-
-            def forward(self, images):
-                """Forward pass with model inference and NMS post-processing."""
-                # model inference
-                outputs = self.model(images)
-
-                boxes, scores = outputs[0], outputs[1]
-                nms_outputs = multiclass_nms_with_indices(
-                    boxes=boxes,
-                    scores=scores,
-                    score_threshold=self.score_threshold,
-                    iou_threshold=self.iou_threshold,
-                    max_detections=self.max_detections,
-                )
-                if self.task == "pose":
-                    kpts = outputs[2]  # (bs, max_detections, kpts 17*3)
-                    out_kpts = torch.gather(kpts, 1, nms_outputs.indices.unsqueeze(-1).expand(-1, -1, kpts.size(-1)))
-                    return nms_outputs.boxes, nms_outputs.scores, nms_outputs.labels, out_kpts
-                return nms_outputs
-
-        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,
-            task=self.model.task,
-        ).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.name).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,
+            if IS_UBUNTU or IS_DEBIAN_TRIXIE:
+                LOGGER.info(f"\n{prefix} installing Java 21 for Ubuntu...")
+                check_apt_requirements(["openjdk-21-jre"])
+            elif IS_RASPBERRYPI or IS_DEBIAN_BOOKWORM:
+                LOGGER.info(f"\n{prefix} installing Java 17 for Raspberry Pi or Debian ...")
+                check_apt_requirements(["openjdk-17-jre"])
+
+        return torch2imx(
+            self.model,
+            self.file,
+            self.args.conf,
+            self.args.iou,
+            self.args.max_det,
+            metadata=self.metadata,
+            dataset=self.get_int8_calibration_dataloader(prefix),
+            prefix=prefix,
         )
 
-        # Needed for imx models.
-        with open(f / "labels.txt", "w", encoding="utf-8") as file:
-            file.writelines([f"{name}\n" for _, name in self.model.names.items()])
-
-        return f, None
-
     def _add_tflite_metadata(self, file):
         """Add metadata to *.tflite models per https://ai.google.dev/edge/litert/models/metadata."""
         import zipfile
@@ -1350,7 +1338,7 @@ class Exporter:
 
     def _pipeline_coreml(self, model, weights_dir=None, prefix=colorstr("CoreML Pipeline:")):
         """Create CoreML pipeline with NMS for YOLO detection models."""
-        import coremltools as ct  # noqa
+        import coremltools as ct
 
         LOGGER.info(f"{prefix} starting pipeline with coremltools {ct.__version__}...")
 
@@ -1365,7 +1353,8 @@ class Exporter:
         names = self.metadata["names"]
         nx, ny = spec.description.input[0].type.imageType.width, spec.description.input[0].type.imageType.height
         nc = outs[0].type.multiArrayType.shape[-1]
-        assert len(names) == nc, f"{len(names)} names found for nc={nc}"  # check
+        if len(names) != nc:  # Hack fix for MLProgram NMS bug https://github.com/ultralytics/ultralytics/issues/22309
+            names = {**names, **{i: str(i) for i in range(len(names), nc)}}
 
         # Model from spec
         model = ct.models.MLModel(spec, weights_dir=weights_dir)
@@ -1442,6 +1431,14 @@ class Exporter:
         LOGGER.info(f"{prefix} pipeline success")
         return model
 
+    @staticmethod
+    def _transform_fn(data_item) -> np.ndarray:
+        """The transformation function for Axelera/OpenVINO quantization preprocessing."""
+        data_item: torch.Tensor = data_item["img"] if isinstance(data_item, dict) else data_item
+        assert data_item.dtype == torch.uint8, "Input image must be uint8 for the quantization preprocessing"
+        im = data_item.numpy().astype(np.float32) / 255.0  # uint8 to fp16/32 and 0 - 255 to 0.0 - 1.0
+        return im[None] if im.ndim == 3 else im
+
     def add_callback(self, event: str, callback):
         """Append the given callback to the specified event."""
         self.callbacks[event].append(callback)
@@ -1455,18 +1452,19 @@ class Exporter:
 class IOSDetectModel(torch.nn.Module):
     """Wrap an Ultralytics YOLO model for Apple iOS CoreML export."""
 
-    def __init__(self, model, im):
-        """
-        Initialize the IOSDetectModel class with a YOLO model and example image.
+    def __init__(self, model, im, mlprogram=True):
+        """Initialize the IOSDetectModel class with a YOLO model and example image.
 
         Args:
             model (torch.nn.Module): The YOLO model to wrap.
             im (torch.Tensor): Example input tensor with shape (B, C, H, W).
+            mlprogram (bool): Whether exporting to MLProgram format to fix NMS bug.
         """
         super().__init__()
         _, _, h, w = im.shape  # batch, channel, height, width
         self.model = model
         self.nc = len(model.names)  # number of classes
+        self.mlprogram = mlprogram
         if w == h:
             self.normalize = 1.0 / w  # scalar
         else:
@@ -1478,15 +1476,18 @@ class IOSDetectModel(torch.nn.Module):
     def forward(self, x):
         """Normalize predictions of object detection model with input size-dependent factors."""
         xywh, cls = self.model(x)[0].transpose(0, 1).split((4, self.nc), 1)
-        return cls, xywh * self.normalize  # confidence (3780, 80), coordinates (3780, 4)
+        if self.mlprogram and self.nc % 80 != 0:  # NMS bug https://github.com/ultralytics/ultralytics/issues/22309
+            pad_length = int(((self.nc + 79) // 80) * 80) - self.nc  # pad class length to multiple of 80
+            cls = torch.nn.functional.pad(cls, (0, pad_length, 0, 0), "constant", 0)
+
+        return cls, xywh * self.normalize
 
 
 class NMSModel(torch.nn.Module):
     """Model wrapper with embedded NMS for Detect, Segment, Pose and OBB."""
 
     def __init__(self, model, args):
-        """
-        Initialize the NMSModel.
+        """Initialize the NMSModel.
 
         Args:
             model (torch.nn.Module): The model to wrap with NMS postprocessing.
@@ -1499,15 +1500,14 @@ class NMSModel(torch.nn.Module):
         self.is_tf = self.args.format in frozenset({"saved_model", "tflite", "tfjs"})
 
     def forward(self, x):
-        """
-        Perform inference with NMS post-processing. Supports Detect, Segment, OBB and Pose.
+        """Perform inference with NMS post-processing. Supports Detect, Segment, OBB and Pose.
 
         Args:
             x (torch.Tensor): The preprocessed tensor with shape (N, 3, H, W).
 
         Returns:
-            (torch.Tensor): List of detections, each an (N, max_det, 4 + 2 + extra_shape) Tensor where N is the
-                number of detections after NMS.
+            (torch.Tensor): List of detections, each an (N, max_det, 4 + 2 + extra_shape) Tensor where N is the number
+                of detections after NMS.
         """
         from functools import partial
 
@@ -1530,7 +1530,7 @@ class NMSModel(torch.nn.Module):
         for i in range(bs):
             box, cls, score, extra = boxes[i], classes[i], scores[i], extras[i]
             mask = score > self.args.conf
-            if self.is_tf:
+            if self.is_tf or (self.args.format == "onnx" and self.obb):
                 # TFLite GatherND error if mask is empty
                 score *= mask
                 # Explicit length otherwise reshape error, hardcoded to `self.args.max_det * 5`
@@ -1538,17 +1538,16 @@ class NMSModel(torch.nn.Module):
             box, score, cls, extra = box[mask], score[mask], cls[mask], extra[mask]
             nmsbox = box.clone()
             # `8` is the minimum value experimented to get correct NMS results for obb
-            multiplier = 8 if self.obb else 1
+            multiplier = 8 if self.obb else 1 / max(len(self.model.names), 1)
             # Normalize boxes for NMS since large values for class offset causes issue with int8 quantization
             if self.args.format == "tflite":  # TFLite is already normalized
                 nmsbox *= multiplier
             else:
-                nmsbox = multiplier * nmsbox / torch.tensor(x.shape[2:], **kwargs).max()
-            if not self.args.agnostic_nms:  # class-specific NMS
+                nmsbox = multiplier * (nmsbox / torch.tensor(x.shape[2:], **kwargs).max())
+            if not self.args.agnostic_nms:  # class-wise NMS
                 end = 2 if self.obb else 4
                 # fully explicit expansion otherwise reshape error
-                # large max_wh causes issues when quantizing
-                cls_offset = cls.reshape(-1, 1).expand(nmsbox.shape[0], end)
+                cls_offset = cls.view(cls.shape[0], 1).expand(cls.shape[0], end)
                 offbox = nmsbox[:, :end] + cls_offset * multiplier
                 nmsbox = torch.cat((offbox, nmsbox[:, end:]), dim=-1)
             nms_fn = (
@@ -1560,6 +1559,7 @@ class NMSModel(torch.nn.Module):
                         or (self.args.format == "openvino" and self.args.int8)  # OpenVINO int8 error with triu
                     ),
                     iou_func=batch_probiou,
+                    exit_early=False,
                 )
                 if self.obb
                 else nms