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

def build_dataset(self, img_path: str, mode: str = "train", batch: int | None = 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 rectangular training.

    Returns:
        (Dataset): YOLO dataset configured for training or validation.
    """
    gs = max(int(unwrap_model(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 get_model(self, cfg=None, weights=None, verbose: bool = True):
    """
    Return a YOLOEModel initialized with the specified configuration and weights.

    Args:
        cfg (dict | str, optional): Model configuration. Can be a dictionary containing a 'yaml_file' key,
            a direct path to a YAML file, or None to use default configuration.
        weights (str | Path, optional): Path to pretrained weights file to load into the model.
        verbose (bool): Whether to display model information during initialization.

    Returns:
        (YOLOEModel): The initialized YOLOE model.

    Notes:
        - The number of classes (nc) is hard-coded to a maximum of 80 following the official configuration.
        - The nc parameter here represents the maximum number of different text samples in one image,
          rather than the actual number of classes.
    """
    # 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=self.data["channels"],
        nc=min(self.data["nc"], 80),
        verbose=verbose and RANK == -1,
    )
    if weights:
        model.load(weights)

    return model

def get_validator(self):
    """Return a YOLOEDetectValidator for YOLOE 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 get_model(self, cfg=None, weights=None, verbose: bool = 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=self.data["channels"],
        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

def build_dataset(self, img_path: list[str] | str, mode: str = "train", batch: int | None = 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.
    """
    return WorldTrainerFromScratch.build_dataset(self, img_path, mode, batch)

def generate_text_embeddings(self, texts: list[str], batch: int, cache_dir: Path):
    """
    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_dir (Path): Directory to save/load cached embeddings.

    Returns:
        (dict): Dictionary mapping text samples to their embeddings.
    """
    model = "mobileclip:blt"
    cache_path = cache_dir / f"text_embeddings_{model.replace(':', '_').replace('/', '_')}.pt"
    if cache_path.exists():
        LOGGER.info(f"Reading existed cache from '{cache_path}'")
        txt_map = torch.load(cache_path, map_location=self.device)
        if sorted(txt_map.keys()) == sorted(texts):
            return txt_map
    LOGGER.info(f"Caching text embeddings to '{cache_path}'")
    assert self.model is not None
    txt_feats = unwrap_model(self.model).get_text_pe(texts, batch, without_reprta=True, cache_clip_model=False)
    txt_map = dict(zip(texts, txt_feats.squeeze(0)))
    torch.save(txt_map, cache_path)
    return txt_map

def get_validator(self):
    """Return 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):
    """Preprocess a batch of images for YOLOE training, adjusting formatting and dimensions as needed."""
    return DetectionTrainer.preprocess_batch(self, batch)

def set_text_embeddings(self, datasets, batch: int):
    """
    Set text embeddings for datasets to accelerate training by caching category names.

    This method collects unique category names from all datasets, generates text embeddings for them,
    and caches these embeddings to improve training efficiency. The embeddings are stored in a file
    in the parent directory of the first dataset's image path.

    Args:
        datasets (list[Dataset]): List of datasets containing category names to process.
        batch (int): Batch size for processing text embeddings.

    Notes:
        The method creates a dictionary mapping text samples to their embeddings and stores it
        at the path specified by 'cache_path'. If the cache file already exists, it will be loaded
        instead of regenerating the embeddings.
    """
    pass

def build_dataset(self, img_path: list[str] | str, mode: str = "train", batch: int | None = 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