# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license import torch from ultralytics.data import YOLODataset from ultralytics.data.augment import Compose, Format, v8_transforms from ultralytics.models.yolo.detect import DetectionValidator from ultralytics.utils import colorstr, ops __all__ = ("RTDETRValidator",) # tuple or list class RTDETRDataset(YOLODataset): """ Real-Time DEtection and TRacking (RT-DETR) dataset class extending the base YOLODataset class. This specialized dataset class is designed for use with the RT-DETR object detection model and is optimized for real-time detection and tracking tasks. """ def __init__(self, *args, data=None, **kwargs): """ Initialize the RTDETRDataset class by inheriting from the YOLODataset class. This constructor sets up a dataset specifically optimized for the RT-DETR (Real-Time DEtection and TRacking) model, building upon the base YOLODataset functionality. Args: *args (Any): Variable length argument list passed to the parent YOLODataset class. data (Dict | None): Dictionary containing dataset information. If None, default values will be used. **kwargs (Any): Additional keyword arguments passed to the parent YOLODataset class. """ super().__init__(*args, data=data, **kwargs) def load_image(self, i, rect_mode=False): """ Load one image from dataset index 'i'. Args: i (int): Index of the image to load. rect_mode (bool, optional): Whether to use rectangular mode for batch inference. Returns: im (numpy.ndarray): The loaded image. resized_hw (tuple): Height and width of the resized image with shape (2,). Examples: >>> dataset = RTDETRDataset(...) >>> image, hw = dataset.load_image(0) """ return super().load_image(i=i, rect_mode=rect_mode) def build_transforms(self, hyp=None): """ Build transformation pipeline for the dataset. Args: hyp (dict, optional): Hyperparameters for transformations. Returns: (Compose): Composition of transformation functions. """ if self.augment: hyp.mosaic = hyp.mosaic if self.augment and not self.rect else 0.0 hyp.mixup = hyp.mixup if self.augment and not self.rect else 0.0 transforms = v8_transforms(self, self.imgsz, hyp, stretch=True) else: # transforms = Compose([LetterBox(new_shape=(self.imgsz, self.imgsz), auto=False, scale_fill=True)]) transforms = Compose([]) transforms.append( Format( bbox_format="xywh", normalize=True, return_mask=self.use_segments, return_keypoint=self.use_keypoints, batch_idx=True, mask_ratio=hyp.mask_ratio, mask_overlap=hyp.overlap_mask, ) ) return transforms class RTDETRValidator(DetectionValidator): """ RTDETRValidator extends the DetectionValidator class to provide validation capabilities specifically tailored for the RT-DETR (Real-Time DETR) object detection model. The class allows building of an RTDETR-specific dataset for validation, applies Non-maximum suppression for post-processing, and updates evaluation metrics accordingly. Examples: >>> from ultralytics.models.rtdetr import RTDETRValidator >>> args = dict(model="rtdetr-l.pt", data="coco8.yaml") >>> validator = RTDETRValidator(args=args) >>> validator() Note: For further details on the attributes and methods, refer to the parent DetectionValidator class. """ def build_dataset(self, img_path, mode="val", batch=None): """ Build an RTDETR Dataset. Args: img_path (str): Path to the folder containing images. mode (str): `train` mode or `val` mode, users are able to customize different augmentations for each mode. batch (int, optional): Size of batches, this is for `rect`. Returns: (RTDETRDataset): Dataset configured for RT-DETR validation. """ return RTDETRDataset( img_path=img_path, imgsz=self.args.imgsz, batch_size=batch, augment=False, # no augmentation hyp=self.args, rect=False, # no rect cache=self.args.cache or None, prefix=colorstr(f"{mode}: "), data=self.data, ) def postprocess(self, preds): """ Apply Non-maximum suppression to prediction outputs. Args: preds (List | Tuple | torch.Tensor): Raw predictions from the model. Returns: (List[torch.Tensor]): List of processed predictions for each image in batch. """ if not isinstance(preds, (list, tuple)): # list for PyTorch inference but list[0] Tensor for export inference preds = [preds, None] bs, _, nd = preds[0].shape bboxes, scores = preds[0].split((4, nd - 4), dim=-1) bboxes *= self.args.imgsz outputs = [torch.zeros((0, 6), device=bboxes.device)] * bs for i, bbox in enumerate(bboxes): # (300, 4) bbox = ops.xywh2xyxy(bbox) score, cls = scores[i].max(-1) # (300, ) pred = torch.cat([bbox, score[..., None], cls[..., None]], dim=-1) # filter # Sort by confidence to correctly get internal metrics pred = pred[score.argsort(descending=True)] outputs[i] = pred[score > self.args.conf] return outputs def _prepare_batch(self, si, batch): """ Prepares a batch for validation by applying necessary transformations. Args: si (int): Batch index. batch (dict): Batch data containing images and annotations. Returns: (dict): Prepared batch with transformed annotations. """ idx = batch["batch_idx"] == si cls = batch["cls"][idx].squeeze(-1) bbox = batch["bboxes"][idx] ori_shape = batch["ori_shape"][si] imgsz = batch["img"].shape[2:] ratio_pad = batch["ratio_pad"][si] if len(cls): bbox = ops.xywh2xyxy(bbox) # target boxes bbox[..., [0, 2]] *= ori_shape[1] # native-space pred bbox[..., [1, 3]] *= ori_shape[0] # native-space pred return {"cls": cls, "bbox": bbox, "ori_shape": ori_shape, "imgsz": imgsz, "ratio_pad": ratio_pad} def _prepare_pred(self, pred, pbatch): """ Prepares predictions by scaling bounding boxes to original image dimensions. Args: pred (torch.Tensor): Raw predictions. pbatch (dict): Prepared batch information. Returns: (torch.Tensor): Predictions scaled to original image dimensions. """ predn = pred.clone() predn[..., [0, 2]] *= pbatch["ori_shape"][1] / self.args.imgsz # native-space pred predn[..., [1, 3]] *= pbatch["ori_shape"][0] / self.args.imgsz # native-space pred return predn.float()