|hfddlZddlZddlmZddlmZddlmZGddeZ Gdde eZ y) N)LoadVisualPrompt)DetectionPredictor)SegmentationPredictorcVeZdZdZd deffd ZdZfdZd fd ZfdZ dZ xZ S) YOLOEVPDetectPredictoraN A mixin class for YOLO-EVP (Enhanced Visual Prompting) predictors. This mixin provides common functionality for YOLO models that use visual prompting, including model setup, prompt handling, and preprocessing transformations. Attributes: model (torch.nn.Module): The YOLO model for inference. device (torch.device): Device to run the model on (CPU or CUDA). prompts (dict | torch.Tensor): Visual prompts containing class indices and bounding boxes or masks. Methods: setup_model: Initialize the YOLO model and set it to evaluation mode. set_prompts: Set the visual prompts for the model. pre_transform: Preprocess images and prompts before inference. inference: Run inference with visual prompts. get_vpe: Process source to get visual prompt embeddings. verbosec6t|||d|_y)z Set up the model for prediction. Args: model (torch.nn.Module): Model to load or use. verbose (bool, optional): If True, provides detailed logging. )rTN)super setup_model done_warmup)selfmodelr __class__s d/var/www/html/test/engine/venv/lib/python3.12/site-packages/ultralytics/models/yolo/yoloe/predict.pyr z"YOLOEVPDetectPredictor.setup_models E73c||_y)z Set the visual prompts for the model. Args: prompts (dict): Dictionary containing class indices and bounding boxes or masks. Must include a 'cls' key with class indices. N)prompts)r rs r set_promptsz"YOLOEVPDetectPredictor.set_prompts*s  rc tt||}|jjdd}|jjdd}|jd}t |dk(rf|j |dj dd|dj dd|||}|jdj|j|_|S| Jd|d t|trtd |Ds Jd |d t|trtd |Ds Jd |d t |t |cxk(rt |k(s.nJd t |dt |dt |d tt |Dcgc]<}|j ||j dd||j dd||||>}}tjj j"j%|dj|j|_|Scc}w)a Preprocess images and prompts before inference. This method applies letterboxing to the input image and transforms the visual prompts (bounding boxes or masks) accordingly. Args: im (list): List containing a single input image. Returns: (list): Preprocessed image ready for model inference. Raises: ValueError: If neither valid bounding boxes nor masks are provided in the prompts. bboxesNmasksclsrzExpected bboxes, but got !c3PK|]}t|tj ywN isinstancenpndarray.0bs r z7YOLOEVPDetectPredictor.pre_transform..Os3^RSJq"**4M3^$&z#Expected List[np.ndarray], but got c3PK|]}t|tj ywrrr"s rr%z7YOLOEVPDetectPredictor.pre_transform..Rs5bTUjBJJ6O5br&z-Expected same length for all inputs, but got vsT) batch_first)r pre_transformrpoplen_process_single_imageshape unsqueezetodevicerlistallrangetorchnnutilsrnn pad_sequence) r imimgrrcategoryvisualsirs rr*z$YOLOEVPDetectPredictor.pre_transform4s2 g#B'!!(D1   $/<<& s8q=00Qbq1A2a5;;rPQ?T\^dfklG",,Q/224;;?DL( #% L)B6(!'L L%fd+3^W]3^0^ 5fXQ? ^h-#5bYa5b2b 5hZqA br7c(m:s6{: ?By3x=/Y[\_`f\g[hhij : s3x**3q6<<+;RU[[!_hWXk[abc[deG!88>>--::7PT:UXXY]YdYdeDL sAH5c@|t|rtj|tj}|jdk(r |dddf}t |d|dz |d|dz }||z}|ddddfxxt |d|d|zz dz dz z cc<|ddddfxxt |d|d|zz dz dz z cc<n:|-t|!|}tj|}d||dk(<n td tj||||S) a Process a single image by resizing bounding boxes or masks and generating visuals. Args: dst_shape (tuple): The target shape (height, width) of the image. src_shape (tuple): The original shape (height, width) of the image. category (str): The category of the image for visual prompts. bboxes (list | np.ndarray, optional): A list of bounding boxes in the format [x1, y1, x2, y2]. masks (np.ndarray, optional): A list of masks corresponding to the image. Returns: (torch.Tensor): The processed visuals for the image. Raises: ValueError: If neither `bboxes` nor `masks` are provided. N)dtyperr.rg?rz$Please provide valid bboxes or masks) r,r arrayfloat32ndimminroundr r*stack ValueErrorr get_visuals) r dst_shape src_shaper<rrgain resized_masksrs rr-z,YOLOEVPDetectPredictor._process_single_image`s>"  #f+XXfBJJ7F{{aay|il2IaL9Q<4OPD dNF 319  ! y|d7J(Ja'ORU'U!V V  319  ! y|d7J(Ja'ORU'U!V V   !G1%8MHH]+E"#E%3, CD D !--h 65QQrcBt||g|d|ji|S)a/ Run inference with visual prompts. Args: im (torch.Tensor): Input image tensor. *args (Any): Variable length argument list. **kwargs (Any): Arbitrary keyword arguments. Returns: (torch.Tensor): Model prediction results. vpe)r inferencer)r r:argskwargsrs rrPz YOLOEVPDetectPredictor.inferences(w GGGGGrc|j|t|jdk(sJd|jD]6\}}}|j|}|j ||j dcSy)a Process the source to get the visual prompt embeddings (VPE). Args: source (str | Path | int | PIL.Image | np.ndarray | torch.Tensor | List | Tuple): The source of the image to make predictions on. Accepts various types including file paths, URLs, PIL images, numpy arrays, and torch tensors. Returns: (torch.Tensor): The visual prompt embeddings (VPE) from the model. rz get_vpe only supports one image!T)rO return_vpeN) setup_sourcer,dataset preprocessrr)r source_im0sr:s rget_vpezYOLOEVPDetectPredictor.get_vpesq &!4<< A%I'II%,, EJAtQ&B::bdllt:D D Er)T)NN) __name__ __module__ __qualname____doc__boolr rr*r-rPr[ __classcell__)rs@rrr s2&  $  *X#RJ HErrceZdZdZy)YOLOEVPSegPredictorz\Predictor for YOLO-EVP segmentation tasks combining detection and segmentation capabilities.N)r\r]r^r_rrrcrcsfrrc) numpyr r5ultralytics.data.augmentrultralytics.models.yolo.detectrultralytics.models.yolo.segmentrrrcrdrrris8 5=AXE/XEv 02G r