Tài liệu tham khảo cho ultralytics/utils/torch_utils.py

@contextmanager
def torch_distributed_zero_first(local_rank: int):
    """Decorator to make all processes in distributed training wait for each local_master to do something."""
    initialized = torch.distributed.is_available() and torch.distributed.is_initialized()
    if initialized and local_rank not in {-1, 0}:
        dist.barrier(device_ids=[local_rank])
    yield
    if initialized and local_rank == 0:
        dist.barrier(device_ids=[0])

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 get_cpu_info():
    """Return a string with system CPU information, i.e. 'Apple M2'."""
    import cpuinfo  # pip install py-cpuinfo

    k = "brand_raw", "hardware_raw", "arch_string_raw"  # info keys sorted by preference (not all keys always available)
    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")
    return string.replace("(R)", "").replace("CPU ", "").replace("@ ", "")

def select_device(device="", batch=0, newline=False, verbose=True):
    """
    Selects 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):
        return device

    s = f"Ultralytics YOLOv{__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"
        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"  # range(torch.cuda.device_count())  # i.e. 0,1,6,7
        n = len(devices)  # device count
        if n > 1 and 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):
            p = torch.cuda.get_device_properties(i)
            s += f"{'' if i == 0 else space}CUDA:{d} ({p.name}, {p.total_memory / (1 << 20):.0f}MiB)\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 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 https://tehnokv.com/posts/fusing-batchnorm-and-conv/."""
    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.clone().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], 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.clone().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):
    """
    Model information.

    imgsz may be int or list, i.e. imgsz=640 or imgsz=[640, 320].
    """
    if not verbose:
        return
    n_p = get_num_params(model)  # number of parameters
    n_g = get_num_gradients(model)  # number of gradients
    n_l = len(list(model.modules()))  # number of layers
    if detailed:
        LOGGER.info(
            f"{'layer':>5} {'name':>40} {'gradient':>9} {'parameters':>12} {'shape':>20} {'mu':>10} {'sigma':>10}"
        )
        for i, (name, p) in enumerate(model.named_parameters()):
            name = name.replace("module_list.", "")
            LOGGER.info(
                "%5g %40s %9s %12g %20s %10.3g %10.3g %10s"
                % (i, name, p.requires_grad, p.numel(), list(p.shape), p.mean(), p.std(), p.dtype)
            )

    flops = get_flops(model, imgsz)
    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.

    Example:
        YOLOv8n info for loggers
        ```python
        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):
    """Return a YOLO model's FLOPs."""
    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:
            # Use stride size for 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:
            # 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
            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 (thop package alternative, but 2-10x slower unfortunately)."""
    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 of shape img(bs,3,y,x) based on given ratio and grid size gs, optionally
    retaining the original shape.
    """
    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 make_divisible(x, divisor):
    """Returns nearest x divisible by divisor."""
    if isinstance(divisor, torch.Tensor):
        divisor = int(divisor.max())  # to int
    return math.ceil(x / divisor) * divisor

def copy_attr(a, b, include=(), exclude=()):
    """Copies attributes from object 'b' to object 'a', with options to include/exclude certain attributes."""
    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."""
    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."""
    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."""
    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."""
    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."""
    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."""
    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("WARNING ⚠️ Upgrade to torch>=2.0.0 for deterministic training.")
    else:
        torch.use_deterministic_algorithms(False)
        torch.backends.cudnn.deterministic = False

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

    Args:
        f (str): file path to model to strip the optimizer from. Default is 'best.pt'.
        s (str): file path to save the model with stripped optimizer to. If not provided, 'f' will be overwritten.

    Returns:
        None

    Example:
        ```python
        from pathlib import Path
        from ultralytics.utils.torch_utils import strip_optimizer

        for f in Path('path/to/weights').rglob('*.pt'):
            strip_optimizer(f)
        ```
    """
    x = torch.load(f, map_location=torch.device("cpu"))
    if "model" not in x:
        LOGGER.info(f"Skipping {f}, not a valid Ultralytics model.")
        return

    if hasattr(x["model"], "args"):
        x["model"].args = dict(x["model"].args)  # convert from IterableSimpleNamespace to dict
    args = {**DEFAULT_CFG_DICT, **x["train_args"]} if "train_args" in x else None  # combine args
    if x.get("ema"):
        x["model"] = x["ema"]  # replace model with ema
    for k in "optimizer", "best_fitness", "ema", "updates":  # keys
        x[k] = None
    x["epoch"] = -1
    x["model"].half()  # to FP16
    for p in x["model"].parameters():
        p.requires_grad = False
    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']
    torch.save(x, s or f)
    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")

def convert_optimizer_state_dict_to_fp16(state_dict):
    """
    Converts the state_dict of a given optimizer to FP16, focusing on the 'state' key for tensor conversions.

    This method aims to reduce storage size without altering 'param_groups' as they contain non-tensor data.
    """
    for state in state_dict["state"].values():
        for k, v in state.items():
            if k != "step" and isinstance(v, torch.Tensor) and v.dtype is torch.float32:
                state[k] = v.half()

    return state_dict

def profile(input, ops, n=10, device=None):
    """
    Ultralytics speed, memory and FLOPs profiler.

    Example:
        ```python
        from ultralytics.utils.torch_utils import profile

        input = torch.randn(16, 3, 640, 640)
        m1 = lambda x: x * torch.sigmoid(x)
        m2 = nn.SiLU()
        profile(input, [m1, m2], n=100)  # profile over 100 iterations
        ```
    """
    results = []
    if not isinstance(device, torch.device):
        device = select_device(device)
    LOGGER.info(
        f"{'Params':>12s}{'GFLOPs':>12s}{'GPU_mem (GB)':>14s}{'forward (ms)':>14s}{'backward (ms)':>14s}"
        f"{'input':>24s}{'output':>24s}"
    )

    for x in input if isinstance(input, list) else [input]:
        x = x.to(device)
        x.requires_grad = True
        for m in ops if isinstance(ops, list) else [ops]:
            m = m.to(device) if hasattr(m, "to") else m  # device
            m = m.half() if hasattr(m, "half") and isinstance(x, torch.Tensor) and x.dtype is torch.float16 else m
            tf, tb, t = 0, 0, [0, 0, 0]  # dt forward, backward
            try:
                flops = thop.profile(m, inputs=[x], verbose=False)[0] / 1e9 * 2 if thop else 0  # GFLOPs
            except Exception:
                flops = 0

            try:
                for _ in range(n):
                    t[0] = time_sync()
                    y = m(x)
                    t[1] = time_sync()
                    try:
                        (sum(yi.sum() for yi in y) if isinstance(y, list) else y).sum().backward()
                        t[2] = time_sync()
                    except Exception:  # no backward method
                        # print(e)  # for debug
                        t[2] = float("nan")
                    tf += (t[1] - t[0]) * 1000 / n  # ms per op forward
                    tb += (t[2] - t[1]) * 1000 / n  # ms per op backward
                mem = torch.cuda.memory_reserved() / 1e9 if torch.cuda.is_available() else 0  # (GB)
                s_in, s_out = (tuple(x.shape) if isinstance(x, torch.Tensor) else "list" for x in (x, y))  # shapes
                p = sum(x.numel() for x in m.parameters()) if isinstance(m, nn.Module) else 0  # parameters
                LOGGER.info(f"{p:12}{flops:12.4g}{mem:>14.3f}{tf:14.4g}{tb:14.4g}{str(s_in):>24s}{str(s_out):>24s}")
                results.append([p, flops, mem, tf, tb, s_in, s_out])
            except Exception as e:
                LOGGER.info(e)
                results.append(None)
            gc.collect()  # attempt to free unused memory
            torch.cuda.empty_cache()
    return results