ultralytics 8.3.98__py3-none-any.whl → 8.3.100__py3-none-any.whl

ultralytics/models/yolo/segment/val.py

@@ -364,29 +364,56 @@ class SegmentationValidator(DetectionValidator):
 
     def eval_json(self, stats):
         """Return COCO-style object detection evaluation metrics."""
-        if self.args.save_json and self.is_coco and len(self.jdict):
-            anno_json = self.data["path"] / "annotations/instances_val2017.json"  # annotations
+        if self.args.save_json and (self.is_lvis or self.is_coco) and len(self.jdict):
             pred_json = self.save_dir / "predictions.json"  # predictions
-            LOGGER.info(f"\nEvaluating pycocotools mAP using {pred_json} and {anno_json}...")
-            try:  # https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb
-                check_requirements("pycocotools>=2.0.6")
-                from pycocotools.coco import COCO  # noqa
-                from pycocotools.cocoeval import COCOeval  # noqa
 
+            anno_json = (
+                self.data["path"]
+                / "annotations"
+                / ("instances_val2017.json" if self.is_coco else f"lvis_v1_{self.args.split}.json")
+            )  # annotations
+
+            pkg = "pycocotools" if self.is_coco else "lvis"
+            LOGGER.info(f"\nEvaluating {pkg} mAP using {pred_json} and {anno_json}...")
+            try:  # https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb
                 for x in anno_json, pred_json:
                     assert x.is_file(), f"{x} file not found"
-                anno = COCO(str(anno_json))  # init annotations api
-                pred = anno.loadRes(str(pred_json))  # init predictions api (must pass string, not Path)
-                for i, eval in enumerate([COCOeval(anno, pred, "bbox"), COCOeval(anno, pred, "segm")]):
-                    if self.is_coco:
-                        eval.params.imgIds = [int(Path(x).stem) for x in self.dataloader.dataset.im_files]  # im to eval
+                check_requirements("pycocotools>=2.0.6" if self.is_coco else "lvis>=0.5.3")
+                if self.is_coco:
+                    from pycocotools.coco import COCO  # noqa
+                    from pycocotools.cocoeval import COCOeval  # noqa
+
+                    anno = COCO(str(anno_json))  # init annotations api
+                    pred = anno.loadRes(str(pred_json))  # init predictions api (must pass string, not Path)
+                    vals = [COCOeval(anno, pred, "bbox"), COCOeval(anno, pred, "segm")]
+                else:
+                    from lvis import LVIS, LVISEval
+
+                    anno = LVIS(str(anno_json))
+                    pred = anno._load_json(str(pred_json))
+                    vals = [LVISEval(anno, pred, "bbox"), LVISEval(anno, pred, "segm")]
+
+                for i, eval in enumerate(vals):
+                    eval.params.imgIds = [int(Path(x).stem) for x in self.dataloader.dataset.im_files]  # im to eval
                     eval.evaluate()
                     eval.accumulate()
                     eval.summarize()
+                    if self.is_lvis:
+                        eval.print_results()
                     idx = i * 4 + 2
-                    stats[self.metrics.keys[idx + 1]], stats[self.metrics.keys[idx]] = eval.stats[
-                        :2
-                    ]  # update mAP50-95 and mAP50
+                    # update mAP50-95 and mAP50
+                    stats[self.metrics.keys[idx + 1]], stats[self.metrics.keys[idx]] = (
+                        eval.stats[:2] if self.is_coco else [eval.results["AP"], eval.results["AP50"]]
+                    )
+                    if self.is_lvis:
+                        tag = "B" if i == 0 else "M"
+                        stats[f"metrics/APr({tag})"] = eval.results["APr"]
+                        stats[f"metrics/APc({tag})"] = eval.results["APc"]
+                        stats[f"metrics/APf({tag})"] = eval.results["APf"]
+
+                if self.is_lvis:
+                    stats["fitness"] = stats["metrics/mAP50-95(B)"]
+
             except Exception as e:
-                LOGGER.warning(f"pycocotools unable to run: {e}")
+                LOGGER.warning(f"{pkg} unable to run: {e}")
         return stats

ultralytics/models/yolo/yoloe/__init__.py

@@ -0,0 +1,21 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+from .predict import YOLOEVPDetectPredictor, YOLOEVPSegPredictor
+from .train import YOLOEPEFreeTrainer, YOLOEPETrainer, YOLOETrainer, YOLOEVPTrainer
+from .train_seg import YOLOEPESegTrainer, YOLOESegTrainer, YOLOESegTrainerFromScratch, YOLOESegVPTrainer
+from .val import YOLOEDetectValidator, YOLOESegValidator
+
+__all__ = [
+    "YOLOETrainer",
+    "YOLOEPETrainer",
+    "YOLOESegTrainer",
+    "YOLOEDetectValidator",
+    "YOLOESegValidator",
+    "YOLOEPESegTrainer",
+    "YOLOESegTrainerFromScratch",
+    "YOLOESegVPTrainer",
+    "YOLOEVPTrainer",
+    "YOLOEPEFreeTrainer",
+    "YOLOEVPDetectPredictor",
+    "YOLOEVPSegPredictor",
+]

ultralytics/models/yolo/yoloe/predict.py

@@ -0,0 +1,170 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+
+import numpy as np
+import torch
+
+from ultralytics.data.augment import LoadVisualPrompt
+from ultralytics.models.yolo.detect import DetectionPredictor
+from ultralytics.models.yolo.segment import SegmentationPredictor
+
+
+class YOLOEVPDetectPredictor(DetectionPredictor):
+    """
+    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): Visual prompts containing class indices and bounding boxes or masks.
+
+    Methods:
+        setup_model: Initialize the YOLO model and set it to evaluation mode.
+        set_return_vpe: Set whether to return visual prompt embeddings.
+        set_prompts: Set the visual prompts for the model.
+        pre_transform: Preprocess images and prompts before inference.
+        inference: Run inference with visual prompts.
+    """
+
+    def setup_model(self, model, verbose=True):
+        """
+        Sets up the model for prediction.
+
+        Args:
+            model (torch.nn.Module): Model to load or use.
+            verbose (bool): If True, provides detailed logging.
+        """
+        super().setup_model(model, verbose=verbose)
+        self.done_warmup = True
+
+    def set_prompts(self, prompts):
+        """
+        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.
+        """
+        self.prompts = prompts
+
+    def pre_transform(self, im):
+        """
+        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.
+        """
+        img = super().pre_transform(im)
+        bboxes = self.prompts.pop("bboxes", None)
+        masks = self.prompts.pop("masks", None)
+        category = self.prompts["cls"]
+        if len(img) == 1:
+            visuals = self._process_single_image(img[0].shape[:2], im[0].shape[:2], category, bboxes, masks)
+            self.prompts = visuals.unsqueeze(0).to(self.device)  # (1, N, H, W)
+        else:
+            # NOTE: only supports bboxes as prompts for now
+            assert bboxes is not None, f"Expected bboxes, but got {bboxes}!"
+            # NOTE: needs List[np.ndarray]
+            assert isinstance(bboxes, list) and all(isinstance(b, np.ndarray) for b in bboxes), (
+                f"Expected List[np.ndarray], but got {bboxes}!"
+            )
+            assert isinstance(category, list) and all(isinstance(b, np.ndarray) for b in category), (
+                f"Expected List[np.ndarray], but got {category}!"
+            )
+            assert len(im) == len(category) == len(bboxes), (
+                f"Expected same length for all inputs, but got {len(im)}vs{len(category)}vs{len(bboxes)}!"
+            )
+            visuals = [
+                self._process_single_image(img[i].shape[:2], im[i].shape[:2], category[i], bboxes[i])
+                for i in range(len(img))
+            ]
+            self.prompts = torch.nn.utils.rnn.pad_sequence(visuals, batch_first=True).to(self.device)
+
+        return img
+
+    def _process_single_image(self, dst_shape, src_shape, category, bboxes=None, masks=None):
+        """
+        Processes 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]. Defaults to None.
+            masks (np.ndarray, optional): A list of masks corresponding to the image. Defaults to None.
+
+        Returns:
+            visuals: The processed visuals for the image.
+
+        Raises:
+            ValueError: If neither `bboxes` nor `masks` are provided.
+        """
+        if bboxes is not None and len(bboxes):
+            bboxes = np.array(bboxes, dtype=np.float32)
+            if bboxes.ndim == 1:
+                bboxes = bboxes[None, :]
+            # Calculate scaling factor and adjust bounding boxes
+            gain = min(dst_shape[0] / src_shape[0], dst_shape[1] / src_shape[1])  # gain = old / new
+            bboxes *= gain
+            bboxes[..., 0::2] += round((dst_shape[1] - src_shape[1] * gain) / 2 - 0.1)
+            bboxes[..., 1::2] += round((dst_shape[0] - src_shape[0] * gain) / 2 - 0.1)
+        elif masks is not None:
+            # Resize and process masks
+            resized_masks = super().pre_transform(masks)
+            masks = np.stack(resized_masks)  # (N, H, W)
+            masks[masks == 114] = 0  # Reset padding values to 0
+        else:
+            raise ValueError("Please provide valid bboxes or masks")
+
+        # Generate visuals using the visual prompt loader
+        return LoadVisualPrompt().get_visuals(category, dst_shape, bboxes, masks)
+
+    def inference(self, im, *args, **kwargs):
+        """
+        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.
+        """
+        return super().inference(im, vpe=self.prompts, *args, **kwargs)
+
+    def get_vpe(self, source):
+        """
+        Processes 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.
+        """
+        self.setup_source(source)
+        assert len(self.dataset) == 1, "get_vpe only supports one image!"
+        for _, im0s, _ in self.dataset:
+            im = self.preprocess(im0s)
+            return self.model(im, vpe=self.prompts, return_vpe=True)
+
+
+class YOLOEVPSegPredictor(YOLOEVPDetectPredictor, SegmentationPredictor):
+    """Predictor for YOLOE VP segmentation."""
+
+    pass

ultralytics/models/yolo/yoloe/train.py

@@ -0,0 +1,355 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+import itertools
+from copy import copy, deepcopy
+from pathlib import Path
+
+import torch
+
+from ultralytics.data import YOLOConcatDataset, build_grounding, build_yolo_dataset
+from ultralytics.data.augment import LoadVisualPrompt
+from ultralytics.data.utils import check_det_dataset
+from ultralytics.models.yolo.detect import DetectionTrainer, DetectionValidator
+from ultralytics.nn.tasks import YOLOEModel
+from ultralytics.utils import DEFAULT_CFG, LOGGER, RANK
+from ultralytics.utils.torch_utils import de_parallel
+
+from .val import YOLOEDetectValidator
+
+
+class YOLOETrainer(DetectionTrainer):
+    """A base trainer for YOLOE training."""
+
+    def __init__(self, cfg=DEFAULT_CFG, overrides=None, _callbacks=None):
+        """
+        Initialize the YOLOE Trainer with specified configurations.
+
+        This method sets up the YOLOE trainer with the provided configuration and overrides, initializing
+        the training environment, model, and callbacks for YOLOE object detection training.
+
+        Args:
+            cfg (dict): Configuration dictionary with default training settings from DEFAULT_CFG.
+            overrides (dict, optional): Dictionary of parameter overrides for the default configuration.
+            _callbacks (list, optional): List of callback functions to be applied during training.
+        """
+        if overrides is None:
+            overrides = {}
+        overrides["overlap_mask"] = False
+        super().__init__(cfg, overrides, _callbacks)
+
+    def get_model(self, cfg=None, weights=None, verbose=True):
+        """Return YOLOEModel initialized with specified config and weights."""
+        # NOTE: This `nc` here is the max number of different text samples in one image, rather than the actual `nc`.
+        # NOTE: Following the official config, nc hard-coded to 80 for now.
+        model = YOLOEModel(
+            cfg["yaml_file"] if isinstance(cfg, dict) else cfg,
+            ch=3,
+            nc=min(self.data["nc"], 80),
+            verbose=verbose and RANK == -1,
+        )
+        if weights:
+            model.load(weights)
+
+        return model
+
+    def get_validator(self):
+        """Returns a DetectionValidator for YOLO model validation."""
+        self.loss_names = "box", "cls", "dfl"
+        return YOLOEDetectValidator(
+            self.test_loader, save_dir=self.save_dir, args=copy(self.args), _callbacks=self.callbacks
+        )
+
+    def build_dataset(self, img_path, mode="train", batch=None):
+        """
+        Build YOLO 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:
+            (Dataset): YOLO dataset configured for training or validation.
+        """
+        gs = max(int(de_parallel(self.model).stride.max() if self.model else 0), 32)
+        return build_yolo_dataset(
+            self.args, img_path, batch, self.data, mode=mode, rect=mode == "val", stride=gs, multi_modal=mode == "train"
+        )
+
+    def preprocess_batch(self, batch):
+        """Process batch for training, moving text features to the appropriate device."""
+        batch = super().preprocess_batch(batch)
+        return batch
+
+
+class YOLOEPETrainer(DetectionTrainer):
+    """Fine-tune YOLOE model in linear probing way."""
+
+    def get_model(self, cfg=None, weights=None, verbose=True):
+        """
+        Return YOLOEModel initialized with specified config and weights.
+
+        Args:
+            cfg (dict | str, optional): Model configuration.
+            weights (str, optional): Path to pretrained weights.
+            verbose (bool): Whether to display model information.
+
+        Returns:
+            (YOLOEModel): Initialized model with frozen layers except for specific projection layers.
+        """
+        # NOTE: This `nc` here is the max number of different text samples in one image, rather than the actual `nc`.
+        # NOTE: Following the official config, nc hard-coded to 80 for now.
+        model = YOLOEModel(
+            cfg["yaml_file"] if isinstance(cfg, dict) else cfg,
+            ch=3,
+            nc=self.data["nc"],
+            verbose=verbose and RANK == -1,
+        )
+
+        del model.model[-1].savpe
+
+        assert weights is not None, "Pretrained weights must be provided for linear probing."
+        if weights:
+            model.load(weights)
+
+        model.eval()
+        names = list(self.data["names"].values())
+        # NOTE: `get_text_pe` related to text model and YOLOEDetect.reprta,
+        # it'd get correct results as long as loading proper pretrained weights.
+        tpe = model.get_text_pe(names)
+        model.set_classes(names, tpe)
+        model.model[-1].fuse(model.pe)  # fuse text embeddings to classify head
+        model.model[-1].cv3[0][2] = deepcopy(model.model[-1].cv3[0][2]).requires_grad_(True)
+        model.model[-1].cv3[1][2] = deepcopy(model.model[-1].cv3[1][2]).requires_grad_(True)
+        model.model[-1].cv3[2][2] = deepcopy(model.model[-1].cv3[2][2]).requires_grad_(True)
+        del model.pe
+        model.train()
+
+        return model
+
+
+class YOLOETrainerFromScratch(YOLOETrainer):
+    """Train YOLOE models from scratch."""
+
+    def __init__(self, cfg=DEFAULT_CFG, overrides=None, _callbacks=None):
+        """
+        Initialize the YOLOETrainerFromScratch class.
+
+        This class extends YOLOETrainer to train YOLOE models from scratch. It inherits all functionality from
+        the parent class while providing specialized initialization for training without pre-trained weights.
+
+        Args:
+            cfg (dict, optional): Configuration dictionary with training parameters. Defaults to DEFAULT_CFG.
+            overrides (dict, optional): Dictionary of parameter overrides for configuration.
+            _callbacks (list, optional): List of callback functions to be executed during training.
+
+        Examples:
+            >>> from ultralytics.models.yoloe.train import YOLOETrainerFromScratch
+            >>> trainer = YOLOETrainerFromScratch()
+            >>> trainer.train()
+        """
+        if overrides is None:
+            overrides = {}
+        super().__init__(cfg, overrides, _callbacks)
+
+    def build_dataset(self, img_path, mode="train", batch=None):
+        """
+        Build YOLO Dataset for training or validation.
+
+        This method constructs appropriate datasets based on the mode and input paths, handling both
+        standard YOLO datasets and grounding datasets with different formats.
+
+        Args:
+            img_path (List[str] | str): Path to the folder containing images or list of paths.
+            mode (str): 'train' mode or 'val' mode, allowing customized augmentations for each mode.
+            batch (int, optional): Size of batches, used for rectangular training/validation.
+
+        Returns:
+            (YOLOConcatDataset | Dataset): The constructed dataset for training or validation.
+        """
+        gs = max(int(de_parallel(self.model).stride.max() if self.model else 0), 32)
+        if mode != "train":
+            return build_yolo_dataset(self.args, img_path, batch, self.data, mode=mode, rect=False, stride=gs)
+        datasets = [
+            build_yolo_dataset(self.args, im_path, batch, self.training_data[im_path], stride=gs, multi_modal=True)
+            if isinstance(im_path, str)
+            else build_grounding(self.args, im_path["img_path"], im_path["json_file"], batch, stride=gs)
+            for im_path in img_path
+        ]
+        self.set_text_embeddings(datasets, batch)  # cache text embeddings to accelerate training
+        return YOLOConcatDataset(datasets) if len(datasets) > 1 else datasets[0]
+
+    def set_text_embeddings(self, datasets, batch):
+        """Set text embeddings for datasets to accelerate training by caching category names."""
+        # TODO: open up an interface to determine whether to do cache
+        category_names = set()
+        for dataset in datasets:
+            if not hasattr(dataset, "category_names"):
+                continue
+            category_names |= dataset.category_names
+
+        # TODO: enable to update the path or use a more general way to get the path
+        img_path = datasets[0].img_path
+        self.text_embeddings = self.generate_text_embeddings(
+            category_names, batch, cache_path=Path(img_path).parent / "text_embeddings.pt"
+        )
+
+    def preprocess_batch(self, batch):
+        """Process batch for training, moving text features to the appropriate device."""
+        batch = super().preprocess_batch(batch)
+
+        texts = list(itertools.chain(*batch["texts"]))
+        txt_feats = torch.stack([self.text_embeddings[text] for text in texts]).to(self.device)
+        txt_feats = txt_feats.reshape(len(batch["texts"]), -1, txt_feats.shape[-1])
+        batch["txt_feats"] = txt_feats
+        return batch
+
+    def generate_text_embeddings(self, texts, batch, cache_path="embeddings.pt"):
+        """
+        Generate text embeddings for a list of text samples.
+
+        Args:
+            texts (List[str]): List of text samples to encode.
+            batch (int): Batch size for processing.
+            cache_path (str | Path): Path to save/load cached embeddings.
+
+        Returns:
+            (dict): Dictionary mapping text samples to their embeddings.
+        """
+        if cache_path.exists():
+            return torch.load(cache_path)
+        assert self.model is not None
+        txt_feats = self.model.get_text_pe(texts, batch, without_reprta=True)
+        txt_map = dict(zip(texts, txt_feats.squeeze(0)))
+        torch.save(txt_map, cache_path)
+        return txt_map
+
+    def get_dataset(self):
+        """
+        Get train and validation paths from data dictionary.
+
+        Processes the data configuration to extract paths for training and validation datasets,
+        handling both YOLO detection datasets and grounding datasets.
+
+        Returns:
+            (str): Train dataset path.
+            (str): Validation dataset path.
+
+        Raises:
+            AssertionError: If train or validation datasets are not found, or if validation has multiple datasets.
+        """
+        final_data = {}
+        data_yaml = self.args.data
+        assert data_yaml.get("train", False), "train dataset not found"  # object365.yaml
+        assert data_yaml.get("val", False), "validation dataset not found"  # lvis.yaml
+        data = {k: [check_det_dataset(d) for d in v.get("yolo_data", [])] for k, v in data_yaml.items()}
+        assert len(data["val"]) == 1, f"Only support validating on 1 dataset for now, but got {len(data['val'])}."
+        val_split = "minival" if "lvis" in data["val"][0]["val"] else "val"
+        for d in data["val"]:
+            if d.get("minival") is None:  # for lvis dataset
+                continue
+            d["minival"] = str(d["path"] / d["minival"])
+        for s in ["train", "val"]:
+            final_data[s] = [d["train" if s == "train" else val_split] for d in data[s]]
+            # save grounding data if there's one
+            grounding_data = data_yaml[s].get("grounding_data")
+            if grounding_data is None:
+                continue
+            grounding_data = grounding_data if isinstance(grounding_data, list) else [grounding_data]
+            for g in grounding_data:
+                assert isinstance(g, dict), f"Grounding data should be provided in dict format, but got {type(g)}"
+            final_data[s] += grounding_data
+        # NOTE: to make training work properly, set `nc` and `names`
+        final_data["nc"] = data["val"][0]["nc"]
+        final_data["names"] = data["val"][0]["names"]
+        # NOTE: add path with lvis path
+        final_data["path"] = data["val"][0]["path"]
+        self.data = final_data
+        if self.args.single_cls:  # consistent with base trainer
+            LOGGER.info("Overriding class names with single class.")
+            self.data["names"] = {0: "object"}
+            self.data["nc"] = 1
+        self.training_data = {}
+        for d in data["train"]:
+            if self.args.single_cls:
+                d["names"] = {0: "object"}
+                d["nc"] = 1
+            self.training_data[d["train"]] = d
+        return final_data["train"], final_data["val"][0]
+
+    def plot_training_labels(self):
+        """Do not plot labels for YOLO-World training."""
+        pass
+
+    def final_eval(self):
+        """
+        Perform final evaluation on the validation dataset.
+
+        Configures the validator with the appropriate dataset and split before running evaluation.
+
+        Returns:
+            (dict): Evaluation metrics.
+        """
+        val = self.args.data["val"]["yolo_data"][0]
+        self.validator.args.data = val
+        self.validator.args.split = "minival" if isinstance(val, str) and "lvis" in val else "val"
+        return super().final_eval()
+
+
+class YOLOEPEFreeTrainer(YOLOEPETrainer, YOLOETrainerFromScratch):
+    """Train prompt-free YOLOE model."""
+
+    def get_validator(self):
+        """Returns a DetectionValidator for YOLO model validation."""
+        self.loss_names = "box", "cls", "dfl"
+        return DetectionValidator(
+            self.test_loader, save_dir=self.save_dir, args=copy(self.args), _callbacks=self.callbacks
+        )
+
+    def preprocess_batch(self, batch):
+        """Preprocesses a batch of images for YOLOE training, adjusting formatting and dimensions as needed."""
+        batch = super(YOLOETrainer, self).preprocess_batch(batch)
+        return batch
+
+    def set_text_embeddings(self, datasets, batch):
+        """No need to set text embeddings for prompt-free fine-tuning."""
+        pass
+
+
+class YOLOEVPTrainer(YOLOETrainerFromScratch):
+    """Train YOLOE model with visual prompts."""
+
+    def build_dataset(self, img_path, mode="train", batch=None):
+        """
+        Build YOLO Dataset for training or validation with visual prompts.
+
+        Args:
+            img_path (List[str] | str): Path to the folder containing images or list of paths.
+            mode (str): 'train' mode or 'val' mode, allowing customized augmentations for each mode.
+            batch (int, optional): Size of batches, used for rectangular training/validation.
+
+        Returns:
+            (Dataset): YOLO dataset configured for training or validation, with visual prompts for training mode.
+        """
+        dataset = super().build_dataset(img_path, mode, batch)
+        if isinstance(dataset, YOLOConcatDataset):
+            for d in dataset.datasets:
+                d.transforms.append(LoadVisualPrompt())
+        else:
+            dataset.transforms.append(LoadVisualPrompt())
+        return dataset
+
+    def _close_dataloader_mosaic(self):
+        """Close mosaic augmentation and add visual prompt loading to the training dataset."""
+        super()._close_dataloader_mosaic()
+        if isinstance(self.train_loader.dataset, YOLOConcatDataset):
+            for d in self.train_loader.dataset.datasets:
+                d.transforms.append(LoadVisualPrompt())
+        else:
+            self.train_loader.dataset.transforms.append(LoadVisualPrompt())
+
+    def preprocess_batch(self, batch):
+        """Preprocesses a batch of images for YOLOE training, moving visual prompts to the appropriate device."""
+        batch = super().preprocess_batch(batch)
+        batch["visuals"] = batch["visuals"].to(self.device)
+        return batch