xdobro23/YOLOv8-16bit
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

ultralytics 8.0.126 Ray Tune refactoring (#3511)

Browse Source 
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
This commit is contained in:
Glenn Jocher
2023-07-05 00:30:17 +02:00
committed by GitHub
parent 69a2d70a78
commit 4c22273198
8 changed files with 223 additions and 151 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
ultralytics/__init__.py
View File

@ -1,6 +1,6 @@
# Ultralytics YOLO 🚀, AGPL-3.0 license
__version__ = '8.0.125'
__version__ = '8.0.126'
from ultralytics.hub import start
from ultralytics.vit.rtdetr import RTDETR

6
ultralytics/yolo/cfg/__init__.py
View File

@ -26,6 +26,12 @@ TASK2MODEL = {
'segment': 'yolov8n-seg.pt',
'classify': 'yolov8n-cls.pt',
'pose': 'yolov8n-pose.pt'}
TASK2METRIC = {
'detect': 'metrics/mAP50-95(B)',
'segment': 'metrics/mAP50-95(M)',
'classify': 'metrics/accuracy_top1',
'pose': 'metrics/mAP50-95(P)'}
CLI_HELP_MSG = \
f"""

75
ultralytics/yolo/engine/model.py
View File

@ -9,8 +9,8 @@ from ultralytics.nn.tasks import (ClassificationModel, DetectionModel, PoseModel
attempt_load_one_weight, guess_model_task, nn, yaml_model_load)
from ultralytics.yolo.cfg import get_cfg
from ultralytics.yolo.engine.exporter import Exporter
from ultralytics.yolo.utils import (DEFAULT_CFG, DEFAULT_CFG_DICT, DEFAULT_CFG_KEYS, LOGGER, NUM_THREADS, RANK, ROOT,
callbacks, is_git_dir, yaml_load)
from ultralytics.yolo.utils import (DEFAULT_CFG, DEFAULT_CFG_DICT, DEFAULT_CFG_KEYS, LOGGER, RANK, ROOT, callbacks,
is_git_dir, yaml_load)
from ultralytics.yolo.utils.checks import check_file, check_imgsz, check_pip_update_available, check_yaml
from ultralytics.yolo.utils.downloads import GITHUB_ASSET_STEMS
from ultralytics.yolo.utils.torch_utils import smart_inference_mode
@ -387,13 +387,7 @@ class YOLO:
self._check_is_pytorch_model()
self.model.to(device)
def tune(self,
data: str,
space: dict = None,
grace_period: int = 10,
gpu_per_trial: int = None,
max_samples: int = 10,
train_args: dict = None):
def tune(self, *args, **kwargs):
"""
Runs hyperparameter tuning using Ray Tune.
@ -411,66 +405,9 @@ class YOLO:
Raises:
ModuleNotFoundError: If Ray Tune is not installed.
"""
if train_args is None:
train_args = {}
try:
from ultralytics.yolo.utils.tuner import (ASHAScheduler, RunConfig, WandbLoggerCallback, default_space,
task_metric_map, tune)
except ImportError:
raise ModuleNotFoundError("Install Ray Tune: `pip install 'ray[tune]'`")
try:
import wandb
from wandb import __version__ # noqa
except ImportError:
wandb = False
def _tune(config):
"""
Trains the YOLO model with the specified hyperparameters and additional arguments.
Args:
config (dict): A dictionary of hyperparameters to use for training.
Returns:
None.
"""
self._reset_callbacks()
config.update(train_args)
self.train(**config)
if not space:
LOGGER.warning('WARNING: search space not provided. Using default search space')
space = default_space
space['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=task_metric_map[self.task],
mode='max',
max_t=train_args.get('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
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, local_dir='./runs'))
# Run the hyperparameter search
tuner.fit()
# Return the results of the hyperparameter search
return tuner.get_results()
self._check_is_pytorch_model()
from ultralytics.yolo.utils.tuner import run_ray_tune
return run_ray_tune(self, *args, **kwargs)
@property
def names(self):

13
ultralytics/yolo/utils/benchmarks.py
View File

@ -33,6 +33,7 @@ import torch.cuda
from tqdm import tqdm
from ultralytics import YOLO
from ultralytics.yolo.cfg import TASK2DATA, TASK2METRIC
from ultralytics.yolo.engine.exporter import export_formats
from ultralytics.yolo.utils import LINUX, LOGGER, MACOS, ROOT, SETTINGS
from ultralytics.yolo.utils.checks import check_requirements, check_yolo
@ -96,6 +97,7 @@ def benchmark(model=Path(SETTINGS['weights_dir']) / 'yolov8n.pt',
emoji = '❎' # indicates export succeeded
# Predict
assert model.task != 'pose' or i != 7, 'GraphDef Pose inference is not supported'
assert i not in (9, 10), 'inference not supported' # Edge TPU and TF.js are unsupported
assert i != 5 or platform.system() == 'Darwin', 'inference only supported on macOS>=10.13' # CoreML
if not (ROOT / 'assets/bus.jpg').exists():
@ -103,15 +105,8 @@ def benchmark(model=Path(SETTINGS['weights_dir']) / 'yolov8n.pt',
export.predict(ROOT / 'assets/bus.jpg', imgsz=imgsz, device=device, half=half)
# Validate
if model.task == 'detect':
data, key = 'coco8.yaml', 'metrics/mAP50-95(B)'
elif model.task == 'segment':
data, key = 'coco8-seg.yaml', 'metrics/mAP50-95(M)'
elif model.task == 'classify':
data, key = 'imagenet100', 'metrics/accuracy_top5'
elif model.task == 'pose':
data, key = 'coco8-pose.yaml', 'metrics/mAP50-95(P)'
data = TASK2DATA[model.task] # task to dataset, i.e. coco8.yaml for task=detect
key = TASK2METRIC[model.task] # task to metric, i.e. metrics/mAP50-95(B) for task=detect
results = export.val(data=data,
batch=1,
imgsz=imgsz,

152
ultralytics/yolo/utils/tuner.py
View File

@ -1,44 +1,120 @@
# Ultralytics YOLO 🚀, AGPL-3.0 license
from ultralytics.yolo.cfg import TASK2DATA, TASK2METRIC
from ultralytics.yolo.utils import DEFAULT_CFG_DICT, LOGGER, NUM_THREADS
from ultralytics.yolo.utils import LOGGER
try:
from ray import tune
from ray.air import RunConfig, session # noqa
from ray.air.integrations.wandb import WandbLoggerCallback # noqa
from ray.tune.schedulers import ASHAScheduler # noqa
from ray.tune.schedulers import AsyncHyperBandScheduler as AHB # noqa
def run_ray_tune(model,
space: dict = None,
grace_period: int = 10,
gpu_per_trial: int = None,
max_samples: int = 10,
**train_args):
"""
Runs hyperparameter tuning using Ray Tune.
except ImportError:
LOGGER.info("Tuning hyperparameters requires ray/tune. Install using `pip install 'ray[tune]'`")
tune = None
Args:
model (YOLO): Model to run the tuner on.
space (dict, optional): The hyperparameter search space. Defaults to None.
grace_period (int, optional): The grace period in epochs of the ASHA scheduler. Defaults to 10.
gpu_per_trial (int, optional): The number of GPUs to allocate per trial. Defaults to None.
max_samples (int, optional): The maximum number of trials to run. Defaults to 10.
train_args (dict, optional): Additional arguments to pass to the `train()` method. Defaults to {}.
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 5e-4
'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)
'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)
Returns:
(dict): A dictionary containing the results of the hyperparameter search.
task_metric_map = {
'detect': 'metrics/mAP50-95(B)',
'segment': 'metrics/mAP50-95(M)',
'classify': 'metrics/accuracy_top1',
'pose': 'metrics/mAP50-95(P)'}
Raises:
ModuleNotFoundError: If Ray Tune is not installed.
"""
if train_args is None:
train_args = {}
try:
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("Tuning hyperparameters requires Ray Tune. Install with: pip install 'ray[tune]'")
try:
import wandb
assert hasattr(wandb, '__version__')
except (ImportError, AssertionError):
wandb = False
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 5e-4
'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)
'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)
def _tune(config):
"""
Trains the YOLO model with the specified hyperparameters and additional arguments.
Args:
config (dict): A dictionary of hyperparameters to use for training.
Returns:
None.
"""
model._reset_callbacks()
config.update(train_args)
model.train(**config)
# 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[model.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[model.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
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='./runs/tune'))
# Run the hyperparameter search
tuner.fit()
# Return the results of the hyperparameter search
return tuner.get_results()