init commit

ultralytics/models/yolo/world/__init__.py

@@ -0,0 +1,5 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+from .train import WorldTrainer
+
+__all__ = ["WorldTrainer"]

ultralytics/models/yolo/world/train.py

@@ -0,0 +1,179 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+from __future__ import annotations
+
+import itertools
+from pathlib import Path
+from typing import Any
+
+import torch
+
+from ultralytics.data import build_yolo_dataset
+from ultralytics.models.yolo.detect import DetectionTrainer
+from ultralytics.nn.tasks import WorldModel
+from ultralytics.utils import DEFAULT_CFG, LOGGER, RANK
+from ultralytics.utils.torch_utils import unwrap_model
+
+
+def on_pretrain_routine_end(trainer) -> None:
+    """Set up model classes and text encoder at the end of the pretrain routine."""
+    if RANK in {-1, 0}:
+        # Set class names for evaluation
+        names = [name.split("/", 1)[0] for name in list(trainer.test_loader.dataset.data["names"].values())]
+        unwrap_model(trainer.ema.ema).set_classes(names, cache_clip_model=False)
+
+
+class WorldTrainer(DetectionTrainer):
+    """
+    A trainer class for fine-tuning YOLO World models on close-set datasets.
+
+    This trainer extends the DetectionTrainer to support training YOLO World models, which combine visual and textual
+    features for improved object detection and understanding. It handles text embedding generation and caching to
+    accelerate training with multi-modal data.
+
+    Attributes:
+        text_embeddings (dict[str, torch.Tensor] | None): Cached text embeddings for category names to accelerate
+            training.
+        model (WorldModel): The YOLO World model being trained.
+        data (dict[str, Any]): Dataset configuration containing class information.
+        args (Any): Training arguments and configuration.
+
+    Methods:
+        get_model: Return WorldModel initialized with specified config and weights.
+        build_dataset: Build YOLO Dataset for training or validation.
+        set_text_embeddings: Set text embeddings for datasets to accelerate training.
+        generate_text_embeddings: Generate text embeddings for a list of text samples.
+        preprocess_batch: Preprocess a batch of images and text for YOLOWorld training.
+
+    Examples:
+        Initialize and train a YOLO World model
+        >>> from ultralytics.models.yolo.world import WorldTrainer
+        >>> args = dict(model="yolov8s-world.pt", data="coco8.yaml", epochs=3)
+        >>> trainer = WorldTrainer(overrides=args)
+        >>> trainer.train()
+    """
+
+    def __init__(self, cfg=DEFAULT_CFG, overrides: dict[str, Any] | None = None, _callbacks=None):
+        """
+        Initialize a WorldTrainer object with given arguments.
+
+        Args:
+            cfg (dict[str, Any]): Configuration for the trainer.
+            overrides (dict[str, Any], optional): Configuration overrides.
+            _callbacks (list[Any], optional): List of callback functions.
+        """
+        if overrides is None:
+            overrides = {}
+        assert not overrides.get("compile"), f"Training with 'model={overrides['model']}' requires 'compile=False'"
+        super().__init__(cfg, overrides, _callbacks)
+        self.text_embeddings = None
+
+    def get_model(self, cfg=None, weights: str | None = None, verbose: bool = True) -> WorldModel:
+        """
+        Return WorldModel initialized with specified config and weights.
+
+        Args:
+            cfg (dict[str, Any] | str, optional): Model configuration.
+            weights (str, optional): Path to pretrained weights.
+            verbose (bool): Whether to display model info.
+
+        Returns:
+            (WorldModel): Initialized WorldModel.
+        """
+        # 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 = WorldModel(
+            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)
+        self.add_callback("on_pretrain_routine_end", on_pretrain_routine_end)
+
+        return model
+
+    def build_dataset(self, img_path: str, mode: str = "train", batch: int | None = None):
+        """
+        Build YOLO Dataset for training or validation.
+
+        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:
+            (Any): YOLO dataset configured for training or validation.
+        """
+        gs = max(int(unwrap_model(self.model).stride.max() if self.model else 0), 32)
+        dataset = build_yolo_dataset(
+            self.args, img_path, batch, self.data, mode=mode, rect=mode == "val", stride=gs, multi_modal=mode == "train"
+        )
+        if mode == "train":
+            self.set_text_embeddings([dataset], batch)  # cache text embeddings to accelerate training
+        return dataset
+
+    def set_text_embeddings(self, datasets: list[Any], batch: int | None) -> None:
+        """
+        Set text embeddings for datasets to accelerate training by caching category names.
+
+        This method collects unique category names from all datasets, then generates and caches text embeddings
+        for these categories to improve training efficiency.
+
+        Args:
+            datasets (list[Any]): List of datasets from which to extract category names.
+            batch (int | None): Batch size used for processing.
+
+        Notes:
+            This method collects category names from datasets that have the 'category_names' attribute,
+            then uses the first dataset's image path to determine where to cache the generated text embeddings.
+        """
+        text_embeddings = {}
+        for dataset in datasets:
+            if not hasattr(dataset, "category_names"):
+                continue
+            text_embeddings.update(
+                self.generate_text_embeddings(
+                    list(dataset.category_names), batch, cache_dir=Path(dataset.img_path).parent
+                )
+            )
+        self.text_embeddings = text_embeddings
+
+    def generate_text_embeddings(self, texts: list[str], batch: int, cache_dir: Path) -> dict[str, torch.Tensor]:
+        """
+        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[str, torch.Tensor]): Dictionary mapping text samples to their embeddings.
+        """
+        model = "clip:ViT-B/32"
+        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, cache_clip_model=False)
+        txt_map = dict(zip(texts, txt_feats.squeeze(0)))
+        torch.save(txt_map, cache_path)
+        return txt_map
+
+    def preprocess_batch(self, batch: dict[str, Any]) -> dict[str, Any]:
+        """Preprocess a batch of images and text for YOLOWorld training."""
+        batch = DetectionTrainer.preprocess_batch(self, batch)
+
+        # Add text features
+        texts = list(itertools.chain(*batch["texts"]))
+        txt_feats = torch.stack([self.text_embeddings[text] for text in texts]).to(
+            self.device, non_blocking=self.device.type == "cuda"
+        )
+        batch["txt_feats"] = txt_feats.reshape(len(batch["texts"]), -1, txt_feats.shape[-1])
+        return batch

ultralytics/models/yolo/world/train_world.py

@@ -0,0 +1,201 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+from pathlib import Path
+
+from ultralytics.data import YOLOConcatDataset, build_grounding, build_yolo_dataset
+from ultralytics.data.utils import check_det_dataset
+from ultralytics.models.yolo.world import WorldTrainer
+from ultralytics.utils import DATASETS_DIR, DEFAULT_CFG, LOGGER
+from ultralytics.utils.torch_utils import unwrap_model
+
+
+class WorldTrainerFromScratch(WorldTrainer):
+    """
+    A class extending the WorldTrainer for training a world model from scratch on open-set datasets.
+
+    This trainer specializes in handling mixed datasets including both object detection and grounding datasets,
+    supporting training YOLO-World models with combined vision-language capabilities.
+
+    Attributes:
+        cfg (dict): Configuration dictionary with default parameters for model training.
+        overrides (dict): Dictionary of parameter overrides to customize the configuration.
+        _callbacks (list): List of callback functions to be executed during different stages of training.
+        data (dict): Final processed data configuration containing train/val paths and metadata.
+        training_data (dict): Dictionary mapping training dataset paths to their configurations.
+
+    Methods:
+        build_dataset: Build YOLO Dataset for training or validation with mixed dataset support.
+        get_dataset: Get train and validation paths from data dictionary.
+        plot_training_labels: Skip label plotting for YOLO-World training.
+        final_eval: Perform final evaluation and validation for the YOLO-World model.
+
+    Examples:
+        >>> from ultralytics.models.yolo.world.train_world import WorldTrainerFromScratch
+        >>> from ultralytics import YOLOWorld
+        >>> data = dict(
+        ...     train=dict(
+        ...         yolo_data=["Objects365.yaml"],
+        ...         grounding_data=[
+        ...             dict(
+        ...                 img_path="flickr30k/images",
+        ...                 json_file="flickr30k/final_flickr_separateGT_train.json",
+        ...             ),
+        ...             dict(
+        ...                 img_path="GQA/images",
+        ...                 json_file="GQA/final_mixed_train_no_coco.json",
+        ...             ),
+        ...         ],
+        ...     ),
+        ...     val=dict(yolo_data=["lvis.yaml"]),
+        ... )
+        >>> model = YOLOWorld("yolov8s-worldv2.yaml")
+        >>> model.train(data=data, trainer=WorldTrainerFromScratch)
+    """
+
+    def __init__(self, cfg=DEFAULT_CFG, overrides=None, _callbacks=None):
+        """
+        Initialize a WorldTrainerFromScratch object.
+
+        This initializes a trainer for YOLO-World models from scratch, supporting mixed datasets including both
+        object detection and grounding datasets for vision-language capabilities.
+
+        Args:
+            cfg (dict): Configuration dictionary with default parameters for model training.
+            overrides (dict, optional): Dictionary of parameter overrides to customize the configuration.
+            _callbacks (list, optional): List of callback functions to be executed during different stages of training.
+
+        Examples:
+            >>> from ultralytics.models.yolo.world.train_world import WorldTrainerFromScratch
+            >>> from ultralytics import YOLOWorld
+            >>> data = dict(
+            ...     train=dict(
+            ...         yolo_data=["Objects365.yaml"],
+            ...         grounding_data=[
+            ...             dict(
+            ...                 img_path="flickr30k/images",
+            ...                 json_file="flickr30k/final_flickr_separateGT_train.json",
+            ...             ),
+            ...         ],
+            ...     ),
+            ...     val=dict(yolo_data=["lvis.yaml"]),
+            ... )
+            >>> model = YOLOWorld("yolov8s-worldv2.yaml")
+            >>> model.train(data=data, trainer=WorldTrainerFromScratch)
+        """
+        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(unwrap_model(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(
+                # assign `nc` from validation set to max number of text samples for training consistency
+                self.args,
+                im_path["img_path"],
+                im_path["json_file"],
+                batch,
+                stride=gs,
+                max_samples=self.data["nc"],
+            )
+            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 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:
+            train_path (str): Train dataset path.
+            val_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)}"
+                for k in {"img_path", "json_file"}:
+                    path = Path(g[k])
+                    if not path.exists() and not path.is_absolute():
+                        g[k] = str((DATASETS_DIR / g[k]).resolve())  # path relative to DATASETS_DIR
+            final_data[s] += grounding_data
+        # assign the first val dataset as currently only one validation set is supported
+        data["val"] = data["val"][0]
+        final_data["val"] = final_data["val"][0]
+        # NOTE: to make training work properly, set `nc` and `names`
+        final_data["nc"] = data["val"]["nc"]
+        final_data["names"] = data["val"]["names"]
+        # NOTE: add path with lvis path
+        final_data["path"] = data["val"]["path"]
+        final_data["channels"] = data["val"]["channels"]
+        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
+
+    def plot_training_labels(self):
+        """Skip label plotting for YOLO-World training."""
+        pass
+
+    def final_eval(self):
+        """
+        Perform final evaluation and validation for the YOLO-World model.
+
+        Configures the validator with appropriate dataset and split information before running evaluation.
+
+        Returns:
+            (dict): Dictionary containing evaluation metrics and results.
+        """
+        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()