diff --git a/docs/en/macros/predict-args.md b/docs/en/macros/predict-args.md index d6ba7205..6bde71e4 100644 --- a/docs/en/macros/predict-args.md +++ b/docs/en/macros/predict-args.md @@ -21,3 +21,4 @@ | `name` | `str` | `None` | Name of the prediction run. Used for creating a subdirectory within the project folder, where prediction outputs are stored if `save` is enabled. | | `stream` | `bool` | `False` | Enables memory-efficient processing for long videos or numerous images by returning a generator of Results objects instead of loading all frames into memory at once. | | `verbose` | `bool` | `True` | Controls whether to display detailed inference logs in the terminal, providing real-time feedback on the prediction process. | +| `compile` | `bool` | `False` | Enables PyTorch 2.x `torch.compile` graph compilation to optimize model execution. Applies to supported devices (CUDA/CPU/MPS). Uses `backend='inductor'`, `mode='max-autotune'`. Falls back to eager mode with a warning if unsupported. | diff --git a/docs/en/macros/train-args.md b/docs/en/macros/train-args.md index 6555f547..7e3734de 100644 --- a/docs/en/macros/train-args.md +++ b/docs/en/macros/train-args.md @@ -48,3 +48,4 @@ | `dropout` | `float` | `0.0` | Dropout rate for regularization in classification tasks, preventing overfitting by randomly omitting units during training. | | `val` | `bool` | `True` | Enables validation during training, allowing for periodic evaluation of model performance on a separate dataset. | | `plots` | `bool` | `False` | Generates and saves plots of training and validation metrics, as well as prediction examples, providing visual insights into model performance and learning progression. | +| `compile` | `bool` | `False` | Enables PyTorch 2.x `torch.compile` graph compilation to optimize model execution. Applies to supported devices (CUDA/CPU/MPS). Uses `backend='inductor'`, `mode='max-autotune'`. Falls back to eager mode with a warning if unsupported. | diff --git a/docs/en/macros/validation-args.md b/docs/en/macros/validation-args.md index c33bbb38..9f41c5cb 100644 --- a/docs/en/macros/validation-args.md +++ b/docs/en/macros/validation-args.md @@ -24,3 +24,4 @@ | `agnostic_nms` | `bool` | `False` | Enables class-agnostic [Non-Maximum Suppression](https://www.ultralytics.com/glossary/non-maximum-suppression-nms), which merges overlapping boxes regardless of their predicted class. Useful for instance-focused applications. | | `single_cls` | `bool` | `False` | Treats all classes as a single class during validation. Useful for evaluating model performance on binary detection tasks or when class distinctions aren't important. | | `visualize` | `bool` | `False` | Visualizes the ground truths, true positives, false positives and false negatives for each image. Useful for debugging and model interpretation. | +| `compile` | `bool` | `False` | Enables PyTorch 2.x `torch.compile` graph compilation to optimize model execution. Applies to supported devices (CUDA/CPU/MPS). Uses `backend='inductor'`, `mode='max-autotune'`. Falls back to eager mode with a warning if unsupported. | diff --git a/docs/en/reference/utils/torch_utils.md b/docs/en/reference/utils/torch_utils.md index e7a7b991..447ed949 100644 --- a/docs/en/reference/utils/torch_utils.md +++ b/docs/en/reference/utils/torch_utils.md @@ -107,7 +107,7 @@ keywords: Ultralytics, torch utils, model optimization, device selection, infere



-## ::: ultralytics.utils.torch_utils.de_parallel +## ::: ultralytics.utils.torch_utils.unwrap_model



@@ -137,4 +137,12 @@ keywords: Ultralytics, torch utils, model optimization, device selection, infere ## ::: ultralytics.utils.torch_utils.profile_ops +



+ +## ::: ultralytics.utils.torch_utils.disable_dynamo + +



+ +## ::: ultralytics.utils.torch_utils.attempt_compile +

diff --git a/ultralytics/__init__.py b/ultralytics/__init__.py index 21e017a4..a2c28781 100644 --- a/ultralytics/__init__.py +++ b/ultralytics/__init__.py @@ -1,6 +1,6 @@ # Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license -__version__ = "8.3.195" +__version__ = "8.3.196" import os diff --git a/ultralytics/cfg/__init__.py b/ultralytics/cfg/__init__.py index 0cc1cb14..fb0b14cd 100644 --- a/ultralytics/cfg/__init__.py +++ b/ultralytics/cfg/__init__.py @@ -237,6 +237,7 @@ CFG_BOOL_KEYS = frozenset( "nms", "profile", "multi_scale", + "compile", } ) diff --git a/ultralytics/cfg/default.yaml b/ultralytics/cfg/default.yaml index b481e75c..b6532b87 100644 --- a/ultralytics/cfg/default.yaml +++ b/ultralytics/cfg/default.yaml @@ -37,6 +37,7 @@ fraction: 1.0 # (float) dataset fraction to train on (default is 1.0, all images profile: False # (bool) profile ONNX and TensorRT speeds during training for loggers freeze: # (int | list, optional) freeze first n layers, or freeze list of layer indices during training multi_scale: False # (bool) Whether to use multiscale during training +compile: False # (bool) Run torch.compile() on the model before train/val/predict # Segmentation overlap_mask: True # (bool) merge object masks into a single image mask during training (segment train only) mask_ratio: 4 # (int) mask downsample ratio (segment train only) diff --git a/ultralytics/data/augment.py b/ultralytics/data/augment.py index 48d9e2ba..d25fba39 100644 --- a/ultralytics/data/augment.py +++ b/ultralytics/data/augment.py @@ -2202,7 +2202,7 @@ class Format: ) labels["masks"] = masks labels["img"] = self._format_img(img) - labels["cls"] = torch.from_numpy(cls) if nl else torch.zeros(nl) + labels["cls"] = torch.from_numpy(cls) if nl else torch.zeros(nl, 1) labels["bboxes"] = torch.from_numpy(instances.bboxes) if nl else torch.zeros((nl, 4)) if self.return_keypoint: labels["keypoints"] = ( diff --git a/ultralytics/data/build.py b/ultralytics/data/build.py index 44b68316..5dfbff12 100644 --- a/ultralytics/data/build.py +++ b/ultralytics/data/build.py @@ -217,7 +217,7 @@ def build_dataloader(dataset, batch: int, workers: int, shuffle: bool = True, ra collate_fn=getattr(dataset, "collate_fn", None), worker_init_fn=seed_worker, generator=generator, - drop_last=drop_last, + drop_last=drop_last and len(dataset) % batch != 0, ) diff --git a/ultralytics/engine/predictor.py b/ultralytics/engine/predictor.py index b358ce0f..771ea589 100644 --- a/ultralytics/engine/predictor.py +++ b/ultralytics/engine/predictor.py @@ -51,7 +51,7 @@ from ultralytics.nn.autobackend import AutoBackend from ultralytics.utils import DEFAULT_CFG, LOGGER, MACOS, WINDOWS, callbacks, colorstr, ops from ultralytics.utils.checks import check_imgsz, check_imshow from ultralytics.utils.files import increment_path -from ultralytics.utils.torch_utils import select_device, smart_inference_mode +from ultralytics.utils.torch_utils import attempt_compile, select_device, smart_inference_mode STREAM_WARNING = """ inference results will accumulate in RAM unless `stream=True` is passed, causing potential out-of-memory @@ -409,6 +409,8 @@ class BasePredictor: if hasattr(self.model, "imgsz") and not getattr(self.model, "dynamic", False): self.args.imgsz = self.model.imgsz # reuse imgsz from export metadata self.model.eval() + if self.args.compile: + self.model = attempt_compile(self.model, device=self.device) def write_results(self, i: int, p: Path, im: torch.Tensor, s: list[str]) -> str: """ diff --git a/ultralytics/engine/trainer.py b/ultralytics/engine/trainer.py index 0869c311..a7460f8e 100644 --- a/ultralytics/engine/trainer.py +++ b/ultralytics/engine/trainer.py @@ -46,6 +46,7 @@ from ultralytics.utils.torch_utils import ( TORCH_2_4, EarlyStopping, ModelEMA, + attempt_compile, autocast, convert_optimizer_state_dict_to_fp16, init_seeds, @@ -54,6 +55,7 @@ from ultralytics.utils.torch_utils import ( strip_optimizer, torch_distributed_zero_first, unset_deterministic, + unwrap_model, ) @@ -256,6 +258,14 @@ class BaseTrainer: self.model = self.model.to(self.device) self.set_model_attributes() + # Initialize loss criterion before compilation for torch.compile compatibility + if hasattr(self.model, "init_criterion"): + self.model.criterion = self.model.init_criterion() + + # Compile model + if self.args.compile: + self.model = attempt_compile(self.model, device=self.device) + # Freeze layers freeze_list = ( self.args.freeze @@ -404,6 +414,7 @@ class BaseTrainer: # Forward with autocast(self.amp): batch = self.preprocess_batch(batch) + metadata = {k: batch.pop(k, None) for k in ["im_file", "ori_shape", "resized_shape"]} loss, self.loss_items = self.model(batch) self.loss = loss.sum() if RANK != -1: @@ -445,6 +456,7 @@ class BaseTrainer: ) self.run_callbacks("on_batch_end") if self.args.plots and ni in self.plot_idx: + batch = {**batch, **metadata} self.plot_training_samples(batch, ni) self.run_callbacks("on_train_batch_end") @@ -565,7 +577,7 @@ class BaseTrainer: "epoch": self.epoch, "best_fitness": self.best_fitness, "model": None, # resume and final checkpoints derive from EMA - "ema": deepcopy(self.ema.ema).half(), + "ema": deepcopy(unwrap_model(self.ema.ema)).half(), "updates": self.ema.updates, "optimizer": convert_optimizer_state_dict_to_fp16(deepcopy(self.optimizer.state_dict())), "train_args": vars(self.args), # save as dict @@ -592,8 +604,6 @@ class BaseTrainer: self.best.write_bytes(serialized_ckpt) # save best.pt if (self.save_period > 0) and (self.epoch % self.save_period == 0): (self.wdir / f"epoch{self.epoch}.pt").write_bytes(serialized_ckpt) # save epoch, i.e. 'epoch3.pt' - # if self.args.close_mosaic and self.epoch == (self.epochs - self.args.close_mosaic - 1): - # (self.wdir / "last_mosaic.pt").write_bytes(serialized_ckpt) # save mosaic checkpoint def get_dataset(self): """ @@ -667,7 +677,7 @@ class BaseTrainer: def validate(self): """ - Run validation on test set using self.validator. + Run validation on val set using self.validator. Returns: metrics (dict): Dictionary of validation metrics. @@ -755,6 +765,7 @@ class BaseTrainer: strip_optimizer(f, updates={k: ckpt[k]} if k in ckpt else None) LOGGER.info(f"\nValidating {f}...") self.validator.args.plots = self.args.plots + self.validator.args.compile = False # disable final val compile as too slow self.metrics = self.validator(model=f) self.metrics.pop("fitness", None) self.run_callbacks("on_fit_epoch_end") diff --git a/ultralytics/engine/validator.py b/ultralytics/engine/validator.py index a7c15219..bcba8459 100644 --- a/ultralytics/engine/validator.py +++ b/ultralytics/engine/validator.py @@ -36,7 +36,7 @@ from ultralytics.nn.autobackend import AutoBackend from ultralytics.utils import LOGGER, TQDM, callbacks, colorstr, emojis from ultralytics.utils.checks import check_imgsz from ultralytics.utils.ops import Profile -from ultralytics.utils.torch_utils import de_parallel, select_device, smart_inference_mode +from ultralytics.utils.torch_utils import attempt_compile, select_device, smart_inference_mode, unwrap_model class BaseValidator: @@ -148,6 +148,8 @@ class BaseValidator: # Force FP16 val during training self.args.half = self.device.type != "cpu" and trainer.amp model = trainer.ema.ema or trainer.model + if trainer.args.compile and hasattr(model, "_orig_mod"): + model = model._orig_mod # validate non-compiled original model to avoid issues model = model.half() if self.args.half else model.float() self.loss = torch.zeros_like(trainer.loss_items, device=trainer.device) self.args.plots &= trainer.stopper.possible_stop or (trainer.epoch == trainer.epochs - 1) @@ -186,6 +188,8 @@ class BaseValidator: self.dataloader = self.dataloader or self.get_dataloader(self.data.get(self.args.split), self.args.batch) model.eval() + if self.args.compile: + model = attempt_compile(model, device=self.device) model.warmup(imgsz=(1 if pt else self.args.batch, self.data["channels"], imgsz, imgsz)) # warmup self.run_callbacks("on_val_start") @@ -196,7 +200,7 @@ class BaseValidator: Profile(device=self.device), ) bar = TQDM(self.dataloader, desc=self.get_desc(), total=len(self.dataloader)) - self.init_metrics(de_parallel(model)) + self.init_metrics(unwrap_model(model)) self.jdict = [] # empty before each val for batch_i, batch in enumerate(bar): self.run_callbacks("on_val_batch_start") diff --git a/ultralytics/models/yolo/classify/train.py b/ultralytics/models/yolo/classify/train.py index 702faee4..06b791dd 100644 --- a/ultralytics/models/yolo/classify/train.py +++ b/ultralytics/models/yolo/classify/train.py @@ -55,20 +55,10 @@ class ClassificationTrainer(BaseTrainer): """ Initialize a ClassificationTrainer object. - This constructor sets up a trainer for image classification tasks, configuring the task type and default - image size if not specified. - Args: cfg (dict[str, Any], optional): Default configuration dictionary containing training parameters. overrides (dict[str, Any], optional): Dictionary of parameter overrides for the default configuration. _callbacks (list[Any], optional): List of callback functions to be executed during training. - - Examples: - Create a trainer with custom configuration - >>> from ultralytics.models.yolo.classify import ClassificationTrainer - >>> args = dict(model="yolo11n-cls.pt", data="imagenet10", epochs=3) - >>> trainer = ClassificationTrainer(overrides=args) - >>> trainer.train() """ if overrides is None: overrides = {} @@ -155,7 +145,7 @@ class ClassificationTrainer(BaseTrainer): with torch_distributed_zero_first(rank): # init dataset *.cache only once if DDP dataset = self.build_dataset(dataset_path, mode) - loader = build_dataloader(dataset, batch_size, self.args.workers, rank=rank) + loader = build_dataloader(dataset, batch_size, self.args.workers, rank=rank, drop_last=self.args.compile) # Attach inference transforms if mode != "train": if is_parallel(self.model): diff --git a/ultralytics/models/yolo/detect/train.py b/ultralytics/models/yolo/detect/train.py index b5f1685d..97d562ec 100644 --- a/ultralytics/models/yolo/detect/train.py +++ b/ultralytics/models/yolo/detect/train.py @@ -15,10 +15,10 @@ from ultralytics.data import build_dataloader, build_yolo_dataset from ultralytics.engine.trainer import BaseTrainer from ultralytics.models import yolo from ultralytics.nn.tasks import DetectionModel -from ultralytics.utils import LOGGER, RANK +from ultralytics.utils import DEFAULT_CFG, LOGGER, RANK from ultralytics.utils.patches import override_configs from ultralytics.utils.plotting import plot_images, plot_labels, plot_results -from ultralytics.utils.torch_utils import de_parallel, torch_distributed_zero_first +from ultralytics.utils.torch_utils import torch_distributed_zero_first, unwrap_model class DetectionTrainer(BaseTrainer): @@ -54,6 +54,18 @@ class DetectionTrainer(BaseTrainer): >>> trainer.train() """ + def __init__(self, cfg=DEFAULT_CFG, overrides: dict[str, Any] | None = None, _callbacks=None): + """ + Initialize a DetectionTrainer object for training YOLO object detection model training. + + Args: + cfg (dict, optional): Default configuration dictionary containing training parameters. + overrides (dict, optional): Dictionary of parameter overrides for the default configuration. + _callbacks (list, optional): List of callback functions to be executed during training. + """ + super().__init__(cfg, overrides, _callbacks) + self.dynamic_tensors = ["batch_idx", "cls", "bboxes"] + def build_dataset(self, img_path: str, mode: str = "train", batch: int | None = None): """ Build YOLO Dataset for training or validation. @@ -66,7 +78,7 @@ class DetectionTrainer(BaseTrainer): Returns: (Dataset): YOLO dataset object configured for the specified mode. """ - gs = max(int(de_parallel(self.model).stride.max() if self.model else 0), 32) + 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) def get_dataloader(self, dataset_path: str, batch_size: int = 16, rank: int = 0, mode: str = "train"): @@ -89,8 +101,14 @@ class DetectionTrainer(BaseTrainer): if getattr(dataset, "rect", False) and shuffle: LOGGER.warning("'rect=True' is incompatible with DataLoader shuffle, setting shuffle=False") shuffle = False - workers = self.args.workers if mode == "train" else self.args.workers * 2 - return build_dataloader(dataset, batch_size, workers, shuffle, rank) # return dataloader + return build_dataloader( + dataset, + batch=batch_size, + workers=self.args.workers if mode == "train" else self.args.workers * 2, + shuffle=shuffle, + rank=rank, + drop_last=self.args.compile and mode == "train", + ) def preprocess_batch(self, batch: dict) -> dict: """ @@ -120,6 +138,10 @@ class DetectionTrainer(BaseTrainer): ] # new shape (stretched to gs-multiple) imgs = nn.functional.interpolate(imgs, size=ns, mode="bilinear", align_corners=False) batch["img"] = imgs + + if self.args.compile: + for k in self.dynamic_tensors: + torch._dynamo.maybe_mark_dynamic(batch[k], 0) return batch def set_model_attributes(self): diff --git a/ultralytics/models/yolo/detect/val.py b/ultralytics/models/yolo/detect/val.py index 3a8112a9..de1d33be 100644 --- a/ultralytics/models/yolo/detect/val.py +++ b/ultralytics/models/yolo/detect/val.py @@ -299,7 +299,9 @@ class DetectionValidator(BaseValidator): (torch.utils.data.DataLoader): Dataloader for validation. """ dataset = self.build_dataset(dataset_path, batch=batch_size, mode="val") - return build_dataloader(dataset, batch_size, self.args.workers, shuffle=False, rank=-1) # return dataloader + return build_dataloader( + dataset, batch_size, self.args.workers, shuffle=False, rank=-1, drop_last=self.args.compile + ) def plot_val_samples(self, batch: dict[str, Any], ni: int) -> None: """ diff --git a/ultralytics/models/yolo/obb/train.py b/ultralytics/models/yolo/obb/train.py index fe9f7920..c1d06ae6 100644 --- a/ultralytics/models/yolo/obb/train.py +++ b/ultralytics/models/yolo/obb/train.py @@ -37,21 +37,12 @@ class OBBTrainer(yolo.detect.DetectionTrainer): """ Initialize an OBBTrainer object for training Oriented Bounding Box (OBB) models. - This trainer extends the DetectionTrainer class to specialize in training models that detect oriented - bounding boxes. It automatically sets the task to 'obb' in the configuration. - Args: cfg (dict, optional): Configuration dictionary for the trainer. Contains training parameters and model configuration. overrides (dict, optional): Dictionary of parameter overrides for the configuration. Any values here will take precedence over those in cfg. _callbacks (list[Any], optional): List of callback functions to be invoked during training. - - Examples: - >>> from ultralytics.models.yolo.obb import OBBTrainer - >>> args = dict(model="yolo11n-obb.pt", data="dota8.yaml", epochs=3) - >>> trainer = OBBTrainer(overrides=args) - >>> trainer.train() """ if overrides is None: overrides = {} diff --git a/ultralytics/models/yolo/pose/train.py b/ultralytics/models/yolo/pose/train.py index ceb280c9..564a23d2 100644 --- a/ultralytics/models/yolo/pose/train.py +++ b/ultralytics/models/yolo/pose/train.py @@ -44,9 +44,6 @@ class PoseTrainer(yolo.detect.DetectionTrainer): """ Initialize a PoseTrainer object for training YOLO pose estimation models. - This initializes a trainer specialized for pose estimation tasks, setting the task to 'pose' and - handling specific configurations needed for keypoint detection models. - Args: cfg (dict, optional): Default configuration dictionary containing training parameters. overrides (dict, optional): Dictionary of parameter overrides for the default configuration. @@ -55,17 +52,12 @@ class PoseTrainer(yolo.detect.DetectionTrainer): Notes: This trainer will automatically set the task to 'pose' regardless of what is provided in overrides. A warning is issued when using Apple MPS device due to known bugs with pose models. - - Examples: - >>> from ultralytics.models.yolo.pose import PoseTrainer - >>> args = dict(model="yolo11n-pose.pt", data="coco8-pose.yaml", epochs=3) - >>> trainer = PoseTrainer(overrides=args) - >>> trainer.train() """ if overrides is None: overrides = {} overrides["task"] = "pose" super().__init__(cfg, overrides, _callbacks) + self.dynamic_tensors = ["batch_idx", "cls", "bboxes", "keypoints"] if isinstance(self.args.device, str) and self.args.device.lower() == "mps": LOGGER.warning( diff --git a/ultralytics/models/yolo/segment/train.py b/ultralytics/models/yolo/segment/train.py index 9d59809e..9badc995 100644 --- a/ultralytics/models/yolo/segment/train.py +++ b/ultralytics/models/yolo/segment/train.py @@ -32,24 +32,16 @@ class SegmentationTrainer(yolo.detect.DetectionTrainer): """ Initialize a SegmentationTrainer object. - This initializes a trainer for segmentation tasks, extending the detection trainer with segmentation-specific - functionality. It sets the task to 'segment' and prepares the trainer for training segmentation models. - Args: cfg (dict): Configuration dictionary with default training settings. overrides (dict, optional): Dictionary of parameter overrides for the default configuration. _callbacks (list, optional): List of callback functions to be executed during training. - - Examples: - >>> from ultralytics.models.yolo.segment import SegmentationTrainer - >>> args = dict(model="yolo11n-seg.pt", data="coco8-seg.yaml", epochs=3) - >>> trainer = SegmentationTrainer(overrides=args) - >>> trainer.train() """ if overrides is None: overrides = {} overrides["task"] = "segment" super().__init__(cfg, overrides, _callbacks) + self.dynamic_tensors = ["batch_idx", "cls", "bboxes", "masks"] def get_model(self, cfg: dict | str | None = None, weights: str | Path | None = None, verbose: bool = True): """ diff --git a/ultralytics/models/yolo/world/train.py b/ultralytics/models/yolo/world/train.py index 1690be7e..825b8859 100644 --- a/ultralytics/models/yolo/world/train.py +++ b/ultralytics/models/yolo/world/train.py @@ -12,7 +12,7 @@ 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 de_parallel +from ultralytics.utils.torch_utils import unwrap_model def on_pretrain_routine_end(trainer) -> None: @@ -20,7 +20,7 @@ def on_pretrain_routine_end(trainer) -> None: 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())] - de_parallel(trainer.ema.ema).set_classes(names, cache_clip_model=False) + unwrap_model(trainer.ema.ema).set_classes(names, cache_clip_model=False) class WorldTrainer(DetectionTrainer): @@ -105,7 +105,7 @@ class WorldTrainer(DetectionTrainer): Returns: (Any): YOLO dataset configured for training or validation. """ - gs = max(int(de_parallel(self.model).stride.max() if self.model else 0), 32) + 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" ) @@ -160,7 +160,7 @@ class WorldTrainer(DetectionTrainer): return txt_map LOGGER.info(f"Caching text embeddings to '{cache_path}'") assert self.model is not None - txt_feats = de_parallel(self.model).get_text_pe(texts, batch, cache_clip_model=False) + 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 diff --git a/ultralytics/models/yolo/world/train_world.py b/ultralytics/models/yolo/world/train_world.py index 9b77603c..558323c5 100644 --- a/ultralytics/models/yolo/world/train_world.py +++ b/ultralytics/models/yolo/world/train_world.py @@ -6,7 +6,7 @@ from ultralytics.data import YOLOConcatDataset, build_grounding, build_yolo_data 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 de_parallel +from ultralytics.utils.torch_utils import unwrap_model class WorldTrainerFromScratch(WorldTrainer): @@ -101,7 +101,7 @@ class WorldTrainerFromScratch(WorldTrainer): 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) + 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 = [ diff --git a/ultralytics/models/yolo/yoloe/train.py b/ultralytics/models/yolo/yoloe/train.py index 395f24db..aaeffd1c 100644 --- a/ultralytics/models/yolo/yoloe/train.py +++ b/ultralytics/models/yolo/yoloe/train.py @@ -13,7 +13,7 @@ from ultralytics.data.augment import LoadVisualPrompt 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 ultralytics.utils.torch_utils import unwrap_model from ..world.train_world import WorldTrainerFromScratch from .val import YOLOEDetectValidator @@ -39,9 +39,6 @@ class YOLOETrainer(DetectionTrainer): """ 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. @@ -102,7 +99,7 @@ class YOLOETrainer(DetectionTrainer): Returns: (Dataset): YOLO dataset configured for training or validation. """ - gs = max(int(de_parallel(self.model).stride.max() if self.model else 0), 32) + 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" ) @@ -223,7 +220,7 @@ class YOLOETrainerFromScratch(YOLOETrainer, WorldTrainerFromScratch): return txt_map LOGGER.info(f"Caching text embeddings to '{cache_path}'") assert self.model is not None - txt_feats = de_parallel(self.model).get_text_pe(texts, batch, without_reprta=True, cache_clip_model=False) + 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 diff --git a/ultralytics/nn/modules/head.py b/ultralytics/nn/modules/head.py index a277aa66..e7708efa 100644 --- a/ultralytics/nn/modules/head.py +++ b/ultralytics/nn/modules/head.py @@ -13,7 +13,7 @@ from torch.nn.init import constant_, xavier_uniform_ from ultralytics.utils import NOT_MACOS14 from ultralytics.utils.tal import TORCH_1_10, dist2bbox, dist2rbox, make_anchors -from ultralytics.utils.torch_utils import fuse_conv_and_bn, smart_inference_mode +from ultralytics.utils.torch_utils import disable_dynamo, fuse_conv_and_bn, smart_inference_mode from .block import DFL, SAVPE, BNContrastiveHead, ContrastiveHead, Proto, Residual, SwiGLUFFN from .conv import Conv, DWConv @@ -149,6 +149,7 @@ class Detect(nn.Module): y = self.postprocess(y.permute(0, 2, 1), self.max_det, self.nc) return y if self.export else (y, {"one2many": x, "one2one": one2one}) + @disable_dynamo def _inference(self, x: list[torch.Tensor]) -> torch.Tensor: """ Decode predicted bounding boxes and class probabilities based on multiple-level feature maps. diff --git a/ultralytics/nn/tasks.py b/ultralytics/nn/tasks.py index 9a044294..e1887794 100644 --- a/ultralytics/nn/tasks.py +++ b/ultralytics/nn/tasks.py @@ -334,7 +334,8 @@ class BaseModel(torch.nn.Module): if getattr(self, "criterion", None) is None: self.criterion = self.init_criterion() - preds = self.forward(batch["img"]) if preds is None else preds + if preds is None: + preds = self.forward(batch["img"]) return self.criterion(preds, batch) def init_criterion(self): @@ -775,7 +776,8 @@ class RTDETRDetectionModel(DetectionModel): "gt_groups": gt_groups, } - preds = self.predict(img, batch=targets) if preds is None else preds + if preds is None: + preds = self.predict(img, batch=targets) dec_bboxes, dec_scores, enc_bboxes, enc_scores, dn_meta = preds if self.training else preds[1] if dn_meta is None: dn_bboxes, dn_scores = None, None diff --git a/ultralytics/utils/callbacks/tensorboard.py b/ultralytics/utils/callbacks/tensorboard.py index f5adc31c..5dbe3e1d 100644 --- a/ultralytics/utils/callbacks/tensorboard.py +++ b/ultralytics/utils/callbacks/tensorboard.py @@ -70,14 +70,14 @@ def _log_tensorboard_graph(trainer) -> None: # Try simple method first (YOLO) try: trainer.model.eval() # place in .eval() mode to avoid BatchNorm statistics changes - WRITER.add_graph(torch.jit.trace(torch_utils.de_parallel(trainer.model), im, strict=False), []) + WRITER.add_graph(torch.jit.trace(torch_utils.unwrap_model(trainer.model), im, strict=False), []) LOGGER.info(f"{PREFIX}model graph visualization added ✅") return except Exception: # Fallback to TorchScript export steps (RTDETR) try: - model = deepcopy(torch_utils.de_parallel(trainer.model)) + model = deepcopy(torch_utils.unwrap_model(trainer.model)) model.eval() model = model.fuse(verbose=False) for m in model.modules(): diff --git a/ultralytics/utils/loss.py b/ultralytics/utils/loss.py index a61037d9..ad0fbb0d 100644 --- a/ultralytics/utils/loss.py +++ b/ultralytics/utils/loss.py @@ -11,7 +11,7 @@ import torch.nn.functional as F from ultralytics.utils.metrics import OKS_SIGMA from ultralytics.utils.ops import crop_mask, xywh2xyxy, xyxy2xywh from ultralytics.utils.tal import RotatedTaskAlignedAssigner, TaskAlignedAssigner, dist2bbox, dist2rbox, make_anchors -from ultralytics.utils.torch_utils import autocast +from ultralytics.utils.torch_utils import autocast, disable_dynamo from .metrics import bbox_iou, probiou from .tal import bbox2dist @@ -215,6 +215,7 @@ class v8DetectionLoss: self.assigner = TaskAlignedAssigner(topk=tal_topk, num_classes=self.nc, alpha=0.5, beta=6.0) self.bbox_loss = BboxLoss(m.reg_max).to(device) self.proj = torch.arange(m.reg_max, dtype=torch.float, device=device) + disable_dynamo(self.__class__) # exclude from compile def preprocess(self, targets: torch.Tensor, batch_size: int, scale_tensor: torch.Tensor) -> torch.Tensor: """Preprocess targets by converting to tensor format and scaling coordinates.""" @@ -287,9 +288,14 @@ class v8DetectionLoss: # Bbox loss if fg_mask.sum(): - target_bboxes /= stride_tensor loss[0], loss[2] = self.bbox_loss( - pred_distri, pred_bboxes, anchor_points, target_bboxes, target_scores, target_scores_sum, fg_mask + pred_distri, + pred_bboxes, + anchor_points, + target_bboxes / stride_tensor, + target_scores, + target_scores_sum, + fg_mask, ) loss[0] *= self.hyp.box # box gain @@ -388,7 +394,7 @@ class v8SegmentationLoss(v8DetectionLoss): loss[2] *= self.hyp.cls # cls gain loss[3] *= self.hyp.dfl # dfl gain - return loss * batch_size, loss.detach() # loss(box, cls, dfl) + return loss * batch_size, loss.detach() # loss(box, seg, cls, dfl) @staticmethod def single_mask_loss( diff --git a/ultralytics/utils/torch_utils.py b/ultralytics/utils/torch_utils.py index e25c5ebb..83f167cd 100644 --- a/ultralytics/utils/torch_utils.py +++ b/ultralytics/utils/torch_utils.py @@ -429,7 +429,7 @@ def get_flops(model, imgsz=640): return 0.0 # if not installed return 0.0 GFLOPs try: - model = de_parallel(model) + model = unwrap_model(model) p = next(model.parameters()) if not isinstance(imgsz, list): imgsz = [imgsz, imgsz] # expand if int/float @@ -460,7 +460,7 @@ def get_flops_with_torch_profiler(model, imgsz=640): """ if not TORCH_2_0: # torch profiler implemented in torch>=2.0 return 0.0 - model = de_parallel(model) + model = unwrap_model(model) p = next(model.parameters()) if not isinstance(imgsz, list): imgsz = [imgsz, imgsz] # expand if int/float @@ -577,17 +577,24 @@ def is_parallel(model): return isinstance(model, (nn.parallel.DataParallel, nn.parallel.DistributedDataParallel)) -def de_parallel(model): +def unwrap_model(m: nn.Module) -> nn.Module: """ - De-parallelize a model: return single-GPU model if model is of type DP or DDP. + Unwrap compiled and parallel models to get the base model. Args: - model (nn.Module): Model to de-parallelize. + m (nn.Module): A model that may be wrapped by torch.compile (._orig_mod) or parallel wrappers such as + DataParallel/DistributedDataParallel (.module). Returns: - (nn.Module): De-parallelized model. + m (nn.Module): The unwrapped base model without compile or parallel wrappers. """ - return model.module if is_parallel(model) else model + while True: + if hasattr(m, "_orig_mod") and isinstance(m._orig_mod, nn.Module): + m = m._orig_mod + elif hasattr(m, "module") and isinstance(m.module, nn.Module): + m = m.module + else: + return m def one_cycle(y1=0.0, y2=1.0, steps=100): @@ -669,7 +676,7 @@ class ModelEMA: tau (int, optional): EMA decay time constant. updates (int, optional): Initial number of updates. """ - self.ema = deepcopy(de_parallel(model)).eval() # FP32 EMA + self.ema = deepcopy(unwrap_model(model)).eval() # FP32 EMA self.updates = updates # number of EMA updates self.decay = lambda x: decay * (1 - math.exp(-x / tau)) # decay exponential ramp (to help early epochs) for p in self.ema.parameters(): @@ -687,7 +694,7 @@ class ModelEMA: self.updates += 1 d = self.decay(self.updates) - msd = de_parallel(model).state_dict() # model state_dict + msd = unwrap_model(model).state_dict() # model state_dict for k, v in self.ema.state_dict().items(): if v.dtype.is_floating_point: # true for FP16 and FP32 v *= d @@ -997,3 +1004,98 @@ class FXModel(nn.Module): x = m(x) # run y.append(x) # save output return x + + +def disable_dynamo(func: Any) -> Any: + """ + Disable torch.compile/dynamo for a callable when available. + + Args: + func (Any): Callable object to wrap. Could be a function, method, or class. + + Returns: + func (Any): Same callable, wrapped by torch._dynamo.disable when available, otherwise unchanged. + + Examples: + >>> @disable_dynamo + ... def fn(x): + ... return x + 1 + >>> # Works even if torch._dynamo is not available + >>> _ = fn(1) + """ + if hasattr(torch, "_dynamo"): + return torch._dynamo.disable(func) + return func + + +def attempt_compile( + model: torch.nn.Module, + device: torch.device, + imgsz: int = 640, + use_autocast: bool = False, + warmup: bool = False, + prefix: str = colorstr("compile:"), +) -> torch.nn.Module: + """ + Compile a model with torch.compile and optionally warm up the graph to reduce first-iteration latency. + + This utility attempts to compile the provided model using the inductor backend with dynamic shapes enabled and an + autotuning mode. If compilation is unavailable or fails, the original model is returned unchanged. An optional + warmup performs a single forward pass on a dummy input to prime the compiled graph and measure compile/warmup time. + + Args: + model (torch.nn.Module): Model to compile. + device (torch.device): Inference device used for warmup and autocast decisions. + imgsz (int, optional): Square input size to create a dummy tensor with shape (1, 3, imgsz, imgsz) for warmup. + use_autocast (bool, optional): Whether to run warmup under autocast on CUDA or MPS devices. + warmup (bool, optional): Whether to execute a single dummy forward pass to warm up the compiled model. + prefix (str, optional): Message prefix for logger output. + + Returns: + model (torch.nn.Module): Compiled model if compilation succeeds, otherwise the original unmodified model. + + Notes: + - If the current PyTorch build does not provide torch.compile, the function returns the input model immediately. + - Warmup runs under torch.inference_mode and may use torch.autocast for CUDA/MPS to align compute precision. + - CUDA devices are synchronized after warmup to account for asynchronous kernel execution. + + Examples: + >>> device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") + >>> # Try to compile and warm up a model with a 640x640 input + >>> model = attempt_compile(model, device=device, imgsz=640, use_autocast=True, warmup=True) + """ + if not hasattr(torch, "compile"): + return model + + LOGGER.info(f"{prefix} starting torch.compile...") + t0 = time.perf_counter() + try: + model = torch.compile(model, mode="max-autotune", backend="inductor") + except Exception as e: + LOGGER.warning(f"{prefix} torch.compile failed, continuing uncompiled: {e}") + return model + t_compile = time.perf_counter() - t0 + + t_warm = 0.0 + if warmup: + # Use a single dummy tensor to build the graph shape state and reduce first-iteration latency + dummy = torch.zeros(1, 3, imgsz, imgsz, device=device) + if use_autocast and device.type == "cuda": + dummy = dummy.half() + t1 = time.perf_counter() + with torch.inference_mode(): + if use_autocast and device.type in {"cuda", "mps"}: + with torch.autocast(device.type): + _ = model(dummy) + else: + _ = model(dummy) + if device.type == "cuda": + torch.cuda.synchronize(device) + t_warm = time.perf_counter() - t1 + + total = t_compile + t_warm + if warmup: + LOGGER.info(f"{prefix} complete in {total:.1f}s (compile {t_compile:.1f}s + warmup {t_warm:.1f}s)") + else: + LOGGER.info(f"{prefix} compile complete in {t_compile:.1f}s (no warmup)") + return model