Reference for ultralytics/models/yolo/yoloe/val.py

@smart_inference_mode()
def __call__(self, trainer=None, model=None, refer_data=None, load_vp=False):
    """
    Run validation on the model using either text or visual prompt embeddings.

    This method validates the model using either text prompts or visual prompts, depending
    on the `load_vp` flag. It supports validation during training (using a trainer object)
    or standalone validation with a provided model.

    Args:
        trainer (object, optional): Trainer object containing the model and device.
        model (YOLOEModel, optional): Model to validate. Required if `trainer` is not provided.
        refer_data (str, optional): Path to reference data for visual prompts.
        load_vp (bool): Whether to load visual prompts. If False, text prompts are used.

    Returns:
        (dict): Validation statistics containing metrics computed during validation.
    """
    if trainer is not None:
        self.device = trainer.device
        model = trainer.ema.ema
        names = [name.split("/")[0] for name in list(self.dataloader.dataset.data["names"].values())]

        if load_vp:
            LOGGER.info("Validate using the visual prompt.")
            self.args.half = False
            # Directly use the same dataloader for visual embeddings extracted during training
            vpe = self.get_visual_pe(self.dataloader, model)
            model.set_classes(names, vpe)
        else:
            LOGGER.info("Validate using the text prompt.")
            tpe = model.get_text_pe(names)
            model.set_classes(names, tpe)
        stats = super().__call__(trainer, model)
    else:
        if refer_data is not None:
            assert load_vp, "Refer data is only used for visual prompt validation."
        self.device = select_device(self.args.device)

        if isinstance(model, str):
            from ultralytics.nn.tasks import attempt_load_weights

            model = attempt_load_weights(model, device=self.device, inplace=True)
        model.eval().to(self.device)
        data = check_det_dataset(refer_data or self.args.data)
        names = [name.split("/")[0] for name in list(data["names"].values())]

        if load_vp:
            LOGGER.info("Validate using the visual prompt.")
            self.args.half = False
            # TODO: need to check if the names from refer data is consistent with the evaluated dataset
            # could use same dataset or refer to extract visual prompt embeddings
            dataloader = self.get_vpe_dataloader(data)
            vpe = self.get_visual_pe(dataloader, model)
            model.set_classes(names, vpe)
            stats = super().__call__(model=deepcopy(model))
        elif isinstance(model.model[-1], YOLOEDetect) and hasattr(model.model[-1], "lrpc"):  # prompt-free
            return super().__call__(trainer, model)
        else:
            LOGGER.info("Validate using the text prompt.")
            tpe = model.get_text_pe(names)
            model.set_classes(names, tpe)
            stats = super().__call__(model=deepcopy(model))
    return stats

@smart_inference_mode()
def get_visual_pe(self, dataloader, model):
    """
    Extract visual prompt embeddings from training samples.

    This function processes a dataloader to compute visual prompt embeddings for each class
    using a YOLOE model. It normalizes the embeddings and handles cases where no samples
    exist for a class.

    Args:
        dataloader (torch.utils.data.DataLoader): The dataloader providing training samples.
        model (YOLOEModel): The YOLOE model from which to extract visual prompt embeddings.

    Returns:
        (torch.Tensor): Visual prompt embeddings with shape (1, num_classes, embed_dim).
    """
    assert isinstance(model, YOLOEModel)
    names = [name.split("/")[0] for name in list(dataloader.dataset.data["names"].values())]
    visual_pe = torch.zeros(len(names), model.model[-1].embed, device=self.device)
    cls_visual_num = torch.zeros(len(names))

    desc = "Get visual prompt embeddings from samples"

    for batch in dataloader:
        cls = batch["cls"].squeeze(-1).to(torch.int).unique()
        count = torch.bincount(cls, minlength=len(names))
        cls_visual_num += count

    cls_visual_num = cls_visual_num.to(self.device)

    pbar = TQDM(dataloader, total=len(dataloader), desc=desc)
    for batch in pbar:
        batch = self.preprocess(batch)
        preds = model.get_visual_pe(batch["img"], visual=batch["visuals"])  # (B, max_n, embed_dim)

        batch_idx = batch["batch_idx"]
        for i in range(preds.shape[0]):
            cls = batch["cls"][batch_idx == i].squeeze(-1).to(torch.int).unique(sorted=True)
            pad_cls = torch.ones(preds.shape[1], device=self.device) * -1
            pad_cls[: len(cls)] = cls
            for c in cls:
                visual_pe[c] += preds[i][pad_cls == c].sum(0) / cls_visual_num[c]

    visual_pe[cls_visual_num != 0] = F.normalize(visual_pe[cls_visual_num != 0], dim=-1, p=2)
    visual_pe[cls_visual_num == 0] = 0
    return visual_pe.unsqueeze(0)

def get_vpe_dataloader(self, data):
    """
    Create a dataloader for LVIS training visual prompt samples.

    This function prepares a dataloader for visual prompt embeddings (VPE) using the LVIS dataset.
    It applies necessary transformations and configurations to the dataset and returns a dataloader
    for validation purposes.

    Args:
        data (dict): Dataset configuration dictionary containing paths and settings.

    Returns:
        (torch.utils.data.DataLoader): The dataLoader for visual prompt samples.
    """
    dataset = build_yolo_dataset(
        self.args,
        data.get(self.args.split, data.get("val")),
        self.args.batch,
        data,
        mode="val",
        rect=False,
    )
    if isinstance(dataset, YOLOConcatDataset):
        for d in dataset.datasets:
            d.transforms.append(LoadVisualPrompt())
    else:
        dataset.transforms.append(LoadVisualPrompt())
    return build_dataloader(
        dataset,
        self.args.batch,
        self.args.workers,
        shuffle=False,
        rank=-1,
    )

def preprocess(self, batch):
    """Preprocess batch data, ensuring visuals are on the same device as images."""
    batch = super().preprocess(batch)
    if "visuals" in batch:
        batch["visuals"] = batch["visuals"].to(batch["img"].device)
    return batch