File size: 5,994 Bytes
e1832f4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license

from ultralytics.cfg import TASK2DATA, TASK2METRIC, get_cfg, get_save_dir
from ultralytics.utils import DEFAULT_CFG, DEFAULT_CFG_DICT, LOGGER, NUM_THREADS, checks


def run_ray_tune(
    model,
    space: dict = None,
    grace_period: int = 10,
    gpu_per_trial: int = None,
    max_samples: int = 10,
    **train_args,
):
    """
    Run hyperparameter tuning using Ray Tune.

    Args:
        model (YOLO): Model to run the tuner on.
        space (dict, optional): The hyperparameter search space.
        grace_period (int, optional): The grace period in epochs of the ASHA scheduler.
        gpu_per_trial (int, optional): The number of GPUs to allocate per trial.
        max_samples (int, optional): The maximum number of trials to run.
        **train_args (Any): Additional arguments to pass to the `train()` method.

    Returns:
        (dict): A dictionary containing the results of the hyperparameter search.

    Examples:
        >>> from ultralytics import YOLO
        >>> model = YOLO("yolo11n.pt")  # Load a YOLO11n model

        Start tuning hyperparameters for YOLO11n training on the COCO8 dataset
        >>> result_grid = model.tune(data="coco8.yaml", use_ray=True)
    """
    LOGGER.info("💡 Learn about RayTune at https://docs.ultralytics.com/integrations/ray-tune")
    if train_args is None:
        train_args = {}

    try:
        checks.check_requirements("ray[tune]")

        import ray
        from ray import tune
        from ray.air import RunConfig
        from ray.air.integrations.wandb import WandbLoggerCallback
        from ray.tune.schedulers import ASHAScheduler
    except ImportError:
        raise ModuleNotFoundError('Ray Tune required but not found. To install run: pip install "ray[tune]"')

    try:
        import wandb

        assert hasattr(wandb, "__version__")
    except (ImportError, AssertionError):
        wandb = False

    checks.check_version(ray.__version__, ">=2.0.0", "ray")
    default_space = {
        # 'optimizer': tune.choice(['SGD', 'Adam', 'AdamW', 'NAdam', 'RAdam', 'RMSProp']),
        "lr0": tune.uniform(1e-5, 1e-1),
        "lrf": tune.uniform(0.01, 1.0),  # final OneCycleLR learning rate (lr0 * lrf)
        "momentum": tune.uniform(0.6, 0.98),  # SGD momentum/Adam beta1
        "weight_decay": tune.uniform(0.0, 0.001),  # optimizer weight decay
        "warmup_epochs": tune.uniform(0.0, 5.0),  # warmup epochs (fractions ok)
        "warmup_momentum": tune.uniform(0.0, 0.95),  # warmup initial momentum
        "box": tune.uniform(0.02, 0.2),  # box loss gain
        "cls": tune.uniform(0.2, 4.0),  # cls loss gain (scale with pixels)
        "hsv_h": tune.uniform(0.0, 0.1),  # image HSV-Hue augmentation (fraction)
        "hsv_s": tune.uniform(0.0, 0.9),  # image HSV-Saturation augmentation (fraction)
        "hsv_v": tune.uniform(0.0, 0.9),  # image HSV-Value augmentation (fraction)
        "degrees": tune.uniform(0.0, 45.0),  # image rotation (+/- deg)
        "translate": tune.uniform(0.0, 0.9),  # image translation (+/- fraction)
        "scale": tune.uniform(0.0, 0.9),  # image scale (+/- gain)
        "shear": tune.uniform(0.0, 10.0),  # image shear (+/- deg)
        "perspective": tune.uniform(0.0, 0.001),  # image perspective (+/- fraction), range 0-0.001
        "flipud": tune.uniform(0.0, 1.0),  # image flip up-down (probability)
        "fliplr": tune.uniform(0.0, 1.0),  # image flip left-right (probability)
        "bgr": tune.uniform(0.0, 1.0),  # image channel BGR (probability)
        "mosaic": tune.uniform(0.0, 1.0),  # image mixup (probability)
        "mixup": tune.uniform(0.0, 1.0),  # image mixup (probability)
        "copy_paste": tune.uniform(0.0, 1.0),  # segment copy-paste (probability)
    }

    # Put the model in ray store
    task = model.task
    model_in_store = ray.put(model)

    def _tune(config):
        """Train the YOLO model with the specified hyperparameters."""
        model_to_train = ray.get(model_in_store)  # get the model from ray store for tuning
        model_to_train.reset_callbacks()
        config.update(train_args)
        results = model_to_train.train(**config)
        return results.results_dict

    # Get search space
    if not space:
        space = default_space
        LOGGER.warning("WARNING ⚠️ search space not provided, using default search space.")

    # Get dataset
    data = train_args.get("data", TASK2DATA[task])
    space["data"] = data
    if "data" not in train_args:
        LOGGER.warning(f'WARNING ⚠️ data not provided, using default "data={data}".')

    # Define the trainable function with allocated resources
    trainable_with_resources = tune.with_resources(_tune, {"cpu": NUM_THREADS, "gpu": gpu_per_trial or 0})

    # Define the ASHA scheduler for hyperparameter search
    asha_scheduler = ASHAScheduler(
        time_attr="epoch",
        metric=TASK2METRIC[task],
        mode="max",
        max_t=train_args.get("epochs") or DEFAULT_CFG_DICT["epochs"] or 100,
        grace_period=grace_period,
        reduction_factor=3,
    )

    # Define the callbacks for the hyperparameter search
    tuner_callbacks = [WandbLoggerCallback(project="YOLOv8-tune")] if wandb else []

    # Create the Ray Tune hyperparameter search tuner
    tune_dir = get_save_dir(
        get_cfg(DEFAULT_CFG, train_args), name=train_args.pop("name", "tune")
    ).resolve()  # must be absolute dir
    tune_dir.mkdir(parents=True, exist_ok=True)
    tuner = tune.Tuner(
        trainable_with_resources,
        param_space=space,
        tune_config=tune.TuneConfig(scheduler=asha_scheduler, num_samples=max_samples),
        run_config=RunConfig(callbacks=tuner_callbacks, storage_path=tune_dir),
    )

    # Run the hyperparameter search
    tuner.fit()

    # Get the results of the hyperparameter search
    results = tuner.get_results()

    # Shut down Ray to clean up workers
    ray.shutdown()

    return results