Reference for ultralytics/utils/torch_utils.py

@contextmanager
def torch_distributed_zero_first(local_rank: int):
    """Ensures all processes in distributed training wait for the local master (rank 0) to complete a task first."""
    initialized = dist.is_available() and dist.is_initialized()
    use_ids = initialized and dist.get_backend() == "nccl"

    if initialized and local_rank not in {-1, 0}:
        dist.barrier(device_ids=[local_rank]) if use_ids else dist.barrier()
    yield
    if initialized and local_rank == 0:
        dist.barrier(device_ids=[local_rank]) if use_ids else dist.barrier()

def smart_inference_mode():
    """Applies torch.inference_mode() decorator if torch>=1.9.0 else torch.no_grad() decorator."""

    def decorate(fn):
        """Applies appropriate torch decorator for inference mode based on torch version."""
        if TORCH_1_9 and torch.is_inference_mode_enabled():
            return fn  # already in inference_mode, act as a pass-through
        else:
            return (torch.inference_mode if TORCH_1_9 else torch.no_grad)()(fn)

    return decorate

def autocast(enabled: bool, device: str = "cuda"):
    """
    Get the appropriate autocast context manager based on PyTorch version and AMP setting.

    This function returns a context manager for automatic mixed precision (AMP) training that is compatible with both
    older and newer versions of PyTorch. It handles the differences in the autocast API between PyTorch versions.

    Args:
        enabled (bool): Whether to enable automatic mixed precision.
        device (str, optional): The device to use for autocast. Defaults to 'cuda'.

    Returns:
        (torch.amp.autocast): The appropriate autocast context manager.

    Notes:
        - For PyTorch versions 1.13 and newer, it uses `torch.amp.autocast`.
        - For older versions, it uses `torch.cuda.autocast`.

    Examples:
        >>> with autocast(enabled=True):
        ...     # Your mixed precision operations here
        ...     pass
    """
    if TORCH_1_13:
        return torch.amp.autocast(device, enabled=enabled)
    else:
        return torch.cuda.amp.autocast(enabled)

def get_cpu_info():
    """Return a string with system CPU information, i.e. 'Apple M2'."""
    from ultralytics.utils import PERSISTENT_CACHE  # avoid circular import error

    if "cpu_info" not in PERSISTENT_CACHE:
        try:
            import cpuinfo  # pip install py-cpuinfo

            k = "brand_raw", "hardware_raw", "arch_string_raw"  # keys sorted by preference
            info = cpuinfo.get_cpu_info()  # info dict
            string = info.get(k[0] if k[0] in info else k[1] if k[1] in info else k[2], "unknown")
            PERSISTENT_CACHE["cpu_info"] = string.replace("(R)", "").replace("CPU ", "").replace("@ ", "")
        except Exception:
            pass
    return PERSISTENT_CACHE.get("cpu_info", "unknown")

def get_gpu_info(index):
    """Return a string with system GPU information, i.e. 'Tesla T4, 15102MiB'."""
    properties = torch.cuda.get_device_properties(index)
    return f"{properties.name}, {properties.total_memory / (1 << 20):.0f}MiB"

def select_device(device="", batch=0, newline=False, verbose=True):
    """
    Select the appropriate PyTorch device based on the provided arguments.

    The function takes a string specifying the device or a torch.device object and returns a torch.device object
    representing the selected device. The function also validates the number of available devices and raises an
    exception if the requested device(s) are not available.

    Args:
        device (str | torch.device, optional): Device string or torch.device object.
            Options are 'None', 'cpu', or 'cuda', or '0' or '0,1,2,3'. Defaults to an empty string, which auto-selects
            the first available GPU, or CPU if no GPU is available.
        batch (int, optional): Batch size being used in your model. Defaults to 0.
        newline (bool, optional): If True, adds a newline at the end of the log string. Defaults to False.
        verbose (bool, optional): If True, logs the device information. Defaults to True.

    Returns:
        (torch.device): Selected device.

    Raises:
        ValueError: If the specified device is not available or if the batch size is not a multiple of the number of
            devices when using multiple GPUs.

    Examples:
        >>> select_device("cuda:0")
        device(type='cuda', index=0)

        >>> select_device("cpu")
        device(type='cpu')

    Note:
        Sets the 'CUDA_VISIBLE_DEVICES' environment variable for specifying which GPUs to use.
    """
    if isinstance(device, torch.device) or str(device).startswith("tpu") or str(device).startswith("intel"):
        return device

    s = f"Ultralytics {__version__} 🚀 Python-{PYTHON_VERSION} torch-{torch.__version__} "
    device = str(device).lower()
    for remove in "cuda:", "none", "(", ")", "[", "]", "'", " ":
        device = device.replace(remove, "")  # to string, 'cuda:0' -> '0' and '(0, 1)' -> '0,1'
    cpu = device == "cpu"
    mps = device in {"mps", "mps:0"}  # Apple Metal Performance Shaders (MPS)
    if cpu or mps:
        os.environ["CUDA_VISIBLE_DEVICES"] = "-1"  # force torch.cuda.is_available() = False
    elif device:  # non-cpu device requested
        if device == "cuda":
            device = "0"
        if "," in device:
            device = ",".join([x for x in device.split(",") if x])  # remove sequential commas, i.e. "0,,1" -> "0,1"
        visible = os.environ.get("CUDA_VISIBLE_DEVICES", None)
        os.environ["CUDA_VISIBLE_DEVICES"] = device  # set environment variable - must be before assert is_available()
        if not (torch.cuda.is_available() and torch.cuda.device_count() >= len(device.split(","))):
            LOGGER.info(s)
            install = (
                "See https://pytorch.org/get-started/locally/ for up-to-date torch install instructions if no "
                "CUDA devices are seen by torch.\n"
                if torch.cuda.device_count() == 0
                else ""
            )
            raise ValueError(
                f"Invalid CUDA 'device={device}' requested."
                f" Use 'device=cpu' or pass valid CUDA device(s) if available,"
                f" i.e. 'device=0' or 'device=0,1,2,3' for Multi-GPU.\n"
                f"\ntorch.cuda.is_available(): {torch.cuda.is_available()}"
                f"\ntorch.cuda.device_count(): {torch.cuda.device_count()}"
                f"\nos.environ['CUDA_VISIBLE_DEVICES']: {visible}\n"
                f"{install}"
            )

    if not cpu and not mps and torch.cuda.is_available():  # prefer GPU if available
        devices = device.split(",") if device else "0"  # i.e. "0,1" -> ["0", "1"]
        n = len(devices)  # device count
        if n > 1:  # multi-GPU
            if batch < 1:
                raise ValueError(
                    "AutoBatch with batch<1 not supported for Multi-GPU training, "
                    "please specify a valid batch size, i.e. batch=16."
                )
            if batch >= 0 and batch % n != 0:  # check batch_size is divisible by device_count
                raise ValueError(
                    f"'batch={batch}' must be a multiple of GPU count {n}. Try 'batch={batch // n * n}' or "
                    f"'batch={batch // n * n + n}', the nearest batch sizes evenly divisible by {n}."
                )
        space = " " * (len(s) + 1)
        for i, d in enumerate(devices):
            s += f"{'' if i == 0 else space}CUDA:{d} ({get_gpu_info(i)})\n"  # bytes to MB
        arg = "cuda:0"
    elif mps and TORCH_2_0 and torch.backends.mps.is_available():
        # Prefer MPS if available
        s += f"MPS ({get_cpu_info()})\n"
        arg = "mps"
    else:  # revert to CPU
        s += f"CPU ({get_cpu_info()})\n"
        arg = "cpu"

    if arg in {"cpu", "mps"}:
        torch.set_num_threads(NUM_THREADS)  # reset OMP_NUM_THREADS for cpu training
    if verbose:
        LOGGER.info(s if newline else s.rstrip())
    return torch.device(arg)

def time_sync():
    """PyTorch-accurate time."""
    if torch.cuda.is_available():
        torch.cuda.synchronize()
    return time.time()

def fuse_conv_and_bn(conv, bn):
    """Fuse Conv2d() and BatchNorm2d() layers."""
    fusedconv = (
        nn.Conv2d(
            conv.in_channels,
            conv.out_channels,
            kernel_size=conv.kernel_size,
            stride=conv.stride,
            padding=conv.padding,
            dilation=conv.dilation,
            groups=conv.groups,
            bias=True,
        )
        .requires_grad_(False)
        .to(conv.weight.device)
    )

    # Prepare filters
    w_conv = conv.weight.view(conv.out_channels, -1)
    w_bn = torch.diag(bn.weight.div(torch.sqrt(bn.eps + bn.running_var)))
    fusedconv.weight.copy_(torch.mm(w_bn, w_conv).view(fusedconv.weight.shape))

    # Prepare spatial bias
    b_conv = (
        torch.zeros(conv.weight.shape[0], dtype=conv.weight.dtype, device=conv.weight.device)
        if conv.bias is None
        else conv.bias
    )
    b_bn = bn.bias - bn.weight.mul(bn.running_mean).div(torch.sqrt(bn.running_var + bn.eps))
    fusedconv.bias.copy_(torch.mm(w_bn, b_conv.reshape(-1, 1)).reshape(-1) + b_bn)

    return fusedconv

def fuse_deconv_and_bn(deconv, bn):
    """Fuse ConvTranspose2d() and BatchNorm2d() layers."""
    fuseddconv = (
        nn.ConvTranspose2d(
            deconv.in_channels,
            deconv.out_channels,
            kernel_size=deconv.kernel_size,
            stride=deconv.stride,
            padding=deconv.padding,
            output_padding=deconv.output_padding,
            dilation=deconv.dilation,
            groups=deconv.groups,
            bias=True,
        )
        .requires_grad_(False)
        .to(deconv.weight.device)
    )

    # Prepare filters
    w_deconv = deconv.weight.view(deconv.out_channels, -1)
    w_bn = torch.diag(bn.weight.div(torch.sqrt(bn.eps + bn.running_var)))
    fuseddconv.weight.copy_(torch.mm(w_bn, w_deconv).view(fuseddconv.weight.shape))

    # Prepare spatial bias
    b_conv = torch.zeros(deconv.weight.shape[1], device=deconv.weight.device) if deconv.bias is None else deconv.bias
    b_bn = bn.bias - bn.weight.mul(bn.running_mean).div(torch.sqrt(bn.running_var + bn.eps))
    fuseddconv.bias.copy_(torch.mm(w_bn, b_conv.reshape(-1, 1)).reshape(-1) + b_bn)

    return fuseddconv

def model_info(model, detailed=False, verbose=True, imgsz=640):
    """
    Print and return detailed model information layer by layer.

    Args:
        model (nn.Module): Model to analyze.
        detailed (bool, optional): Whether to print detailed layer information. Defaults to False.
        verbose (bool, optional): Whether to print model information. Defaults to True.
        imgsz (int | List, optional): Input image size. Defaults to 640.

    Returns:
        (Tuple[int, int, int, float]): Number of layers, parameters, gradients, and GFLOPs.
    """
    if not verbose:
        return
    n_p = get_num_params(model)  # number of parameters
    n_g = get_num_gradients(model)  # number of gradients
    layers = __import__("collections").OrderedDict((n, m) for n, m in model.named_modules() if len(m._modules) == 0)
    n_l = len(layers)  # number of layers
    if detailed:
        h = f"{'layer':>5}{'name':>40}{'type':>20}{'gradient':>10}{'parameters':>12}{'shape':>20}{'mu':>10}{'sigma':>10}"
        LOGGER.info(h)
        for i, (mn, m) in enumerate(layers.items()):
            mn = mn.replace("module_list.", "")
            mt = m.__class__.__name__
            if len(m._parameters):
                for pn, p in m.named_parameters():
                    LOGGER.info(
                        f"{i:>5g}{f'{mn}.{pn}':>40}{mt:>20}{p.requires_grad!r:>10}{p.numel():>12g}{str(list(p.shape)):>20}{p.mean():>10.3g}{p.std():>10.3g}{str(p.dtype).replace('torch.', ''):>15}"
                    )
            else:  # layers with no learnable params
                LOGGER.info(f"{i:>5g}{mn:>40}{mt:>20}{False!r:>10}{0:>12g}{str([]):>20}{'-':>10}{'-':>10}{'-':>15}")

    flops = get_flops(model, imgsz)  # imgsz may be int or list, i.e. imgsz=640 or imgsz=[640, 320]
    fused = " (fused)" if getattr(model, "is_fused", lambda: False)() else ""
    fs = f", {flops:.1f} GFLOPs" if flops else ""
    yaml_file = getattr(model, "yaml_file", "") or getattr(model, "yaml", {}).get("yaml_file", "")
    model_name = Path(yaml_file).stem.replace("yolo", "YOLO") or "Model"
    LOGGER.info(f"{model_name} summary{fused}: {n_l:,} layers, {n_p:,} parameters, {n_g:,} gradients{fs}")
    return n_l, n_p, n_g, flops

def get_num_params(model):
    """Return the total number of parameters in a YOLO model."""
    return sum(x.numel() for x in model.parameters())

def get_num_gradients(model):
    """Return the total number of parameters with gradients in a YOLO model."""
    return sum(x.numel() for x in model.parameters() if x.requires_grad)

def model_info_for_loggers(trainer):
    """
    Return model info dict with useful model information.

    Args:
        trainer (ultralytics.engine.trainer.BaseTrainer): The trainer object containing model and validation data.

    Returns:
        (dict): Dictionary containing model parameters, GFLOPs, and inference speeds.

    Examples:
        YOLOv8n info for loggers
        >>> results = {
        ...    "model/parameters": 3151904,
        ...    "model/GFLOPs": 8.746,
        ...    "model/speed_ONNX(ms)": 41.244,
        ...    "model/speed_TensorRT(ms)": 3.211,
        ...    "model/speed_PyTorch(ms)": 18.755,
        ...}
    """
    if trainer.args.profile:  # profile ONNX and TensorRT times
        from ultralytics.utils.benchmarks import ProfileModels

        results = ProfileModels([trainer.last], device=trainer.device).profile()[0]
        results.pop("model/name")
    else:  # only return PyTorch times from most recent validation
        results = {
            "model/parameters": get_num_params(trainer.model),
            "model/GFLOPs": round(get_flops(trainer.model), 3),
        }
    results["model/speed_PyTorch(ms)"] = round(trainer.validator.speed["inference"], 3)
    return results

def get_flops(model, imgsz=640):
    """
    Calculate FLOPs (floating point operations) for a model in billions.

    Attempts two calculation methods: first with a stride-based tensor for efficiency,
    then falls back to full image size if needed (e.g., for RTDETR models). Returns 0.0
    if thop library is unavailable or calculation fails.

    Args:
        model (nn.Module): The model to calculate FLOPs for.
        imgsz (int | List[int], optional): Input image size. Defaults to 640.

    Returns:
        (float): The model FLOPs in billions.
    """
    if not thop:
        return 0.0  # if not installed return 0.0 GFLOPs

    try:
        model = de_parallel(model)
        p = next(model.parameters())
        if not isinstance(imgsz, list):
            imgsz = [imgsz, imgsz]  # expand if int/float
        try:
            # Method 1: Use stride-based input tensor
            stride = max(int(model.stride.max()), 32) if hasattr(model, "stride") else 32  # max stride
            im = torch.empty((1, p.shape[1], stride, stride), device=p.device)  # input image in BCHW format
            flops = thop.profile(deepcopy(model), inputs=[im], verbose=False)[0] / 1e9 * 2  # stride GFLOPs
            return flops * imgsz[0] / stride * imgsz[1] / stride  # imgsz GFLOPs
        except Exception:
            # Method 2: Use actual image size (required for RTDETR models)
            im = torch.empty((1, p.shape[1], *imgsz), device=p.device)  # input image in BCHW format
            return thop.profile(deepcopy(model), inputs=[im], verbose=False)[0] / 1e9 * 2  # imgsz GFLOPs
    except Exception:
        return 0.0

def get_flops_with_torch_profiler(model, imgsz=640):
    """
    Compute model FLOPs using torch profiler (alternative to thop package, but 2-10x slower).

    Args:
        model (nn.Module): The model to calculate FLOPs for.
        imgsz (int | List[int], optional): Input image size. Defaults to 640.

    Returns:
        (float): The model's FLOPs in billions.
    """
    if not TORCH_2_0:  # torch profiler implemented in torch>=2.0
        return 0.0
    model = de_parallel(model)
    p = next(model.parameters())
    if not isinstance(imgsz, list):
        imgsz = [imgsz, imgsz]  # expand if int/float
    try:
        # Use stride size for input tensor
        stride = (max(int(model.stride.max()), 32) if hasattr(model, "stride") else 32) * 2  # max stride
        im = torch.empty((1, p.shape[1], stride, stride), device=p.device)  # input image in BCHW format
        with torch.profiler.profile(with_flops=True) as prof:
            model(im)
        flops = sum(x.flops for x in prof.key_averages()) / 1e9
        flops = flops * imgsz[0] / stride * imgsz[1] / stride  # 640x640 GFLOPs
    except Exception:
        # Use actual image size for input tensor (i.e. required for RTDETR models)
        im = torch.empty((1, p.shape[1], *imgsz), device=p.device)  # input image in BCHW format
        with torch.profiler.profile(with_flops=True) as prof:
            model(im)
        flops = sum(x.flops for x in prof.key_averages()) / 1e9
    return flops

def initialize_weights(model):
    """Initialize model weights to random values."""
    for m in model.modules():
        t = type(m)
        if t is nn.Conv2d:
            pass  # nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
        elif t is nn.BatchNorm2d:
            m.eps = 1e-3
            m.momentum = 0.03
        elif t in {nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6, nn.SiLU}:
            m.inplace = True

def scale_img(img, ratio=1.0, same_shape=False, gs=32):
    """
    Scales and pads an image tensor, optionally maintaining aspect ratio and padding to gs multiple.

    Args:
        img (torch.Tensor): Input image tensor.
        ratio (float, optional): Scaling ratio. Defaults to 1.0.
        same_shape (bool, optional): Whether to maintain the same shape. Defaults to False.
        gs (int, optional): Grid size for padding. Defaults to 32.

    Returns:
        (torch.Tensor): Scaled and padded image tensor.
    """
    if ratio == 1.0:
        return img
    h, w = img.shape[2:]
    s = (int(h * ratio), int(w * ratio))  # new size
    img = F.interpolate(img, size=s, mode="bilinear", align_corners=False)  # resize
    if not same_shape:  # pad/crop img
        h, w = (math.ceil(x * ratio / gs) * gs for x in (h, w))
    return F.pad(img, [0, w - s[1], 0, h - s[0]], value=0.447)  # value = imagenet mean

def copy_attr(a, b, include=(), exclude=()):
    """
    Copies attributes from object 'b' to object 'a', with options to include/exclude certain attributes.

    Args:
        a (object): Destination object to copy attributes to.
        b (object): Source object to copy attributes from.
        include (tuple, optional): Attributes to include. If empty, all attributes are included. Defaults to ().
        exclude (tuple, optional): Attributes to exclude. Defaults to ().
    """
    for k, v in b.__dict__.items():
        if (len(include) and k not in include) or k.startswith("_") or k in exclude:
            continue
        else:
            setattr(a, k, v)

def get_latest_opset():
    """
    Return the second-most recent ONNX opset version supported by this version of PyTorch, adjusted for maturity.

    Returns:
        (int): The ONNX opset version.
    """
    if TORCH_1_13:
        # If the PyTorch>=1.13, dynamically compute the latest opset minus one using 'symbolic_opset'
        return max(int(k[14:]) for k in vars(torch.onnx) if "symbolic_opset" in k) - 1
    # Otherwise for PyTorch<=1.12 return the corresponding predefined opset
    version = torch.onnx.producer_version.rsplit(".", 1)[0]  # i.e. '2.3'
    return {"1.12": 15, "1.11": 14, "1.10": 13, "1.9": 12, "1.8": 12}.get(version, 12)

def intersect_dicts(da, db, exclude=()):
    """
    Returns a dictionary of intersecting keys with matching shapes, excluding 'exclude' keys, using da values.

    Args:
        da (dict): First dictionary.
        db (dict): Second dictionary.
        exclude (tuple, optional): Keys to exclude. Defaults to ().

    Returns:
        (dict): Dictionary of intersecting keys with matching shapes.
    """
    return {k: v for k, v in da.items() if k in db and all(x not in k for x in exclude) and v.shape == db[k].shape}

def is_parallel(model):
    """
    Returns True if model is of type DP or DDP.

    Args:
        model (nn.Module): Model to check.

    Returns:
        (bool): True if model is DataParallel or DistributedDataParallel.
    """
    return isinstance(model, (nn.parallel.DataParallel, nn.parallel.DistributedDataParallel))

def de_parallel(model):
    """
    De-parallelize a model: returns single-GPU model if model is of type DP or DDP.

    Args:
        model (nn.Module): Model to de-parallelize.

    Returns:
        (nn.Module): De-parallelized model.
    """
    return model.module if is_parallel(model) else model

def one_cycle(y1=0.0, y2=1.0, steps=100):
    """
    Returns a lambda function for sinusoidal ramp from y1 to y2 https://arxiv.org/pdf/1812.01187.pdf.

    Args:
        y1 (float, optional): Initial value. Defaults to 0.0.
        y2 (float, optional): Final value. Defaults to 1.0.
        steps (int, optional): Number of steps. Defaults to 100.

    Returns:
        (function): Lambda function for computing the sinusoidal ramp.
    """
    return lambda x: max((1 - math.cos(x * math.pi / steps)) / 2, 0) * (y2 - y1) + y1

def init_seeds(seed=0, deterministic=False):
    """
    Initialize random number generator (RNG) seeds https://pytorch.org/docs/stable/notes/randomness.html.

    Args:
        seed (int, optional): Random seed. Defaults to 0.
        deterministic (bool, optional): Whether to set deterministic algorithms. Defaults to False.
    """
    random.seed(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)
    torch.cuda.manual_seed(seed)
    torch.cuda.manual_seed_all(seed)  # for Multi-GPU, exception safe
    # torch.backends.cudnn.benchmark = True  # AutoBatch problem https://github.com/ultralytics/yolov5/issues/9287
    if deterministic:
        if TORCH_2_0:
            torch.use_deterministic_algorithms(True, warn_only=True)  # warn if deterministic is not possible
            torch.backends.cudnn.deterministic = True
            os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8"
            os.environ["PYTHONHASHSEED"] = str(seed)
        else:
            LOGGER.warning("Upgrade to torch>=2.0.0 for deterministic training.")
    else:
        unset_deterministic()

def unset_deterministic():
    """Unsets all the configurations applied for deterministic training."""
    torch.use_deterministic_algorithms(False)
    torch.backends.cudnn.deterministic = False
    os.environ.pop("CUBLAS_WORKSPACE_CONFIG", None)
    os.environ.pop("PYTHONHASHSEED", None)

def strip_optimizer(f: Union[str, Path] = "best.pt", s: str = "", updates: dict = None) -> dict:
    """
    Strip optimizer from 'f' to finalize training, optionally save as 's'.

    Args:
        f (str | Path): File path to model to strip the optimizer from. Defaults to 'best.pt'.
        s (str, optional): File path to save the model with stripped optimizer to. If not provided, 'f' will be overwritten.
        updates (dict, optional): A dictionary of updates to overlay onto the checkpoint before saving.

    Returns:
        (dict): The combined checkpoint dictionary.

    Examples:
        >>> from pathlib import Path
        >>> from ultralytics.utils.torch_utils import strip_optimizer
        >>> for f in Path("path/to/model/checkpoints").rglob("*.pt"):
        >>>    strip_optimizer(f)
    """
    try:
        x = torch.load(f, map_location=torch.device("cpu"))
        assert isinstance(x, dict), "checkpoint is not a Python dictionary"
        assert "model" in x, "'model' missing from checkpoint"
    except Exception as e:
        LOGGER.warning(f"Skipping {f}, not a valid Ultralytics model: {e}")
        return {}

    metadata = {
        "date": datetime.now().isoformat(),
        "version": __version__,
        "license": "AGPL-3.0 License (https://ultralytics.com/license)",
        "docs": "https://docs.ultralytics.com",
    }

    # Update model
    if x.get("ema"):
        x["model"] = x["ema"]  # replace model with EMA
    if hasattr(x["model"], "args"):
        x["model"].args = dict(x["model"].args)  # convert from IterableSimpleNamespace to dict
    if hasattr(x["model"], "criterion"):
        x["model"].criterion = None  # strip loss criterion
    x["model"].half()  # to FP16
    for p in x["model"].parameters():
        p.requires_grad = False

    # Update other keys
    args = {**DEFAULT_CFG_DICT, **x.get("train_args", {})}  # combine args
    for k in "optimizer", "best_fitness", "ema", "updates":  # keys
        x[k] = None
    x["epoch"] = -1
    x["train_args"] = {k: v for k, v in args.items() if k in DEFAULT_CFG_KEYS}  # strip non-default keys
    # x['model'].args = x['train_args']

    # Save
    combined = {**metadata, **x, **(updates or {})}
    torch.save(combined, s or f)  # combine dicts (prefer to the right)
    mb = os.path.getsize(s or f) / 1e6  # file size
    LOGGER.info(f"Optimizer stripped from {f},{f' saved as {s},' if s else ''} {mb:.1f}MB")
    return combined