参考 ultralytics/nn/tasks.py

class BaseModel(nn.Module):
    """The BaseModel class serves as a base class for all the models in the Ultralytics YOLO family."""

    def forward(self, x, *args, **kwargs):
        """
        Forward pass of the model on a single scale. Wrapper for `_forward_once` method.

        Args:
            x (torch.Tensor | dict): The input image tensor or a dict including image tensor and gt labels.

        Returns:
            (torch.Tensor): The output of the network.
        """
        if isinstance(x, dict):  # for cases of training and validating while training.
            return self.loss(x, *args, **kwargs)
        return self.predict(x, *args, **kwargs)

    def predict(self, x, profile=False, visualize=False, augment=False, embed=None):
        """
        Perform a forward pass through the network.

        Args:
            x (torch.Tensor): The input tensor to the model.
            profile (bool):  Print the computation time of each layer if True, defaults to False.
            visualize (bool): Save the feature maps of the model if True, defaults to False.
            augment (bool): Augment image during prediction, defaults to False.
            embed (list, optional): A list of feature vectors/embeddings to return.

        Returns:
            (torch.Tensor): The last output of the model.
        """
        if augment:
            return self._predict_augment(x)
        return self._predict_once(x, profile, visualize, embed)

    def _predict_once(self, x, profile=False, visualize=False, embed=None):
        """
        Perform a forward pass through the network.

        Args:
            x (torch.Tensor): The input tensor to the model.
            profile (bool):  Print the computation time of each layer if True, defaults to False.
            visualize (bool): Save the feature maps of the model if True, defaults to False.
            embed (list, optional): A list of feature vectors/embeddings to return.

        Returns:
            (torch.Tensor): The last output of the model.
        """
        y, dt, embeddings = [], [], []  # outputs
        for m in self.model:
            if m.f != -1:  # if not from previous layer
                x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f]  # from earlier layers
            if profile:
                self._profile_one_layer(m, x, dt)
            x = m(x)  # run
            y.append(x if m.i in self.save else None)  # save output
            if visualize:
                feature_visualization(x, m.type, m.i, save_dir=visualize)
            if embed and m.i in embed:
                embeddings.append(nn.functional.adaptive_avg_pool2d(x, (1, 1)).squeeze(-1).squeeze(-1))  # flatten
                if m.i == max(embed):
                    return torch.unbind(torch.cat(embeddings, 1), dim=0)
        return x

    def _predict_augment(self, x):
        """Perform augmentations on input image x and return augmented inference."""
        LOGGER.warning(
            f"WARNING ⚠️ {self.__class__.__name__} does not support augmented inference yet. "
            f"Reverting to single-scale inference instead."
        )
        return self._predict_once(x)

    def _profile_one_layer(self, m, x, dt):
        """
        Profile the computation time and FLOPs of a single layer of the model on a given input. Appends the results to
        the provided list.

        Args:
            m (nn.Module): The layer to be profiled.
            x (torch.Tensor): The input data to the layer.
            dt (list): A list to store the computation time of the layer.

        Returns:
            None
        """
        c = m == self.model[-1] and isinstance(x, list)  # is final layer list, copy input as inplace fix
        flops = thop.profile(m, inputs=[x.copy() if c else x], verbose=False)[0] / 1e9 * 2 if thop else 0  # FLOPs
        t = time_sync()
        for _ in range(10):
            m(x.copy() if c else x)
        dt.append((time_sync() - t) * 100)
        if m == self.model[0]:
            LOGGER.info(f"{'time (ms)':>10s} {'GFLOPs':>10s} {'params':>10s}  module")
        LOGGER.info(f"{dt[-1]:10.2f} {flops:10.2f} {m.np:10.0f}  {m.type}")
        if c:
            LOGGER.info(f"{sum(dt):10.2f} {'-':>10s} {'-':>10s}  Total")

    def fuse(self, verbose=True):
        """
        Fuse the `Conv2d()` and `BatchNorm2d()` layers of the model into a single layer, in order to improve the
        computation efficiency.

        Returns:
            (nn.Module): The fused model is returned.
        """
        if not self.is_fused():
            for m in self.model.modules():
                if isinstance(m, (Conv, Conv2, DWConv)) and hasattr(m, "bn"):
                    if isinstance(m, Conv2):
                        m.fuse_convs()
                    m.conv = fuse_conv_and_bn(m.conv, m.bn)  # update conv
                    delattr(m, "bn")  # remove batchnorm
                    m.forward = m.forward_fuse  # update forward
                if isinstance(m, ConvTranspose) and hasattr(m, "bn"):
                    m.conv_transpose = fuse_deconv_and_bn(m.conv_transpose, m.bn)
                    delattr(m, "bn")  # remove batchnorm
                    m.forward = m.forward_fuse  # update forward
                if isinstance(m, RepConv):
                    m.fuse_convs()
                    m.forward = m.forward_fuse  # update forward
            self.info(verbose=verbose)

        return self

    def is_fused(self, thresh=10):
        """
        Check if the model has less than a certain threshold of BatchNorm layers.

        Args:
            thresh (int, optional): The threshold number of BatchNorm layers. Default is 10.

        Returns:
            (bool): True if the number of BatchNorm layers in the model is less than the threshold, False otherwise.
        """
        bn = tuple(v for k, v in nn.__dict__.items() if "Norm" in k)  # normalization layers, i.e. BatchNorm2d()
        return sum(isinstance(v, bn) for v in self.modules()) < thresh  # True if < 'thresh' BatchNorm layers in model

    def info(self, detailed=False, verbose=True, imgsz=640):
        """
        Prints model information.

        Args:
            detailed (bool): if True, prints out detailed information about the model. Defaults to False
            verbose (bool): if True, prints out the model information. Defaults to False
            imgsz (int): the size of the image that the model will be trained on. Defaults to 640
        """
        return model_info(self, detailed=detailed, verbose=verbose, imgsz=imgsz)

    def _apply(self, fn):
        """
        Applies a function to all the tensors in the model that are not parameters or registered buffers.

        Args:
            fn (function): the function to apply to the model

        Returns:
            (BaseModel): An updated BaseModel object.
        """
        self = super()._apply(fn)
        m = self.model[-1]  # Detect()
        if isinstance(m, Detect):  # includes all Detect subclasses like Segment, Pose, OBB, WorldDetect
            m.stride = fn(m.stride)
            m.anchors = fn(m.anchors)
            m.strides = fn(m.strides)
        return self

    def load(self, weights, verbose=True):
        """
        Load the weights into the model.

        Args:
            weights (dict | torch.nn.Module): The pre-trained weights to be loaded.
            verbose (bool, optional): Whether to log the transfer progress. Defaults to True.
        """
        model = weights["model"] if isinstance(weights, dict) else weights  # torchvision models are not dicts
        csd = model.float().state_dict()  # checkpoint state_dict as FP32
        csd = intersect_dicts(csd, self.state_dict())  # intersect
        self.load_state_dict(csd, strict=False)  # load
        if verbose:
            LOGGER.info(f"Transferred {len(csd)}/{len(self.model.state_dict())} items from pretrained weights")

    def loss(self, batch, preds=None):
        """
        Compute loss.

        Args:
            batch (dict): Batch to compute loss on
            preds (torch.Tensor | List[torch.Tensor]): Predictions.
        """
        if not hasattr(self, "criterion"):
            self.criterion = self.init_criterion()

        preds = self.forward(batch["img"]) if preds is None else preds
        return self.criterion(preds, batch)

    def init_criterion(self):
        """Initialize the loss criterion for the BaseModel."""
        raise NotImplementedError("compute_loss() needs to be implemented by task heads")

class DetectionModel(BaseModel):
    """YOLOv8 detection model."""

    def __init__(self, cfg="yolov8n.yaml", ch=3, nc=None, verbose=True):  # model, input channels, number of classes
        """Initialize the YOLOv8 detection model with the given config and parameters."""
        super().__init__()
        self.yaml = cfg if isinstance(cfg, dict) else yaml_model_load(cfg)  # cfg dict

        # Define model
        ch = self.yaml["ch"] = self.yaml.get("ch", ch)  # input channels
        if nc and nc != self.yaml["nc"]:
            LOGGER.info(f"Overriding model.yaml nc={self.yaml['nc']} with nc={nc}")
            self.yaml["nc"] = nc  # override YAML value
        self.model, self.save = parse_model(deepcopy(self.yaml), ch=ch, verbose=verbose)  # model, savelist
        self.names = {i: f"{i}" for i in range(self.yaml["nc"])}  # default names dict
        self.inplace = self.yaml.get("inplace", True)

        # Build strides
        m = self.model[-1]  # Detect()
        if isinstance(m, Detect):  # includes all Detect subclasses like Segment, Pose, OBB, WorldDetect
            s = 256  # 2x min stride
            m.inplace = self.inplace
            forward = lambda x: self.forward(x)[0] if isinstance(m, (Segment, Pose, OBB)) else self.forward(x)
            m.stride = torch.tensor([s / x.shape[-2] for x in forward(torch.zeros(1, ch, s, s))])  # forward
            self.stride = m.stride
            m.bias_init()  # only run once
        else:
            self.stride = torch.Tensor([32])  # default stride for i.e. RTDETR

        # Init weights, biases
        initialize_weights(self)
        if verbose:
            self.info()
            LOGGER.info("")

    def _predict_augment(self, x):
        """Perform augmentations on input image x and return augmented inference and train outputs."""
        img_size = x.shape[-2:]  # height, width
        s = [1, 0.83, 0.67]  # scales
        f = [None, 3, None]  # flips (2-ud, 3-lr)
        y = []  # outputs
        for si, fi in zip(s, f):
            xi = scale_img(x.flip(fi) if fi else x, si, gs=int(self.stride.max()))
            yi = super().predict(xi)[0]  # forward
            yi = self._descale_pred(yi, fi, si, img_size)
            y.append(yi)
        y = self._clip_augmented(y)  # clip augmented tails
        return torch.cat(y, -1), None  # augmented inference, train

    @staticmethod
    def _descale_pred(p, flips, scale, img_size, dim=1):
        """De-scale predictions following augmented inference (inverse operation)."""
        p[:, :4] /= scale  # de-scale
        x, y, wh, cls = p.split((1, 1, 2, p.shape[dim] - 4), dim)
        if flips == 2:
            y = img_size[0] - y  # de-flip ud
        elif flips == 3:
            x = img_size[1] - x  # de-flip lr
        return torch.cat((x, y, wh, cls), dim)

    def _clip_augmented(self, y):
        """Clip YOLO augmented inference tails."""
        nl = self.model[-1].nl  # number of detection layers (P3-P5)
        g = sum(4**x for x in range(nl))  # grid points
        e = 1  # exclude layer count
        i = (y[0].shape[-1] // g) * sum(4**x for x in range(e))  # indices
        y[0] = y[0][..., :-i]  # large
        i = (y[-1].shape[-1] // g) * sum(4 ** (nl - 1 - x) for x in range(e))  # indices
        y[-1] = y[-1][..., i:]  # small
        return y

    def init_criterion(self):
        """Initialize the loss criterion for the DetectionModel."""
        return v8DetectionLoss(self)

class OBBModel(DetectionModel):
    """YOLOv8 Oriented Bounding Box (OBB) model."""

    def __init__(self, cfg="yolov8n-obb.yaml", ch=3, nc=None, verbose=True):
        """Initialize YOLOv8 OBB model with given config and parameters."""
        super().__init__(cfg=cfg, ch=ch, nc=nc, verbose=verbose)

    def init_criterion(self):
        """Initialize the loss criterion for the model."""
        return v8OBBLoss(self)

class SegmentationModel(DetectionModel):
    """YOLOv8 segmentation model."""

    def __init__(self, cfg="yolov8n-seg.yaml", ch=3, nc=None, verbose=True):
        """Initialize YOLOv8 segmentation model with given config and parameters."""
        super().__init__(cfg=cfg, ch=ch, nc=nc, verbose=verbose)

    def init_criterion(self):
        """Initialize the loss criterion for the SegmentationModel."""
        return v8SegmentationLoss(self)

class PoseModel(DetectionModel):
    """YOLOv8 pose model."""

    def __init__(self, cfg="yolov8n-pose.yaml", ch=3, nc=None, data_kpt_shape=(None, None), verbose=True):
        """Initialize YOLOv8 Pose model."""
        if not isinstance(cfg, dict):
            cfg = yaml_model_load(cfg)  # load model YAML
        if any(data_kpt_shape) and list(data_kpt_shape) != list(cfg["kpt_shape"]):
            LOGGER.info(f"Overriding model.yaml kpt_shape={cfg['kpt_shape']} with kpt_shape={data_kpt_shape}")
            cfg["kpt_shape"] = data_kpt_shape
        super().__init__(cfg=cfg, ch=ch, nc=nc, verbose=verbose)

    def init_criterion(self):
        """Initialize the loss criterion for the PoseModel."""
        return v8PoseLoss(self)

class ClassificationModel(BaseModel):
    """YOLOv8 classification model."""

    def __init__(self, cfg="yolov8n-cls.yaml", ch=3, nc=None, verbose=True):
        """Init ClassificationModel with YAML, channels, number of classes, verbose flag."""
        super().__init__()
        self._from_yaml(cfg, ch, nc, verbose)

    def _from_yaml(self, cfg, ch, nc, verbose):
        """Set YOLOv8 model configurations and define the model architecture."""
        self.yaml = cfg if isinstance(cfg, dict) else yaml_model_load(cfg)  # cfg dict

        # Define model
        ch = self.yaml["ch"] = self.yaml.get("ch", ch)  # input channels
        if nc and nc != self.yaml["nc"]:
            LOGGER.info(f"Overriding model.yaml nc={self.yaml['nc']} with nc={nc}")
            self.yaml["nc"] = nc  # override YAML value
        elif not nc and not self.yaml.get("nc", None):
            raise ValueError("nc not specified. Must specify nc in model.yaml or function arguments.")
        self.model, self.save = parse_model(deepcopy(self.yaml), ch=ch, verbose=verbose)  # model, savelist
        self.stride = torch.Tensor([1])  # no stride constraints
        self.names = {i: f"{i}" for i in range(self.yaml["nc"])}  # default names dict
        self.info()

    @staticmethod
    def reshape_outputs(model, nc):
        """Update a TorchVision classification model to class count 'n' if required."""
        name, m = list((model.model if hasattr(model, "model") else model).named_children())[-1]  # last module
        if isinstance(m, Classify):  # YOLO Classify() head
            if m.linear.out_features != nc:
                m.linear = nn.Linear(m.linear.in_features, nc)
        elif isinstance(m, nn.Linear):  # ResNet, EfficientNet
            if m.out_features != nc:
                setattr(model, name, nn.Linear(m.in_features, nc))
        elif isinstance(m, nn.Sequential):
            types = [type(x) for x in m]
            if nn.Linear in types:
                i = types.index(nn.Linear)  # nn.Linear index
                if m[i].out_features != nc:
                    m[i] = nn.Linear(m[i].in_features, nc)
            elif nn.Conv2d in types:
                i = types.index(nn.Conv2d)  # nn.Conv2d index
                if m[i].out_channels != nc:
                    m[i] = nn.Conv2d(m[i].in_channels, nc, m[i].kernel_size, m[i].stride, bias=m[i].bias is not None)

    def init_criterion(self):
        """Initialize the loss criterion for the ClassificationModel."""
        return v8ClassificationLoss()

class RTDETRDetectionModel(DetectionModel):
    """
    RTDETR (Real-time DEtection and Tracking using Transformers) Detection Model class.

    This class is responsible for constructing the RTDETR architecture, defining loss functions, and facilitating both
    the training and inference processes. RTDETR is an object detection and tracking model that extends from the
    DetectionModel base class.

    Attributes:
        cfg (str): The configuration file path or preset string. Default is 'rtdetr-l.yaml'.
        ch (int): Number of input channels. Default is 3 (RGB).
        nc (int, optional): Number of classes for object detection. Default is None.
        verbose (bool): Specifies if summary statistics are shown during initialization. Default is True.

    Methods:
        init_criterion: Initializes the criterion used for loss calculation.
        loss: Computes and returns the loss during training.
        predict: Performs a forward pass through the network and returns the output.
    """

    def __init__(self, cfg="rtdetr-l.yaml", ch=3, nc=None, verbose=True):
        """
        Initialize the RTDETRDetectionModel.

        Args:
            cfg (str): Configuration file name or path.
            ch (int): Number of input channels.
            nc (int, optional): Number of classes. Defaults to None.
            verbose (bool, optional): Print additional information during initialization. Defaults to True.
        """
        super().__init__(cfg=cfg, ch=ch, nc=nc, verbose=verbose)

    def init_criterion(self):
        """Initialize the loss criterion for the RTDETRDetectionModel."""
        from ultralytics.models.utils.loss import RTDETRDetectionLoss

        return RTDETRDetectionLoss(nc=self.nc, use_vfl=True)

    def loss(self, batch, preds=None):
        """
        Compute the loss for the given batch of data.

        Args:
            batch (dict): Dictionary containing image and label data.
            preds (torch.Tensor, optional): Precomputed model predictions. Defaults to None.

        Returns:
            (tuple): A tuple containing the total loss and main three losses in a tensor.
        """
        if not hasattr(self, "criterion"):
            self.criterion = self.init_criterion()

        img = batch["img"]
        # NOTE: preprocess gt_bbox and gt_labels to list.
        bs = len(img)
        batch_idx = batch["batch_idx"]
        gt_groups = [(batch_idx == i).sum().item() for i in range(bs)]
        targets = {
            "cls": batch["cls"].to(img.device, dtype=torch.long).view(-1),
            "bboxes": batch["bboxes"].to(device=img.device),
            "batch_idx": batch_idx.to(img.device, dtype=torch.long).view(-1),
            "gt_groups": gt_groups,
        }

        preds = self.predict(img, batch=targets) if preds is None else preds
        dec_bboxes, dec_scores, enc_bboxes, enc_scores, dn_meta = preds if self.training else preds[1]
        if dn_meta is None:
            dn_bboxes, dn_scores = None, None
        else:
            dn_bboxes, dec_bboxes = torch.split(dec_bboxes, dn_meta["dn_num_split"], dim=2)
            dn_scores, dec_scores = torch.split(dec_scores, dn_meta["dn_num_split"], dim=2)

        dec_bboxes = torch.cat([enc_bboxes.unsqueeze(0), dec_bboxes])  # (7, bs, 300, 4)
        dec_scores = torch.cat([enc_scores.unsqueeze(0), dec_scores])

        loss = self.criterion(
            (dec_bboxes, dec_scores), targets, dn_bboxes=dn_bboxes, dn_scores=dn_scores, dn_meta=dn_meta
        )
        # NOTE: There are like 12 losses in RTDETR, backward with all losses but only show the main three losses.
        return sum(loss.values()), torch.as_tensor(
            [loss[k].detach() for k in ["loss_giou", "loss_class", "loss_bbox"]], device=img.device
        )

    def predict(self, x, profile=False, visualize=False, batch=None, augment=False, embed=None):
        """
        Perform a forward pass through the model.

        Args:
            x (torch.Tensor): The input tensor.
            profile (bool, optional): If True, profile the computation time for each layer. Defaults to False.
            visualize (bool, optional): If True, save feature maps for visualization. Defaults to False.
            batch (dict, optional): Ground truth data for evaluation. Defaults to None.
            augment (bool, optional): If True, perform data augmentation during inference. Defaults to False.
            embed (list, optional): A list of feature vectors/embeddings to return.

        Returns:
            (torch.Tensor): Model's output tensor.
        """
        y, dt, embeddings = [], [], []  # outputs
        for m in self.model[:-1]:  # except the head part
            if m.f != -1:  # if not from previous layer
                x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f]  # from earlier layers
            if profile:
                self._profile_one_layer(m, x, dt)
            x = m(x)  # run
            y.append(x if m.i in self.save else None)  # save output
            if visualize:
                feature_visualization(x, m.type, m.i, save_dir=visualize)
            if embed and m.i in embed:
                embeddings.append(nn.functional.adaptive_avg_pool2d(x, (1, 1)).squeeze(-1).squeeze(-1))  # flatten
                if m.i == max(embed):
                    return torch.unbind(torch.cat(embeddings, 1), dim=0)
        head = self.model[-1]
        x = head([y[j] for j in head.f], batch)  # head inference
        return x